vllm-cpu-amxbf16 0.11.2.post2__cp310-cp310-manylinux_2_17_x86_64.whl

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (1536) hide show
  1. vllm/_C.abi3.so +0 -0
  2. vllm/__init__.py +225 -0
  3. vllm/_aiter_ops.py +983 -0
  4. vllm/_bc_linter.py +54 -0
  5. vllm/_custom_ops.py +2863 -0
  6. vllm/_ipex_ops.py +457 -0
  7. vllm/_version.py +34 -0
  8. vllm/assets/__init__.py +0 -0
  9. vllm/assets/audio.py +43 -0
  10. vllm/assets/base.py +40 -0
  11. vllm/assets/image.py +59 -0
  12. vllm/assets/video.py +149 -0
  13. vllm/attention/__init__.py +18 -0
  14. vllm/attention/backends/__init__.py +0 -0
  15. vllm/attention/backends/abstract.py +391 -0
  16. vllm/attention/backends/registry.py +195 -0
  17. vllm/attention/backends/utils.py +33 -0
  18. vllm/attention/layer.py +1052 -0
  19. vllm/attention/layers/__init__.py +0 -0
  20. vllm/attention/layers/chunked_local_attention.py +121 -0
  21. vllm/attention/layers/cross_attention.py +178 -0
  22. vllm/attention/layers/encoder_only_attention.py +103 -0
  23. vllm/attention/ops/__init__.py +0 -0
  24. vllm/attention/ops/chunked_prefill_paged_decode.py +401 -0
  25. vllm/attention/ops/common.py +414 -0
  26. vllm/attention/ops/flashmla.py +251 -0
  27. vllm/attention/ops/merge_attn_states.py +47 -0
  28. vllm/attention/ops/paged_attn.py +262 -0
  29. vllm/attention/ops/pallas_kv_cache_update.py +130 -0
  30. vllm/attention/ops/prefix_prefill.py +814 -0
  31. vllm/attention/ops/rocm_aiter_paged_attn.py +123 -0
  32. vllm/attention/ops/triton_decode_attention.py +712 -0
  33. vllm/attention/ops/triton_merge_attn_states.py +105 -0
  34. vllm/attention/ops/triton_reshape_and_cache_flash.py +184 -0
  35. vllm/attention/ops/triton_unified_attention.py +941 -0
  36. vllm/attention/ops/vit_attn_wrappers.py +178 -0
  37. vllm/attention/selector.py +231 -0
  38. vllm/attention/utils/__init__.py +0 -0
  39. vllm/attention/utils/fa_utils.py +109 -0
  40. vllm/attention/utils/kv_sharing_utils.py +33 -0
  41. vllm/attention/utils/kv_transfer_utils.py +60 -0
  42. vllm/beam_search.py +88 -0
  43. vllm/benchmarks/__init__.py +0 -0
  44. vllm/benchmarks/datasets.py +3222 -0
  45. vllm/benchmarks/latency.py +172 -0
  46. vllm/benchmarks/lib/__init__.py +3 -0
  47. vllm/benchmarks/lib/endpoint_request_func.py +777 -0
  48. vllm/benchmarks/lib/ready_checker.py +72 -0
  49. vllm/benchmarks/lib/utils.py +79 -0
  50. vllm/benchmarks/serve.py +1531 -0
  51. vllm/benchmarks/sweep/__init__.py +0 -0
  52. vllm/benchmarks/sweep/cli.py +38 -0
  53. vllm/benchmarks/sweep/param_sweep.py +91 -0
  54. vllm/benchmarks/sweep/plot.py +580 -0
  55. vllm/benchmarks/sweep/serve.py +416 -0
  56. vllm/benchmarks/sweep/serve_sla.py +492 -0
  57. vllm/benchmarks/sweep/server.py +114 -0
  58. vllm/benchmarks/sweep/sla_sweep.py +132 -0
  59. vllm/benchmarks/sweep/utils.py +4 -0
  60. vllm/benchmarks/throughput.py +799 -0
  61. vllm/collect_env.py +857 -0
  62. vllm/compilation/__init__.py +0 -0
  63. vllm/compilation/activation_quant_fusion.py +209 -0
  64. vllm/compilation/backends.py +759 -0
  65. vllm/compilation/base_static_graph.py +57 -0
  66. vllm/compilation/caching.py +178 -0
  67. vllm/compilation/collective_fusion.py +1234 -0
  68. vllm/compilation/compiler_interface.py +639 -0
  69. vllm/compilation/counter.py +48 -0
  70. vllm/compilation/cuda_graph.py +208 -0
  71. vllm/compilation/decorators.py +571 -0
  72. vllm/compilation/fix_functionalization.py +253 -0
  73. vllm/compilation/fusion.py +374 -0
  74. vllm/compilation/fusion_attn.py +359 -0
  75. vllm/compilation/fx_utils.py +91 -0
  76. vllm/compilation/inductor_pass.py +133 -0
  77. vllm/compilation/matcher_utils.py +317 -0
  78. vllm/compilation/monitor.py +62 -0
  79. vllm/compilation/noop_elimination.py +134 -0
  80. vllm/compilation/partition_rules.py +72 -0
  81. vllm/compilation/pass_manager.py +135 -0
  82. vllm/compilation/piecewise_backend.py +121 -0
  83. vllm/compilation/post_cleanup.py +21 -0
  84. vllm/compilation/qk_norm_rope_fusion.py +238 -0
  85. vllm/compilation/sequence_parallelism.py +363 -0
  86. vllm/compilation/torch25_custom_graph_pass.py +44 -0
  87. vllm/compilation/vllm_inductor_pass.py +173 -0
  88. vllm/compilation/wrapper.py +238 -0
  89. vllm/config/__init__.py +102 -0
  90. vllm/config/cache.py +207 -0
  91. vllm/config/compilation.py +975 -0
  92. vllm/config/device.py +75 -0
  93. vllm/config/ec_transfer.py +110 -0
  94. vllm/config/kv_events.py +56 -0
  95. vllm/config/kv_transfer.py +114 -0
  96. vllm/config/load.py +124 -0
  97. vllm/config/lora.py +112 -0
  98. vllm/config/model.py +2162 -0
  99. vllm/config/multimodal.py +248 -0
  100. vllm/config/observability.py +123 -0
  101. vllm/config/parallel.py +655 -0
  102. vllm/config/pooler.py +122 -0
  103. vllm/config/scheduler.py +298 -0
  104. vllm/config/speculative.py +654 -0
  105. vllm/config/speech_to_text.py +38 -0
  106. vllm/config/structured_outputs.py +92 -0
  107. vllm/config/utils.py +178 -0
  108. vllm/config/vllm.py +1166 -0
  109. vllm/connections.py +189 -0
  110. vllm/device_allocator/__init__.py +0 -0
  111. vllm/device_allocator/cumem.py +327 -0
  112. vllm/distributed/__init__.py +6 -0
  113. vllm/distributed/communication_op.py +43 -0
  114. vllm/distributed/device_communicators/__init__.py +0 -0
  115. vllm/distributed/device_communicators/all2all.py +490 -0
  116. vllm/distributed/device_communicators/all_reduce_utils.py +344 -0
  117. vllm/distributed/device_communicators/base_device_communicator.py +297 -0
  118. vllm/distributed/device_communicators/cpu_communicator.py +209 -0
  119. vllm/distributed/device_communicators/cuda_communicator.py +340 -0
  120. vllm/distributed/device_communicators/cuda_wrapper.py +216 -0
  121. vllm/distributed/device_communicators/custom_all_reduce.py +326 -0
  122. vllm/distributed/device_communicators/mnnvl_compat.py +27 -0
  123. vllm/distributed/device_communicators/pynccl.py +386 -0
  124. vllm/distributed/device_communicators/pynccl_allocator.py +191 -0
  125. vllm/distributed/device_communicators/pynccl_wrapper.py +564 -0
  126. vllm/distributed/device_communicators/quick_all_reduce.py +290 -0
  127. vllm/distributed/device_communicators/ray_communicator.py +259 -0
  128. vllm/distributed/device_communicators/shm_broadcast.py +733 -0
  129. vllm/distributed/device_communicators/shm_object_storage.py +660 -0
  130. vllm/distributed/device_communicators/symm_mem.py +156 -0
  131. vllm/distributed/device_communicators/tpu_communicator.py +107 -0
  132. vllm/distributed/device_communicators/xpu_communicator.py +95 -0
  133. vllm/distributed/ec_transfer/__init__.py +14 -0
  134. vllm/distributed/ec_transfer/ec_connector/__init__.py +0 -0
  135. vllm/distributed/ec_transfer/ec_connector/base.py +247 -0
  136. vllm/distributed/ec_transfer/ec_connector/factory.py +88 -0
  137. vllm/distributed/ec_transfer/ec_connector/shared_storage_connector.py +201 -0
  138. vllm/distributed/ec_transfer/ec_transfer_state.py +42 -0
  139. vllm/distributed/eplb/__init__.py +8 -0
  140. vllm/distributed/eplb/eplb_state.py +837 -0
  141. vllm/distributed/eplb/rebalance_algo.py +260 -0
  142. vllm/distributed/eplb/rebalance_execute.py +431 -0
  143. vllm/distributed/kv_events.py +371 -0
  144. vllm/distributed/kv_transfer/README.md +29 -0
  145. vllm/distributed/kv_transfer/__init__.py +20 -0
  146. vllm/distributed/kv_transfer/disagg_prefill_workflow.jpg +0 -0
  147. vllm/distributed/kv_transfer/kv_connector/__init__.py +0 -0
  148. vllm/distributed/kv_transfer/kv_connector/base.py +10 -0
  149. vllm/distributed/kv_transfer/kv_connector/factory.py +192 -0
  150. vllm/distributed/kv_transfer/kv_connector/utils.py +268 -0
  151. vllm/distributed/kv_transfer/kv_connector/v1/__init__.py +19 -0
  152. vllm/distributed/kv_transfer/kv_connector/v1/base.py +546 -0
  153. vllm/distributed/kv_transfer/kv_connector/v1/decode_bench_connector.py +419 -0
  154. vllm/distributed/kv_transfer/kv_connector/v1/lmcache_connector.py +216 -0
  155. vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/__init__.py +18 -0
  156. vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/multi_process_adapter.py +379 -0
  157. vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/utils.py +221 -0
  158. vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/vllm_v1_adapter.py +1411 -0
  159. vllm/distributed/kv_transfer/kv_connector/v1/lmcache_mp_connector.py +867 -0
  160. vllm/distributed/kv_transfer/kv_connector/v1/metrics.py +189 -0
  161. vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py +454 -0
  162. vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py +2440 -0
  163. vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py +504 -0
  164. vllm/distributed/kv_transfer/kv_connector/v1/p2p/__init__.py +0 -0
  165. vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py +531 -0
  166. vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py +632 -0
  167. vllm/distributed/kv_transfer/kv_connector/v1/p2p/tensor_memory_pool.py +273 -0
  168. vllm/distributed/kv_transfer/kv_connector/v1/shared_storage_connector.py +450 -0
  169. vllm/distributed/kv_transfer/kv_lookup_buffer/__init__.py +0 -0
  170. vllm/distributed/kv_transfer/kv_lookup_buffer/base.py +179 -0
  171. vllm/distributed/kv_transfer/kv_lookup_buffer/mooncake_store.py +164 -0
  172. vllm/distributed/kv_transfer/kv_lookup_buffer/simple_buffer.py +242 -0
  173. vllm/distributed/kv_transfer/kv_pipe/__init__.py +0 -0
  174. vllm/distributed/kv_transfer/kv_pipe/base.py +66 -0
  175. vllm/distributed/kv_transfer/kv_pipe/mooncake_pipe.py +295 -0
  176. vllm/distributed/kv_transfer/kv_pipe/pynccl_pipe.py +285 -0
  177. vllm/distributed/kv_transfer/kv_transfer_state.py +78 -0
  178. vllm/distributed/parallel_state.py +1759 -0
  179. vllm/distributed/tpu_distributed_utils.py +188 -0
  180. vllm/distributed/utils.py +543 -0
  181. vllm/engine/__init__.py +0 -0
  182. vllm/engine/arg_utils.py +2144 -0
  183. vllm/engine/async_llm_engine.py +6 -0
  184. vllm/engine/llm_engine.py +6 -0
  185. vllm/engine/protocol.py +170 -0
  186. vllm/entrypoints/__init__.py +0 -0
  187. vllm/entrypoints/anthropic/__init__.py +0 -0
  188. vllm/entrypoints/anthropic/protocol.py +162 -0
  189. vllm/entrypoints/anthropic/serving_messages.py +460 -0
  190. vllm/entrypoints/api_server.py +184 -0
  191. vllm/entrypoints/chat_utils.py +1690 -0
  192. vllm/entrypoints/cli/__init__.py +13 -0
  193. vllm/entrypoints/cli/benchmark/__init__.py +0 -0
  194. vllm/entrypoints/cli/benchmark/base.py +25 -0
  195. vllm/entrypoints/cli/benchmark/latency.py +21 -0
  196. vllm/entrypoints/cli/benchmark/main.py +56 -0
  197. vllm/entrypoints/cli/benchmark/serve.py +21 -0
  198. vllm/entrypoints/cli/benchmark/sweep.py +21 -0
  199. vllm/entrypoints/cli/benchmark/throughput.py +21 -0
  200. vllm/entrypoints/cli/collect_env.py +38 -0
  201. vllm/entrypoints/cli/main.py +79 -0
  202. vllm/entrypoints/cli/openai.py +256 -0
  203. vllm/entrypoints/cli/run_batch.py +68 -0
  204. vllm/entrypoints/cli/serve.py +249 -0
  205. vllm/entrypoints/cli/types.py +29 -0
  206. vllm/entrypoints/constants.py +10 -0
  207. vllm/entrypoints/context.py +572 -0
  208. vllm/entrypoints/dynamic_lora.py +57 -0
  209. vllm/entrypoints/harmony_utils.py +535 -0
  210. vllm/entrypoints/launcher.py +175 -0
  211. vllm/entrypoints/llm.py +1768 -0
  212. vllm/entrypoints/logger.py +84 -0
  213. vllm/entrypoints/openai/__init__.py +0 -0
  214. vllm/entrypoints/openai/api_server.py +2096 -0
  215. vllm/entrypoints/openai/cli_args.py +302 -0
  216. vllm/entrypoints/openai/orca_metrics.py +120 -0
  217. vllm/entrypoints/openai/protocol.py +3299 -0
  218. vllm/entrypoints/openai/run_batch.py +547 -0
  219. vllm/entrypoints/openai/serving_chat.py +1772 -0
  220. vllm/entrypoints/openai/serving_classification.py +235 -0
  221. vllm/entrypoints/openai/serving_completion.py +715 -0
  222. vllm/entrypoints/openai/serving_embedding.py +695 -0
  223. vllm/entrypoints/openai/serving_engine.py +1433 -0
  224. vllm/entrypoints/openai/serving_models.py +304 -0
  225. vllm/entrypoints/openai/serving_pooling.py +346 -0
  226. vllm/entrypoints/openai/serving_responses.py +2021 -0
  227. vllm/entrypoints/openai/serving_score.py +503 -0
  228. vllm/entrypoints/openai/serving_tokenization.py +203 -0
  229. vllm/entrypoints/openai/serving_tokens.py +269 -0
  230. vllm/entrypoints/openai/serving_transcription.py +148 -0
  231. vllm/entrypoints/openai/speech_to_text.py +405 -0
  232. vllm/entrypoints/openai/tool_parsers/__init__.py +142 -0
  233. vllm/entrypoints/openai/tool_parsers/abstract_tool_parser.py +273 -0
  234. vllm/entrypoints/openai/tool_parsers/deepseekv31_tool_parser.py +390 -0
  235. vllm/entrypoints/openai/tool_parsers/deepseekv3_tool_parser.py +390 -0
  236. vllm/entrypoints/openai/tool_parsers/ernie45_tool_parser.py +210 -0
  237. vllm/entrypoints/openai/tool_parsers/glm4_moe_tool_parser.py +200 -0
  238. vllm/entrypoints/openai/tool_parsers/granite_20b_fc_tool_parser.py +273 -0
  239. vllm/entrypoints/openai/tool_parsers/granite_tool_parser.py +253 -0
  240. vllm/entrypoints/openai/tool_parsers/hermes_tool_parser.py +494 -0
  241. vllm/entrypoints/openai/tool_parsers/hunyuan_a13b_tool_parser.py +420 -0
  242. vllm/entrypoints/openai/tool_parsers/internlm2_tool_parser.py +227 -0
  243. vllm/entrypoints/openai/tool_parsers/jamba_tool_parser.py +323 -0
  244. vllm/entrypoints/openai/tool_parsers/kimi_k2_tool_parser.py +590 -0
  245. vllm/entrypoints/openai/tool_parsers/llama4_pythonic_tool_parser.py +341 -0
  246. vllm/entrypoints/openai/tool_parsers/llama_tool_parser.py +290 -0
  247. vllm/entrypoints/openai/tool_parsers/longcat_tool_parser.py +37 -0
  248. vllm/entrypoints/openai/tool_parsers/minimax_m2_tool_parser.py +643 -0
  249. vllm/entrypoints/openai/tool_parsers/minimax_tool_parser.py +849 -0
  250. vllm/entrypoints/openai/tool_parsers/mistral_tool_parser.py +390 -0
  251. vllm/entrypoints/openai/tool_parsers/olmo3_tool_parser.py +366 -0
  252. vllm/entrypoints/openai/tool_parsers/openai_tool_parser.py +97 -0
  253. vllm/entrypoints/openai/tool_parsers/phi4mini_tool_parser.py +120 -0
  254. vllm/entrypoints/openai/tool_parsers/pythonic_tool_parser.py +332 -0
  255. vllm/entrypoints/openai/tool_parsers/qwen3coder_tool_parser.py +781 -0
  256. vllm/entrypoints/openai/tool_parsers/qwen3xml_tool_parser.py +1316 -0
  257. vllm/entrypoints/openai/tool_parsers/seed_oss_tool_parser.py +744 -0
  258. vllm/entrypoints/openai/tool_parsers/step3_tool_parser.py +303 -0
  259. vllm/entrypoints/openai/tool_parsers/utils.py +229 -0
  260. vllm/entrypoints/openai/tool_parsers/xlam_tool_parser.py +556 -0
  261. vllm/entrypoints/renderer.py +409 -0
  262. vllm/entrypoints/responses_utils.py +77 -0
  263. vllm/entrypoints/sagemaker/__init__.py +4 -0
  264. vllm/entrypoints/sagemaker/routes.py +72 -0
  265. vllm/entrypoints/score_utils.py +242 -0
  266. vllm/entrypoints/ssl.py +78 -0
  267. vllm/entrypoints/tool.py +143 -0
  268. vllm/entrypoints/tool_server.py +209 -0
  269. vllm/entrypoints/utils.py +319 -0
  270. vllm/env_override.py +378 -0
  271. vllm/envs.py +1659 -0
  272. vllm/forward_context.py +356 -0
  273. vllm/inputs/__init__.py +44 -0
  274. vllm/inputs/data.py +359 -0
  275. vllm/inputs/parse.py +137 -0
  276. vllm/inputs/preprocess.py +727 -0
  277. vllm/logger.py +267 -0
  278. vllm/logging_utils/__init__.py +10 -0
  279. vllm/logging_utils/dump_input.py +83 -0
  280. vllm/logging_utils/formatter.py +77 -0
  281. vllm/logging_utils/log_time.py +34 -0
  282. vllm/logits_process.py +121 -0
  283. vllm/logprobs.py +208 -0
  284. vllm/lora/__init__.py +0 -0
  285. vllm/lora/layers/__init__.py +41 -0
  286. vllm/lora/layers/base.py +67 -0
  287. vllm/lora/layers/base_linear.py +164 -0
  288. vllm/lora/layers/column_parallel_linear.py +578 -0
  289. vllm/lora/layers/fused_moe.py +472 -0
  290. vllm/lora/layers/logits_processor.py +252 -0
  291. vllm/lora/layers/replicated_linear.py +70 -0
  292. vllm/lora/layers/row_parallel_linear.py +181 -0
  293. vllm/lora/layers/utils.py +65 -0
  294. vllm/lora/layers/vocal_parallel_embedding.py +166 -0
  295. vllm/lora/lora_weights.py +198 -0
  296. vllm/lora/models.py +890 -0
  297. vllm/lora/ops/__init__.py +0 -0
  298. vllm/lora/ops/ipex_ops/__init__.py +6 -0
  299. vllm/lora/ops/ipex_ops/lora_ops.py +57 -0
  300. vllm/lora/ops/torch_ops/__init__.py +20 -0
  301. vllm/lora/ops/torch_ops/lora_ops.py +128 -0
  302. vllm/lora/ops/triton_ops/README_TUNING.md +60 -0
  303. vllm/lora/ops/triton_ops/__init__.py +21 -0
  304. vllm/lora/ops/triton_ops/fused_moe_lora_op.py +641 -0
  305. vllm/lora/ops/triton_ops/kernel_utils.py +340 -0
  306. vllm/lora/ops/triton_ops/lora_expand_op.py +310 -0
  307. vllm/lora/ops/triton_ops/lora_kernel_metadata.py +154 -0
  308. vllm/lora/ops/triton_ops/lora_shrink_op.py +287 -0
  309. vllm/lora/ops/triton_ops/utils.py +295 -0
  310. vllm/lora/ops/xla_ops/__init__.py +6 -0
  311. vllm/lora/ops/xla_ops/lora_ops.py +141 -0
  312. vllm/lora/peft_helper.py +128 -0
  313. vllm/lora/punica_wrapper/__init__.py +10 -0
  314. vllm/lora/punica_wrapper/punica_base.py +492 -0
  315. vllm/lora/punica_wrapper/punica_cpu.py +351 -0
  316. vllm/lora/punica_wrapper/punica_gpu.py +411 -0
  317. vllm/lora/punica_wrapper/punica_selector.py +21 -0
  318. vllm/lora/punica_wrapper/punica_tpu.py +359 -0
  319. vllm/lora/punica_wrapper/punica_xpu.py +279 -0
  320. vllm/lora/punica_wrapper/utils.py +150 -0
  321. vllm/lora/request.py +100 -0
  322. vllm/lora/resolver.py +88 -0
  323. vllm/lora/utils.py +293 -0
  324. vllm/lora/worker_manager.py +279 -0
  325. vllm/model_executor/__init__.py +11 -0
  326. vllm/model_executor/custom_op.py +194 -0
  327. vllm/model_executor/layers/__init__.py +0 -0
  328. vllm/model_executor/layers/activation.py +569 -0
  329. vllm/model_executor/layers/attention_layer_base.py +35 -0
  330. vllm/model_executor/layers/batch_invariant.py +854 -0
  331. vllm/model_executor/layers/conv.py +236 -0
  332. vllm/model_executor/layers/fla/__init__.py +8 -0
  333. vllm/model_executor/layers/fla/ops/__init__.py +17 -0
  334. vllm/model_executor/layers/fla/ops/chunk.py +240 -0
  335. vllm/model_executor/layers/fla/ops/chunk_delta_h.py +344 -0
  336. vllm/model_executor/layers/fla/ops/chunk_o.py +183 -0
  337. vllm/model_executor/layers/fla/ops/chunk_scaled_dot_kkt.py +154 -0
  338. vllm/model_executor/layers/fla/ops/cumsum.py +280 -0
  339. vllm/model_executor/layers/fla/ops/fused_recurrent.py +390 -0
  340. vllm/model_executor/layers/fla/ops/index.py +41 -0
  341. vllm/model_executor/layers/fla/ops/kda.py +1351 -0
  342. vllm/model_executor/layers/fla/ops/l2norm.py +146 -0
  343. vllm/model_executor/layers/fla/ops/layernorm_guard.py +396 -0
  344. vllm/model_executor/layers/fla/ops/op.py +60 -0
  345. vllm/model_executor/layers/fla/ops/solve_tril.py +556 -0
  346. vllm/model_executor/layers/fla/ops/utils.py +194 -0
  347. vllm/model_executor/layers/fla/ops/wy_fast.py +158 -0
  348. vllm/model_executor/layers/fused_moe/__init__.py +106 -0
  349. vllm/model_executor/layers/fused_moe/all2all_utils.py +160 -0
  350. vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py +406 -0
  351. vllm/model_executor/layers/fused_moe/batched_triton_or_deep_gemm_moe.py +180 -0
  352. vllm/model_executor/layers/fused_moe/config.py +916 -0
  353. vllm/model_executor/layers/fused_moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  354. vllm/model_executor/layers/fused_moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  355. vllm/model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  356. vllm/model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  357. vllm/model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  358. vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  359. vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  360. vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json +218 -0
  361. vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H200,dtype=int8_w8a16.json +146 -0
  362. vllm/model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  363. vllm/model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  364. vllm/model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  365. vllm/model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  366. vllm/model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  367. vllm/model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  368. vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  369. vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI300X.json +200 -0
  370. vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  371. vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H100,dtype=fp8_w8a8.json +123 -0
  372. vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  373. vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H200.json +146 -0
  374. vllm/model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_H100_80GB_HBM3.json +147 -0
  375. vllm/model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_L40S.json +147 -0
  376. vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  377. vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  378. vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20-3e.json +146 -0
  379. vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20.json +146 -0
  380. vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H200.json +146 -0
  381. vllm/model_executor/layers/fused_moe/configs/E=128,N=352,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +122 -0
  382. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  383. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  384. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  385. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  386. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  387. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20-3e.json +146 -0
  388. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20.json +146 -0
  389. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  390. vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200.json +146 -0
  391. vllm/model_executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  392. vllm/model_executor/layers/fused_moe/configs/E=128,N=704,device_name=NVIDIA_B200,dtype=fp8_w8a8.json +146 -0
  393. vllm/model_executor/layers/fused_moe/configs/E=128,N=704,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +114 -0
  394. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  395. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI308X.json +213 -0
  396. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  397. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  398. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  399. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  400. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20.json +146 -0
  401. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  402. vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200.json +146 -0
  403. vllm/model_executor/layers/fused_moe/configs/E=128,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=bf16.json +82 -0
  404. vllm/model_executor/layers/fused_moe/configs/E=128,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +82 -0
  405. vllm/model_executor/layers/fused_moe/configs/E=128,N=928,device_name=NVIDIA_H100_80GB_HBM3.json +147 -0
  406. vllm/model_executor/layers/fused_moe/configs/E=128,N=928,device_name=NVIDIA_L40S.json +147 -0
  407. vllm/model_executor/layers/fused_moe/configs/E=128,N=96,device_name=NVIDIA_H20.json +146 -0
  408. vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=AMD_Instinct_MI300X.json +200 -0
  409. vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_B200,dtype=fp8_w8a8.json +147 -0
  410. vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_B200.json +146 -0
  411. vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H100.json +146 -0
  412. vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  413. vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H200.json +146 -0
  414. vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  415. vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  416. vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  417. vllm/model_executor/layers/fused_moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  418. vllm/model_executor/layers/fused_moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  419. vllm/model_executor/layers/fused_moe/configs/E=16,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  420. vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  421. vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  422. vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  423. vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  424. vllm/model_executor/layers/fused_moe/configs/E=16,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  425. vllm/model_executor/layers/fused_moe/configs/E=16,N=2048,device_name=NVIDIA_H200.json +146 -0
  426. vllm/model_executor/layers/fused_moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  427. vllm/model_executor/layers/fused_moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  428. vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  429. vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json +146 -0
  430. vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  431. vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_H200,dtype=int8_w8a16.json +146 -0
  432. vllm/model_executor/layers/fused_moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  433. vllm/model_executor/layers/fused_moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  434. vllm/model_executor/layers/fused_moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  435. vllm/model_executor/layers/fused_moe/configs/E=16,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  436. vllm/model_executor/layers/fused_moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  437. vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  438. vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  439. vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json +146 -0
  440. vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  441. vllm/model_executor/layers/fused_moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  442. vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=AMD_Instinct_MI300X.json +201 -0
  443. vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=AMD_Instinct_MI350_OAM,dtype=fp8_w8a8.json +164 -0
  444. vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  445. vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=NVIDIA_H20-3e.json +146 -0
  446. vllm/model_executor/layers/fused_moe/configs/E=160,N=320,device_name=NVIDIA_H20-3e.json +146 -0
  447. vllm/model_executor/layers/fused_moe/configs/E=160,N=384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  448. vllm/model_executor/layers/fused_moe/configs/E=160,N=384,device_name=AMD_Instinct_MI350_OAM,dtype=fp8_w8a8.json +164 -0
  449. vllm/model_executor/layers/fused_moe/configs/E=160,N=384,device_name=AMD_Instinct_MI355_OAM,dtype=fp8_w8a8.json +164 -0
  450. vllm/model_executor/layers/fused_moe/configs/E=160,N=640,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  451. vllm/model_executor/layers/fused_moe/configs/E=160,N=640,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  452. vllm/model_executor/layers/fused_moe/configs/E=160,N=640,device_name=NVIDIA_H100,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  453. vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  454. vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  455. vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_H100,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  456. vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  457. vllm/model_executor/layers/fused_moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json +200 -0
  458. vllm/model_executor/layers/fused_moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +200 -0
  459. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  460. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  461. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  462. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  463. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  464. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  465. vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  466. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  467. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +200 -0
  468. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +200 -0
  469. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  470. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  471. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  472. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  473. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  474. vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  475. vllm/model_executor/layers/fused_moe/configs/E=256,N=384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +147 -0
  476. vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +200 -0
  477. vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  478. vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  479. vllm/model_executor/layers/fused_moe/configs/E=32,N=1408,device_name=NVIDIA_B200.json +147 -0
  480. vllm/model_executor/layers/fused_moe/configs/E=32,N=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +147 -0
  481. vllm/model_executor/layers/fused_moe/configs/E=32,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +147 -0
  482. vllm/model_executor/layers/fused_moe/configs/E=384,N=128,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  483. vllm/model_executor/layers/fused_moe/configs/E=384,N=128,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  484. vllm/model_executor/layers/fused_moe/configs/E=384,N=128,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  485. vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  486. vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  487. vllm/model_executor/layers/fused_moe/configs/E=40,N=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8.json +147 -0
  488. vllm/model_executor/layers/fused_moe/configs/E=40,N=2560,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  489. vllm/model_executor/layers/fused_moe/configs/E=40,N=2560,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  490. vllm/model_executor/layers/fused_moe/configs/E=40,N=2560,device_name=NVIDIA_H100,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  491. vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_A100-SXM4-80GB.json +147 -0
  492. vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_B200.json +146 -0
  493. vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_GB200,dtype=fp8_w8a8.json +147 -0
  494. vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  495. vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H20-3e.json +146 -0
  496. vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H200.json +146 -0
  497. vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_B200.json +146 -0
  498. vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_GB200,dtype=fp8_w8a8.json +146 -0
  499. vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +147 -0
  500. vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  501. vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H20-3e.json +146 -0
  502. vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H200.json +146 -0
  503. vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_B200.json +146 -0
  504. vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_GB200,dtype=fp8_w8a8.json +146 -0
  505. vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  506. vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_H20-3e.json +146 -0
  507. vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_H200.json +146 -0
  508. vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_A100-SXM4-80GB.json +147 -0
  509. vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_B200.json +146 -0
  510. vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_H20-3e.json +146 -0
  511. vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_H200.json +146 -0
  512. vllm/model_executor/layers/fused_moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json +200 -0
  513. vllm/model_executor/layers/fused_moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json +200 -0
  514. vllm/model_executor/layers/fused_moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json +200 -0
  515. vllm/model_executor/layers/fused_moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json +200 -0
  516. vllm/model_executor/layers/fused_moe/configs/E=62,N=128,device_name=AMD_Instinct_MI300X.json +200 -0
  517. vllm/model_executor/layers/fused_moe/configs/E=62,N=256,device_name=AMD_Instinct_MI300X.json +200 -0
  518. vllm/model_executor/layers/fused_moe/configs/E=62,N=256,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  519. vllm/model_executor/layers/fused_moe/configs/E=62,N=512,device_name=AMD_Instinct_MI300X.json +200 -0
  520. vllm/model_executor/layers/fused_moe/configs/E=62,N=512,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  521. vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  522. vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  523. vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  524. vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  525. vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  526. vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json +146 -0
  527. vllm/model_executor/layers/fused_moe/configs/E=64,N=1408,device_name=NVIDIA_B200.json +147 -0
  528. vllm/model_executor/layers/fused_moe/configs/E=64,N=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8.json +146 -0
  529. vllm/model_executor/layers/fused_moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  530. vllm/model_executor/layers/fused_moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  531. vllm/model_executor/layers/fused_moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json +146 -0
  532. vllm/model_executor/layers/fused_moe/configs/E=64,N=3072,device_name=NVIDIA_H20,dtype=fp8_w8a8.json +146 -0
  533. vllm/model_executor/layers/fused_moe/configs/E=64,N=3072,device_name=NVIDIA_H20.json +146 -0
  534. vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  535. vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  536. vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  537. vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H200.json +146 -0
  538. vllm/model_executor/layers/fused_moe/configs/E=64,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8.json +146 -0
  539. vllm/model_executor/layers/fused_moe/configs/E=64,N=384,device_name=NVIDIA_H20.json +146 -0
  540. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  541. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  542. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json +146 -0
  543. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  544. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  545. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  546. vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H200.json +146 -0
  547. vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H100_PCIe,dtype=fp8_w8a8,block_shape=[128,128].json +147 -0
  548. vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8.json +146 -0
  549. vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H20.json +146 -0
  550. vllm/model_executor/layers/fused_moe/configs/E=64,N=896,device_name=NVIDIA_H20.json +146 -0
  551. vllm/model_executor/layers/fused_moe/configs/E=64,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=bf16.json +82 -0
  552. vllm/model_executor/layers/fused_moe/configs/E=64,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +82 -0
  553. vllm/model_executor/layers/fused_moe/configs/E=72,N=192,device_name=AMD_Instinct_MI300X.json +200 -0
  554. vllm/model_executor/layers/fused_moe/configs/E=72,N=384,device_name=AMD_Instinct_MI300X.json +200 -0
  555. vllm/model_executor/layers/fused_moe/configs/E=72,N=384,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  556. vllm/model_executor/layers/fused_moe/configs/E=72,N=768,device_name=AMD_Instinct_MI300X.json +200 -0
  557. vllm/model_executor/layers/fused_moe/configs/E=72,N=768,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  558. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  559. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json +200 -0
  560. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  561. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json +200 -0
  562. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +138 -0
  563. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  564. vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json +146 -0
  565. vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  566. vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json +200 -0
  567. vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  568. vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json +200 -0
  569. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  570. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json +200 -0
  571. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  572. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json +200 -0
  573. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  574. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  575. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  576. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  577. vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json +146 -0
  578. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  579. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json +200 -0
  580. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  581. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json +200 -0
  582. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  583. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  584. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  585. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +154 -0
  586. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  587. vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json +146 -0
  588. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  589. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json +200 -0
  590. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  591. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json +200 -0
  592. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  593. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  594. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json +146 -0
  595. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  596. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  597. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  598. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json +146 -0
  599. vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json +173 -0
  600. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  601. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json +200 -0
  602. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  603. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json +200 -0
  604. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  605. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  606. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  607. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  608. vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json +146 -0
  609. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  610. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json +200 -0
  611. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  612. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json +200 -0
  613. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  614. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  615. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  616. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  617. vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json +146 -0
  618. vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +164 -0
  619. vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json +200 -0
  620. vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +164 -0
  621. vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json +200 -0
  622. vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  623. vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  624. vllm/model_executor/layers/fused_moe/configs/README +12 -0
  625. vllm/model_executor/layers/fused_moe/cpu_fused_moe.py +354 -0
  626. vllm/model_executor/layers/fused_moe/cutlass_moe.py +1052 -0
  627. vllm/model_executor/layers/fused_moe/deep_gemm_moe.py +387 -0
  628. vllm/model_executor/layers/fused_moe/deep_gemm_utils.py +416 -0
  629. vllm/model_executor/layers/fused_moe/deepep_ht_prepare_finalize.py +420 -0
  630. vllm/model_executor/layers/fused_moe/deepep_ll_prepare_finalize.py +367 -0
  631. vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py +307 -0
  632. vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py +362 -0
  633. vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py +192 -0
  634. vllm/model_executor/layers/fused_moe/fused_batched_moe.py +1012 -0
  635. vllm/model_executor/layers/fused_moe/fused_marlin_moe.py +792 -0
  636. vllm/model_executor/layers/fused_moe/fused_moe.py +2175 -0
  637. vllm/model_executor/layers/fused_moe/fused_moe_method_base.py +112 -0
  638. vllm/model_executor/layers/fused_moe/fused_moe_modular_method.py +164 -0
  639. vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py +316 -0
  640. vllm/model_executor/layers/fused_moe/layer.py +1944 -0
  641. vllm/model_executor/layers/fused_moe/modular_kernel.py +1222 -0
  642. vllm/model_executor/layers/fused_moe/moe_align_block_size.py +174 -0
  643. vllm/model_executor/layers/fused_moe/moe_pallas.py +83 -0
  644. vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py +229 -0
  645. vllm/model_executor/layers/fused_moe/moe_torch_iterative.py +60 -0
  646. vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py +362 -0
  647. vllm/model_executor/layers/fused_moe/prepare_finalize.py +77 -0
  648. vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py +265 -0
  649. vllm/model_executor/layers/fused_moe/routing_simulator.py +310 -0
  650. vllm/model_executor/layers/fused_moe/shared_fused_moe.py +97 -0
  651. vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py +171 -0
  652. vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py +163 -0
  653. vllm/model_executor/layers/fused_moe/trtllm_moe.py +143 -0
  654. vllm/model_executor/layers/fused_moe/unquantized_fused_moe_method.py +578 -0
  655. vllm/model_executor/layers/fused_moe/utils.py +332 -0
  656. vllm/model_executor/layers/kda.py +448 -0
  657. vllm/model_executor/layers/layernorm.py +442 -0
  658. vllm/model_executor/layers/lightning_attn.py +729 -0
  659. vllm/model_executor/layers/linear.py +1424 -0
  660. vllm/model_executor/layers/logits_processor.py +106 -0
  661. vllm/model_executor/layers/mamba/__init__.py +0 -0
  662. vllm/model_executor/layers/mamba/abstract.py +71 -0
  663. vllm/model_executor/layers/mamba/linear_attn.py +402 -0
  664. vllm/model_executor/layers/mamba/mamba_mixer.py +535 -0
  665. vllm/model_executor/layers/mamba/mamba_mixer2.py +928 -0
  666. vllm/model_executor/layers/mamba/mamba_utils.py +225 -0
  667. vllm/model_executor/layers/mamba/ops/__init__.py +0 -0
  668. vllm/model_executor/layers/mamba/ops/causal_conv1d.py +1240 -0
  669. vllm/model_executor/layers/mamba/ops/layernorm_gated.py +172 -0
  670. vllm/model_executor/layers/mamba/ops/mamba_ssm.py +478 -0
  671. vllm/model_executor/layers/mamba/ops/ssd_bmm.py +211 -0
  672. vllm/model_executor/layers/mamba/ops/ssd_chunk_scan.py +456 -0
  673. vllm/model_executor/layers/mamba/ops/ssd_chunk_state.py +700 -0
  674. vllm/model_executor/layers/mamba/ops/ssd_combined.py +230 -0
  675. vllm/model_executor/layers/mamba/ops/ssd_state_passing.py +157 -0
  676. vllm/model_executor/layers/mamba/short_conv.py +264 -0
  677. vllm/model_executor/layers/mla.py +168 -0
  678. vllm/model_executor/layers/pooler.py +817 -0
  679. vllm/model_executor/layers/quantization/__init__.py +174 -0
  680. vllm/model_executor/layers/quantization/auto_round.py +454 -0
  681. vllm/model_executor/layers/quantization/awq.py +277 -0
  682. vllm/model_executor/layers/quantization/awq_marlin.py +659 -0
  683. vllm/model_executor/layers/quantization/awq_triton.py +337 -0
  684. vllm/model_executor/layers/quantization/base_config.py +170 -0
  685. vllm/model_executor/layers/quantization/bitblas.py +502 -0
  686. vllm/model_executor/layers/quantization/bitsandbytes.py +658 -0
  687. vllm/model_executor/layers/quantization/compressed_tensors/__init__.py +3 -0
  688. vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py +914 -0
  689. vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py +2284 -0
  690. vllm/model_executor/layers/quantization/compressed_tensors/schemes/__init__.py +35 -0
  691. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py +392 -0
  692. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py +55 -0
  693. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py +176 -0
  694. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_nvfp4.py +124 -0
  695. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4.py +218 -0
  696. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a8_fp8.py +183 -0
  697. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a8_int.py +153 -0
  698. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a16_fp8.py +138 -0
  699. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +200 -0
  700. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py +125 -0
  701. vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py +219 -0
  702. vllm/model_executor/layers/quantization/compressed_tensors/transform/__init__.py +0 -0
  703. vllm/model_executor/layers/quantization/compressed_tensors/transform/linear.py +260 -0
  704. vllm/model_executor/layers/quantization/compressed_tensors/transform/module.py +173 -0
  705. vllm/model_executor/layers/quantization/compressed_tensors/transform/schemes/__init__.py +0 -0
  706. vllm/model_executor/layers/quantization/compressed_tensors/transform/schemes/linear_qutlass_nvfp4.py +64 -0
  707. vllm/model_executor/layers/quantization/compressed_tensors/transform/utils.py +13 -0
  708. vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py +224 -0
  709. vllm/model_executor/layers/quantization/compressed_tensors/utils.py +216 -0
  710. vllm/model_executor/layers/quantization/deepspeedfp.py +218 -0
  711. vllm/model_executor/layers/quantization/experts_int8.py +240 -0
  712. vllm/model_executor/layers/quantization/fbgemm_fp8.py +195 -0
  713. vllm/model_executor/layers/quantization/fp8.py +1333 -0
  714. vllm/model_executor/layers/quantization/fp_quant.py +420 -0
  715. vllm/model_executor/layers/quantization/gguf.py +643 -0
  716. vllm/model_executor/layers/quantization/gptq.py +393 -0
  717. vllm/model_executor/layers/quantization/gptq_bitblas.py +482 -0
  718. vllm/model_executor/layers/quantization/gptq_marlin.py +789 -0
  719. vllm/model_executor/layers/quantization/gptq_marlin_24.py +320 -0
  720. vllm/model_executor/layers/quantization/hqq_marlin.py +371 -0
  721. vllm/model_executor/layers/quantization/inc.py +65 -0
  722. vllm/model_executor/layers/quantization/input_quant_fp8.py +171 -0
  723. vllm/model_executor/layers/quantization/ipex_quant.py +467 -0
  724. vllm/model_executor/layers/quantization/kernels/__init__.py +0 -0
  725. vllm/model_executor/layers/quantization/kernels/mixed_precision/MPLinearKernel.py +94 -0
  726. vllm/model_executor/layers/quantization/kernels/mixed_precision/__init__.py +105 -0
  727. vllm/model_executor/layers/quantization/kernels/mixed_precision/allspark.py +115 -0
  728. vllm/model_executor/layers/quantization/kernels/mixed_precision/bitblas.py +323 -0
  729. vllm/model_executor/layers/quantization/kernels/mixed_precision/conch.py +98 -0
  730. vllm/model_executor/layers/quantization/kernels/mixed_precision/cutlass.py +119 -0
  731. vllm/model_executor/layers/quantization/kernels/mixed_precision/dynamic_4bit.py +111 -0
  732. vllm/model_executor/layers/quantization/kernels/mixed_precision/exllama.py +161 -0
  733. vllm/model_executor/layers/quantization/kernels/mixed_precision/machete.py +159 -0
  734. vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py +166 -0
  735. vllm/model_executor/layers/quantization/kernels/scaled_mm/ScaledMMLinearKernel.py +73 -0
  736. vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py +97 -0
  737. vllm/model_executor/layers/quantization/kernels/scaled_mm/aiter.py +120 -0
  738. vllm/model_executor/layers/quantization/kernels/scaled_mm/cpu.py +219 -0
  739. vllm/model_executor/layers/quantization/kernels/scaled_mm/cutlass.py +140 -0
  740. vllm/model_executor/layers/quantization/kernels/scaled_mm/triton.py +42 -0
  741. vllm/model_executor/layers/quantization/kernels/scaled_mm/xla.py +105 -0
  742. vllm/model_executor/layers/quantization/kv_cache.py +146 -0
  743. vllm/model_executor/layers/quantization/modelopt.py +1788 -0
  744. vllm/model_executor/layers/quantization/moe_wna16.py +541 -0
  745. vllm/model_executor/layers/quantization/mxfp4.py +1162 -0
  746. vllm/model_executor/layers/quantization/petit.py +320 -0
  747. vllm/model_executor/layers/quantization/ptpc_fp8.py +137 -0
  748. vllm/model_executor/layers/quantization/quark/__init__.py +0 -0
  749. vllm/model_executor/layers/quantization/quark/quark.py +528 -0
  750. vllm/model_executor/layers/quantization/quark/quark_moe.py +683 -0
  751. vllm/model_executor/layers/quantization/quark/schemes/__init__.py +9 -0
  752. vllm/model_executor/layers/quantization/quark/schemes/quark_ocp_mx.py +306 -0
  753. vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py +55 -0
  754. vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py +179 -0
  755. vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py +139 -0
  756. vllm/model_executor/layers/quantization/quark/utils.py +105 -0
  757. vllm/model_executor/layers/quantization/qutlass_utils.py +185 -0
  758. vllm/model_executor/layers/quantization/rtn.py +652 -0
  759. vllm/model_executor/layers/quantization/schema.py +90 -0
  760. vllm/model_executor/layers/quantization/torchao.py +380 -0
  761. vllm/model_executor/layers/quantization/tpu_int8.py +139 -0
  762. vllm/model_executor/layers/quantization/utils/__init__.py +6 -0
  763. vllm/model_executor/layers/quantization/utils/allspark_utils.py +67 -0
  764. vllm/model_executor/layers/quantization/utils/bitblas_utils.py +229 -0
  765. vllm/model_executor/layers/quantization/utils/configs/N=12288,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  766. vllm/model_executor/layers/quantization/utils/configs/N=12288,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  767. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  768. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  769. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  770. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  771. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  772. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  773. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  774. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  775. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  776. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  777. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  778. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  779. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  780. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  781. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  782. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  783. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  784. vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  785. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  786. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  787. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  788. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  789. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  790. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  791. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  792. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  793. vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  794. vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  795. vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  796. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  797. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  798. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  799. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  800. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  801. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  802. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  803. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  804. vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  805. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  806. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  807. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  808. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  809. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  810. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  811. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  812. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  813. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  814. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  815. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  816. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  817. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  818. vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  819. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  820. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  821. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  822. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  823. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  824. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  825. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  826. vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  827. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  828. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  829. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  830. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  831. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  832. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  833. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  834. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  835. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  836. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  837. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  838. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  839. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  840. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  841. vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  842. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  843. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  844. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  845. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  846. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  847. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  848. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  849. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  850. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  851. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  852. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  853. vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  854. vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  855. vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  856. vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  857. vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  858. vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  859. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  860. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  861. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  862. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  863. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  864. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  865. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  866. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  867. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  868. vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  869. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  870. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  871. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  872. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  873. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  874. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  875. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  876. vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  877. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  878. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  879. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  880. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  881. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  882. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  883. vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  884. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  885. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  886. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  887. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  888. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  889. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  890. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  891. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  892. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  893. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  894. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +18 -0
  895. vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  896. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  897. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  898. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  899. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  900. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  901. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  902. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  903. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  904. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  905. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  906. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  907. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  908. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  909. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  910. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  911. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  912. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  913. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  914. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  915. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  916. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  917. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  918. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  919. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  920. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  921. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  922. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  923. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  924. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  925. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  926. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  927. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  928. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  929. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  930. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  931. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  932. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  933. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  934. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  935. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  936. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  937. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  938. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  939. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  940. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  941. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  942. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  943. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  944. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  945. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  946. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  947. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  948. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  949. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  950. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  951. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  952. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  953. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  954. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  955. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  956. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  957. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  958. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  959. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  960. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  961. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  962. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  963. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  964. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  965. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  966. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json +146 -0
  967. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json +146 -0
  968. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json +26 -0
  969. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  970. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  971. vllm/model_executor/layers/quantization/utils/configs/N=7168,K=8192,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  972. vllm/model_executor/layers/quantization/utils/configs/N=8192,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  973. vllm/model_executor/layers/quantization/utils/configs/N=8192,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  974. vllm/model_executor/layers/quantization/utils/configs/N=8192,K=1536,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json +164 -0
  975. vllm/model_executor/layers/quantization/utils/configs/README.md +3 -0
  976. vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py +89 -0
  977. vllm/model_executor/layers/quantization/utils/flashinfer_utils.py +298 -0
  978. vllm/model_executor/layers/quantization/utils/fp8_utils.py +1203 -0
  979. vllm/model_executor/layers/quantization/utils/gptq_utils.py +158 -0
  980. vllm/model_executor/layers/quantization/utils/int8_utils.py +489 -0
  981. vllm/model_executor/layers/quantization/utils/layer_utils.py +41 -0
  982. vllm/model_executor/layers/quantization/utils/machete_utils.py +56 -0
  983. vllm/model_executor/layers/quantization/utils/marlin_utils.py +575 -0
  984. vllm/model_executor/layers/quantization/utils/marlin_utils_fp4.py +397 -0
  985. vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py +351 -0
  986. vllm/model_executor/layers/quantization/utils/marlin_utils_test.py +161 -0
  987. vllm/model_executor/layers/quantization/utils/marlin_utils_test_24.py +467 -0
  988. vllm/model_executor/layers/quantization/utils/mxfp4_utils.py +181 -0
  989. vllm/model_executor/layers/quantization/utils/mxfp6_utils.py +142 -0
  990. vllm/model_executor/layers/quantization/utils/mxfp8_utils.py +24 -0
  991. vllm/model_executor/layers/quantization/utils/nvfp4_emulation_utils.py +142 -0
  992. vllm/model_executor/layers/quantization/utils/nvfp4_moe_support.py +63 -0
  993. vllm/model_executor/layers/quantization/utils/ocp_mx_utils.py +51 -0
  994. vllm/model_executor/layers/quantization/utils/petit_utils.py +124 -0
  995. vllm/model_executor/layers/quantization/utils/quant_utils.py +687 -0
  996. vllm/model_executor/layers/quantization/utils/w8a8_utils.py +516 -0
  997. vllm/model_executor/layers/resampler.py +283 -0
  998. vllm/model_executor/layers/rotary_embedding/__init__.py +278 -0
  999. vllm/model_executor/layers/rotary_embedding/base.py +235 -0
  1000. vllm/model_executor/layers/rotary_embedding/common.py +188 -0
  1001. vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py +165 -0
  1002. vllm/model_executor/layers/rotary_embedding/dual_chunk_rope.py +215 -0
  1003. vllm/model_executor/layers/rotary_embedding/dynamic_ntk_alpha_rope.py +43 -0
  1004. vllm/model_executor/layers/rotary_embedding/dynamic_ntk_scaling_rope.py +68 -0
  1005. vllm/model_executor/layers/rotary_embedding/ernie45_vl_rope.py +75 -0
  1006. vllm/model_executor/layers/rotary_embedding/linear_scaling_rope.py +115 -0
  1007. vllm/model_executor/layers/rotary_embedding/llama3_rope.py +54 -0
  1008. vllm/model_executor/layers/rotary_embedding/llama4_vision_rope.py +80 -0
  1009. vllm/model_executor/layers/rotary_embedding/mrope.py +397 -0
  1010. vllm/model_executor/layers/rotary_embedding/ntk_scaling_rope.py +47 -0
  1011. vllm/model_executor/layers/rotary_embedding/phi3_long_rope_scaled_rope.py +159 -0
  1012. vllm/model_executor/layers/rotary_embedding/yarn_scaling_rope.py +81 -0
  1013. vllm/model_executor/layers/utils.py +251 -0
  1014. vllm/model_executor/layers/vocab_parallel_embedding.py +558 -0
  1015. vllm/model_executor/model_loader/__init__.py +148 -0
  1016. vllm/model_executor/model_loader/base_loader.py +57 -0
  1017. vllm/model_executor/model_loader/bitsandbytes_loader.py +822 -0
  1018. vllm/model_executor/model_loader/default_loader.py +327 -0
  1019. vllm/model_executor/model_loader/dummy_loader.py +28 -0
  1020. vllm/model_executor/model_loader/gguf_loader.py +176 -0
  1021. vllm/model_executor/model_loader/online_quantization.py +224 -0
  1022. vllm/model_executor/model_loader/runai_streamer_loader.py +116 -0
  1023. vllm/model_executor/model_loader/sharded_state_loader.py +206 -0
  1024. vllm/model_executor/model_loader/tensorizer.py +790 -0
  1025. vllm/model_executor/model_loader/tensorizer_loader.py +151 -0
  1026. vllm/model_executor/model_loader/tpu.py +118 -0
  1027. vllm/model_executor/model_loader/utils.py +288 -0
  1028. vllm/model_executor/model_loader/weight_utils.py +1084 -0
  1029. vllm/model_executor/models/__init__.py +44 -0
  1030. vllm/model_executor/models/adapters.py +543 -0
  1031. vllm/model_executor/models/afmoe.py +711 -0
  1032. vllm/model_executor/models/aimv2.py +247 -0
  1033. vllm/model_executor/models/apertus.py +587 -0
  1034. vllm/model_executor/models/arcee.py +439 -0
  1035. vllm/model_executor/models/arctic.py +635 -0
  1036. vllm/model_executor/models/aria.py +655 -0
  1037. vllm/model_executor/models/aya_vision.py +450 -0
  1038. vllm/model_executor/models/baichuan.py +496 -0
  1039. vllm/model_executor/models/bailing_moe.py +646 -0
  1040. vllm/model_executor/models/bamba.py +522 -0
  1041. vllm/model_executor/models/bee.py +157 -0
  1042. vllm/model_executor/models/bert.py +925 -0
  1043. vllm/model_executor/models/bert_with_rope.py +732 -0
  1044. vllm/model_executor/models/blip.py +349 -0
  1045. vllm/model_executor/models/blip2.py +695 -0
  1046. vllm/model_executor/models/bloom.py +390 -0
  1047. vllm/model_executor/models/chameleon.py +1120 -0
  1048. vllm/model_executor/models/chatglm.py +498 -0
  1049. vllm/model_executor/models/clip.py +965 -0
  1050. vllm/model_executor/models/cohere2_vision.py +472 -0
  1051. vllm/model_executor/models/commandr.py +473 -0
  1052. vllm/model_executor/models/config.py +503 -0
  1053. vllm/model_executor/models/dbrx.py +482 -0
  1054. vllm/model_executor/models/deepencoder.py +673 -0
  1055. vllm/model_executor/models/deepseek_eagle.py +260 -0
  1056. vllm/model_executor/models/deepseek_mtp.py +360 -0
  1057. vllm/model_executor/models/deepseek_ocr.py +593 -0
  1058. vllm/model_executor/models/deepseek_v2.py +1649 -0
  1059. vllm/model_executor/models/deepseek_vl2.py +655 -0
  1060. vllm/model_executor/models/dots1.py +574 -0
  1061. vllm/model_executor/models/dots_ocr.py +900 -0
  1062. vllm/model_executor/models/ernie45.py +53 -0
  1063. vllm/model_executor/models/ernie45_moe.py +759 -0
  1064. vllm/model_executor/models/ernie45_vl.py +1742 -0
  1065. vllm/model_executor/models/ernie45_vl_moe.py +803 -0
  1066. vllm/model_executor/models/ernie_mtp.py +279 -0
  1067. vllm/model_executor/models/exaone.py +545 -0
  1068. vllm/model_executor/models/exaone4.py +531 -0
  1069. vllm/model_executor/models/fairseq2_llama.py +154 -0
  1070. vllm/model_executor/models/falcon.py +545 -0
  1071. vllm/model_executor/models/falcon_h1.py +685 -0
  1072. vllm/model_executor/models/flex_olmo.py +155 -0
  1073. vllm/model_executor/models/fuyu.py +373 -0
  1074. vllm/model_executor/models/gemma.py +426 -0
  1075. vllm/model_executor/models/gemma2.py +439 -0
  1076. vllm/model_executor/models/gemma3.py +571 -0
  1077. vllm/model_executor/models/gemma3_mm.py +741 -0
  1078. vllm/model_executor/models/gemma3n.py +1165 -0
  1079. vllm/model_executor/models/gemma3n_mm.py +811 -0
  1080. vllm/model_executor/models/glm.py +23 -0
  1081. vllm/model_executor/models/glm4.py +305 -0
  1082. vllm/model_executor/models/glm4_1v.py +1821 -0
  1083. vllm/model_executor/models/glm4_moe.py +747 -0
  1084. vllm/model_executor/models/glm4_moe_mtp.py +359 -0
  1085. vllm/model_executor/models/glm4v.py +784 -0
  1086. vllm/model_executor/models/gpt2.py +397 -0
  1087. vllm/model_executor/models/gpt_bigcode.py +339 -0
  1088. vllm/model_executor/models/gpt_j.py +346 -0
  1089. vllm/model_executor/models/gpt_neox.py +344 -0
  1090. vllm/model_executor/models/gpt_oss.py +738 -0
  1091. vllm/model_executor/models/granite.py +516 -0
  1092. vllm/model_executor/models/granite_speech.py +913 -0
  1093. vllm/model_executor/models/granitemoe.py +569 -0
  1094. vllm/model_executor/models/granitemoehybrid.py +709 -0
  1095. vllm/model_executor/models/granitemoeshared.py +333 -0
  1096. vllm/model_executor/models/gritlm.py +245 -0
  1097. vllm/model_executor/models/grok1.py +558 -0
  1098. vllm/model_executor/models/h2ovl.py +554 -0
  1099. vllm/model_executor/models/hunyuan_v1.py +1053 -0
  1100. vllm/model_executor/models/hyperclovax_vision.py +1166 -0
  1101. vllm/model_executor/models/idefics2_vision_model.py +426 -0
  1102. vllm/model_executor/models/idefics3.py +717 -0
  1103. vllm/model_executor/models/interfaces.py +1092 -0
  1104. vllm/model_executor/models/interfaces_base.py +214 -0
  1105. vllm/model_executor/models/intern_vit.py +453 -0
  1106. vllm/model_executor/models/internlm2.py +460 -0
  1107. vllm/model_executor/models/internlm2_ve.py +142 -0
  1108. vllm/model_executor/models/interns1.py +830 -0
  1109. vllm/model_executor/models/interns1_vit.py +432 -0
  1110. vllm/model_executor/models/internvl.py +1452 -0
  1111. vllm/model_executor/models/jais.py +397 -0
  1112. vllm/model_executor/models/jamba.py +610 -0
  1113. vllm/model_executor/models/jina_vl.py +147 -0
  1114. vllm/model_executor/models/keye.py +1761 -0
  1115. vllm/model_executor/models/keye_vl1_5.py +726 -0
  1116. vllm/model_executor/models/kimi_linear.py +663 -0
  1117. vllm/model_executor/models/kimi_vl.py +578 -0
  1118. vllm/model_executor/models/lfm2.py +532 -0
  1119. vllm/model_executor/models/lfm2_moe.py +762 -0
  1120. vllm/model_executor/models/lightonocr.py +195 -0
  1121. vllm/model_executor/models/llama.py +732 -0
  1122. vllm/model_executor/models/llama4.py +859 -0
  1123. vllm/model_executor/models/llama4_eagle.py +223 -0
  1124. vllm/model_executor/models/llama_eagle.py +218 -0
  1125. vllm/model_executor/models/llama_eagle3.py +367 -0
  1126. vllm/model_executor/models/llava.py +842 -0
  1127. vllm/model_executor/models/llava_next.py +583 -0
  1128. vllm/model_executor/models/llava_next_video.py +467 -0
  1129. vllm/model_executor/models/llava_onevision.py +923 -0
  1130. vllm/model_executor/models/longcat_flash.py +749 -0
  1131. vllm/model_executor/models/longcat_flash_mtp.py +349 -0
  1132. vllm/model_executor/models/mamba.py +276 -0
  1133. vllm/model_executor/models/mamba2.py +289 -0
  1134. vllm/model_executor/models/medusa.py +179 -0
  1135. vllm/model_executor/models/midashenglm.py +827 -0
  1136. vllm/model_executor/models/mimo.py +188 -0
  1137. vllm/model_executor/models/mimo_mtp.py +294 -0
  1138. vllm/model_executor/models/minicpm.py +664 -0
  1139. vllm/model_executor/models/minicpm3.py +242 -0
  1140. vllm/model_executor/models/minicpm_eagle.py +389 -0
  1141. vllm/model_executor/models/minicpmo.py +768 -0
  1142. vllm/model_executor/models/minicpmv.py +1745 -0
  1143. vllm/model_executor/models/minimax_m2.py +552 -0
  1144. vllm/model_executor/models/minimax_text_01.py +1012 -0
  1145. vllm/model_executor/models/minimax_vl_01.py +396 -0
  1146. vllm/model_executor/models/mistral3.py +637 -0
  1147. vllm/model_executor/models/mixtral.py +621 -0
  1148. vllm/model_executor/models/mllama4.py +1147 -0
  1149. vllm/model_executor/models/mlp_speculator.py +235 -0
  1150. vllm/model_executor/models/modernbert.py +450 -0
  1151. vllm/model_executor/models/module_mapping.py +74 -0
  1152. vllm/model_executor/models/molmo.py +1555 -0
  1153. vllm/model_executor/models/moonvit.py +677 -0
  1154. vllm/model_executor/models/mpt.py +335 -0
  1155. vllm/model_executor/models/nano_nemotron_vl.py +1740 -0
  1156. vllm/model_executor/models/nemotron.py +518 -0
  1157. vllm/model_executor/models/nemotron_h.py +852 -0
  1158. vllm/model_executor/models/nemotron_nas.py +491 -0
  1159. vllm/model_executor/models/nemotron_vl.py +653 -0
  1160. vllm/model_executor/models/nvlm_d.py +216 -0
  1161. vllm/model_executor/models/olmo.py +414 -0
  1162. vllm/model_executor/models/olmo2.py +454 -0
  1163. vllm/model_executor/models/olmoe.py +498 -0
  1164. vllm/model_executor/models/openpangu.py +1062 -0
  1165. vllm/model_executor/models/openpangu_mtp.py +265 -0
  1166. vllm/model_executor/models/opt.py +426 -0
  1167. vllm/model_executor/models/orion.py +372 -0
  1168. vllm/model_executor/models/ouro.py +516 -0
  1169. vllm/model_executor/models/ovis.py +559 -0
  1170. vllm/model_executor/models/ovis2_5.py +673 -0
  1171. vllm/model_executor/models/paddleocr_vl.py +1407 -0
  1172. vllm/model_executor/models/paligemma.py +412 -0
  1173. vllm/model_executor/models/persimmon.py +377 -0
  1174. vllm/model_executor/models/phi.py +374 -0
  1175. vllm/model_executor/models/phi3.py +18 -0
  1176. vllm/model_executor/models/phi3v.py +737 -0
  1177. vllm/model_executor/models/phi4_multimodal.py +1447 -0
  1178. vllm/model_executor/models/phi4mm.py +1253 -0
  1179. vllm/model_executor/models/phi4mm_audio.py +1296 -0
  1180. vllm/model_executor/models/phi4mm_utils.py +1907 -0
  1181. vllm/model_executor/models/phimoe.py +675 -0
  1182. vllm/model_executor/models/pixtral.py +1352 -0
  1183. vllm/model_executor/models/plamo2.py +981 -0
  1184. vllm/model_executor/models/qwen.py +368 -0
  1185. vllm/model_executor/models/qwen2.py +541 -0
  1186. vllm/model_executor/models/qwen2_5_omni_thinker.py +1246 -0
  1187. vllm/model_executor/models/qwen2_5_vl.py +1613 -0
  1188. vllm/model_executor/models/qwen2_audio.py +473 -0
  1189. vllm/model_executor/models/qwen2_moe.py +596 -0
  1190. vllm/model_executor/models/qwen2_rm.py +123 -0
  1191. vllm/model_executor/models/qwen2_vl.py +1670 -0
  1192. vllm/model_executor/models/qwen3.py +336 -0
  1193. vllm/model_executor/models/qwen3_moe.py +744 -0
  1194. vllm/model_executor/models/qwen3_next.py +1395 -0
  1195. vllm/model_executor/models/qwen3_next_mtp.py +296 -0
  1196. vllm/model_executor/models/qwen3_omni_moe_thinker.py +1721 -0
  1197. vllm/model_executor/models/qwen3_vl.py +1673 -0
  1198. vllm/model_executor/models/qwen3_vl_moe.py +415 -0
  1199. vllm/model_executor/models/qwen_vl.py +802 -0
  1200. vllm/model_executor/models/radio.py +555 -0
  1201. vllm/model_executor/models/registry.py +1155 -0
  1202. vllm/model_executor/models/roberta.py +259 -0
  1203. vllm/model_executor/models/rvl.py +107 -0
  1204. vllm/model_executor/models/seed_oss.py +497 -0
  1205. vllm/model_executor/models/siglip.py +1174 -0
  1206. vllm/model_executor/models/siglip2navit.py +724 -0
  1207. vllm/model_executor/models/skyworkr1v.py +953 -0
  1208. vllm/model_executor/models/smolvlm.py +38 -0
  1209. vllm/model_executor/models/solar.py +502 -0
  1210. vllm/model_executor/models/stablelm.py +359 -0
  1211. vllm/model_executor/models/starcoder2.py +367 -0
  1212. vllm/model_executor/models/step3_text.py +559 -0
  1213. vllm/model_executor/models/step3_vl.py +1148 -0
  1214. vllm/model_executor/models/swin.py +514 -0
  1215. vllm/model_executor/models/tarsier.py +619 -0
  1216. vllm/model_executor/models/telechat2.py +153 -0
  1217. vllm/model_executor/models/teleflm.py +78 -0
  1218. vllm/model_executor/models/terratorch.py +319 -0
  1219. vllm/model_executor/models/transformers/__init__.py +127 -0
  1220. vllm/model_executor/models/transformers/base.py +464 -0
  1221. vllm/model_executor/models/transformers/causal.py +65 -0
  1222. vllm/model_executor/models/transformers/legacy.py +90 -0
  1223. vllm/model_executor/models/transformers/moe.py +318 -0
  1224. vllm/model_executor/models/transformers/multimodal.py +411 -0
  1225. vllm/model_executor/models/transformers/pooling.py +119 -0
  1226. vllm/model_executor/models/transformers/utils.py +207 -0
  1227. vllm/model_executor/models/ultravox.py +681 -0
  1228. vllm/model_executor/models/utils.py +877 -0
  1229. vllm/model_executor/models/vision.py +552 -0
  1230. vllm/model_executor/models/voxtral.py +845 -0
  1231. vllm/model_executor/models/whisper.py +959 -0
  1232. vllm/model_executor/models/zamba2.py +986 -0
  1233. vllm/model_executor/parameter.py +642 -0
  1234. vllm/model_executor/utils.py +94 -0
  1235. vllm/model_executor/warmup/__init__.py +0 -0
  1236. vllm/model_executor/warmup/deep_gemm_warmup.py +314 -0
  1237. vllm/model_executor/warmup/kernel_warmup.py +98 -0
  1238. vllm/multimodal/__init__.py +40 -0
  1239. vllm/multimodal/audio.py +118 -0
  1240. vllm/multimodal/base.py +26 -0
  1241. vllm/multimodal/cache.py +755 -0
  1242. vllm/multimodal/evs.py +294 -0
  1243. vllm/multimodal/hasher.py +106 -0
  1244. vllm/multimodal/image.py +130 -0
  1245. vllm/multimodal/inputs.py +1036 -0
  1246. vllm/multimodal/parse.py +544 -0
  1247. vllm/multimodal/processing.py +2186 -0
  1248. vllm/multimodal/profiling.py +369 -0
  1249. vllm/multimodal/registry.py +360 -0
  1250. vllm/multimodal/utils.py +512 -0
  1251. vllm/multimodal/video.py +306 -0
  1252. vllm/outputs.py +345 -0
  1253. vllm/platforms/__init__.py +277 -0
  1254. vllm/platforms/cpu.py +414 -0
  1255. vllm/platforms/cuda.py +657 -0
  1256. vllm/platforms/interface.py +639 -0
  1257. vllm/platforms/rocm.py +466 -0
  1258. vllm/platforms/tpu.py +276 -0
  1259. vllm/platforms/xpu.py +274 -0
  1260. vllm/plugins/__init__.py +78 -0
  1261. vllm/plugins/io_processors/__init__.py +68 -0
  1262. vllm/plugins/io_processors/interface.py +77 -0
  1263. vllm/plugins/lora_resolvers/__init__.py +0 -0
  1264. vllm/plugins/lora_resolvers/filesystem_resolver.py +52 -0
  1265. vllm/pooling_params.py +228 -0
  1266. vllm/profiler/__init__.py +0 -0
  1267. vllm/profiler/gpu_profiler.py +37 -0
  1268. vllm/profiler/layerwise_profile.py +392 -0
  1269. vllm/profiler/utils.py +151 -0
  1270. vllm/py.typed +2 -0
  1271. vllm/ray/__init__.py +0 -0
  1272. vllm/ray/lazy_utils.py +26 -0
  1273. vllm/ray/ray_env.py +79 -0
  1274. vllm/reasoning/__init__.py +92 -0
  1275. vllm/reasoning/abs_reasoning_parsers.py +290 -0
  1276. vllm/reasoning/basic_parsers.py +162 -0
  1277. vllm/reasoning/deepseek_r1_reasoning_parser.py +67 -0
  1278. vllm/reasoning/deepseek_v3_reasoning_parser.py +62 -0
  1279. vllm/reasoning/ernie45_reasoning_parser.py +165 -0
  1280. vllm/reasoning/glm4_moe_reasoning_parser.py +171 -0
  1281. vllm/reasoning/gptoss_reasoning_parser.py +173 -0
  1282. vllm/reasoning/granite_reasoning_parser.py +363 -0
  1283. vllm/reasoning/hunyuan_a13b_reasoning_parser.py +237 -0
  1284. vllm/reasoning/identity_reasoning_parser.py +58 -0
  1285. vllm/reasoning/minimax_m2_reasoning_parser.py +67 -0
  1286. vllm/reasoning/mistral_reasoning_parser.py +55 -0
  1287. vllm/reasoning/olmo3_reasoning_parser.py +302 -0
  1288. vllm/reasoning/qwen3_reasoning_parser.py +67 -0
  1289. vllm/reasoning/seedoss_reasoning_parser.py +27 -0
  1290. vllm/reasoning/step3_reasoning_parser.py +107 -0
  1291. vllm/sampling_params.py +669 -0
  1292. vllm/scalar_type.py +355 -0
  1293. vllm/scripts.py +17 -0
  1294. vllm/sequence.py +98 -0
  1295. vllm/tasks.py +13 -0
  1296. vllm/third_party/__init__.py +0 -0
  1297. vllm/third_party/pynvml.py +6140 -0
  1298. vllm/tracing.py +135 -0
  1299. vllm/transformers_utils/__init__.py +26 -0
  1300. vllm/transformers_utils/chat_templates/__init__.py +5 -0
  1301. vllm/transformers_utils/chat_templates/registry.py +73 -0
  1302. vllm/transformers_utils/chat_templates/template_basic.jinja +3 -0
  1303. vllm/transformers_utils/chat_templates/template_blip2.jinja +11 -0
  1304. vllm/transformers_utils/chat_templates/template_chatml.jinja +10 -0
  1305. vllm/transformers_utils/chat_templates/template_deepseek_ocr.jinja +14 -0
  1306. vllm/transformers_utils/chat_templates/template_deepseek_vl2.jinja +23 -0
  1307. vllm/transformers_utils/chat_templates/template_fuyu.jinja +3 -0
  1308. vllm/transformers_utils/chat_templates/template_minicpmv45.jinja +93 -0
  1309. vllm/transformers_utils/config.py +1203 -0
  1310. vllm/transformers_utils/config_parser_base.py +20 -0
  1311. vllm/transformers_utils/configs/__init__.py +70 -0
  1312. vllm/transformers_utils/configs/afmoe.py +84 -0
  1313. vllm/transformers_utils/configs/arctic.py +206 -0
  1314. vllm/transformers_utils/configs/chatglm.py +75 -0
  1315. vllm/transformers_utils/configs/deepseek_vl2.py +126 -0
  1316. vllm/transformers_utils/configs/dotsocr.py +71 -0
  1317. vllm/transformers_utils/configs/eagle.py +84 -0
  1318. vllm/transformers_utils/configs/falcon.py +89 -0
  1319. vllm/transformers_utils/configs/flex_olmo.py +77 -0
  1320. vllm/transformers_utils/configs/jais.py +243 -0
  1321. vllm/transformers_utils/configs/kimi_linear.py +144 -0
  1322. vllm/transformers_utils/configs/kimi_vl.py +38 -0
  1323. vllm/transformers_utils/configs/lfm2_moe.py +159 -0
  1324. vllm/transformers_utils/configs/medusa.py +65 -0
  1325. vllm/transformers_utils/configs/midashenglm.py +103 -0
  1326. vllm/transformers_utils/configs/mistral.py +174 -0
  1327. vllm/transformers_utils/configs/mlp_speculator.py +69 -0
  1328. vllm/transformers_utils/configs/moonvit.py +33 -0
  1329. vllm/transformers_utils/configs/nemotron.py +212 -0
  1330. vllm/transformers_utils/configs/nemotron_h.py +282 -0
  1331. vllm/transformers_utils/configs/olmo3.py +79 -0
  1332. vllm/transformers_utils/configs/ovis.py +182 -0
  1333. vllm/transformers_utils/configs/qwen3_next.py +274 -0
  1334. vllm/transformers_utils/configs/radio.py +89 -0
  1335. vllm/transformers_utils/configs/speculators/__init__.py +2 -0
  1336. vllm/transformers_utils/configs/speculators/algos.py +38 -0
  1337. vllm/transformers_utils/configs/speculators/base.py +114 -0
  1338. vllm/transformers_utils/configs/step3_vl.py +174 -0
  1339. vllm/transformers_utils/configs/ultravox.py +118 -0
  1340. vllm/transformers_utils/detokenizer_utils.py +198 -0
  1341. vllm/transformers_utils/dynamic_module.py +59 -0
  1342. vllm/transformers_utils/processor.py +402 -0
  1343. vllm/transformers_utils/processors/__init__.py +15 -0
  1344. vllm/transformers_utils/processors/deepseek_ocr.py +438 -0
  1345. vllm/transformers_utils/processors/deepseek_vl2.py +406 -0
  1346. vllm/transformers_utils/processors/ovis.py +453 -0
  1347. vllm/transformers_utils/processors/ovis2_5.py +468 -0
  1348. vllm/transformers_utils/runai_utils.py +104 -0
  1349. vllm/transformers_utils/s3_utils.py +95 -0
  1350. vllm/transformers_utils/tokenizer.py +293 -0
  1351. vllm/transformers_utils/tokenizer_base.py +155 -0
  1352. vllm/transformers_utils/tokenizers/__init__.py +16 -0
  1353. vllm/transformers_utils/tokenizers/mistral.py +502 -0
  1354. vllm/transformers_utils/utils.py +130 -0
  1355. vllm/triton_utils/__init__.py +19 -0
  1356. vllm/triton_utils/importing.py +103 -0
  1357. vllm/usage/__init__.py +0 -0
  1358. vllm/usage/usage_lib.py +294 -0
  1359. vllm/utils/__init__.py +82 -0
  1360. vllm/utils/argparse_utils.py +487 -0
  1361. vllm/utils/async_utils.py +303 -0
  1362. vllm/utils/cache.py +214 -0
  1363. vllm/utils/collection_utils.py +139 -0
  1364. vllm/utils/counter.py +45 -0
  1365. vllm/utils/deep_gemm.py +391 -0
  1366. vllm/utils/flashinfer.py +490 -0
  1367. vllm/utils/func_utils.py +236 -0
  1368. vllm/utils/gc_utils.py +147 -0
  1369. vllm/utils/hashing.py +63 -0
  1370. vllm/utils/import_utils.py +411 -0
  1371. vllm/utils/jsontree.py +165 -0
  1372. vllm/utils/math_utils.py +32 -0
  1373. vllm/utils/mem_constants.py +13 -0
  1374. vllm/utils/mem_utils.py +232 -0
  1375. vllm/utils/nccl.py +64 -0
  1376. vllm/utils/network_utils.py +331 -0
  1377. vllm/utils/platform_utils.py +59 -0
  1378. vllm/utils/profiling.py +56 -0
  1379. vllm/utils/registry.py +49 -0
  1380. vllm/utils/serial_utils.py +169 -0
  1381. vllm/utils/system_utils.py +229 -0
  1382. vllm/utils/tensor_schema.py +255 -0
  1383. vllm/utils/torch_utils.py +657 -0
  1384. vllm/v1/__init__.py +0 -0
  1385. vllm/v1/attention/__init__.py +0 -0
  1386. vllm/v1/attention/backends/__init__.py +0 -0
  1387. vllm/v1/attention/backends/cpu_attn.py +496 -0
  1388. vllm/v1/attention/backends/flash_attn.py +1028 -0
  1389. vllm/v1/attention/backends/flashinfer.py +1572 -0
  1390. vllm/v1/attention/backends/flex_attention.py +926 -0
  1391. vllm/v1/attention/backends/gdn_attn.py +387 -0
  1392. vllm/v1/attention/backends/linear_attn.py +74 -0
  1393. vllm/v1/attention/backends/mamba1_attn.py +165 -0
  1394. vllm/v1/attention/backends/mamba2_attn.py +354 -0
  1395. vllm/v1/attention/backends/mamba_attn.py +115 -0
  1396. vllm/v1/attention/backends/mla/__init__.py +0 -0
  1397. vllm/v1/attention/backends/mla/common.py +2031 -0
  1398. vllm/v1/attention/backends/mla/cutlass_mla.py +275 -0
  1399. vllm/v1/attention/backends/mla/flashattn_mla.py +337 -0
  1400. vllm/v1/attention/backends/mla/flashinfer_mla.py +171 -0
  1401. vllm/v1/attention/backends/mla/flashmla.py +314 -0
  1402. vllm/v1/attention/backends/mla/flashmla_sparse.py +548 -0
  1403. vllm/v1/attention/backends/mla/indexer.py +362 -0
  1404. vllm/v1/attention/backends/mla/rocm_aiter_mla.py +294 -0
  1405. vllm/v1/attention/backends/mla/triton_mla.py +171 -0
  1406. vllm/v1/attention/backends/pallas.py +436 -0
  1407. vllm/v1/attention/backends/rocm_aiter_fa.py +816 -0
  1408. vllm/v1/attention/backends/rocm_aiter_unified_attn.py +196 -0
  1409. vllm/v1/attention/backends/rocm_attn.py +362 -0
  1410. vllm/v1/attention/backends/short_conv_attn.py +105 -0
  1411. vllm/v1/attention/backends/tree_attn.py +425 -0
  1412. vllm/v1/attention/backends/triton_attn.py +373 -0
  1413. vllm/v1/attention/backends/utils.py +1116 -0
  1414. vllm/v1/attention/backends/xformers.py +417 -0
  1415. vllm/v1/core/__init__.py +0 -0
  1416. vllm/v1/core/block_pool.py +428 -0
  1417. vllm/v1/core/encoder_cache_manager.py +343 -0
  1418. vllm/v1/core/kv_cache_coordinator.py +480 -0
  1419. vllm/v1/core/kv_cache_manager.py +420 -0
  1420. vllm/v1/core/kv_cache_utils.py +1340 -0
  1421. vllm/v1/core/sched/__init__.py +0 -0
  1422. vllm/v1/core/sched/async_scheduler.py +62 -0
  1423. vllm/v1/core/sched/interface.py +181 -0
  1424. vllm/v1/core/sched/output.py +202 -0
  1425. vllm/v1/core/sched/request_queue.py +221 -0
  1426. vllm/v1/core/sched/scheduler.py +1617 -0
  1427. vllm/v1/core/sched/utils.py +72 -0
  1428. vllm/v1/core/single_type_kv_cache_manager.py +736 -0
  1429. vllm/v1/cudagraph_dispatcher.py +148 -0
  1430. vllm/v1/engine/__init__.py +206 -0
  1431. vllm/v1/engine/async_llm.py +797 -0
  1432. vllm/v1/engine/coordinator.py +377 -0
  1433. vllm/v1/engine/core.py +1420 -0
  1434. vllm/v1/engine/core_client.py +1400 -0
  1435. vllm/v1/engine/detokenizer.py +351 -0
  1436. vllm/v1/engine/exceptions.py +18 -0
  1437. vllm/v1/engine/llm_engine.py +408 -0
  1438. vllm/v1/engine/logprobs.py +182 -0
  1439. vllm/v1/engine/output_processor.py +642 -0
  1440. vllm/v1/engine/parallel_sampling.py +145 -0
  1441. vllm/v1/engine/processor.py +621 -0
  1442. vllm/v1/engine/utils.py +1072 -0
  1443. vllm/v1/executor/__init__.py +6 -0
  1444. vllm/v1/executor/abstract.py +352 -0
  1445. vllm/v1/executor/multiproc_executor.py +877 -0
  1446. vllm/v1/executor/ray_distributed_executor.py +8 -0
  1447. vllm/v1/executor/ray_executor.py +626 -0
  1448. vllm/v1/executor/ray_utils.py +465 -0
  1449. vllm/v1/executor/uniproc_executor.py +183 -0
  1450. vllm/v1/kv_cache_interface.py +403 -0
  1451. vllm/v1/kv_offload/__init__.py +0 -0
  1452. vllm/v1/kv_offload/abstract.py +161 -0
  1453. vllm/v1/kv_offload/arc_manager.py +237 -0
  1454. vllm/v1/kv_offload/backend.py +97 -0
  1455. vllm/v1/kv_offload/backends/__init__.py +0 -0
  1456. vllm/v1/kv_offload/backends/cpu.py +62 -0
  1457. vllm/v1/kv_offload/cpu.py +93 -0
  1458. vllm/v1/kv_offload/factory.py +56 -0
  1459. vllm/v1/kv_offload/lru_manager.py +139 -0
  1460. vllm/v1/kv_offload/mediums.py +39 -0
  1461. vllm/v1/kv_offload/spec.py +62 -0
  1462. vllm/v1/kv_offload/worker/__init__.py +0 -0
  1463. vllm/v1/kv_offload/worker/cpu_gpu.py +185 -0
  1464. vllm/v1/kv_offload/worker/worker.py +144 -0
  1465. vllm/v1/metrics/__init__.py +0 -0
  1466. vllm/v1/metrics/loggers.py +1238 -0
  1467. vllm/v1/metrics/prometheus.py +82 -0
  1468. vllm/v1/metrics/ray_wrappers.py +169 -0
  1469. vllm/v1/metrics/reader.py +257 -0
  1470. vllm/v1/metrics/stats.py +420 -0
  1471. vllm/v1/outputs.py +249 -0
  1472. vllm/v1/pool/__init__.py +0 -0
  1473. vllm/v1/pool/metadata.py +82 -0
  1474. vllm/v1/request.py +259 -0
  1475. vllm/v1/sample/__init__.py +0 -0
  1476. vllm/v1/sample/logits_processor/__init__.py +352 -0
  1477. vllm/v1/sample/logits_processor/builtin.py +274 -0
  1478. vllm/v1/sample/logits_processor/interface.py +106 -0
  1479. vllm/v1/sample/logits_processor/state.py +165 -0
  1480. vllm/v1/sample/metadata.py +44 -0
  1481. vllm/v1/sample/ops/__init__.py +0 -0
  1482. vllm/v1/sample/ops/bad_words.py +52 -0
  1483. vllm/v1/sample/ops/logprobs.py +25 -0
  1484. vllm/v1/sample/ops/penalties.py +57 -0
  1485. vllm/v1/sample/ops/topk_topp_sampler.py +290 -0
  1486. vllm/v1/sample/rejection_sampler.py +793 -0
  1487. vllm/v1/sample/sampler.py +316 -0
  1488. vllm/v1/sample/tpu/__init__.py +0 -0
  1489. vllm/v1/sample/tpu/metadata.py +120 -0
  1490. vllm/v1/sample/tpu/sampler.py +215 -0
  1491. vllm/v1/serial_utils.py +532 -0
  1492. vllm/v1/spec_decode/__init__.py +0 -0
  1493. vllm/v1/spec_decode/eagle.py +1225 -0
  1494. vllm/v1/spec_decode/medusa.py +73 -0
  1495. vllm/v1/spec_decode/metadata.py +66 -0
  1496. vllm/v1/spec_decode/metrics.py +224 -0
  1497. vllm/v1/spec_decode/ngram_proposer.py +291 -0
  1498. vllm/v1/spec_decode/suffix_decoding.py +103 -0
  1499. vllm/v1/spec_decode/utils.py +16 -0
  1500. vllm/v1/structured_output/__init__.py +338 -0
  1501. vllm/v1/structured_output/backend_guidance.py +265 -0
  1502. vllm/v1/structured_output/backend_lm_format_enforcer.py +177 -0
  1503. vllm/v1/structured_output/backend_outlines.py +324 -0
  1504. vllm/v1/structured_output/backend_types.py +136 -0
  1505. vllm/v1/structured_output/backend_xgrammar.py +362 -0
  1506. vllm/v1/structured_output/request.py +94 -0
  1507. vllm/v1/structured_output/utils.py +469 -0
  1508. vllm/v1/utils.py +414 -0
  1509. vllm/v1/worker/__init__.py +0 -0
  1510. vllm/v1/worker/block_table.py +327 -0
  1511. vllm/v1/worker/cpu_model_runner.py +122 -0
  1512. vllm/v1/worker/cpu_worker.py +206 -0
  1513. vllm/v1/worker/dp_utils.py +230 -0
  1514. vllm/v1/worker/ec_connector_model_runner_mixin.py +87 -0
  1515. vllm/v1/worker/gpu_input_batch.py +975 -0
  1516. vllm/v1/worker/gpu_model_runner.py +5102 -0
  1517. vllm/v1/worker/gpu_ubatch_wrapper.py +466 -0
  1518. vllm/v1/worker/gpu_worker.py +894 -0
  1519. vllm/v1/worker/kv_connector_model_runner_mixin.py +144 -0
  1520. vllm/v1/worker/lora_model_runner_mixin.py +213 -0
  1521. vllm/v1/worker/tpu_input_batch.py +593 -0
  1522. vllm/v1/worker/tpu_model_runner.py +2173 -0
  1523. vllm/v1/worker/tpu_worker.py +355 -0
  1524. vllm/v1/worker/ubatch_utils.py +73 -0
  1525. vllm/v1/worker/ubatching.py +231 -0
  1526. vllm/v1/worker/utils.py +366 -0
  1527. vllm/v1/worker/worker_base.py +375 -0
  1528. vllm/v1/worker/xpu_model_runner.py +55 -0
  1529. vllm/v1/worker/xpu_worker.py +189 -0
  1530. vllm/version.py +39 -0
  1531. vllm/vllm_flash_attn/.gitkeep +0 -0
  1532. vllm_cpu_amxbf16-0.11.2.post2.dist-info/METADATA +345 -0
  1533. vllm_cpu_amxbf16-0.11.2.post2.dist-info/RECORD +1536 -0
  1534. vllm_cpu_amxbf16-0.11.2.post2.dist-info/WHEEL +5 -0
  1535. vllm_cpu_amxbf16-0.11.2.post2.dist-info/entry_points.txt +5 -0
  1536. vllm_cpu_amxbf16-0.11.2.post2.dist-info/top_level.txt +1 -0
@@ -0,0 +1,1536 @@
1
+ vllm/_C.abi3.so,sha256=l88KN9UUOX9dUFSQ6N_lkqqe-CxBLPGEjkW7M6Ey_Ro,79481696
2
+ vllm/__init__.py,sha256=xSE2I9WhdAKvVQw2EPoRoyah9Z3ETDTul7kg-YAF6eQ,8548
3
+ vllm/_aiter_ops.py,sha256=QD3hNuAJcyBXLBl0R-4c_TVp-sttr_EjXVsye08O6wQ,29538
4
+ vllm/_bc_linter.py,sha256=2F7ZE_cba80XNDfbrZKVBunazE7ZDxVMtpOUdEFdH3w,1105
5
+ vllm/_custom_ops.py,sha256=8M4JwuOxvSXbRSFWRRACIhuYWPhANA0uQKszwmakvMQ,85072
6
+ vllm/_ipex_ops.py,sha256=L8DTX14Xfc7WFfckWdCM5X6B3oMJSJ63UrC7Fj9g50Q,14935
7
+ vllm/_version.py,sha256=QITrAxOn_9noDdp43khkCNh930uwdmWt4uaZBbaaSXM,721
8
+ vllm/beam_search.py,sha256=XUQRXltedVFtJRp43dKHKn87VqKeG5tr494D0yHemHM,2593
9
+ vllm/collect_env.py,sha256=7peWZtL6DVINkwdvI59rI7mHs0WfT3WEvlbsSjzeNow,28050
10
+ vllm/connections.py,sha256=LpQUe9gGRiBCIGY6RWUGi_7d8ql6tO2jYNecU1EJR6c,5350
11
+ vllm/env_override.py,sha256=A_B0QvbQffPFle5vBbS8aTaXCsRWRgWxOC4h5WKqaVI,13796
12
+ vllm/envs.py,sha256=sCKanWRijRtPpAboEzMlMQfwGyWiyLjEEYy1BX18YWI,75831
13
+ vllm/forward_context.py,sha256=NKIuaTWhy9-Gy6HZY7Tqi1Jpk81SNDSHSxN7G2-4t1Q,13220
14
+ vllm/logger.py,sha256=GeKwrLgMEZwUYedxG0XENsBQkTO-bXyA2Dihq4-S7Zw,9052
15
+ vllm/logits_process.py,sha256=nslFzzBxPogpW7_FKQx7jibxLdgh9TS7XKz7JYtat1E,4356
16
+ vllm/logprobs.py,sha256=JX3la0iot1HG_rnfd_QsP7C1AE-vuAFcD7Eh8pPeJaY,7915
17
+ vllm/outputs.py,sha256=qhKihxRd0aVm8rL8DvViVzBiuoFrOX24HH_Kn31k_CI,12563
18
+ vllm/pooling_params.py,sha256=o6i3geWBQRvfNh4SGGAffJLL9Oy87OcRpdrQQQc-2S4,8636
19
+ vllm/py.typed,sha256=F5LUrt0voM87SNuuOky2X9veCVDqJUgRg_VohYqDigY,65
20
+ vllm/sampling_params.py,sha256=SBYHWtft5u7KM6trFV3z8BWZgVdSKhjmXkwLy02ZxO0,27443
21
+ vllm/scalar_type.py,sha256=s8CZpZru6-4vEexIofKsJzhw5OBjVuvueG6b6yb6A94,12573
22
+ vllm/scripts.py,sha256=d2RU3e6Qi_xY7QUD44uqsdz8DdBR-_zQnFf4rYUW_iw,504
23
+ vllm/sequence.py,sha256=1_wfrC7M7xMnerklKjmf2AcQe9ttQPhMkq6_quGa7kw,3489
24
+ vllm/tasks.py,sha256=8s7iKgpdisoALAsu98kCxHNMYP6ebdp8uRpWvd_Cu68,437
25
+ vllm/tracing.py,sha256=BaBpIGie6_A-ee_XEzu_-aGWkGt_stgeq_b1M3Q8kb0,5043
26
+ vllm/version.py,sha256=vXpCd2tjC2EeL1-kaYvJfh9PJYPDXMskbsLnvUPIMxM,1339
27
+ vllm/assets/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
28
+ vllm/assets/audio.py,sha256=gDEsheV3FDEIX9U5xGdllwMYB1qF5_QYpU8Fj36zkP4,1185
29
+ vllm/assets/base.py,sha256=SErfS4p7zJUe_sJBBwR539_1dy49vsBwl0tAi5aPRHY,1215
30
+ vllm/assets/image.py,sha256=W2qpY_yGQa3x67Oag1u3Hazpu-rPxB3SEJcOB0qA4Bk,1483
31
+ vllm/assets/video.py,sha256=7ulapb3jq2MqKxEVrhzPWgIbcv-CDEfjYMA1cSWPa4E,4487
32
+ vllm/attention/__init__.py,sha256=nqBMrlX4n__QjIO6qcCOYl3-qqn81QyS57O3W753VmA,443
33
+ vllm/attention/layer.py,sha256=EdblYAno3LhM4KwROTpFvHDMMKvyWWZ9uApXUuNuGtg,37992
34
+ vllm/attention/selector.py,sha256=dHnub4G689WSsKjWhGDqvWPqALdr-3_mzy2tniDgB_Y,7328
35
+ vllm/attention/backends/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
36
+ vllm/attention/backends/abstract.py,sha256=CA16Ca9IBJtB6YcXPtnZtGespRE9LQid_7QNitjITrM,12375
37
+ vllm/attention/backends/registry.py,sha256=NF-CcGDtLtBAhAph3EnVcgLYVA_j0_6T2SHU-kGEb9Y,7036
38
+ vllm/attention/backends/utils.py,sha256=_Z7h3rbFlbw6zEmd9wOTtvrx3OF8gZ-7N2wo0yp81oo,857
39
+ vllm/attention/layers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
40
+ vllm/attention/layers/chunked_local_attention.py,sha256=ZVSQlZTLZpExjn29ktITegNQa5IJBZMcNvAnf0tUuaQ,4225
41
+ vllm/attention/layers/cross_attention.py,sha256=A-i_WfJLRIEFBvnQTqmxLeCMgsqZLUgBE_IEw4aPBdk,5893
42
+ vllm/attention/layers/encoder_only_attention.py,sha256=p7rH1yRLdPI5oMvnewxrBwX0vK9an18U2v6JYvMrplQ,3073
43
+ vllm/attention/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
44
+ vllm/attention/ops/chunked_prefill_paged_decode.py,sha256=iT-VbCwzjyaNLH98vnqo3H-ASyNOQYcSLu8BoGs4BiY,13068
45
+ vllm/attention/ops/common.py,sha256=2F18M1Glj5dWEoLYGlWCM4uRNEJcE738AiSrN0ko7xA,12379
46
+ vllm/attention/ops/flashmla.py,sha256=Zthte13CPZ8HjtPYmqMKjHFtPW9kUZXsu7htIRkR9P8,8314
47
+ vllm/attention/ops/merge_attn_states.py,sha256=SEVz4_z-x2DC1hqtGdkjkYs8olYmGjZmpwnV-oo81Ns,1657
48
+ vllm/attention/ops/paged_attn.py,sha256=p_8NFuHoCSqkagp1DUkkxjAl3w-QOmIuULRPqS_qQNA,8492
49
+ vllm/attention/ops/pallas_kv_cache_update.py,sha256=wbl56j_Wv9PEseDhkjZcKurNO-CV785P8fRsjVR10ho,4190
50
+ vllm/attention/ops/prefix_prefill.py,sha256=yYrV900OQWKvYBDPhuq2lUci7gXnqJxfErAugrRwA0M,25201
51
+ vllm/attention/ops/rocm_aiter_paged_attn.py,sha256=K1XeEKeOoqc8k-gcOJ3YJM5PSW71QOVeeLPob3Rx_YY,4112
52
+ vllm/attention/ops/triton_decode_attention.py,sha256=9v1XfYlUcrX7j4aYDYKEud8DCVsdltKavp3ucYe1bco,19515
53
+ vllm/attention/ops/triton_merge_attn_states.py,sha256=eLCLBPmKzZ4hAnoIfxRgRaXlUVZzYoe5G7gWwhfZfBs,3536
54
+ vllm/attention/ops/triton_reshape_and_cache_flash.py,sha256=R62WxvWZeCdOOHDWl4Fl5565xWz5fdXN9aay2aJ6Vhs,5968
55
+ vllm/attention/ops/triton_unified_attention.py,sha256=0m-vgM-ZCSd-HddgTGZs-U9nBoQthk0uqsINFdw-sGg,32453
56
+ vllm/attention/ops/vit_attn_wrappers.py,sha256=wMMDzh3c4Hu2Ez69okRLCgGeNdbwLWogrHD7KQC8CwI,5351
57
+ vllm/attention/utils/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
58
+ vllm/attention/utils/fa_utils.py,sha256=jpJCzHL0aLgBkBMpSK0YqHDA0l3a6dqiKFRVGHkA3Ew,3402
59
+ vllm/attention/utils/kv_sharing_utils.py,sha256=o4YaEd_bjYU6fQq0L5dEdhz9r4aZP1uJ6RBrVwtdX4M,1615
60
+ vllm/attention/utils/kv_transfer_utils.py,sha256=-dUtRw8Zu1hBTcMuemUe3zQjzjQ5UeN2gndokVENtuU,2009
61
+ vllm/benchmarks/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
62
+ vllm/benchmarks/datasets.py,sha256=tIwCYJmlx8PsLHWlsIRzn63-3IlPs9zdliqfQUKHIz0,120079
63
+ vllm/benchmarks/latency.py,sha256=ubZUzbP14GlHMIVJETQq9YFyC_4ROGX2Oo2Kkq4QU9w,5869
64
+ vllm/benchmarks/serve.py,sha256=6NGg91AqB-puVJR2_olvU-czEYGUk7N-sEt-Dc-jwQc,55514
65
+ vllm/benchmarks/throughput.py,sha256=EWszDhrlwSVGUkvVJ7VlSmu8y7Q3yQzuJf_NcbSljF8,28814
66
+ vllm/benchmarks/lib/__init__.py,sha256=BlbNrv7ga0LXYeHLKKiey2woiGh5MkW36g7QG6L6BN0,142
67
+ vllm/benchmarks/lib/endpoint_request_func.py,sha256=6JoTGH3VnTu6IaFusduLvZhBS8dQcnkQmwlxM7ojbQQ,26362
68
+ vllm/benchmarks/lib/ready_checker.py,sha256=QXyydJaq4m1EFdHtpuobeHykIKnwgOawe-ZZbS-jweM,2346
69
+ vllm/benchmarks/lib/utils.py,sha256=YMFxp0A81__5oZJ8wp-PyqOlbYGuxovcVMZzR6v9XJo,2449
70
+ vllm/benchmarks/sweep/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
71
+ vllm/benchmarks/sweep/cli.py,sha256=AIQZNRmte_I3bv3fEP790cmsUk8hoYk6a-y9vgCBQEo,1197
72
+ vllm/benchmarks/sweep/param_sweep.py,sha256=Iz6RzAe08vJ1B7ig1ncGf4SF-q2nnG2c1KCT0pGzmrA,3035
73
+ vllm/benchmarks/sweep/plot.py,sha256=ZaW0MIImMqB6NDFmHdvNLE1Ud2L0HuSJJSWVmDDw6Wg,16469
74
+ vllm/benchmarks/sweep/serve.py,sha256=f17M5WkB0Hr0b_yPePQeu24OXmHz3x9VUKiI-dOWQOk,11927
75
+ vllm/benchmarks/sweep/serve_sla.py,sha256=GVq10xa6quOr_BJi1u4FYXib1LcXEFTksJCBih6lrU4,13978
76
+ vllm/benchmarks/sweep/server.py,sha256=QAT4ixDPla1lCbWcKMPksQj5fM5uiAOq7HWsrU94Bpw,3588
77
+ vllm/benchmarks/sweep/sla_sweep.py,sha256=AuxpVkWFrfJPFFcFYMAa_5RdIHNvdwpW-X2WRtsD0rE,3674
78
+ vllm/benchmarks/sweep/utils.py,sha256=YzhiqE4jmEN7TzUnzQXnlm8TLERkNjgoQ_SF6jJ8hB4,232
79
+ vllm/compilation/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
80
+ vllm/compilation/activation_quant_fusion.py,sha256=HcKJLon0bK7VSPHQV-fxSTDbefeW_-G7dAH_9fP-tjo,6492
81
+ vllm/compilation/backends.py,sha256=p3_Ab_apqmZ-YiBDCewGqlnk94kcPTERniaeKdsUxls,28376
82
+ vllm/compilation/base_static_graph.py,sha256=XP6OBtuEqkLNQc3ZPr6TrwCzokLIARBCBXUw4pIrUjo,2007
83
+ vllm/compilation/caching.py,sha256=1DlTxx7zfIzUZ1yLhE-ib7kkJYBdW1Vu1WdhZ4og388,6722
84
+ vllm/compilation/collective_fusion.py,sha256=p15SUc9wMrkJo3mSG3BOaBnVpqyofmEKzkjQMDgnFp4,44308
85
+ vllm/compilation/compiler_interface.py,sha256=-Y6CoefQpBhhk136g995Wk1wrawpilpuf5hxMy9qhiE,24450
86
+ vllm/compilation/counter.py,sha256=Nr75qGYyNzHhCvpWa1SWxPxZme5h1sSNNNEtYYJxknM,1638
87
+ vllm/compilation/cuda_graph.py,sha256=a1GPTU6cN4hshys5WlCVkyJej6n5XFsYgswaX7oStHo,9050
88
+ vllm/compilation/decorators.py,sha256=A1ifvZ-3CRfuB7Fmh9Jbl9-C4oaArqepvKvSzRSOiwA,21823
89
+ vllm/compilation/fix_functionalization.py,sha256=gW4rX9NbVnqOgqC62svzHAtaza6MEeMrEdBzHz0hRAs,10545
90
+ vllm/compilation/fusion.py,sha256=pmFZXc0ckZIAEffmNMyF3kbiFIFD1clxw9WkuIp0E54,12481
91
+ vllm/compilation/fusion_attn.py,sha256=UyjAlsOqp4AIRf2czPnwmjqewLcZTfQHEiiyz4tdLwg,12227
92
+ vllm/compilation/fx_utils.py,sha256=w89dD5KnsAMXJmIiQOceZfEkySmnXgObNRsfXJgbvIE,3185
93
+ vllm/compilation/inductor_pass.py,sha256=UlOj0qwyB6FJdvUTOh22Iw5oaUbGFWAI4dGuBjDHzfE,3983
94
+ vllm/compilation/matcher_utils.py,sha256=8GWpVGJTDj0nE2rdtP20wP4oyzGAuTbvXRfeD9EVClE,10124
95
+ vllm/compilation/monitor.py,sha256=tmsTH8SVaKV2TpMbfcVEqBxxKawBON-pgACP0VGRLDE,2100
96
+ vllm/compilation/noop_elimination.py,sha256=_RnMhU5GCjo18ISBwBkarAPS0SUfJSAigbcPIcHFJ_Q,5351
97
+ vllm/compilation/partition_rules.py,sha256=wVUBgr73FzDfyIVhAIt2wgpnEZhGsaA8Unuw_B8_xRI,2213
98
+ vllm/compilation/pass_manager.py,sha256=WhDkUk87SNk3cy4KuyNJ-c3j3GkTP98ZdwKoOG38t1s,5046
99
+ vllm/compilation/piecewise_backend.py,sha256=IRt7MXwk1KLWgYT484-xjDhCrFQ7UOtxs8RqMuUVlpQ,4436
100
+ vllm/compilation/post_cleanup.py,sha256=z0Jbq_dQDFzRTrUZUcvE5prvDUxQ7ItogQrEnYdxVzE,758
101
+ vllm/compilation/qk_norm_rope_fusion.py,sha256=RTANrwJPH_CRo4PBEJZuh4L7aM_OMinT7Vh7J3qvN5c,8518
102
+ vllm/compilation/sequence_parallelism.py,sha256=uZtBmxeqvgObqdthHieCFvHM7MWUvw54pW5A_CIx7vE,14028
103
+ vllm/compilation/torch25_custom_graph_pass.py,sha256=9r4T74gV7PxjalJDxyDOOGPyCc8xUWMPAe_gYdTWgwI,1413
104
+ vllm/compilation/vllm_inductor_pass.py,sha256=v_bRjbaI1G8Wn4UkSPULHmECLvuPCT44Nphmsf-NP40,6490
105
+ vllm/compilation/wrapper.py,sha256=mz09zseDk0dfuzydo8zjTFolTqFYgOcrOvImaurGgI4,9510
106
+ vllm/config/__init__.py,sha256=zv74dRyPFfyx1m-EHUiEYqpGlOTdpe1NdbEthZgK-kI,3002
107
+ vllm/config/cache.py,sha256=qmGaiO1L4SwERkzuEsXl-oKVdqG9Ut_Zy8r10Ps2rIU,9822
108
+ vllm/config/compilation.py,sha256=hE-oWB8ysocgLpaTi5TIicMDHL6Vd5v5jwXP_9GilRM,41983
109
+ vllm/config/device.py,sha256=qKm02-gy8hl_3r29MKjLbDYx8Tmj0Y2UAJUACbmDTyU,2749
110
+ vllm/config/ec_transfer.py,sha256=_D5VNYNy8CU7IrTkDm9IB5PF9Y4CBCqzyijl-kYgWpk,3863
111
+ vllm/config/kv_events.py,sha256=-JSREpsV49ohD7songzjiHiLhJguL2v1r4_cx7po--0,1645
112
+ vllm/config/kv_transfer.py,sha256=XVXBEg23qraatFQHUW5mhSg9Dk_R6iU_vBbTsQFB1Jw,4018
113
+ vllm/config/load.py,sha256=eiv6rsHwRgQcK2tXvHnI-GX5NxpZ_MAfd_jvO_Dztis,5727
114
+ vllm/config/lora.py,sha256=eMMHR8Tlx-R1_svkTPm2_xcvWEz2iuQd_CER8qEuhNo,4383
115
+ vllm/config/model.py,sha256=NeqJfJA6tGjq7ApxcvKjukr8Ak9T6tmfAzh0qSpw_fQ,87525
116
+ vllm/config/multimodal.py,sha256=-RlSu-0qR3RGGojiJgizGnc0Cw-W9u3lWNGtrwGUa7Q,9926
117
+ vllm/config/observability.py,sha256=Szd91-NCMn9W-m-qe-q1CXX8UuFJorZLsseuwWvAhx4,4940
118
+ vllm/config/parallel.py,sha256=muS1VMaBmlMgIfgizVEoZ-NgBp2FgysaA3UuzaH75no,28637
119
+ vllm/config/pooler.py,sha256=5AgoMBWM8p6oLt8-jEP4VbQjSFxdLGdjv5WNyOr3uMU,4093
120
+ vllm/config/scheduler.py,sha256=ZGs_Gk1zRmyY7uEhL1p57myhTShPAiSVqqqro5QOL_M,12289
121
+ vllm/config/speculative.py,sha256=nc_xLpn3Vm08k_6uqDrSHhDri1R3m6eKi660tdu1udA,29000
122
+ vllm/config/speech_to_text.py,sha256=nNeiXvk1d16si_w_w8eEREEKeylq1b_gSyNFfH_OFJs,1462
123
+ vllm/config/structured_outputs.py,sha256=CzECe1UjVkFv6_odTw078X_Bbi_cjL_Vs1vJp2XOeZE,3933
124
+ vllm/config/utils.py,sha256=EzboU2lucMI8Br2zHwxlSpniA-vyU3_cHD85x1ZBfH8,5930
125
+ vllm/config/vllm.py,sha256=borKRPGvspRmEATOCPQLwi2Dwm7sUdarM6bUn9i-H4o,52120
126
+ vllm/device_allocator/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
127
+ vllm/device_allocator/cumem.py,sha256=tIx6IZI6r4DMS4MbhuAdrcLM5hN-BCFjoGLSsXBX6hs,12483
128
+ vllm/distributed/__init__.py,sha256=l_KMMLMmYq-MqxY5OeTuQjA38RYwdS4TtVy8JafQq3E,191
129
+ vllm/distributed/communication_op.py,sha256=PD-Kxg2zjs6JHTncHi9vGUetQ4aFiT0uV3fQcjIL0bg,1323
130
+ vllm/distributed/kv_events.py,sha256=3z_NXsaQoCKWInLhKzQlAJHKhv6UIsyrtSJ-4NzhP-Q,12437
131
+ vllm/distributed/parallel_state.py,sha256=a6ZZ9qujAjoWT2PIHCRq5IN7qr1AdXemVejBMX77pi8,64192
132
+ vllm/distributed/tpu_distributed_utils.py,sha256=LHc2lCK_zeODmV5loJV7Eq9Pg8hfplEh4aWN-A-nawo,7585
133
+ vllm/distributed/utils.py,sha256=CgQc26tnQuBg_4oU86hpg6WSzBYZGLiYFNuX6nEt6Zw,20612
134
+ vllm/distributed/device_communicators/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
135
+ vllm/distributed/device_communicators/all2all.py,sha256=RUfFZpEIxXmCKrNkIKZnVQysUQf7uY24MB24fw9q--g,16263
136
+ vllm/distributed/device_communicators/all_reduce_utils.py,sha256=FXN49WKCqJM3WKxMlQntN1xVwrlss5IjrGRXjfIaoYY,12324
137
+ vllm/distributed/device_communicators/base_device_communicator.py,sha256=mwMTx-DdFEdPyTEUbaPghlCuodhNboqRGKaJ9QQmM0I,10711
138
+ vllm/distributed/device_communicators/cpu_communicator.py,sha256=IyRrYtDIw9suqbBxpKBYgvKOKdEk7egpnOeQtk3keQ0,6966
139
+ vllm/distributed/device_communicators/cuda_communicator.py,sha256=y5o76GD9r6WGsMCrdxz9MxQ085sTWvaI1VXF6ZBfVOg,13439
140
+ vllm/distributed/device_communicators/cuda_wrapper.py,sha256=Yj-IHZt2nWAksvsmi88JAC8XNYht8Ylz748guEujWMc,8411
141
+ vllm/distributed/device_communicators/custom_all_reduce.py,sha256=qwIMODFiNFmXGFWjr9wyzXZ6dl6WASbya1wEqAEmXP0,12857
142
+ vllm/distributed/device_communicators/mnnvl_compat.py,sha256=A5UvRWaIyPRQkw4dSd4Mq6sy2I655Wf0bHkKM7aPMZc,830
143
+ vllm/distributed/device_communicators/pynccl.py,sha256=IZRiSwMlWK-yS5cwGBPr7kn4Ov4XSVhA61aVP5ggpxs,13902
144
+ vllm/distributed/device_communicators/pynccl_allocator.py,sha256=jwxg-zmQXjkiobgJKPCtY14TUM4U7XZKYT1pEyfnNFc,6246
145
+ vllm/distributed/device_communicators/pynccl_wrapper.py,sha256=iG602w2gqZOwuhULKP_PF2ywJBN-ZDcE11jJEQdhAZI,18840
146
+ vllm/distributed/device_communicators/quick_all_reduce.py,sha256=IxoLQ89rCXr_SbJPLWDK7t-w23EUzv2zWEO7C_QrdkA,10887
147
+ vllm/distributed/device_communicators/ray_communicator.py,sha256=Myw39lSVtgVk1h21m8LTDKAjatwzRC0BRkeQvemuyJc,9122
148
+ vllm/distributed/device_communicators/shm_broadcast.py,sha256=YnHR3eB7_aYl4KxNrgyHuNDQYMZM4IJnaG_eiAO4xXQ,30277
149
+ vllm/distributed/device_communicators/shm_object_storage.py,sha256=Ql3bO9mG2ECSdwHtfQAPV9kVtCOzqUtDPhXhG8NUTUg,26099
150
+ vllm/distributed/device_communicators/symm_mem.py,sha256=fgc15QF836cFul6AQ8dJDI-xw8P5_wK8qLnczYOoHV0,5395
151
+ vllm/distributed/device_communicators/tpu_communicator.py,sha256=v4WhiR_EONizzUVldECDHe8RpLVC8o0wXc45YYiepCM,4237
152
+ vllm/distributed/device_communicators/xpu_communicator.py,sha256=eHZdPFturEqH1WUJC-Yy56jY9jsY0CRiUY8B3JzarSo,3423
153
+ vllm/distributed/ec_transfer/__init__.py,sha256=TfKcHCVJ27cpYAxvVMFyrBZe40qJfxEX6GeWdCOcIoc,348
154
+ vllm/distributed/ec_transfer/ec_transfer_state.py,sha256=60Iu-N4qVnjSa2kXaysANPa77HX-cYomPTu7GRk5Evw,1152
155
+ vllm/distributed/ec_transfer/ec_connector/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
156
+ vllm/distributed/ec_transfer/ec_connector/base.py,sha256=L6ZTjzL6kIXUyRVsA6ldI4AahpRSbEL3NAMTqwJJf60,8029
157
+ vllm/distributed/ec_transfer/ec_connector/factory.py,sha256=a7N3PRZ1DKwurRgt18UrY16Lwgn8kaXFE1oNUwlTPj8,3121
158
+ vllm/distributed/ec_transfer/ec_connector/shared_storage_connector.py,sha256=NhXdHEnwqcrihsc2P_bRLyPzeQDgneMAwluJqkUNeyY,7322
159
+ vllm/distributed/eplb/__init__.py,sha256=84KX6CIcar6LmWgwmA7huMJaUKlCdFQ6sCEnCQJ8geY,213
160
+ vllm/distributed/eplb/eplb_state.py,sha256=cWTfWMYG78AyGslRpNchopxsJ8OoncFnZp6LYYF7LeQ,32136
161
+ vllm/distributed/eplb/rebalance_algo.py,sha256=L3Nu7_zjzRKLsJsdLz-Cz1BKC6pOpoMX-_e6AhDmPxE,9230
162
+ vllm/distributed/eplb/rebalance_execute.py,sha256=bvkQ-44tW3Vl3MdX_CjcnsL1lF6IubouWSDaq28Pdjc,15006
163
+ vllm/distributed/kv_transfer/README.md,sha256=cKIw6vXYSxBlf0wWwO7haP82CX2oB2QzW0-RxZE5mT4,2007
164
+ vllm/distributed/kv_transfer/__init__.py,sha256=wb5OWOxpI7rmB6NVrSKJHHQPBWpp_3NYseEOo_0mgOo,552
165
+ vllm/distributed/kv_transfer/disagg_prefill_workflow.jpg,sha256=fOFUEx-2Fm1uxHCGopvCREaRqdvR87Z7C0bMqEVH3Iw,142656
166
+ vllm/distributed/kv_transfer/kv_transfer_state.py,sha256=jFrtqPkifgETZZpoqUzdJhK74GMRWnxRCv_mZC4C99w,2290
167
+ vllm/distributed/kv_transfer/kv_connector/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
168
+ vllm/distributed/kv_transfer/kv_connector/base.py,sha256=KuKixI9XMfNTMWanVED-kedkMyMFtdcT34QO26lweJ0,370
169
+ vllm/distributed/kv_transfer/kv_connector/factory.py,sha256=xK_kbFs_YdMkpHnKmsvNPLW94yArS-gUlYqmCG7ACW8,6942
170
+ vllm/distributed/kv_transfer/kv_connector/utils.py,sha256=TY10RE9_B9AuPuckcE8u9-FB9x4Mj7Gorr4OuNwwQWs,10553
171
+ vllm/distributed/kv_transfer/kv_connector/v1/__init__.py,sha256=XGVYgVwmenrlulLD3EgNx0CglBC1YgbOFRPKN-cB7yA,508
172
+ vllm/distributed/kv_transfer/kv_connector/v1/base.py,sha256=zBwJnWrlYfWX3GY94hUYH7WJLrdKIXHB-_ZMwGl1rvU,19129
173
+ vllm/distributed/kv_transfer/kv_connector/v1/decode_bench_connector.py,sha256=JoWX7fharL5nLe323aUbm5hkINOp5nk-IGz-s8uS1Fc,15557
174
+ vllm/distributed/kv_transfer/kv_connector/v1/lmcache_connector.py,sha256=R77uNkdDJcOrK7E7hYP7PMOIx8RG4DsL99xLvxMuUa4,7687
175
+ vllm/distributed/kv_transfer/kv_connector/v1/lmcache_mp_connector.py,sha256=9w618Q0TMPSLYyqVuWh5jkhzTB_8GqxInjd57ElEDqs,30910
176
+ vllm/distributed/kv_transfer/kv_connector/v1/metrics.py,sha256=8bxS4egu3_Ucltnc6dC_HPqFNKW0K_-SaddjRDFglFo,7036
177
+ vllm/distributed/kv_transfer/kv_connector/v1/multi_connector.py,sha256=iapwbyI5WSK2yH_Zrz0smT8E0UGf8E6UJ-ZciI43GoM,17731
178
+ vllm/distributed/kv_transfer/kv_connector/v1/nixl_connector.py,sha256=zLmholM9hLgHl84ZMwPOSJivE-Kf4XicoBKhNNT-ENE,101733
179
+ vllm/distributed/kv_transfer/kv_connector/v1/offloading_connector.py,sha256=SmF_JaHgVNm4NrUaN-IXlXmcTneMe_k9aA_TBB1vSoM,19343
180
+ vllm/distributed/kv_transfer/kv_connector/v1/shared_storage_connector.py,sha256=kcXTAsmBDikc4j86m0lLCuPhswjOF20W3vhk7EyZRGs,17094
181
+ vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/__init__.py,sha256=k3tOaUj8P0IjtWcZ0CGxO3iE8IuALduzLpW5ejYJAFE,426
182
+ vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/multi_process_adapter.py,sha256=qvfxY3TUyOXnuWvodyBm9kjPrX10HV8LlZfOcY-g3xI,12527
183
+ vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/utils.py,sha256=aSe-GgW3-PvuMFkabphG9wB-LDVMXuYn5GqLFv2dONo,8206
184
+ vllm/distributed/kv_transfer/kv_connector/v1/lmcache_integration/vllm_v1_adapter.py,sha256=O7r3fwYNiT9FA_UllZ1et6RKA8NRuw7lcHtIcklY27A,51921
185
+ vllm/distributed/kv_transfer/kv_connector/v1/p2p/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
186
+ vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_connector.py,sha256=6A36M4ebHBjZcigJhIscyE_39yOJd_Dzdoa-RzTtj-s,19314
187
+ vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py,sha256=gdHYw55Qm1z8cz5cWoq76siI8tL6tewzYldcxNCG2s4,23390
188
+ vllm/distributed/kv_transfer/kv_connector/v1/p2p/tensor_memory_pool.py,sha256=A2mMTFKJ1zBYEo0gPhe166-ms1S_gAaL4Zs9tULMbTc,9225
189
+ vllm/distributed/kv_transfer/kv_lookup_buffer/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
190
+ vllm/distributed/kv_transfer/kv_lookup_buffer/base.py,sha256=Tm4t3_GYEW4x4QgWq3asuIPHfoStJmxnWagwyKloMxc,6191
191
+ vllm/distributed/kv_transfer/kv_lookup_buffer/mooncake_store.py,sha256=b1SD1nrKU5LmdGCHE2oDl_qk_Ywz0sJxyUxaqNCN2VE,5614
192
+ vllm/distributed/kv_transfer/kv_lookup_buffer/simple_buffer.py,sha256=sV2v7QElLTyuGZBuaOF1-U9IFjLkzE9Swvxunucims8,9023
193
+ vllm/distributed/kv_transfer/kv_pipe/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
194
+ vllm/distributed/kv_transfer/kv_pipe/base.py,sha256=nU-YDmWz3DYq3-PyJd6mMymC_1itJxBLeLIv-Ely7Zc,2102
195
+ vllm/distributed/kv_transfer/kv_pipe/mooncake_pipe.py,sha256=ua69E5af9jBvkv85FOB32GEs3oM9azHuTE83WvriMs8,12108
196
+ vllm/distributed/kv_transfer/kv_pipe/pynccl_pipe.py,sha256=ZtwzoNwntpeIqeX3lVweUCl6AY3ZfLB5NYUzVcOcpk8,9604
197
+ vllm/engine/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
198
+ vllm/engine/arg_utils.py,sha256=N7zqm-BI5pQyHQDNvQaLjS998lPvzqqiNZs03xLRZ_g,89389
199
+ vllm/engine/async_llm_engine.py,sha256=B4Xf26qGXnF7yc06F_OM2CH8k5CLUU34tjUEafnIg2U,197
200
+ vllm/engine/llm_engine.py,sha256=h0FBz_1vf8ThMr1PHUQVus1sEQH-N8XZsiQwNZrs1W0,212
201
+ vllm/engine/protocol.py,sha256=AjHYOXx5dbBijlXqGIBBE1qelMw2Cp8_GHpbD5JwNXM,4707
202
+ vllm/entrypoints/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
203
+ vllm/entrypoints/api_server.py,sha256=oYUrcvccxfZ-Wh2ALlA-3E7GT83I6L1TIwmOzJMvpQs,5783
204
+ vllm/entrypoints/chat_utils.py,sha256=dUctDnGT6Hw06zKWKn74TovTgppw_VdJo_h7KPOcK60,58681
205
+ vllm/entrypoints/constants.py,sha256=ZX2zFNKkQjQM46StHeaSFaQIJEToSYcyJ7EBr98ZpWs,335
206
+ vllm/entrypoints/context.py,sha256=JlaNWzq_CDe1k8FzB0PlRhcZg3-FR2owWP44EdiFXQ8,21116
207
+ vllm/entrypoints/dynamic_lora.py,sha256=SFQD4z0XPKwtKvldOxikG1v8xj2uyWz52fPgVh3zCvs,2119
208
+ vllm/entrypoints/harmony_utils.py,sha256=DMNEkTlz1bbSH8FPYzaTKQQAsgTeW5rGSj3IU5HCRc0,20139
209
+ vllm/entrypoints/launcher.py,sha256=AKCjD0uDH5IHLEnyUI8qjhihRIowyMJ8k77rvUmtiZI,6064
210
+ vllm/entrypoints/llm.py,sha256=LRxXt0Dxguu-_jefKugeRIRlETY7aNgyqszMP52XUy0,74170
211
+ vllm/entrypoints/logger.py,sha256=KYayEwEnv1ImBsZTcQ_7okyNqvkKjFqpZcv8SJYT5sg,2537
212
+ vllm/entrypoints/renderer.py,sha256=Jx5rWN1w3Oaww0banMOwAIOh-RAG8mHwZWvON4eSEfg,15047
213
+ vllm/entrypoints/responses_utils.py,sha256=Sm2ozv-M0nChqFWZG-eoOhruzIy0CXm4WHbwm7IKFzc,2490
214
+ vllm/entrypoints/score_utils.py,sha256=-FKLSjuQNzSboneZ9YVmXW5qlcRu7j9tqR-bL5f9WxU,7872
215
+ vllm/entrypoints/ssl.py,sha256=P6ApVGAQFLHERnJeoLNpZk-HQI1nNqWdlF8UAQpnC48,2712
216
+ vllm/entrypoints/tool.py,sha256=6WuSW4x0IVzh4ZM6hbu5ESMeJkwwIsF7_XYOYqnvN-I,4634
217
+ vllm/entrypoints/tool_server.py,sha256=nwu8rKpaZRUOloT2v0usUxm-bvL5bGPBQnWkqtsx-GY,7051
218
+ vllm/entrypoints/utils.py,sha256=ZqdfqYq4oPozDOXcxtC2W7RqPE3H6QmLfVAk7hX8XIY,12264
219
+ vllm/entrypoints/anthropic/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
220
+ vllm/entrypoints/anthropic/protocol.py,sha256=nGkTR2TQtcy52uJJM8cJNSQtASDCChhgpQ6_-z_A59s,4284
221
+ vllm/entrypoints/anthropic/serving_messages.py,sha256=MlZJzclRBGYg06wdNkpcqTq2UMnndobPAJoShoycChY,19965
222
+ vllm/entrypoints/cli/__init__.py,sha256=7yL1JNi3jwUtykSxiULk93tN1lqykU-XdMbtBWtjU0A,582
223
+ vllm/entrypoints/cli/collect_env.py,sha256=H9qpqasc2FRmAQUvmRV5WuCxJqdj-ouYAQ2CMpN_ihk,1106
224
+ vllm/entrypoints/cli/main.py,sha256=vHRDDzGjKK2ZE9EWyGn--jpfBWewDtiYFDab9srDc78,2468
225
+ vllm/entrypoints/cli/openai.py,sha256=J28gD13rwUqI_IwDh0VGB3Jp5E2g67ILsKjPyYc6Bzs,7864
226
+ vllm/entrypoints/cli/run_batch.py,sha256=8ChYNc3UbbCVEbtp_tTbgR32_jVq3WRJgO1RzGhP8yI,2242
227
+ vllm/entrypoints/cli/serve.py,sha256=sjfXeDQW_zaWE6yX1YnhgHfNuWPa2jVOpmofHWqbywE,9001
228
+ vllm/entrypoints/cli/types.py,sha256=On70PbJbmYXWTQZaL3EbI7I-fd1eIE21FAhKuHnTcfo,812
229
+ vllm/entrypoints/cli/benchmark/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
230
+ vllm/entrypoints/cli/benchmark/base.py,sha256=3KVCkfdN9VsNhP4RniGqJi02pzIEW4IMIktMFgwU1Y4,674
231
+ vllm/entrypoints/cli/benchmark/latency.py,sha256=ruUvsQxW_zv19NxZe4dhdVRjmBl2Gipa7o0aGp3VN-I,653
232
+ vllm/entrypoints/cli/benchmark/main.py,sha256=aoHKd2EJPtmCcZv3G7TBBWXbYR_bi1-qUDpn9I0AFeY,1847
233
+ vllm/entrypoints/cli/benchmark/serve.py,sha256=v735WpqjlGE846KvCqAlrb6eAdysihhNZNX5SDL2Tac,635
234
+ vllm/entrypoints/cli/benchmark/sweep.py,sha256=kHcCsBsGQwNC_bdhIdKs9ykJawLbRCrZF-HT4NFCjkE,629
235
+ vllm/entrypoints/cli/benchmark/throughput.py,sha256=Sb73dvXCAs_P7rk3sYkEbKIezlA8cCF7piMX1HA-xEc,652
236
+ vllm/entrypoints/openai/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
237
+ vllm/entrypoints/openai/api_server.py,sha256=Y5ZLVRJ8fu4kgJKIkB1jtU9UDxczr_DD0T4y0elYhnw,74766
238
+ vllm/entrypoints/openai/cli_args.py,sha256=vunq2FOygqUjzwov6lew12OKkMesb584kJWzQy9otjo,12943
239
+ vllm/entrypoints/openai/orca_metrics.py,sha256=qLJmaC4SWLSfvGtwOECR8DexnCXBVqYZuENYINORYgI,4509
240
+ vllm/entrypoints/openai/protocol.py,sha256=3wsZyipyXK7L9CR_msnHymsZcZNwtgEhxAqipGEdBCU,116004
241
+ vllm/entrypoints/openai/run_batch.py,sha256=5Q6KXLIfg1asWHj10vnCFPK1vecla-Q1uDnbPBAdNXs,18910
242
+ vllm/entrypoints/openai/serving_chat.py,sha256=TqBuLBtbMk3xYYTK0INWq3s9oBc_r5yTMfp_haGZy70,78438
243
+ vllm/entrypoints/openai/serving_classification.py,sha256=zx2A14xQMTOX_onoXzl4G1u3F7-jquEioa7t1Mg9OI8,8226
244
+ vllm/entrypoints/openai/serving_completion.py,sha256=_osBguTL6nZSY480w2LcYMOwaK6sLgWh9vEn1BoqbQo,29414
245
+ vllm/entrypoints/openai/serving_embedding.py,sha256=1nAGHGyJ_P8uRs540b4GVFKfCHWA7EKTBjWTqbDkw3o,26989
246
+ vllm/entrypoints/openai/serving_engine.py,sha256=KDomMXttpG2kHmjHLEKBPyx_Ib9se56ZsjR3egItxR0,50846
247
+ vllm/entrypoints/openai/serving_models.py,sha256=DVlvTvkMUpJkribhp62RyNQ1nOCUL_UzeVNQvDo-N24,11395
248
+ vllm/entrypoints/openai/serving_pooling.py,sha256=qadtp5ZPSG8nXW786SRkyg1nXaEE6-iynfek3GJsKgg,12974
249
+ vllm/entrypoints/openai/serving_responses.py,sha256=rS8E6R9mEQBiAgsnQ6gcRVJeLmiizv1LLZeFNhKVJ68,86339
250
+ vllm/entrypoints/openai/serving_score.py,sha256=jNf7N84L1xLBiT_i_yglhiLvtT7PDsGx5Y6M7p4iNcw,17159
251
+ vllm/entrypoints/openai/serving_tokenization.py,sha256=_1r4Cd73aa9Wcs2mQKDAXuXDZzovaOhjDle_IXxScGU,7469
252
+ vllm/entrypoints/openai/serving_tokens.py,sha256=tp-KzybaWkE_vItiRde-2x4B9i71vH3zWgCghPHMzcY,10496
253
+ vllm/entrypoints/openai/serving_transcription.py,sha256=EcOrd43U_GVR1pOiT4PdVKzC9FPoQAXwRcBUvDQKAqw,5496
254
+ vllm/entrypoints/openai/speech_to_text.py,sha256=9MxkDSn1MZnD62QoVt5O5NEAPECZbr5knFdocIhTH_o,15959
255
+ vllm/entrypoints/openai/tool_parsers/__init__.py,sha256=eVVn57wdfoV3PLrJcODkMQW6RErjyOhnAbio3GJurkU,3260
256
+ vllm/entrypoints/openai/tool_parsers/abstract_tool_parser.py,sha256=8f1UVUBfvmPg864mt4gKjjzJ8LMIM2AcgHS0uz3oeSA,9695
257
+ vllm/entrypoints/openai/tool_parsers/deepseekv31_tool_parser.py,sha256=V8wv7AdlGu2GVZz_Uzge6j_6614Z1CkBNqt3o_T3il4,16472
258
+ vllm/entrypoints/openai/tool_parsers/deepseekv3_tool_parser.py,sha256=qin4qzP-YKFcFRR7-UFewECuMWWAQPi_MQ-T3yhXbEw,16570
259
+ vllm/entrypoints/openai/tool_parsers/ernie45_tool_parser.py,sha256=GxhJuHpCRru8BQ51aQB2OlvHxB1IJyswfI9dJ6YYJ3s,8373
260
+ vllm/entrypoints/openai/tool_parsers/glm4_moe_tool_parser.py,sha256=Zp1uW_ldQEr10R59D5VKU_dgbQJXZZ8BNkC-Kf31pAY,7525
261
+ vllm/entrypoints/openai/tool_parsers/granite_20b_fc_tool_parser.py,sha256=FV0xqBAcvEy2VvW3StKe0umy_WxB7QQNZITnnRj5-0A,11040
262
+ vllm/entrypoints/openai/tool_parsers/granite_tool_parser.py,sha256=6uRr4SIrp33TY2Rzm4KSj8rmC1jxedRiqofoz53w0_g,10296
263
+ vllm/entrypoints/openai/tool_parsers/hermes_tool_parser.py,sha256=MUqW-mU2tg32IrEsGFDP7Sew-8GHHhsGW3i2YnHjDPA,21228
264
+ vllm/entrypoints/openai/tool_parsers/hunyuan_a13b_tool_parser.py,sha256=K81FODcrWrWsLs8GJEKXmOB1YFUV0oAZQlupM6CfYnI,16730
265
+ vllm/entrypoints/openai/tool_parsers/internlm2_tool_parser.py,sha256=enqsdUG9QoFrsCXGfzpxEPcYk40UyJ2Ikz1_ro9pTYA,9123
266
+ vllm/entrypoints/openai/tool_parsers/jamba_tool_parser.py,sha256=qZ4bou7o6HMdAmi1vLegWhGN6PvfOVp7xndxGfc4lpk,13773
267
+ vllm/entrypoints/openai/tool_parsers/kimi_k2_tool_parser.py,sha256=z8MloKvSuoNaqHKhdWRo09XokPSqE2-RcrnMY4ljIEQ,26186
268
+ vllm/entrypoints/openai/tool_parsers/llama4_pythonic_tool_parser.py,sha256=G6ID7HObcZmdY2SUHW_gguXoDlR7bdXrBeqmth9xlnU,12834
269
+ vllm/entrypoints/openai/tool_parsers/llama_tool_parser.py,sha256=mgz5q0F0Y5nCcmNiX5Oxzk5SODe-5mH8-1apAn6ZHbI,12095
270
+ vllm/entrypoints/openai/tool_parsers/longcat_tool_parser.py,sha256=phvLp6Ws0U9hOAnTIdyWyy3LvURBbTrbb9FHPSs9gkw,1322
271
+ vllm/entrypoints/openai/tool_parsers/minimax_m2_tool_parser.py,sha256=AlUDTydFt7zvDS4wKQYRyjzqNhpcrgJRst_PXOewKUw,26699
272
+ vllm/entrypoints/openai/tool_parsers/minimax_tool_parser.py,sha256=ORiKysTxJL91Glch9DIizbYt0nsQ_afovOphECyCDEk,28535
273
+ vllm/entrypoints/openai/tool_parsers/mistral_tool_parser.py,sha256=sUOAEvevQlr-B_QYtOFS_OGc97biLfRolFeBQWiCbpk,16449
274
+ vllm/entrypoints/openai/tool_parsers/olmo3_tool_parser.py,sha256=ZjqzgSCjWDRJXKznS1LYkma1vA6TEPxcq3llqtwGo64,13937
275
+ vllm/entrypoints/openai/tool_parsers/openai_tool_parser.py,sha256=5FlaBNONmqC0yh2F7FfoOZq49Rmwv5bf91Ch01R_tI8,3473
276
+ vllm/entrypoints/openai/tool_parsers/phi4mini_tool_parser.py,sha256=z7FS-G-FYgTqxJwKgY_2GkK29v-R6lwIv2Ly5XhATZM,4031
277
+ vllm/entrypoints/openai/tool_parsers/pythonic_tool_parser.py,sha256=IpITSRHECxaOkNqn1UBuLUogpiOLGIu6BxnvL-MGh6Y,12216
278
+ vllm/entrypoints/openai/tool_parsers/qwen3coder_tool_parser.py,sha256=k7xddaFPRjHkqe65CuCC7Tb2RqojivwAqLASWm6xKMs,32781
279
+ vllm/entrypoints/openai/tool_parsers/qwen3xml_tool_parser.py,sha256=Djksii-diFirgYt6PZ2g0h1MoltOSVpIdzBBO7Zs8mc,53820
280
+ vllm/entrypoints/openai/tool_parsers/seed_oss_tool_parser.py,sha256=oTJz-adrMNI8v_1vzmf_L3aaNRAWP-5MNgQU6Jzv3qI,31367
281
+ vllm/entrypoints/openai/tool_parsers/step3_tool_parser.py,sha256=KzjYrAIz4CQJiKVeKeC16lg3dnVXpJcjh9lMNpqewDs,12122
282
+ vllm/entrypoints/openai/tool_parsers/utils.py,sha256=5VZCLxbxuOP8FtiFjYneEb7PKtQyLJBVp247IaxsHyQ,7444
283
+ vllm/entrypoints/openai/tool_parsers/xlam_tool_parser.py,sha256=wD2tZZccD-NqHdk2vLb-vCeO7lUBdWWUpndXzkJ1OoE,24653
284
+ vllm/entrypoints/sagemaker/__init__.py,sha256=Fh71s9Oigag26xVp42xT7JCAr_dvDRGnjoRhEV18cwk,155
285
+ vllm/entrypoints/sagemaker/routes.py,sha256=h4D4htgTCSOOd4UXnHG2AjX37BeJKac3ZMF8boZRr1o,2694
286
+ vllm/inputs/__init__.py,sha256=mOYaFF0JBFeIgf4cs5_kS5d7zk4yA_TRBxhfOncKhGw,959
287
+ vllm/inputs/data.py,sha256=vCFClf9AbX4bNhc6_3jit-xg5rRbdRplEPaxxHloRcE,11465
288
+ vllm/inputs/parse.py,sha256=kIpN6UrnUM1sDxip3f5TkaeiFe_Gl1mUr2h0cFCmfcs,4257
289
+ vllm/inputs/preprocess.py,sha256=HkXWM8IyXpH90Yp5XmtD_6kCX96rJTdH3KmbtEVo1Rs,24980
290
+ vllm/logging_utils/__init__.py,sha256=MU0sAgeBaQM4PJW5VAFq2WsnszAGljJ79cCmXvihA0c,268
291
+ vllm/logging_utils/dump_input.py,sha256=jqfKTqU2b2O4g5I_P2snzPuuIFlSTYg33BjZ5P87-JE,2914
292
+ vllm/logging_utils/formatter.py,sha256=EwkKqUbitmwO_04kqDX2Bd_7VjXDAIT475EOv5HfVO8,2821
293
+ vllm/logging_utils/log_time.py,sha256=ea9dQUzQAu_f5juFfGVeOX403d4yGVqiBmI35gmpqdQ,864
294
+ vllm/lora/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
295
+ vllm/lora/lora_weights.py,sha256=-tKk1jNxSET_0YOPAS6LRMvQ46T1aMl3fkwJLb4MYl0,5805
296
+ vllm/lora/models.py,sha256=u8b3xXVwdfEMrLZ_eeE5jeR69yl9wNgs_senKFB9Pf8,35579
297
+ vllm/lora/peft_helper.py,sha256=rLbCLj-UsjXFD2wXD18hOAtzP_389FwcpozeItzgRvk,4678
298
+ vllm/lora/request.py,sha256=OIU4IXeoLkewhawohV77tXhDTBzaklm0ELvPPXaNbQE,3149
299
+ vllm/lora/resolver.py,sha256=SvNATAduYkVIuFA59Q5pPTuPjmLQCg-IH-_35u2mn98,2880
300
+ vllm/lora/utils.py,sha256=V1M6-OYjvdPlB36oFEKkYGpnw_Ivc2yjZuXZxBQYRC0,10210
301
+ vllm/lora/worker_manager.py,sha256=125xoWx_FIAEfdy6Wd4D4mX39QaNEOsB3LJQcwdf_tI,11361
302
+ vllm/lora/layers/__init__.py,sha256=2uI1_QzXelNWyWclhsgxKhq7CAQA7aX2E0xsCI-S_tg,1564
303
+ vllm/lora/layers/base.py,sha256=j3OexlQVNZxAOXzS-wdSn_7hdEBQmGbKju0fGVWz9rI,1824
304
+ vllm/lora/layers/base_linear.py,sha256=FEGcvPPs8unnC-l_kR6sXI73ygLwRz20FiK9fYq-ItU,5644
305
+ vllm/lora/layers/column_parallel_linear.py,sha256=xPzCQ-bIgiJ96egTlx_V3xLbQxNU8cheSAqwGzF1sBo,19879
306
+ vllm/lora/layers/fused_moe.py,sha256=bBn1vEzfafaiuKD2xwUMgCuZ4OlN2xmp1VJhbpp6UlI,17958
307
+ vllm/lora/layers/logits_processor.py,sha256=rtLXmRZM_Uw7AQjniWrbYpZmCCEujFO3_FMtul2JaiA,8247
308
+ vllm/lora/layers/replicated_linear.py,sha256=VayTCHJ-sWDMOGbpD9agjegPEWo1s5mVRHN0rYkVZ4U,2198
309
+ vllm/lora/layers/row_parallel_linear.py,sha256=BhjVYUR_rpeCikzhG1w9cs3Sy_no9KSl0ZyJNPKx_oI,6259
310
+ vllm/lora/layers/utils.py,sha256=1PQgqQGhlTUeGj452uOCMQJJ_LNE3FMdoZ4Uzqg2j_I,1963
311
+ vllm/lora/layers/vocal_parallel_embedding.py,sha256=L11eDMbS-M0RCql3Q6XcGzcCpn6ew_nlDBQ4gpiOiZ8,6272
312
+ vllm/lora/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
313
+ vllm/lora/ops/ipex_ops/__init__.py,sha256=GoE4WbwmiV3lmZ9BvAzF6HribyiAa62lEsRdAOORl3g,259
314
+ vllm/lora/ops/ipex_ops/lora_ops.py,sha256=sQTAhgirQaxWxhc3xTdSlDBfZKUUJaY4MwISeBBXOEg,1344
315
+ vllm/lora/ops/torch_ops/__init__.py,sha256=T8KI4Uy6j2AJOjjfe_9HYZmuSO9xp0G_GTCdvHyJiNE,426
316
+ vllm/lora/ops/torch_ops/lora_ops.py,sha256=M7wqynKevDeeFyXDwcqHDnCqOyePCtPSpiHm7I7pnCA,3888
317
+ vllm/lora/ops/triton_ops/README_TUNING.md,sha256=vRdibcgJphUomT22SAtQSrEf7LVROHwHgugGjqpBqY8,3078
318
+ vllm/lora/ops/triton_ops/__init__.py,sha256=-IQeOBrAW7uCUQQVUt5zjjL4PyqB1in1AJei7ezUpNM,598
319
+ vllm/lora/ops/triton_ops/fused_moe_lora_op.py,sha256=6OiSxZZUEjbHoiBH6dannpRvSffPAm9wYgxIJseNuG0,18976
320
+ vllm/lora/ops/triton_ops/kernel_utils.py,sha256=YiE1F7sojPFFF06QsVSeaDN19Fpo4X1MhFSoX4F8EXA,10375
321
+ vllm/lora/ops/triton_ops/lora_expand_op.py,sha256=0T8MQkWC1A-AgUudjjTZPsGmzaj5mamEXDyZBytUqH8,9417
322
+ vllm/lora/ops/triton_ops/lora_kernel_metadata.py,sha256=_7kOuoEYIx7lGOzFPwp6idbExJm6tN__CBUATMr41JY,5407
323
+ vllm/lora/ops/triton_ops/lora_shrink_op.py,sha256=Olu1MoQ24II09hKTLUGhG33BQtoKpRmLhdOj9LM8ZYk,8871
324
+ vllm/lora/ops/triton_ops/utils.py,sha256=sCiYwaSL9CGh6rVT70wpeqh6vsozJu-Z8XBEDKpYoHM,9744
325
+ vllm/lora/ops/xla_ops/__init__.py,sha256=EbrQj03YWghRsJGDiwWghqJ_yNeYsczL8Z9UuizHxgU,258
326
+ vllm/lora/ops/xla_ops/lora_ops.py,sha256=ZkV2n8L8MVQWCVqNUtFtqPaIcy-wkQk7HAsXRDDoV3M,4245
327
+ vllm/lora/punica_wrapper/__init__.py,sha256=A5cDJmdCPRBN23kwLfemRlWI4YA-t_7qIxeoeimCkT8,313
328
+ vllm/lora/punica_wrapper/punica_base.py,sha256=k1rSUwXYFOxBdF-tj_10tyvU6VxKs1OJL3ctSvogt_0,15611
329
+ vllm/lora/punica_wrapper/punica_cpu.py,sha256=qJLetoy4aPyQazVqbEq0pOOr848ox_hXvO3AOIH4XR4,10817
330
+ vllm/lora/punica_wrapper/punica_gpu.py,sha256=cvV4aSKN8LtEG2YsJHAlmBvkyfdaHlgwZlBaX7iDojI,12859
331
+ vllm/lora/punica_wrapper/punica_selector.py,sha256=-IyHqLvTPs5onKUXacOqS9J03otlfaT8J7yj4iD_YRs,818
332
+ vllm/lora/punica_wrapper/punica_tpu.py,sha256=orV_J-7DCZFKfPsuxKwdodGhhskughhokIzQ8He8qaI,12374
333
+ vllm/lora/punica_wrapper/punica_xpu.py,sha256=3AgxXUxlX-wBX_6ZdTcYEjraGB1r0Ho6VYiZDnr3wKA,8841
334
+ vllm/lora/punica_wrapper/utils.py,sha256=omm684H29vAOXrC4ny19kmKFRk0_-pIAMTR87CmRMXw,5517
335
+ vllm/model_executor/__init__.py,sha256=HldmuFnAddB5kxjNzrqSLCTBGzvOHJ3VPdN4xI13sEo,333
336
+ vllm/model_executor/custom_op.py,sha256=y0I1l7C9UhZ3Q1PgWxiQZpkteam64_U249j12FR2bWY,7441
337
+ vllm/model_executor/parameter.py,sha256=zo3JRXFPq62Dg8624hZygs3zAq-W8MwJirmY1qwHkTY,22854
338
+ vllm/model_executor/utils.py,sha256=KKvNUXXff9cvoU_EXGvaeSB-KTmTjj9-kAKoONE2BZQ,3477
339
+ vllm/model_executor/layers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
340
+ vllm/model_executor/layers/activation.py,sha256=nanZDWwPsteYk9B54gIHshJSYzSvPEgVsfTJngFE1Io,20411
341
+ vllm/model_executor/layers/attention_layer_base.py,sha256=8X1VLYLaZMe5yJcI1ljz3s_Y1ScmALimOzwcO2hlTb0,1038
342
+ vllm/model_executor/layers/batch_invariant.py,sha256=VkF_8mTgwjBW6FC-HcG99yAPM9Pb7R8xTo8IM_tY8FI,27107
343
+ vllm/model_executor/layers/conv.py,sha256=Fe7VP_SLP6FzoW77mB6IsMf9NFYIUz65y4xpAIP_RnY,7214
344
+ vllm/model_executor/layers/kda.py,sha256=81d4LHHMHWftwO56RL2o8Fa_ZY7CacHecMBZJY6wwfY,15724
345
+ vllm/model_executor/layers/layernorm.py,sha256=ly_FRF4zymmim4o8xfv-huqd1q_DjLM58j5QQ2HPwTE,13887
346
+ vllm/model_executor/layers/lightning_attn.py,sha256=MLwa9XbQynBNiIv3rGhhuPWQLCp3XDVRNZbJTa7xZT8,20637
347
+ vllm/model_executor/layers/linear.py,sha256=lWucztYwZxBgRELBK0G-Zm_YKJWvmOw0g3ED_oopOvY,56807
348
+ vllm/model_executor/layers/logits_processor.py,sha256=yUVukCx0lBtFP7-2rfEsSSw-po25YD0GW-WF6beyHGo,3840
349
+ vllm/model_executor/layers/mla.py,sha256=JIjoOzw1kHd3mdNg71GXGyWplBfdtV_Q2mIvwJCPpG4,6084
350
+ vllm/model_executor/layers/pooler.py,sha256=iZjEZXwJ7p-uO8SJVJTXp-f-Gff7JVBM-5TS2Xt78iU,26626
351
+ vllm/model_executor/layers/resampler.py,sha256=kK_gbASYVbvMoPPN3petUTfNGK8DkdNdihYnFx-JLn4,9877
352
+ vllm/model_executor/layers/utils.py,sha256=FI3DqvZAPjDcmESgiQv5MHWDwx8Nw_x_nWmEehneprc,8024
353
+ vllm/model_executor/layers/vocab_parallel_embedding.py,sha256=6Btumz_IkeCVvpgVICMEeG55wnQizWJl7ViPMooWpaE,22034
354
+ vllm/model_executor/layers/fla/__init__.py,sha256=Xbv6P8KG7bLpcOF17zPO5PYYHImweE9aiYYgGyMS9U4,391
355
+ vllm/model_executor/layers/fla/ops/__init__.py,sha256=Tx9ajmbsCCptuJofX-QpCC_Ut2SwZBTT56SUjGXGNs0,642
356
+ vllm/model_executor/layers/fla/ops/chunk.py,sha256=NJFnrsO_TdmXwAsyAmyUptSMxrVCH69T_xHjDp6DRko,8925
357
+ vllm/model_executor/layers/fla/ops/chunk_delta_h.py,sha256=rSBPoldJnPs9MHO_XOCw_E-fy-R5H8vA38zsSQjI9CA,11864
358
+ vllm/model_executor/layers/fla/ops/chunk_o.py,sha256=ucfFMons_gjoof5N36dImFVcWDZpeBjHVxdFMI1XOTw,5095
359
+ vllm/model_executor/layers/fla/ops/chunk_scaled_dot_kkt.py,sha256=_i7MMGO66MjD1RYQFDqNpa9yGknPyHBOTEDwChc9hFw,4653
360
+ vllm/model_executor/layers/fla/ops/cumsum.py,sha256=qhS_Lny-roiECMEDvtWBE2RpiFYDKjqWxwQONRdL2gA,8224
361
+ vllm/model_executor/layers/fla/ops/fused_recurrent.py,sha256=jZyUUmgmaq75r_Mg4nYSewk3pK-YP8tWcWhFTktQ5TA,13064
362
+ vllm/model_executor/layers/fla/ops/index.py,sha256=cg2IV5Tn7cvDzPw8fTdI96sIYFy87tk960b0nPJFXSE,1221
363
+ vllm/model_executor/layers/fla/ops/kda.py,sha256=7eRv3Hs8oOWM1_nnrBC1kERDSUixaax_-RHDKjQz5xE,38297
364
+ vllm/model_executor/layers/fla/ops/l2norm.py,sha256=HfUKWKl1EvrUKcowbDu4E5QPQCgZbPFumK43eli0Q6Q,3978
365
+ vllm/model_executor/layers/fla/ops/layernorm_guard.py,sha256=nfQwy29Wu6Pt3X3ZL3oPVqxhzITarzf7AmxjBzaKDWI,12449
366
+ vllm/model_executor/layers/fla/ops/op.py,sha256=atkZKsyTmBUTa68GnRk0VfEVfKAbmRoFLYtQlf8_JBM,1712
367
+ vllm/model_executor/layers/fla/ops/solve_tril.py,sha256=ARfbUgKOHld2Zbq3VlgjP07bjczFSCCQc0MRLmABcV8,19385
368
+ vllm/model_executor/layers/fla/ops/utils.py,sha256=7zOakRwlL0cVsyqpZeWcknuuIs76WvuR4-2TzdlL9AQ,6357
369
+ vllm/model_executor/layers/fla/ops/wy_fast.py,sha256=TqyZCDUrAUo0ZFdtHPRY4ntUxavxsYsAGGj0m23oyiI,4372
370
+ vllm/model_executor/layers/fused_moe/__init__.py,sha256=yvxdIrRCBRyyaolU_seol4tQSY6yx4S3tj02lnsMB_k,3199
371
+ vllm/model_executor/layers/fused_moe/all2all_utils.py,sha256=iFoXKhR7nDzFu706YK_vFbpY_1g9nSOREFwP3bJoLT0,5424
372
+ vllm/model_executor/layers/fused_moe/batched_deep_gemm_moe.py,sha256=uAQeX9mm5c-mi_0WHMWrQKYC-b-t4JB7xcAklowvvBE,14815
373
+ vllm/model_executor/layers/fused_moe/batched_triton_or_deep_gemm_moe.py,sha256=3IsUrQEa-xv6i4kcrQKY8txc59p1q2C-s9qv6lF0JMw,6070
374
+ vllm/model_executor/layers/fused_moe/config.py,sha256=t5lnTX9NzB2Ydx5Pjyv4B8_5fYbJ1oyclywiQzM8EEw,30961
375
+ vllm/model_executor/layers/fused_moe/cpu_fused_moe.py,sha256=yew4WsXt6L8-q3iGk0OoGVUD2PkXWrOyPq9Zq0MRt9U,12610
376
+ vllm/model_executor/layers/fused_moe/cutlass_moe.py,sha256=M1LBG0_OS2h5DK8Z1mf8KNvxQlCKKMue-aATKuDFDQU,34903
377
+ vllm/model_executor/layers/fused_moe/deep_gemm_moe.py,sha256=K9CRhD4ua90OSgHO6vagg0KFQIfp5obUihRFbY6v8Cg,12957
378
+ vllm/model_executor/layers/fused_moe/deep_gemm_utils.py,sha256=4QdWHqornCJ6vQqa7m_neRlKUYs58xawpGcG3XYiNLc,12824
379
+ vllm/model_executor/layers/fused_moe/deepep_ht_prepare_finalize.py,sha256=3jXq-FQ7m1aASduI-35O_isUZcPe-3jHtkbPv4h038s,14791
380
+ vllm/model_executor/layers/fused_moe/deepep_ll_prepare_finalize.py,sha256=0kfUcceMmnytr9fx1fgamGNLkcLZqS2OZfkFygP_igM,12062
381
+ vllm/model_executor/layers/fused_moe/flashinfer_cutlass_moe.py,sha256=scSk_NF4-pfll7jEkt8qdDAtUSpxlRxYO8G8kZfSmpI,10765
382
+ vllm/model_executor/layers/fused_moe/flashinfer_cutlass_prepare_finalize.py,sha256=6N_KQtguaJVJK8n46l-geXAwi4vPFxTEpnK1kyHuLDk,12210
383
+ vllm/model_executor/layers/fused_moe/flashinfer_trtllm_moe.py,sha256=YMv1Opt0C_LfIDEsk1FsQ0gJkY8btWk1R1wi5R4ORpg,6481
384
+ vllm/model_executor/layers/fused_moe/fused_batched_moe.py,sha256=frQD9kNlwawFbMBYI1lh6HW8VUTZULgoLdZPwOh4o_Q,32100
385
+ vllm/model_executor/layers/fused_moe/fused_marlin_moe.py,sha256=ODktcbnK7-h3gXi41_BX2nONNYLuiTiwZA6v8CKzJS4,27151
386
+ vllm/model_executor/layers/fused_moe/fused_moe.py,sha256=_JFIOLIb7knLkTIoQK7qvFM0l199LTqtx8kvEzGTuCk,75441
387
+ vllm/model_executor/layers/fused_moe/fused_moe_method_base.py,sha256=SSg7uPmlwtT3Cn419FWD4bPYccWhNZCdoxJOgLgpLfg,3571
388
+ vllm/model_executor/layers/fused_moe/fused_moe_modular_method.py,sha256=0I6Bp4x1Hv-eL1uJ8F2UxsLJwyuxqZfVhtxAhzg0k2U,5847
389
+ vllm/model_executor/layers/fused_moe/gpt_oss_triton_kernels_moe.py,sha256=VB3A1KkJiw6EmAVGNih_b6SbUYJ9RypK6umw_BmRctc,10293
390
+ vllm/model_executor/layers/fused_moe/layer.py,sha256=MlcCzIcomQ3ihxu0I96qmo_gvMG3weN7qfuGrxVLjSU,77456
391
+ vllm/model_executor/layers/fused_moe/modular_kernel.py,sha256=1Uh2YVlTbxPbNfLG82HSagX9etRaPNeG7Qlg0ng9XzU,43849
392
+ vllm/model_executor/layers/fused_moe/moe_align_block_size.py,sha256=RCLkxgqt5g5ouCUZHg7IwLzCPXw8LwCmQ9tuZVqOarY,7605
393
+ vllm/model_executor/layers/fused_moe/moe_pallas.py,sha256=7WcySwfOiFkyw-udjrWKjIdzYgcB47NGK3reblL0wrI,3118
394
+ vllm/model_executor/layers/fused_moe/moe_permute_unpermute.py,sha256=WftEGpSesMQmjg5sbvko3LGHuDl8EpySliW5X10Xiwo,8196
395
+ vllm/model_executor/layers/fused_moe/moe_torch_iterative.py,sha256=u_cFfZnz34AliNgZWfQUltCTXP_yTg4NjFxKb0r7zU4,2154
396
+ vllm/model_executor/layers/fused_moe/pplx_prepare_finalize.py,sha256=cEP8YRKHP4l4lcf2m-dNuQmgF5jCRIqEexp2aC44oe4,11776
397
+ vllm/model_executor/layers/fused_moe/prepare_finalize.py,sha256=yDShWGR9Wa1Vt4wQDwHWNGeECsFfZg2jgJHpgiZS5o8,2580
398
+ vllm/model_executor/layers/fused_moe/rocm_aiter_fused_moe.py,sha256=9_d-7nwKC3PJtyfLSJHvDH6fAq8_KI2-4lV2gu9UWGc,9216
399
+ vllm/model_executor/layers/fused_moe/routing_simulator.py,sha256=TY1mPAGcG6SKjRh2kt-wvM1sn6Xd6bHmYmuXjAovBbA,10724
400
+ vllm/model_executor/layers/fused_moe/shared_fused_moe.py,sha256=ePpQVvqAgh0fIglOHCWEI1P9U_s9R1umDdlfyzC3jys,3434
401
+ vllm/model_executor/layers/fused_moe/topk_weight_and_reduce.py,sha256=sLKn82zPW7Yij8Bcrt4sXW78SyuwIwDELfiqsXaTxsw,5801
402
+ vllm/model_executor/layers/fused_moe/triton_deep_gemm_moe.py,sha256=zO6a9hD6tkx0dcohItDS5hEjUCPlQq0q9P_-rNZ__N0,5215
403
+ vllm/model_executor/layers/fused_moe/trtllm_moe.py,sha256=2OUD7vd-ptjUFtIT-9QNdsxT6IKldivr_T5Rr1_JSoQ,4660
404
+ vllm/model_executor/layers/fused_moe/unquantized_fused_moe_method.py,sha256=HxOcqZ0m4CLosBAfsQPR3co_zNbCpfg_Cb14nV0uxwE,22297
405
+ vllm/model_executor/layers/fused_moe/utils.py,sha256=Y_J7w47cIXgx51N1SeG1KJEUfUbut6CZiYScsyv6xLc,10855
406
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=iNGsE2ZeVnQEnN4A8UJ9Jv0d3hbRF2MJ9oBgjup5Szk,2737
407
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=hH5rRN9Wtyv35azxMzyUMHWtiKgOHev5tNjIG8j6dsE,2751
408
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=qPumkNxaHMvVBnEjPe_Xiuz9ICb6Hqc-9I1DAR8s3gA,4130
409
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=s47lb8VLnyxMgWlqcIR4BdPBsjKWL4olXF49uZvygzQ,4140
410
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=1792,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=SkuSt9_xoUgmwy-EIOMJ5DT3pEm7g4dMh78fZSyCaKU,3244
411
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=gzfjrYDcS0vsACq7ONGVkNA3FqVjr3e89q9fO9kokkg,4133
412
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=sIaUmqgtlHW1lvmycafqWUAKyyLuf1L4t_ixzOiVmlA,2744
413
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=yl2lRcoGAKs1OQVKR2JX2C0VkpwUiCXGc_BTvh6h-r8,4146
414
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3072,device_name=NVIDIA_H200,dtype=int8_w8a16.json",sha256=NZUb5kQi0HyBrzhRyPl6pdKwB4F6KqHcJepwHWbxsYY,3248
415
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=H0K4_O1CMbNLi-srcycT3lSl4JaBl3EGF89GY5Rj9MU,4130
416
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=arPqstZMzZjz8BNpY3alKT4vGCJyUj5I2hEeK02aq98,4152
417
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=XcoSAxPTaefnIozjfToNebc2wdctRQQRc0Ju0d-QglY,3243
418
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=wjnQ4v-dflJMR3iFDHBuZI_1R0xXjsNoWc2kHu6C8JI,4135
419
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=7WHPz_0fxeI3Ed0D9VIpZVoeN9RtJVVARvptfcmQu40,4146
420
+ "vllm/model_executor/layers/fused_moe/configs/E=1,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=KPpfnhkIzT9JiiHDLakJlQvoNIzmCCq5NLpoKjZxaOU,3253
421
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=2kWS9Qvy5Q3mvUFmbPVures5iZAriAXsy8WrtE5wu00,3727
422
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI300X.json",sha256=D2dn9vXyN4FCKsZCf7VYgAWLedCx8XpPjbkQVVAvwAA,4737
423
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=2kWS9Qvy5Q3mvUFmbPVures5iZAriAXsy8WrtE5wu00,3727
424
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H100,dtype=fp8_w8a8.json",sha256=IAD1itR3hNQyHzj6AyfNj7G974sVQEaYafewGm_OKdU,2747
425
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=l3y3uQkhlA2_7cbrDHsRslWXR9NzgtkD_BEQckJ-Of8,2746
426
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1024,device_name=NVIDIA_H200.json",sha256=2j-E88dNQghnJuIrCLwUn7QfUI66tieoCt6TIYjMvHQ,2743
427
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=mFxdW1jiRGxoW09LjL8rRHxgiJXXPgtGCCtjmj83UII,3272
428
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=1856,device_name=NVIDIA_L40S.json",sha256=-wcihTQoV82En7NC9eU2QNlm46fjjdAW9ywahwBLzus,3281
429
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=5QqFljwwA8OaPlFnXy1zogl5oi6aE0OqN39xk2IUC64,3245
430
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=I3k416HbXU_rYb8scD8gAI4fuBlElHl06PM347Qa11w,3253
431
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20-3e.json",sha256=CoC3pMKx0vkjI9T6rqRLTIwbDskxljTj31fCHI34B5w,3232
432
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H20.json",sha256=RgV8C4F1LO09h01YsgF_eqX6GNoBtC7ulPfJRUUbg_g,3241
433
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=192,device_name=NVIDIA_H200.json",sha256=nsNEuDNks0tVLfQfIm7xxFwEeptTfQcoa9fJy0NS8xQ,3247
434
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=352,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=LCN-mYQ8xqL_ewIGV95k6EC_dfJtcdfQZp8uZR9Air4,2721
435
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=PvRpT_brUV3Y3zMfWEcsXMmdrYKjiq2qI9iHejPhhsU,3743
436
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=MCTOhQ01id6YjPtWbG6Mw5dlU1xtilsiq3HAstGn36w,3258
437
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=3o_aYn580j2L0ZPdKSTLqrYnginfBOFNhCksS5qxeNA,3258
438
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=qbqjisJ4oKmcYzumHPRk5UyOzsdi8J6xas82UWHMeAI,3263
439
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=PflsK6VWRhwGaDKSIQ9vD7NMHcKLg3B4wENarR0aIq4,3252
440
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20-3e.json",sha256=gkimxy2r78McckKxx4U4R3xahTI1KMH2pMOdUFOUdu8,3234
441
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H20.json",sha256=vS2DRIDOqWyiBvbG6H746ownfkD1F8Aj2YZ0ET9xll8,3232
442
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=MlpzcrkZo78kFYr6cqmh4lBdpxKcEvlzqvRf0bmeduQ,3264
443
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=384,device_name=NVIDIA_H200.json",sha256=xqhl748it8GV2KXX0XixitE_ywnsKksqK8AGL7tAgT8,3254
444
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=FsWbV4Q6AzAtgegVuENBDz2ZcSJsqNiwUIVfQbpP7hQ,3244
445
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=704,device_name=NVIDIA_B200,dtype=fp8_w8a8.json",sha256=W8C1GtP4K43SK9128U52DD5WWofvPleAJE4us2Qju1k,3251
446
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=704,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=ZrC8J_AmZWNx30UDrXG6sHWtFY6FNVPsNywLhEBQdi0,2530
447
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=rN55MyeJ8U6VGNRg7lwC3aa8BgjxdzVg-CofcZ7LTyk,3743
448
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=AMD_Instinct_MI308X.json",sha256=SDxGApRr0Nl8Thz-99Y3vKEz2Aw60juCGo7zSQriMKg,5021
449
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=HqskM2MV6SPZ5LskeOY50lOjFP0DFdYrgRpZFmTpWTo,3256
450
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=L7y3Ev8CbIF-68FWhoMvK9HH72bj6r6-09__zxK-fvo,3257
451
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=IuvyC8TNhCVAmUZfLSoETsyCKsmejKXrs_0zuwFLPAU,3265
452
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Nd2qn9-7yjO6vAQSHAuetP4ytysnd6UxekL8UADiLKg,3254
453
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H20.json",sha256=10Ntu2aVD5vGLonx-jW0qNw-tgZWdZmzMGx7utDVeng,3237
454
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RFH5FcN2ZCPk6DsxviTti1Q8JU5jzBRFXvUQNgOvnmI,3265
455
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=768,device_name=NVIDIA_H200.json",sha256=JraM-Nvbg5V_TJkSl6UPFYZN1zHHoIbr2pAcksenoTY,3248
456
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=bf16.json",sha256=PxeEH6PFy8bFefjyrPS1ikD_-kywMALWnSrGbpSjMQo,1804
457
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=-DrgCsN-PW9kWkr-bK4817zIuyxUa-CQxlONN1-QQAs,1798
458
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=928,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=iplHjPj366nNBu44ZHlCRe7aMaV2T4QthXINmkWi218,3269
459
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=928,device_name=NVIDIA_L40S.json",sha256=l59WyUnnymmJQykoJzEvvkzwlhIlmyEeVtvCFRzXGRM,3279
460
+ "vllm/model_executor/layers/fused_moe/configs/E=128,N=96,device_name=NVIDIA_H20.json",sha256=JtcHRlPz8xQEAqJ9EWI63oYvdmjQFG6VTHqtt85VOSA,3221
461
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=AMD_Instinct_MI300X.json",sha256=f3iM3xm8hGUirJ4ilAIPO6Pe9bs4sm3qaRKMswN9SKE,4731
462
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_B200,dtype=fp8_w8a8.json",sha256=Pux4G7sjfPL22uOJU6t35TXe-VU3OaoPA97TDj9_wZQ,3251
463
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_B200.json",sha256=he873aoOy7KfFg-uMoTFV4IP7Yk0Dk7mOTuLCTqwZZc,3250
464
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H100.json",sha256=Bq57MPQXuSib06u6OwiEmSzOr3XvPYoD6ohYDJaBnII,3244
465
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=95zLafBGPI4zCLqdsIjH7mgblYiBBTzGCItFRtKiQQw,2749
466
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1024,device_name=NVIDIA_H200.json",sha256=h63zG698KnzeDHL20MfxJ3cKiROnBd7GVZ-a6WgbcCw,2738
467
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json",sha256=pCCKkdUzzuBVtljyk7AEIAbeDf12DUiieXaODZXzm5E,3254
468
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=trX2-c4N6hTTD6zFNi6A2bT3FkhxKjkM2rPl-o1K9ss,3250
469
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=I4d56uD7E1JMXD9RAxq3FebdPquDsnNEkVaIY9Ctm9w,3246
470
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=ypuAxMQ7JESPXLBltt68wly2wTrJzlnobhUMip6xAmc,2751
471
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=tUptlureu5QgyAEedtx5sm7CFudXAE6fIXepOb9gfas,2745
472
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=HUWDnLXTE_gHIeMSNJtgxrx5QK3ctTZYRtNim1TJUYg,2750
473
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=h57svdmDlZC_D8w9XWjPRS8ciYVkJiPEYfhrD2NRVVY,4127
474
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=JmXhUnhX6YOy8RsmT0zFLGyNCpRBPV2q2Db9Y9ctZeE,4144
475
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=k7JkvPualha73fUM4GeVI822m-Zfl0P1c8DIw2b4R5A,3253
476
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=vvf0vmcIHMqctQ_X8j1LqlmkuImr46rDUCnl-apb6r0,3039
477
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=_4U3pXWYERfwF0Ig8UTSCmeEI1vKH2irHMEBov6ZNqA,2751
478
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=2048,device_name=NVIDIA_H200.json",sha256=Lc8Wquo4hmXGkUq_HCxkxW1fzDnvE6jHZSEc28-i7lA,2741
479
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=G4PKqWxh0MlBhg7QHKj0m--_fP3Ll0gs7VJaeg-NIDM,3254
480
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=bKX9AvcxN6k-i3RUmHSchZZ3rjoYRYb4iBqhCI4L3MY,3257
481
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=bWR6XBZ4nJ_ROg8rEgrQGc04I3BDbwILDHMZxATO-H4,2740
482
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json",sha256=iUuPfTcmuEDLjHm-lUfJp_-b0mhnf1_RxRglKLJ83Jk,3259
483
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=AP6-GeoO8OtIcGA2kvJseui10HsvmFSjg5bHAaZ_aQ4,2753
484
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3072,device_name=NVIDIA_H200,dtype=int8_w8a16.json",sha256=KVACMsHRnkumqwKwAGIFFvogtqoPZ7XC6iM5KOZD6uU,3248
485
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=_9HO7SaR6aQeh6vqCDpo3kjHnGJ9BVKLiMwYYgd3SmQ,2913
486
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=2ONiQSa9odzdPe1dIgBpP24l5z-5wB1eos06xOj0V_Q,2738
487
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=Twkm9DVNxijpowfvioJ_4cKwIIlAWdyNWO9TA3gxAHs,4149
488
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=1xLr3PpmuUJD912yjzGoMmgvJQlazSao0V4f8nIJBT8,3254
489
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=THQWP1o2bWhnJh0rq3ZIVvs_sagIJgoK4x3pJbiFbHk,2910
490
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json",sha256=o1pR3rNpO1eW4BHOKpPIQLjviw4P2X5Fr4HQBcdHA-I,2747
491
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=iySqae0zI_PRBLqV-vfSCwDS4Jxcl5QjWa2NnhndL0U,2752
492
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=float8.json",sha256=tDqh_env5r2Z3r9WLMH8fJ1oCGb83L7RrHk2eoH02-M,3258
493
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json",sha256=Ha6zrvORRUe8TH4DmVVnHKQIbxX6BnUmPER5rtWJz-U,2748
494
+ "vllm/model_executor/layers/fused_moe/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=ydsFUdXdVE_ZSScVhUxvxOFwKG-nkTraNeN69wqzxIM,2903
495
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=AMD_Instinct_MI300X.json",sha256=AFLeub_dwZHc4qEtvkQW03mfG9fdgeuqTTiP4qC8i1M,4767
496
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=AMD_Instinct_MI350_OAM,dtype=fp8_w8a8.json",sha256=RmeFztUNysKjmg89emtghZlGWl9uKqmlMupww44Q4j4,3735
497
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json",sha256=TtDngG7ljrU5RtWZ7g-xxdBT3uEuawiKhP8EwPr97XM,3254
498
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=192,device_name=NVIDIA_H20-3e.json",sha256=u09XGUdUQqSDasrUgOQeu7ydp5ft5US1oSM0iT-BT3M,3235
499
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=320,device_name=NVIDIA_H20-3e.json",sha256=JBO_l8hfxCbiQ_PvoqS3Xxqlz9i5PP7QFfXjGKLf7bw,3237
500
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=Hjpi5BF5UjT0R0uuluIeZMWihlAM9zyWdV5Y2BJu0_w,3730
501
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=384,device_name=AMD_Instinct_MI350_OAM,dtype=fp8_w8a8.json",sha256=KudLj1Md-XBFWsivs_7aGg-UQno7H6tzxyYqHQKoNg0,3734
502
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=384,device_name=AMD_Instinct_MI355_OAM,dtype=fp8_w8a8.json",sha256=6QmGz-7RZzAT1-XS2twHGAbEpjm2Tpkn7qjfBMqm76g,3736
503
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=640,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=_HrEta1vlWll_2c0v6RpEIUQirMD1QaOMU81VaJh9Nc,3254
504
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=640,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=HvSQqJi8Lb7NBGmvp9YeF8k3nB0t94DJ-LvCX6Spjyk,3255
505
+ "vllm/model_executor/layers/fused_moe/configs/E=160,N=640,device_name=NVIDIA_H100,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=EZV3ffICZEq6X16UNFpHL7kOuV_qmj1L8fJu_F1c-DM,3260
506
+ "vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=rKQwdgkictjg9DjBFVzHq9LOMlX_Ul27jllTABjaUtU,3252
507
+ "vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=qKcB2Ka8jkahKCNF215Ec6rCoy31OrElDL1ZHNgCs3M,3252
508
+ "vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_H100,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=K1aeYxtOM4niB6auzCEBVudmKIKNzHPjMoFeQt_KD-A,3263
509
+ "vllm/model_executor/layers/fused_moe/configs/E=20,N=2560,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=rtOl3v_tR1iCF-Em0KihtLFzK5qLtndPK-jKCERLbNg,3264
510
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325X,block_shape=[128,128].json",sha256=fT7fwjuit4HbbyREYV3ECJ9Rm88FW-V54e27nG9nA_Q,4741
511
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=fT7fwjuit4HbbyREYV3ECJ9Rm88FW-V54e27nG9nA_Q,4741
512
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=HNvrgcXxV-eVMLwb7zY_R5KgJ7uBz-YIyQsKq1lWnWA,3263
513
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json",sha256=bHJEVy-CeImiY9JBRCMlHfHPAUi5xO7ENxgVVboN2Yo,3258
514
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=PnNmKSRFznCIUzZ4ZfaYTrMHeF2_kCQr4_bsEy_9Zu8,3259
515
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json",sha256=0Vlxxzp4wrvkFj-NF4OAsJAaPkm-hhisJg0tgNl-W9g,3254
516
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=0aSYzpv_cBAlpWCPrfGgNTCfae1KdKQnT56E8XFQl7A,3262
517
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Lqom_VMIPduSZTZQdeL2Wl_x3r9q6RmI9bojJrYwQZ4,3255
518
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=fd2p65T9OboKIgw7MQc4IdKaJsoO73Nu3VQiKjV6Ffk,3261
519
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=FUGuYbs_QhqKfErofvbTUplhAVN465A7NR_-ryXvebE,3741
520
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=bpDPbTyrXLyCSy-o0diveVVeVUF_xj-fdSzCzWmEcKA,4733
521
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=bpDPbTyrXLyCSy-o0diveVVeVUF_xj-fdSzCzWmEcKA,4733
522
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=skSJdv0Pr4rba5ODxp-fHZ6dpxn8KkvACGzNf74j81I,3257
523
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=wMt0NyoRSdACdmS1Qi3qFiu6GiFX-4lVvbGEno1W4zE,3252
524
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=mtm7RgEBEJJkHsOis9BtAFo1OCk3vBbt7l7eumDzd7k,3263
525
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=kfEjBrcwt21QRzimrgY_SQ0syJVJzx0AcWQcKQ3j5sQ,3254
526
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=R4B2n2vGt4pPo6jS4Bmnx8AYtcfF9qQJE5bD7OhmXHs,3265
527
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=JnqtO0t2HBcQECdYavi18mu9_MwblGr4zfRcW4zU7_c,3265
528
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=b9odBjWC1mjQRIIZIFV8UPMTVE4axTP8vkmOXfdUids,3296
529
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=bpDPbTyrXLyCSy-o0diveVVeVUF_xj-fdSzCzWmEcKA,4733
530
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=512,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=l8y606GTM4Xnd9CBdhjt7LuA_1KLQS41PInHKNfZAZM,3242
531
+ "vllm/model_executor/layers/fused_moe/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json",sha256=rVORXxNsxy4WmO5SJR8Sd4k7vozKqhYf50wZNCMeQzs,3239
532
+ "vllm/model_executor/layers/fused_moe/configs/E=32,N=1408,device_name=NVIDIA_B200.json",sha256=PoVwSGuqyF_0DrRTrggdmlfhUn2dmSbc-pP4PjMEF4k,3278
533
+ "vllm/model_executor/layers/fused_moe/configs/E=32,N=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=r0vxILYelmq7KarzO-3yjv9fA0Z4ORzCo7u_UgWbdXQ,3292
534
+ "vllm/model_executor/layers/fused_moe/configs/E=32,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=IOYqUudflMQ7PBGNxFYihtuVqb-J4eVp5BmL_OnIFS4,3287
535
+ "vllm/model_executor/layers/fused_moe/configs/E=384,N=128,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=ZvsFOizI7EY5V_V5gdjyTgbdJfeKrbxhISxcppSfduo,3255
536
+ "vllm/model_executor/layers/fused_moe/configs/E=384,N=128,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=1XFStvhg3yV7cMwJcpbLfaCY2B3nqZNrJgohJn0ma5g,3254
537
+ "vllm/model_executor/layers/fused_moe/configs/E=384,N=128,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=iXoa10iZi9EQ5jAiMIu3hCemsvjiWVBtn4rKVarVCCA,3256
538
+ "vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=BdVjlDZWnX6_rduBTQWfc6Gb4o9z8JRlI5XF00CnJC8,3255
539
+ "vllm/model_executor/layers/fused_moe/configs/E=384,N=256,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=KPI0AqxGlxtJ18tj-NiFjBleS_ihwmLSt5pg3GXGJ-c,3255
540
+ "vllm/model_executor/layers/fused_moe/configs/E=40,N=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8.json",sha256=pwoiPdk-veNV7Um7os7sG5UrRjKgOsbyxMfQG4z4r9g,3285
541
+ "vllm/model_executor/layers/fused_moe/configs/E=40,N=2560,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=S3nHQ7okmQlkuz7gQnHmZWcQQnFDacB4Lz1isbdZ49k,3258
542
+ "vllm/model_executor/layers/fused_moe/configs/E=40,N=2560,device_name=NVIDIA_GB200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=86ajYfCiq1urzJxcPrvyx884njCMKizCrS6hfx69WPM,3252
543
+ "vllm/model_executor/layers/fused_moe/configs/E=40,N=2560,device_name=NVIDIA_H100,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=86HDa0gqUZkPaOJLmuax-j-HBC7t5P48NuKMpH5Uj3g,3262
544
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=qZgSX927zWhvdQc7KjSBZt-Wk11LTwFIK8UtEU9AFJM,3274
545
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_B200.json",sha256=NBj9yhhUWHpNcvwYhPtBKEn1It7BbrnRMRN6Ji2Iazo,3234
546
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_GB200,dtype=fp8_w8a8.json",sha256=XP1z8IdGklVXVkMIbCtAzWj8nUz65GIZaqZvBvD7v4M,3242
547
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=7FM4MDavNVOcEOFldsZs5H-E5O6EIAHDY9n-aiLY_Kg,3238
548
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H20-3e.json",sha256=DaChAo79RafXGjpAOeO1CQ9DazPk7Vazb5bAEyDOYSA,3233
549
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=128,device_name=NVIDIA_H200.json",sha256=BB3qPMREaACzWHmhvcETQHLDYGX7GXTKr1E5cm29uC0,3235
550
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_B200.json",sha256=0avMzgfoHRzULbxIdzI1SVCWUSBP10kKvxy-3Mv_y_M,3243
551
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_GB200,dtype=fp8_w8a8.json",sha256=yIEwME0J6ZNfGq13Vj7OA9DmUyKdFZzFj8uaLVb2Yp8,3246
552
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=YopFXEwMf6qpCwrQ5OdJtGVsDvvaQFUcEuGGIFvmS4w,3286
553
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=pnK1HQ2B5ECqCBolRa5Yb4qAClVoJ2pidBnPy3GBLeY,3244
554
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H20-3e.json",sha256=q5AAAH8gIaU--3mXhSF1VdFTFHNAy5c-gUvYcm9qhEg,3235
555
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=256,device_name=NVIDIA_H200.json",sha256=he8rleYTT40kpshJW1FsdiyR0nRU367CqytS-9-UZNs,3243
556
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_B200.json",sha256=LcMD2JddiX488ddRCsh0cXaf3my0VT0IweqQDyTWZwc,3245
557
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_GB200,dtype=fp8_w8a8.json",sha256=44VdfyrLo-gl23cMeSqSfoGrsqWcjv5-lXwU5y5spE8,3250
558
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=SyMqNl2wL11cbj95x14YynB4YwyQlaOf1yPPIK0jcJQ,3249
559
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_H20-3e.json",sha256=npr855kvWlHILzi5M0sjYjofPAO9bf6WCF2oZZ4u3to,3236
560
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=512,device_name=NVIDIA_H200.json",sha256=xR_v4wy8_ae9fGyuTnhWY0d29NwC9ChPKvdCK5_PS2Y,3244
561
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=9M3IlMbbIrCvfhQQHsY08NAOoaR8pdd57E30JYPKlRk,3264
562
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_B200.json",sha256=femFBZsNptZ6DlQ32dpBu4zhFxaG-kcs4MFPTxipui0,3234
563
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_H20-3e.json",sha256=iVnryOC43Rirq38PwPxzIHrWvf6pA4wHZtf1HOgGtlI,3232
564
+ "vllm/model_executor/layers/fused_moe/configs/E=512,N=64,device_name=NVIDIA_H200.json",sha256=K1HkIsowgSbvwtAppZZPudYIIni_waoSxDkARTCEQso,3238
565
+ "vllm/model_executor/layers/fused_moe/configs/E=60,N=1408,device_name=AMD_Instinct_MI300X.json",sha256=4UXbsSNHmrSWnD85SdRMLp4cFGRufndzJjB6hoQPclU,4736
566
+ "vllm/model_executor/layers/fused_moe/configs/E=60,N=176,device_name=AMD_Instinct_MI300X.json",sha256=p6TKUp-KDeLB9E9LqThR1e7J2-ogSXPJojISdHgCxaY,4727
567
+ "vllm/model_executor/layers/fused_moe/configs/E=60,N=352,device_name=AMD_Instinct_MI300X.json",sha256=gHxtmO_uvpueLVlsJgXBVE3_pS1S9EeRxNmHG_ZQszg,4729
568
+ "vllm/model_executor/layers/fused_moe/configs/E=60,N=704,device_name=AMD_Instinct_MI300X.json",sha256=tVdpbIU1scsylx6oz3IADhkcwvZaNqw-_QVb7a6oVX8,4732
569
+ "vllm/model_executor/layers/fused_moe/configs/E=62,N=128,device_name=AMD_Instinct_MI300X.json",sha256=_Mn_3ZuKGqNnXtcUUZA0rr29c_lcp-b32uw57oYHgVM,4721
570
+ "vllm/model_executor/layers/fused_moe/configs/E=62,N=256,device_name=AMD_Instinct_MI300X.json",sha256=2eFdAHTU_YTTVad_OEcYZcSbtm5gtlHg4ChxU_v1a_k,4725
571
+ "vllm/model_executor/layers/fused_moe/configs/E=62,N=256,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=sCmfFCa-I5Nft9ap-ScL0PVEKZFibkhtVslbFs_NLQ8,3234
572
+ "vllm/model_executor/layers/fused_moe/configs/E=62,N=512,device_name=AMD_Instinct_MI300X.json",sha256=kKu6HpksA-NKVS4CA2ivjlAQEshl1HEufQ0Qt-lTZGM,4732
573
+ "vllm/model_executor/layers/fused_moe/configs/E=62,N=512,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=KmYAPOgz-2qn2c5lY8u9XRy8i8HNkmOR5m45TIuwt4s,3235
574
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=6QPLhZurIqcMVdy3w0Dd7gLViKxsyJRBz-qd8agpi6Q,3248
575
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json",sha256=WPu80-OWyEJBy1hdnewLN1H1neFW8UVJrqyeDGegXc0,3250
576
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=PaFLbT5ftJiiVSOVkq_DH01EcbIs0sBVkCd9PdYYmw4,3253
577
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=ozS2ECxk-Dsd4Y9DgCGGwDwJlCf5T20ANf5gnTUMuSc,3252
578
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=KEN6xt8pgPH_FbLT2fsAD4s03_V-Z9GXuEC4IKe3cPg,3262
579
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1280,device_name=NVIDIA_H200.json",sha256=w18R3eHB4oUhfbcCXjHyDvp0RiDSeCrfM-VFESim2hQ,3253
580
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1408,device_name=NVIDIA_B200.json",sha256=4MqYiUxVPLgkiaTl1vZ5Q5wCwsLYNp7O2-kbOSYmKjI,3277
581
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8.json",sha256=Nm3LPD8JxBYGflI42Cy-xyxZlBLrGvVbnkf9NUmx92U,3250
582
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=iz4W1UAV1fcz1ZFh4hNQSLJ_F1MdXW-V3msy7t0WrRM,3262
583
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=dYpKgvuG7Jji0W0zg_E9NfIojStBAdBcKd4B3nhimqk,3263
584
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=2560,device_name=NVIDIA_H200.json",sha256=CXiHlGpea5cEGmFi28Jec34uxEZITF2XldVFcJteZX0,3251
585
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=3072,device_name=NVIDIA_H20,dtype=fp8_w8a8.json",sha256=hLFfmEnpHDZlwBlx7TzfxbjCEywqHEuhjllVR7g9yMY,3259
586
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=3072,device_name=NVIDIA_H20.json",sha256=g3M9w0Noi3DyQB7fcr3jUM62_LKZvTwbY2GtzDGd9_o,3251
587
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=W1q4PfievvgJ_SiPsDhOsR0Q0eJKb4o8JZhMcVhC-_4,3264
588
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=tku4-yTbIr0H5TNrm1Pq3tJJFYTXqHpdzJDSEF3bk9A,3238
589
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=HJcV-Tzt-yojzNQkPCgi84B44F_RppXxOIicRyg20-U,3264
590
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=320,device_name=NVIDIA_H200.json",sha256=bM9g-XpirsThO3Q2x8ChSx3PPtHuHRXLvVMnTWt8jLI,3243
591
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8.json",sha256=9vB_5KLq88zBCNpyz6LE7qAo2eS_sgTNsOh6awGnjT0,3235
592
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=384,device_name=NVIDIA_H20.json",sha256=b6lqVlPt68GZ1wlYu3UtC6zkXnnnKwh4tKGltXNvTVY,3235
593
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=oxOKFDrgmw1YmgxTtRa1uoe3p09ylTLrkj_jOTqNh1Q,3249
594
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json",sha256=-B6gZAEYLwMJZOnpO81pTxqs-YVKs_144Nn9BSLaMh0,3247
595
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json",sha256=GPjPHicomrS7ntHu7nnvgNXcHCoUw9vhyTUewkXpppo,3252
596
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=ObHUCUAgHTnld8Cq9Dy1n3ilmbBzyNC4jZcz6YYhMXA,3264
597
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=WegYsHl39QVlHu_4EZJSrgA4LQ5fYxSVNWFhoL6W2Rc,3251
598
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=Hrlas0Nt7d3JMr1vTpI3OVgkzxqcRziSMfFf_U5pQ58,3267
599
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=640,device_name=NVIDIA_H200.json",sha256=J59rmqF8NQWkqmay__ahA3t3IwaPXNu5AVNLnTaDfYA,3252
600
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H100_PCIe,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=d_tIZcrWnGVvgmXBpIbsFMAKKC9-1eP6nbZSqWgzOm8,3288
601
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8.json",sha256=0bH7NZ6VmBBycMus6v7sDXCXtort9awuEpttLLMTZms,3242
602
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=768,device_name=NVIDIA_H20.json",sha256=VptbMpckYKHKYMJZS-EaO9G0cSL6nciL9XyeHiZN4Fg,3237
603
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=896,device_name=NVIDIA_H20.json",sha256=GNbp4W4MBoHHN4-0sXJovY0lX6rHfZzGyKicrumupGQ,3225
604
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=bf16.json",sha256=AbwYfuQiAbPJb3Okwd-BCTWs2P2Ao9REEh2sFOkoMuo,1804
605
+ "vllm/model_executor/layers/fused_moe/configs/E=64,N=8960,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=FiUQOl86K_LlCd_RmJR3BpDM8whX07CE6Q-6pp1rR9w,1797
606
+ "vllm/model_executor/layers/fused_moe/configs/E=72,N=192,device_name=AMD_Instinct_MI300X.json",sha256=r36dUUfR5rv7yS_I8W56z99VAa8LIfaNRQko_97R14c,4720
607
+ "vllm/model_executor/layers/fused_moe/configs/E=72,N=384,device_name=AMD_Instinct_MI300X.json",sha256=XMMCswkzdQxVfOBDM0KXG_ryA9JnX4hFKhF2QspifsY,4724
608
+ "vllm/model_executor/layers/fused_moe/configs/E=72,N=384,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=56iA4xNOSuaGWgz2Qg-NiROhpdoBMto4NVUnisUsyD8,3238
609
+ "vllm/model_executor/layers/fused_moe/configs/E=72,N=768,device_name=AMD_Instinct_MI300X.json",sha256=68SI7XVkxHS0k9ApTW1-nSnIqOtbe8rmatNexWRFyqU,4736
610
+ "vllm/model_executor/layers/fused_moe/configs/E=72,N=768,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=ICdOnZoM8waiOVeVYP18QiPDSbtg9Q6yO-OrO0-2xtI,3242
611
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=X8FVPE7rLblDs_Dw_Iu-KDw9H7PaC417EHyVclYjfv8,3733
612
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json",sha256=FsIv5bqSpkWbxK2dBfg1N6tX9epZ55ZhgkJCD7hENlY,4733
613
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=CnjQX3SlQn6fIGsX6P_dbNO0TYgAd-sVUb1FfDcDFUo,3732
614
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json",sha256=fnO-v4YqBz0vUo0UtOTTD0n7VDG_ivczeQ1tR6Qm9f0,4734
615
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=V_sgDtEtGEuBsGVa0maYJHhhGqe1NE7l-1ek2ed9WP8,3082
616
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=QaITFIJU4UsrOBXaGdPYJwTmYJ0nT9kiiqeUiZzvd1k,3270
617
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=14336,device_name=NVIDIA_H200.json",sha256=CC_jsMhXzrYne7eIOroDa0fCBKNnffiaVW2TKd4P-ek,3260
618
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=LgHbxG1kQV36zZPkJcnurHYzwAjMh04lvEHEsfzS1t0,3732
619
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI300X.json",sha256=_fcdkmWvdMqHiH8ZAGke-zXhH7qVPQx5CmKELW5hRCA,4735
620
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=mVH8Rl4sLATinf7_0A9lTS83kv1E7Cm9oC0BL-pc9n4,3732
621
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=16384,device_name=AMD_Instinct_MI325X.json",sha256=JKYW21c0CzR0fgE5ZnYp6C1sY_tVRlm8L_lgak5V5zE,4736
622
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=yTf2R9cngSf4OafucAYlDDn4-bftaMFKaY7qhaBZPqQ,3739
623
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json",sha256=_1eVE7ok935L2V43-3D3bVNWSVaoViia19sh0VrXmXM,4735
624
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=5exlPUKvZxGDR0UT4_Dn5fp-_ZETJ6_Dbw_Vk1u8bbE,3735
625
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json",sha256=18v6YruKbQ95pXPV8ocV4VdM1zNw3aZFp3WByeUkNSM,4736
626
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json",sha256=AffDc0_51ML8HiA3757zbD10TZJdUsUDIYIqO4g0yUw,3250
627
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=IEYBNjt9HGnzoOVSWvL0A0jUqq926QD0_BvVYR4RA1Y,3252
628
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=Ns9Y12aZbJnFhcG3nwb67bDqqiQAo9tdTAIe8K2Ajz4,3255
629
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=uGSLFPZXK_JQ3GTDUAEiIecDor1yjbC3bJvMolF0Xl8,3267
630
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=1792,device_name=NVIDIA_H200.json",sha256=8q6ol5JQBWj6yVfzFOn7Gz5MSXTaW9javL7qQmYVOwg,3245
631
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=6jRC0oOpVpq5c1xePFKNRy-Xtmb038i4LE9N2zao2W4,3730
632
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI300X.json",sha256=cFWeyNJtEbs-Bfohgzclxo1rcYGU863oV0BzJyQ4T0w,4734
633
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=SMtsqtQeqcyy8aNwl9hPxRvx_XQdT7I3SBDNJ3OIvwY,3728
634
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=AMD_Instinct_MI325X.json",sha256=ZyOFJB6GUgGZsAjjT43XJwG8P-QrZ5yTvmgzQP7ThQY,4734
635
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=HOxWmCI2ifHmWc0or2y8nEen86jDeLDov1-tuMzuhxo,3256
636
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=csHezh0HGWaNwrblGzMgcE95hqbqjWS8HImLRJYr_ts,3266
637
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=_5weLBinQCDzyV75hHKIT95Y0ce94KWft2_5BC6EkbQ,3254
638
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=BTpwe2RgMbzP9MTtbcJ16I1IAK0ghD0rauWEea8TOKE,3446
639
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=DlatRLPaSr8HJuO50gRZ2lzXoelx55EP3SDUdgIT2v4,3269
640
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=2048,device_name=NVIDIA_H200.json",sha256=TXSOoqvi-x8H13xPqrB9qz2T3opEGA-2D0v_4n5BEG4,3259
641
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=ro3drDpWAdeXH7IjMvx8wYGhIuDPOl0bpbJaIB5Msns,3732
642
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json",sha256=w_R2LL8k5jNVUARcqvSgGLvNoQiQC0Mh73ciqSIAz54,4734
643
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=hjDoTXRmEFLKhhmBFEjPowQus_z23ISonxFljql3c9k,3732
644
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json",sha256=AdOTy7ASetdAXUhNM8buoU8_rLLjcUYF0m8RGFrLWRo,4733
645
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json",sha256=Ru460ZgnUP4U8OsJfwF8n-AI-gfcolNR3_qzoxG6DtY,3254
646
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=K6BGrKw_oHTAtHjsZldcjp-BUM1dIecKXrrRn9OpRGs,3254
647
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json",sha256=4aK_plqztXcJ-hs5_PsAvM0jclMzcO3hd3zTo0FhDro,3251
648
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=qqFoMaObuO8pFWcSb9q0wYsdC4eSCO7B-_ruQhR1N9M,3264
649
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=-5nkLIunjG1ghPoUEtt2AXEQw9oGiilP7K3UvQv9CqE,3252
650
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=WKzddrIXo-KavpuXuouW3aLLAptu5Q4XJUb5K2PLgDM,3262
651
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_H200.json",sha256=ad1ZkkSyLJwRGb4Kf24qg5hW_DPmt0BXrKR85oAiV34,3257
652
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=3584,device_name=NVIDIA_L40S.json",sha256=qX5_yErBEwDRzhv2FvxrS3pEMa8zn0GHzLp5TUMX90g,3872
653
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=ysRCWmxV20K2BYD9XEUtxwREFGtA3QHI191vHRA0k_Q,3733
654
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X.json",sha256=L8VA1sfygHoyLJ-Ybfs8DP5c0YWFmMkwxHT8yJ9PEFM,4732
655
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=FJWpDLr13XF3hHiHfJykpjbLiP7Ccu2en3U6BL-QwXw,3732
656
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X.json",sha256=FnVcfzf5gXkQRt0XgsRzIQVbDPaUDOwWJX_9qOlyvRc,4731
657
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=DxYu8regZOSFu8ugFGA_QbwWK4g8xwQUZF9a_nNY4Cs,3255
658
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=obzfE_9XgsbFNfC9biYOHxR-V_Bgc7PKT8qZZJaiJJc,3262
659
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=qwKy8oaMsd3QrXgQbM_x9xcfYiHK_Ou1CEwDPL5Gbgo,3259
660
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=qUifbWbE4cOKZbIHWmmLx68VRaslQX69eZHwRIQx-7I,3269
661
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=4096,device_name=NVIDIA_H200.json",sha256=JT-ZMLhAqqzSkqivOW5ATTKRlyyaFQkqQDnaPS4DE10,3262
662
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=QsR-Xr9vyuiArMTSo-dX-1DFgATfqwIGOzFuQJAuE_Y,3734
663
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json",sha256=EtVorGY4khTEuimlqZu0AAlPz84PH3ZkDZmVpxLtgQw,4735
664
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=D3wX0_s_ylo3nLIUfaWZmGYtMvX7oiieOLMdQ9k7mng,3734
665
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json",sha256=JPdO0azlh4yUvbpC9dEHYpRT11ELEr5LXBSb5XP4E_4,4735
666
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json",sha256=BAJnXTZoewwCtzJLUPJ0oYuALv640MvDuLseGcsYaaw,3252
667
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=-Tj7ImS6ZFDof_0VTyq7kVm8XD9B54RD6CUOPSf3Jjg,3265
668
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json",sha256=tme0ydWzIxdABZLk4tU8G_X2dJUYGGZNkQzNGcmcvUc,3261
669
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=g6Ivy4wvadaCAMJ4ZElbUU-CwyTMdbaa49M7IVQhVjk,3273
670
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=7168,device_name=NVIDIA_H200.json",sha256=GstQosPPHUn_I2DV3eMGtn3xXOw6kl1hb8L0EvRsbEU,3261
671
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json",sha256=kF4Fx0yHUmiMSLFNXT6xqAEA4AgCaHOoy_3irv4dNss,3732
672
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X.json",sha256=uOlVzTdJl_4VrRK4wmxIb8JKfveFZRjO9syjw_oEeL0,4732
673
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json",sha256=plnx7r9jkcYXkhvapbeeNvUg3NMGdGsIgIPSrfVy2qU,3733
674
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X.json",sha256=UC-iTgh8_dUSXRaYHOIhDH31KOiJmcfqM_Bv_UBf3ks,4733
675
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json",sha256=sY2nWMPh9lsIkhPCjkHO245wpnfFbrHmzdcZDVFPVww,3265
676
+ "vllm/model_executor/layers/fused_moe/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json",sha256=WQLKugnKzlQ0avf1N-41lRHtG6wJ56DfVPv_nip6NBc,3273
677
+ vllm/model_executor/layers/fused_moe/configs/README,sha256=W2yIZkP9O8GGlg97We9BJfTtWUtPbuz5ZH3esrrjBX0,572
678
+ vllm/model_executor/layers/mamba/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
679
+ vllm/model_executor/layers/mamba/abstract.py,sha256=C3tXQm9WU1XBmYJBfEIOSK805-Iml8kS76svt0kLQ2g,2410
680
+ vllm/model_executor/layers/mamba/linear_attn.py,sha256=phy1X6MlklahGfw7XYlcizfHsbFY_zcJusgsrHUKBFU,14798
681
+ vllm/model_executor/layers/mamba/mamba_mixer.py,sha256=5srTbHZ5IisVC_hSAXYMZX44w8JAEPjMMD7yvToBZNQ,20125
682
+ vllm/model_executor/layers/mamba/mamba_mixer2.py,sha256=GD1aJOsyGfzs1d6hiRqmtDz4qtpnDZ7QJN2efL-459Y,37398
683
+ vllm/model_executor/layers/mamba/mamba_utils.py,sha256=UevD33ksJN7oTSto0uUpHUZQfv2EVmMMW8YQgtKuoFA,7511
684
+ vllm/model_executor/layers/mamba/short_conv.py,sha256=emxa6pH0an67sITdP4h7dUkmetmGXHxrNo_3ZoZSwPg,8752
685
+ vllm/model_executor/layers/mamba/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
686
+ vllm/model_executor/layers/mamba/ops/causal_conv1d.py,sha256=q-isreyIw522evOyXtC2ij80eK4IyejfhsLOZxIu9P4,49540
687
+ vllm/model_executor/layers/mamba/ops/layernorm_gated.py,sha256=KA5JRI8jEStDEkN6U6dqaOBQSCWugxUelAhZISZyf7A,5228
688
+ vllm/model_executor/layers/mamba/ops/mamba_ssm.py,sha256=uD__R1MjEURDEiubfvKIUqftP9Z-COIsSBvmVHHj-Js,15494
689
+ vllm/model_executor/layers/mamba/ops/ssd_bmm.py,sha256=mlQ258Jsexnb5vHGh09YW0l-rF3SZM4H1R29o3hb2-s,6842
690
+ vllm/model_executor/layers/mamba/ops/ssd_chunk_scan.py,sha256=Yow2h10PSWWKvrwT3TRz7UZTCq--99j7MZToXN59aCc,15550
691
+ vllm/model_executor/layers/mamba/ops/ssd_chunk_state.py,sha256=6-0Dl2-frs5av7ra-H0xHWPupVaSo0Vex4d5vZVq41Q,24226
692
+ vllm/model_executor/layers/mamba/ops/ssd_combined.py,sha256=oKPphPh0wjU3XT3k5U2t9cHb4jDLuYzIQuWK9w0sOcs,7353
693
+ vllm/model_executor/layers/mamba/ops/ssd_state_passing.py,sha256=OWIEJ6WjdYHbX_7HRzhPt50oejUjvzhBTST6bQPhjas,5275
694
+ vllm/model_executor/layers/quantization/__init__.py,sha256=41XSA8W_77Ab3xCaN-SYZxU55xsPPlW40aBbgFEyKBg,5560
695
+ vllm/model_executor/layers/quantization/auto_round.py,sha256=5-kg2W5BSF_scMM85-EFlS3-qbhzQRpXEUu9-HPS5Ko,16822
696
+ vllm/model_executor/layers/quantization/awq.py,sha256=abPEn2kk_OUtOHsRdwMIxJgc1SJYkakBxsbKwDp6NsM,10010
697
+ vllm/model_executor/layers/quantization/awq_marlin.py,sha256=nGhAcgIDeY-grWyqsmqbO1yR0hCWxneBAC-9WM2LQeQ,23834
698
+ vllm/model_executor/layers/quantization/awq_triton.py,sha256=fRH9rX5jDSqQtdmcrXwxtPtQEiKxTUdPkAJmtvW_zng,11620
699
+ vllm/model_executor/layers/quantization/base_config.py,sha256=1iCxfCdxsz4ZJ64SzDK6QUsDsW50jJgb8Q3K4K8C0yk,5765
700
+ vllm/model_executor/layers/quantization/bitblas.py,sha256=QGNzQWDLRRvohawjWVUVMLoK9_GyKDEEjVDm70qQLgY,17358
701
+ vllm/model_executor/layers/quantization/bitsandbytes.py,sha256=TwbUzwjNEqrIs7MmlIq3uv0N1z8KNi5bM3QG11cf_QU,22393
702
+ vllm/model_executor/layers/quantization/deepspeedfp.py,sha256=f8fKA4ZPKhEFCuv7jcZIrC8IQd6wXDecddoWbMSkCzE,7278
703
+ vllm/model_executor/layers/quantization/experts_int8.py,sha256=_LGcXs2FPQMTmagfy1a6F_um34VzVIRnOjXZagPWQvY,8319
704
+ vllm/model_executor/layers/quantization/fbgemm_fp8.py,sha256=2SZDxFbn1T_FVhR6hGj-c-6oyGw7fJQqTHbpSkY4PCI,6621
705
+ vllm/model_executor/layers/quantization/fp8.py,sha256=0nfH8mCxP86zW8W37Xmr2eBB84ELkDZ6ovoKXWweR-4,55121
706
+ vllm/model_executor/layers/quantization/fp_quant.py,sha256=lp6PRY-Hwu5QfUChi4yJ3z9shstEnQzUHLlkEnDNN6c,12858
707
+ vllm/model_executor/layers/quantization/gguf.py,sha256=VN3a3zm1TS3PQQjYlxdkvMH6sE3gQFcQVcJxvchHRDo,22103
708
+ vllm/model_executor/layers/quantization/gptq.py,sha256=ZuszLML4Og7f8eTZmdEEwCZAAeNj9YjA9wz46wIR46E,14814
709
+ vllm/model_executor/layers/quantization/gptq_bitblas.py,sha256=rPSt1jT8dyqdslSyGHd_hDB3uIUVRalCbA3SHMs0R3I,16626
710
+ vllm/model_executor/layers/quantization/gptq_marlin.py,sha256=SPJJ5mMxGkO2XhxGjWkzpFpWaL8AOCYD_FVkQDci4QE,29518
711
+ vllm/model_executor/layers/quantization/gptq_marlin_24.py,sha256=PNVtvV2PCMPU_J1UtXlzrZmeOMfEXEfR5_KSYLZlUoY,10559
712
+ vllm/model_executor/layers/quantization/hqq_marlin.py,sha256=pHHb3gAokVXbxvmB4n08Pw7YYQdXbKelPyFb_RWJews,12484
713
+ vllm/model_executor/layers/quantization/inc.py,sha256=mJJ0pQCm-rFafcbsv-0BBK9AooCtQ4nK5OdpPQfg2Ds,2252
714
+ vllm/model_executor/layers/quantization/input_quant_fp8.py,sha256=6ep_71E7Yc0Y3-1DCOWWcRI-Kz32MsITjulFLpQf3Jk,6484
715
+ vllm/model_executor/layers/quantization/ipex_quant.py,sha256=er3Kwl6wzOVq2zJ3NRhM4eWOv6ytldwoyDVs9kiH6io,17442
716
+ vllm/model_executor/layers/quantization/kv_cache.py,sha256=K0Kn8peguLbNHvkl25ZK8WH-_hCMn8OS1O7D4XZMxjg,6210
717
+ vllm/model_executor/layers/quantization/modelopt.py,sha256=PAJbhwA2w0bicqfikdyr6tI6UNlaEooFEA4Ff2VIsrU,69701
718
+ vllm/model_executor/layers/quantization/moe_wna16.py,sha256=dvPpBQoR7HIybOlchOpgrkyj6ExFXZavDik3fPByboY,21345
719
+ vllm/model_executor/layers/quantization/mxfp4.py,sha256=KRNCCdoYN8-8oOSZr5SSnAI093fBPZLQjmzOw7r-P3o,46830
720
+ vllm/model_executor/layers/quantization/petit.py,sha256=6mQCXY7L5jAthFNwfRLhWujxcXJIw1mP-pTLI_BOY7U,11594
721
+ vllm/model_executor/layers/quantization/ptpc_fp8.py,sha256=2jXCOzOePM1dVAc0c3S-bKyYf0nOZP9MY6q5dni4934,5122
722
+ vllm/model_executor/layers/quantization/qutlass_utils.py,sha256=lLu6gVZoWiitJh0VXCPIAXpizT472Eyqi0GcSriVFx8,5641
723
+ vllm/model_executor/layers/quantization/rtn.py,sha256=NEIUwViEQOq2vEvYyA6oTPh2P6ZuyWOnqsZhr1upKSE,21201
724
+ vllm/model_executor/layers/quantization/schema.py,sha256=bOHc_0KCkTYR4BhpgGG6Sc1g0legFhDyt8S9HjjFItw,3807
725
+ vllm/model_executor/layers/quantization/torchao.py,sha256=y1F-T5v7SFyT7ktl9HH-q5rW92oa7DZY8Q2MK8CON_A,13454
726
+ vllm/model_executor/layers/quantization/tpu_int8.py,sha256=NG9vRjqpPFdPPmH8VJagcqInBYTKNslUNBVWfGQwq40,4681
727
+ vllm/model_executor/layers/quantization/compressed_tensors/__init__.py,sha256=PRLyXFLYk26eH3Y5R_HtJSmz_eK5Fw7C-1liuvMlz3I,108
728
+ vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py,sha256=ttoclKgFOmrNXtlc8wkZds8aCwPNbTBqugXCrK4DeuQ,35958
729
+ vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py,sha256=Vm6SIo6Bu7hvt7QsHqdrEm9hAtksLtCFauJz_jY1ekc,89278
730
+ vllm/model_executor/layers/quantization/compressed_tensors/triton_scaled_mm.py,sha256=TkQU56VWuQL9mzvskUDn65wqCGo2jDDfHl3O-WSniK8,7113
731
+ vllm/model_executor/layers/quantization/compressed_tensors/utils.py,sha256=MGBWAxDxZxB37rkc2nvxCg-PPhz--akasm86cdODMcc,7563
732
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/__init__.py,sha256=e-cfknR2aOVB5hznA58BE2_zqXxBoAJXCmozfrc-ook,1380
733
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_24.py,sha256=T9Q0a6UphFHPh1zD0SU2VFFuipS7Il0gAmIsPT6wyXQ,14327
734
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py,sha256=UfHG3dCxyrZjzdqdgDkmlcsUkbKWqwmEzQrwWGAfq_c,1552
735
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_24.py,sha256=p4tFtmPGqUj8UCMxn7r4VoO0UBXlSs8HC9PPR6bqUyg,5610
736
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a16_nvfp4.py,sha256=93toDuZyglexh_fkhjRoSnSRoXKwYHmB-mgbX8wvYnA,4262
737
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a4_nvfp4.py,sha256=t4VDwbsGwTfZYnnC6Fl3kijO7-VhvcXxfACdYMnCmHM,8041
738
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a8_fp8.py,sha256=b09jYa6gRK33_nSl88Y9khXbxNnoLByP-DnIIWQuTzw,6396
739
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w4a8_int.py,sha256=RTtbNo3-YEN8AsLeoXSinr8PxJ1eKV2HGas7rVBRG1c,5153
740
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a16_fp8.py,sha256=iuFaXLeGOoPVyzGXX7ZgXCIs9BmcKXtkKOyg3Zn0NmY,4995
741
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py,sha256=TUOn6G4VqLFcNOzmX-fuyj15lipAp8_ghgX-T8ygRoM,7289
742
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py,sha256=4C5gkE5jf1UGSxcFV6_3eYlmasxkXHG2VZ4bg-SKeW0,4534
743
+ vllm/model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py,sha256=c3SgsaVP2T36Bd4S0cqi88P5ma01U5kIBvORe2VCL1g,7560
744
+ vllm/model_executor/layers/quantization/compressed_tensors/transform/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
745
+ vllm/model_executor/layers/quantization/compressed_tensors/transform/linear.py,sha256=5-7G8oPqHqKUB-rPxL6zJl4zxDBqp_42FeRJpwbczGM,9548
746
+ vllm/model_executor/layers/quantization/compressed_tensors/transform/module.py,sha256=J1QrXE8c4orA-UC3lhz8c9Qj85JT1-JfsI3Y9Qfn7aA,6277
747
+ vllm/model_executor/layers/quantization/compressed_tensors/transform/utils.py,sha256=zOt-O0AsZ978entPoKuAqg2gzEgMVK5FmW14GVBLfso,349
748
+ vllm/model_executor/layers/quantization/compressed_tensors/transform/schemes/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
749
+ vllm/model_executor/layers/quantization/compressed_tensors/transform/schemes/linear_qutlass_nvfp4.py,sha256=LeSahYO4dwi5fkLE-q-6-lZ4Uw7j8w4pzEae0uUpFUY,1877
750
+ vllm/model_executor/layers/quantization/kernels/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
751
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/MPLinearKernel.py,sha256=dvSBv9nYGrOiSEN3udSv2ElQYpf7vQorIjA61b32D2g,2825
752
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/__init__.py,sha256=aUTIpuzLXH0BSeIiMlJvqzxJ9mFtjJQJtM4aMlbf2CM,3778
753
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/allspark.py,sha256=qzAUeUHkdQaV5AgjKZjbARrkHo8j9O1lGzkWUAg6jB4,4167
754
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/bitblas.py,sha256=Ur0Qfg5Eh1CUnIW9g8NvHAbvST-kFNyFfFyACJvwknA,12137
755
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/conch.py,sha256=hpCmB90lYFC2P-HTX-fznDIs_MbpYxhDEAw57Ky7mdU,3266
756
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/cutlass.py,sha256=rxqQNpCBceXBmpvBkknbLgbqnqePhCmkzingRswcwts,4270
757
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/dynamic_4bit.py,sha256=HDtGt3LLwLkNFOsoYmhjnis5h7e-dGHPdCCeAY56-FE,3938
758
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/exllama.py,sha256=VGmhBP9IuV5xSoUB_O_IkFMyt0KGpqiDo5XmPsPwEFY,6213
759
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/machete.py,sha256=gdBdme4JUaZ_YcrKTzJ1JfdoeX_oG9e0SSQa204AvTI,5726
760
+ vllm/model_executor/layers/quantization/kernels/mixed_precision/marlin.py,sha256=TgqJHjOAjy6jgFCTZpKG3SYQyDLjUuanOLZ-Z88xW3U,6115
761
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/ScaledMMLinearKernel.py,sha256=e6znbeey3aanpeg61C6gO5euhwDuzDy6pH-A0rOCMeY,2005
762
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/__init__.py,sha256=jpd4yNfgQXIFHVy-J3Zxr3X1KIDSmW3qg8ictrm-Hhw,3615
763
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/aiter.py,sha256=wpb2ssbeewS29Ayb8eBkW7ozdRTVkLfIkJYN-8kcaz4,4600
764
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/cpu.py,sha256=Kp8rLPuuFdgq3b7xs8q55kdz9DWJcggGF3gedqvIb00,8003
765
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/cutlass.py,sha256=HVQmITW0w08icugokpqhfe5mSyjLVN1I7NrUNMAgmGE,5422
766
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/triton.py,sha256=Xjrl8P1PiuP13Jc_7GrJKoWvYXnW3V3pGUSG1EIjwjI,1291
767
+ vllm/model_executor/layers/quantization/kernels/scaled_mm/xla.py,sha256=6f4RJLPwQ5XgjBs47Wx3j8_qiihzJseThDYE_MtPm-w,3889
768
+ vllm/model_executor/layers/quantization/quark/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
769
+ vllm/model_executor/layers/quantization/quark/quark.py,sha256=DfzS_GZyR0IB_c97MXJ1HG1wcPJdv0jWofwR8G3ODuM,19982
770
+ vllm/model_executor/layers/quantization/quark/quark_moe.py,sha256=APUtWKhTTQG7oXoHkG9xWPTTrXK01-yipDQ_UEta8ts,26645
771
+ vllm/model_executor/layers/quantization/quark/utils.py,sha256=lOvhDC7p5Oe21iNAz2gAt2g_msZE9L7UhmPx7T_NIJY,3510
772
+ vllm/model_executor/layers/quantization/quark/schemes/__init__.py,sha256=A699VuIbNQcvuANSmPjZSu9SmSHuqKfSQhvn9Y7GlpE,343
773
+ vllm/model_executor/layers/quantization/quark/schemes/quark_ocp_mx.py,sha256=wjwovMMaCmtNExWyw7KD-gaGm0iiiWt66nVqf2GzqXk,11113
774
+ vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py,sha256=VMPNUJAQ6PqGlGmVD0wUx3a2NstTp2aC5Dx6w5ZTHn0,1516
775
+ vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py,sha256=AcxbIOddblmpdnjCd_xtexp9Ea8qnAz0K4C0LpR5LYQ,6757
776
+ vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py,sha256=147BnUsGz5tKSVecKUn9BxfA7aWc_UXGSaiIaqhGKXY,5036
777
+ vllm/model_executor/layers/quantization/utils/__init__.py,sha256=ifoIA36ObRUZOBfRBb2Y9xEGV5lLuc0xe_6HYLxV8Y0,232
778
+ vllm/model_executor/layers/quantization/utils/allspark_utils.py,sha256=-VRplxaZTs52TJRNpTtpjm0pp9f51_S1t1muD76v2bk,2285
779
+ vllm/model_executor/layers/quantization/utils/bitblas_utils.py,sha256=TiKCRtacJQAEroE63vfmeeaGFLEsVQnEtG5lOv16dlk,7640
780
+ vllm/model_executor/layers/quantization/utils/flashinfer_fp4_moe.py,sha256=n4XgKtM1vvp4S636Ws0eNoU99N_rYP0UfZZWg_hLFe0,3104
781
+ vllm/model_executor/layers/quantization/utils/flashinfer_utils.py,sha256=zwgRn7pYacko_xkgrX_5Zjw3nUu_w08NTLT1txk5RH4,10767
782
+ vllm/model_executor/layers/quantization/utils/fp8_utils.py,sha256=zhc7nfPzxm3E2nFXUCMZOZYIYd9kRVlAuzx-MnQZevo,39025
783
+ vllm/model_executor/layers/quantization/utils/gptq_utils.py,sha256=-0GnKH51q6vGtt42RhlbGRkvDI4Yhta4QBG1FlWWV5w,5886
784
+ vllm/model_executor/layers/quantization/utils/int8_utils.py,sha256=tlXWgz7xy2vJSX-z-yH8kpznoMNXYCZqqc0UcYcyczk,14753
785
+ vllm/model_executor/layers/quantization/utils/layer_utils.py,sha256=CvEUZEj2ZFt7oxcDaRklOJo0IFc-llCiA2bVjybRKFc,1574
786
+ vllm/model_executor/layers/quantization/utils/machete_utils.py,sha256=Wv5uFC2UuuoBC_X1YGFOORBf23Zl_Yp5Ck_hFQIbRUA,1699
787
+ vllm/model_executor/layers/quantization/utils/marlin_utils.py,sha256=dfZbCn5pRXkDBAOvfpG55FrieF7qVyfBm7XZ4_EMM50,18450
788
+ vllm/model_executor/layers/quantization/utils/marlin_utils_fp4.py,sha256=b6CCW2hpGnY4awO-Bd6diADMWT6fLmRvlqpKGsLZGBE,13652
789
+ vllm/model_executor/layers/quantization/utils/marlin_utils_fp8.py,sha256=Tl0E_WfNMKMaDABIoo1d0QhZjB0HUhbLrk733wWUpkg,12559
790
+ vllm/model_executor/layers/quantization/utils/marlin_utils_test.py,sha256=yrFBPjJG5KHxFf7iCZaN1yQQeATURT2mu5uwWhsf3-U,4947
791
+ vllm/model_executor/layers/quantization/utils/marlin_utils_test_24.py,sha256=DXy1fCTnTEhQtMg7C-AKwL5oTDvwdDwM7iZjSZ1NbMU,17248
792
+ vllm/model_executor/layers/quantization/utils/mxfp4_utils.py,sha256=0kQ8diANXek2KTSvOk6_uM6YohT-IQOGlMu7DqxUUzg,6054
793
+ vllm/model_executor/layers/quantization/utils/mxfp6_utils.py,sha256=3ccQ0AmjiDMVtIyZ-G5WZ778QDKbpbRqjv8vfNE9u5I,4533
794
+ vllm/model_executor/layers/quantization/utils/mxfp8_utils.py,sha256=KyD72E8jsbF-bHSHdYFkY6Aseqsa9aiTmlQfIjcnHnw,754
795
+ vllm/model_executor/layers/quantization/utils/nvfp4_emulation_utils.py,sha256=rFpYWNmA1DD4I536b0U0FgjB4zBVDSpN6cpeldjssvw,4633
796
+ vllm/model_executor/layers/quantization/utils/nvfp4_moe_support.py,sha256=6a0Pkw8K46wjDLtYXRYjAAvoAvWcj9j8F3GY60aVA10,1979
797
+ vllm/model_executor/layers/quantization/utils/ocp_mx_utils.py,sha256=WGqf-4DcMRs65l6wZsZNixc9KzRJSRl3fRpk_XIA4v0,1770
798
+ vllm/model_executor/layers/quantization/utils/petit_utils.py,sha256=8oiFeO9uCAaHeV0QrKPs-GbmJok66L9jylexITD7Xpk,3967
799
+ vllm/model_executor/layers/quantization/utils/quant_utils.py,sha256=AdmOm_XyLwyr8I1WTbF3UbNH74ONgKi3P9IBJrH9lc4,21860
800
+ vllm/model_executor/layers/quantization/utils/w8a8_utils.py,sha256=aY42AALhUYziOh-7bpAqXCKFaPpUx6T-ylXDZhlrh0k,17175
801
+ "vllm/model_executor/layers/quantization/utils/configs/N=12288,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Szg1W2xH7h5U-UOH8vHbDV_xs1xO3AM_wITHbJtITgU,3264
802
+ "vllm/model_executor/layers/quantization/utils/configs/N=12288,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Ta8XgWb_aVhJwqJ59i1zzY45NhCGazJ75whDUmOfyVw,3259
803
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=t8TaODfMF2Nq0qg6KOc8NSTs7m90Jcu6Ih3BXUvFb04,3799
804
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=CNI-I9ncqHJ7ukpzgyxdJtz0bd29vsgC38tvMM6TV1U,3803
805
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=CNI-I9ncqHJ7ukpzgyxdJtz0bd29vsgC38tvMM6TV1U,3803
806
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=-j7Xyk4xFaiAD90FeH4AqRSnS82f4owKRGMHbObrrHQ,3250
807
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=sW_T-BdLbjJoFqlr-B5f9emF8E0IdKfy_1wUSIEi55g,3253
808
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=tkLjwLC_aVXhzuvo-2QHkojXZauPJsf3jNHFn1S7uRA,3244
809
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=akDNAjUZ3EXBznF9w6qUcpXxaLWq7oXnX5jy-R9cleI,3246
810
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=1536,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=JAycl7EaUZtmCoXMjq4JwKXCeXxZ6S4Ts_DricRUw_o,549
811
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=q5KZyi9T-l07P3r1u9i6-Dpw89Upjw1gpTp3f1CluEo,3799
812
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RTnTPFQNg5JULbPLWJDTRNRZHI7FsrTxqSDkZfSbmzw,3806
813
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RTnTPFQNg5JULbPLWJDTRNRZHI7FsrTxqSDkZfSbmzw,3806
814
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=DLCfW5tQ9k74AGZ2yER1etP-HgUGglPp_woJiaPuxgQ,3249
815
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=8v9mdWPs1eXczo3iwFrNnRo2LF9wPU4Scm-r9bL7Fz8,3251
816
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Qoj9rLLRDbKM4IKBCXvN8RcxzSmNPd0TQUiM7CXDqHI,3241
817
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=7OFCbBqqEA7vQ1oiygfW-7Tqqx8OJATaLujtcQIgyTU,3247
818
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4D3Ku4y7BCVEJzueKvQC_KvOR026w3ONWsxfsA_YrEc,3249
819
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=iJZ_tAzoYGUmg9ltil4e8vzKlKi980yTmswEMWqV1Jw,546
820
+ "vllm/model_executor/layers/quantization/utils/configs/N=1536,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=fDomA7uBQKX8kbO_4MFcoBwHhIR_7sOkngQPv6cQq4Y,548
821
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=ucrZBIN_ivmmfMAvkT40xQpH87LdQK38lZbeLWMyV4M,3806
822
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=zDnVqBqgT-nLkz_Cou-KTPsNIVh-YbTBno9L2MgdRTM,3803
823
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=zDnVqBqgT-nLkz_Cou-KTPsNIVh-YbTBno9L2MgdRTM,3803
824
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=zd5cMYrxQ6PD0jKpd3YF6ThT9RGdqgEQnCW6F4W-r4E,3249
825
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=CjO6dh_qt1iTu5kYRs98tTLL-W6FOzLO4AESMUFHz5s,3254
826
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=7v4tp0RaT4vxF4urSBrkK5FR_5ikeFQ1htF3DwDl1lk,3249
827
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=M5F5wzSmFokEm0X8__ogLvdE1QVC6EW8atqq-kp3rVA,3253
828
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=0J2MFgaLkv-mfVE5x363lgVKYU6miLG_xRO3tJUga_M,3249
829
+ "vllm/model_executor/layers/quantization/utils/configs/N=2048,K=512,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=983yfFeeo-BClL_H1g-owXwbA6t0l-kREiy7kLURUMw,550
830
+ "vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RzrnY_xKo39wZ4nO8zUorLp1ivTFabB8ZQOFRx5JcMc,3251
831
+ "vllm/model_executor/layers/quantization/utils/configs/N=2112,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=f7upf6kaHS5_3EqQYDxtSsbgb4D1iTvjCiC4auzbx3Q,3254
832
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=C2eM8RudmP-qXEf_Apg-qcB5n2Ugxf8-7uG8hQDSt1g,3801
833
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=llI6PWlSDgQf-ouTDXkFYOoSz9u3bzklwBtZYY_fWVM,3807
834
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=llI6PWlSDgQf-ouTDXkFYOoSz9u3bzklwBtZYY_fWVM,3807
835
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=q9HUcoL0cdZCOWZ8MKbcpR8NSy5iNEBq6NPTaHLgRB0,3242
836
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=uJu6Gv4e80vxVrDyBo8_y47tOV03RmWVsMIWQ-bbW6Q,3251
837
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4ubbhwSFX_XbefRLEkLoWxJkcetFWPzsszPu0X3_Wrw,3242
838
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=euiKvhb3DXkvPPQJLqNE_xN2evsTOoZnVIiquyN2Cm4,3246
839
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=FhyniGTx5QeCuVrBSVTQys6q05Pr5lPEcPykpAX7Iyo,3247
840
+ "vllm/model_executor/layers/quantization/utils/configs/N=2304,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=pLQvMaVvlet_JenEz25mxxplAaHNisl6SFTSZ7lYP2w,548
841
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=irQUBLd3KHNd8JNX8eFe2fBB3ZZ3zMl3aAF22uxJ65Q,3266
842
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=gklD55iBvg488-PecvtcEypwCDZ2lCi8c5o9bqgEEeI,3266
843
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=uAa-ZQmASwlqZbr1l1CM6FyJI9irNdLBzc1U5Hdyw1E,3802
844
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RnN7lfu15CE-4ywMjAbEz8wWV743AP-1Fq5U_j8EQeI,3812
845
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RnN7lfu15CE-4ywMjAbEz8wWV743AP-1Fq5U_j8EQeI,3812
846
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=cE3BscS_zEtF_m_jr51IPfpaZZgIEojmhTHsrb9jABM,3260
847
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=SScyo-oYCBxJR9C7ZIKu_pJJNiXdpT13kYe26rddvPQ,3261
848
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=0v17v78pETXv6S2ZoibekxOVhiTmCm807DYG4DONUck,3259
849
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=AOuovUsPAHqZlbr4G3_CnCNE__fgxCz6RuOhOxCwWv4,3258
850
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=ulvOEAFO8c-UOa34FEZrjOkCR6ovhJlfFFDhmaKIBiU,3245
851
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=BiZowqExbvXftuE37SYcheOdtYX7Z5BEXyykJ6GbYSk,3254
852
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=W-xd0c2ziC5YbC96TXlc0xkj2cmbfcdXclW453PsLpI,3258
853
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=7ok0uooTihvRSckZMNd6jInRvht_xkC5posHO66ejqc,552
854
+ "vllm/model_executor/layers/quantization/utils/configs/N=24576,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=h_Z6wBKdSGBEo5BfQKaxuFlxztrnbbZR0pkcYKv92sk,551
855
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=k63VgniyN3Rl_-h1hYmT_q9QZtSFqQmXBqhEXJQkxqE,3800
856
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=icswqRYUsUdoQMrv4YIqO46GG9BzepmBJmnTre9-VjU,3800
857
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=icswqRYUsUdoQMrv4YIqO46GG9BzepmBJmnTre9-VjU,3800
858
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=sL0E4zZzb01g6GHaTCXltg20uSbthXHSJFQ0SaxZ7PU,3245
859
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=MZcJz7BjwVOHHHxvYqGrWw77WnxslYhwW80bZw-jSKQ,3249
860
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=GsLoYkaZ2p4Qu0Coj-X90s7JWyfZBOloIHPlyNKSIes,3246
861
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4--7YWnJYUK4XmQ2zZ4M1ZYdKvUkET0VkNgIBn6xaOA,3247
862
+ "vllm/model_executor/layers/quantization/utils/configs/N=256,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=NjEA2QjOVXyOaVSMPch5qa1Dq3igbW7MmE986-7taW0,547
863
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=21Buh2aiGTHjpW45Rm-TwZD8MSaAy8NMUrK5l_hGT5k,3803
864
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=P8p-dZZt_D61G6k3PgUetF01xzTRmCDJAnqCIsSDW8I,3805
865
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=P8p-dZZt_D61G6k3PgUetF01xzTRmCDJAnqCIsSDW8I,3805
866
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=8zuJhFdd6aXREpiqPFhIKEFWA5lgLVGrG0-a9UXcBqk,3262
867
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=_42bDZX4VODErI6OL-NrWja36iNHC4DzgF1l5Mk67-c,3248
868
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Zn1TvhAoPOv0zQBYHOZhwdDw3oqyxm0zIa7IJkTCHpo,3247
869
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=1536,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=H9wONEU0XXSxOJfkx5UkS8Ss3A2QCp9G0XNoJEqE9nQ,548
870
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=2T2TYZhXgC97slH92HQ8GvZS3KuUt1ZiC3RtudPVEPA,3802
871
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=b6_bhUuQrI9HYvvwmAvUYh4v1GZ8w0sjApOmwuj_t8Y,3806
872
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=b6_bhUuQrI9HYvvwmAvUYh4v1GZ8w0sjApOmwuj_t8Y,3806
873
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=yqjO7zML7EseBJw6Bn5MTyHeAitkPsl1dndXeL6Rn6A,3257
874
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=-nQIhKAOVCQrxLV6HDlcD0V8HMWvqrv-vyiORVU7qls,3244
875
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=KKmCvNh5T_qfD8v7JijMqXxQ5L6-gRX7oc6c5re6EF0,3248
876
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=M3nwpZd2-0w263ywZt9gaw53z7MN673T5tl4tc43Ntk,3249
877
+ "vllm/model_executor/layers/quantization/utils/configs/N=3072,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=H9wONEU0XXSxOJfkx5UkS8Ss3A2QCp9G0XNoJEqE9nQ,548
878
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=KmEgJ7zP2Sr_7GsAfL-12_g2S2a2wVpnxgCiF5dFiLI,3802
879
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=J4SXwpsioBRdTXOaj2OjrdNrEuW1NF43cLds65UWzCY,3808
880
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=J4SXwpsioBRdTXOaj2OjrdNrEuW1NF43cLds65UWzCY,3808
881
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=UjBOmVqYynBH3dJVuMJXjKnuZ6LssohzzEBpLBG4_G4,3256
882
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=8BJsjc2UUYdotrIqwyzisjrq0wcyW4jnTo_M8J3qYwA,3263
883
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=vLoV3JMtvHOKpR5D1BeCQPMuYlWUAlrXu54gByNkwKY,3266
884
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=BmXTZvWk7kphfBmGmuSIuolAK-3qCGdmcPhD4FcKd3g,3265
885
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=N0kCPHvybNK-HvMO2EqNDLkj7m7WrHTl-3AD32LBD4k,3248
886
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=mjh-AgJN_IoWAc1uwhUiB1lE3ufAPDf-KPP6vUTrDKw,3251
887
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=h0qz-pNlC9ZGNbyeFsESFdowFPfTTK3rh8SK4NH2Css,3259
888
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=mcF12eQTtGxocrVIA3I98NHd1NLd0-8EyfXtqDgv0PM,549
889
+ "vllm/model_executor/layers/quantization/utils/configs/N=32768,K=512,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=AThoa7FUcGdNXYB_v9iMpBh2X8C0iLfc7y-C0xy2cRY,548
890
+ "vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=MJgIvZHf01ju8IWEVO6vyMedy5OTZxDpzv6A7_8W-Tg,3813
891
+ "vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=AT2yrMoTvmoizi4sxwLtiULZ57P1CBhKGg9-6Gxnuc4,3819
892
+ "vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=AT2yrMoTvmoizi4sxwLtiULZ57P1CBhKGg9-6Gxnuc4,3819
893
+ "vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=T60CKtM9YhIEZs8F9Lljrdqqc4ReloR7Xl9IYsfex-E,3261
894
+ "vllm/model_executor/layers/quantization/utils/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=kk8WqNCGmjh8-tOMHBP8sv_5fW81Xkdzdf8-2WDm0RQ,3263
895
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=cPtr1UJq_B-dTqgMrVm8ptiYXA6qOy_F8rs2f7ljuEI,3811
896
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=cobt_ZhR3dt2CySr12bGPVwn1oS98YvGLdIh9H8BDQ0,3801
897
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=cobt_ZhR3dt2CySr12bGPVwn1oS98YvGLdIh9H8BDQ0,3801
898
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=6Z7kIa14RjVq3ek_C15q5mUu1IrY2r0OP8S-_pm-MYU,3252
899
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=r63SZkUJJV87B00hAX074_uaC7wwQXdurlJsB1jUA0I,3254
900
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=hL3doX7zzxld3UcS8p9ACSadDaE6t3xXlYwM7X3GOeI,3252
901
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=xBhxdCFf3waTUsLxJxA54R90zODbC_DKI3XXBVKjKRw,3252
902
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=512,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=2ks7TQUULAD-Zn5i69YHo_2hpmsmxlocdYmJccSh2No,552
903
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=zm2eqlVlPWlP-5o944QL40OCzMpUHGkPJwduy8HOV8A,3259
904
+ "vllm/model_executor/layers/quantization/utils/configs/N=4096,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Yfg4GDiXIYLyzL-334YirvDbcChz-Ep_atCghEZSntU,3257
905
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=eiI8X2fFNknJmiT0uHbzSaEKQwwZk5bxn676gNvcyg0,3802
906
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=fQQDJMlLdYsY5Cosg5HkRzvrJ4asjQmc0WGgoD4bC20,3810
907
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=fQQDJMlLdYsY5Cosg5HkRzvrJ4asjQmc0WGgoD4bC20,3810
908
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=O_SV2vo_oaABfT6Mxqcmo12pnhKtfX4TnXfe02OcHJk,3254
909
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=g12Xkurat7oUS7LdS9pHLKFlur4_FaMGiGBvdq-iBCs,3242
910
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=EWLxbWncwGJyL-dV6EO-s8kk25wfYrESa0STjCnzD64,3244
911
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=tFdrY5nADmXUlShdN8w8Jzkxuj_RPLXCRceX9FhQ35E,3251
912
+ "vllm/model_executor/layers/quantization/utils/configs/N=4608,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=M-ewEHbgHLBLYLi1Hgz5Pp4kypnUiCRo0ut2scNnvDw,550
913
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=zTzLbdff09HwMuWlWpoAIgQZ6NEjsFXSF0Y5z4Be7Ig,3802
914
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=dcPHbYEbz8T9SM5-a5sP_K_npDkhH7u0KM9aiLn9esE,3806
915
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=dcPHbYEbz8T9SM5-a5sP_K_npDkhH7u0KM9aiLn9esE,3806
916
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=TO2qRGmp37v53Zqu8Joeq_BSbtwM_mpVoozGyoNg0-o,3254
917
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=QqijmgLqIoBUxRPnuUQGsoQASRFRMsCVQKTjEjGecVo,3247
918
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=0xquf00fgfrDODpaxyre0VDcjqfzqExj939rzeJ8pMo,3244
919
+ "vllm/model_executor/layers/quantization/utils/configs/N=512,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=ipg8iK8w2ySRe1Z08YJUWAHX43rvkrXpR6svxRhSnFE,548
920
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=-wuzdNXf3K0jfFQGB8nFSyoSZ4BfAvIkY10k6FdjnLY,3800
921
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=-o9QqqQQ-9kRVCuDOUGBuKXHRTd0asGTzrDcHGGYJLQ,3799
922
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=-o9QqqQQ-9kRVCuDOUGBuKXHRTd0asGTzrDcHGGYJLQ,3799
923
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=DbemSQdo2h5vGjSNB6Fovnn-aAGfjti04Bp-5KxLALk,3246
924
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=6glWpljtfiuspJv_Esg_LWCDDQ57d2HETsOIv0zr2Ec,3249
925
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=qG6v3n3qF6LE2DdGT-mDIXecZ1a7vg7p3QqXYCMX85k,3254
926
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=l-9r2k1gcKB8UXlBXVuzkoa1JDLgJVTBQ_OaQK80z-k,3252
927
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4--7YWnJYUK4XmQ2zZ4M1ZYdKvUkET0VkNgIBn6xaOA,3247
928
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=ZfPPlx0qcuR4WjaFAE-W1QZgSPAMf3NyGcpvQIvyFMs,3245
929
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=9w-sicV97vSQxkRcEKnFKFjkzBOx-VOHlrh6b1hhQ1g,3254
930
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=QgSlDAhlB2W4bzTd2O98UL-C_IKfJm_cVmQz8FqsLF0,361
931
+ "vllm/model_executor/layers/quantization/utils/configs/N=576,K=7168,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=i3wy_CBO7BQQVhKReRC2F0PaRIQDdN9F5lJ7kD0xe1I,548
932
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=QpkqpJnyjuHH8Zo4U4QZgehUF2F2uQDZFb8fdhixXWI,3794
933
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=wv5GjGAA-NyJ41SYdYG3tPAgwf6JK7Zf6SaWALQ5c3Y,3806
934
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=wv5GjGAA-NyJ41SYdYG3tPAgwf6JK7Zf6SaWALQ5c3Y,3806
935
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=RRMNeM_qiHvlUTOAeqwgs7ukSoAZSlK8XN4z8hgWl0k,3258
936
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=brB0-FFr-Sv2bdrz4DQJ_NaFhETctf1g4Yzwj_Fcczc,3251
937
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=bPQWtvaJrzOOIgI-R-MIxs_f4yC_FobkDydu3OkOFtg,3252
938
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RYLh-Uim9U2_djLkFwwpV0rNQHik0tZHzecuj1_hPLw,3248
939
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=ZRgiuHZ2SFC6u-WV5DGwau4k1RiPLI67eENO0e-5Ylg,3253
940
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1024,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4EzbnLWHVwrjyKYPMcDxbxM2o-krjlT0YXvM8oPH5Cg,549
941
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=OFgOtRkUHwyOT7Hk_BQft_WzuZOwbhMSLP65Fbr4goA,3799
942
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=AOu05da2LZbCzD9SKsrgnzH-ih3CdXsRIdJc_4J1lps,3807
943
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=AOu05da2LZbCzD9SKsrgnzH-ih3CdXsRIdJc_4J1lps,3807
944
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=qzmFm2pqxphir1LBrycDZp5JA4It8OdQeQ5iTrTwLNE,3253
945
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=2UyOMRMdbvHt6WlZdOKALm3Or0eMCx7vvwgLiCYyoOs,3259
946
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=-hP_P8NM0K04mGzTmpGBNibQ5xxh5gPz5WtoMXhoz1E,3253
947
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=vEU4_YOMnLdYFf1BkBEdFbGRMG8KLhsO_t0gv7vaO4Y,3244
948
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=FB5Le4obvPoCgFSnC_3-Uh59n-Mt4Rol8saXVcK3RPw,3252
949
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=1152,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=k1rzpgm9m19AHf_HPQcNCuSBtAwFgMePUYB1jZeFyYY,549
950
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=9IbzTwLRgTCfFLSvjEWKiajCjG81R-wTljIV2zUYUA8,3809
951
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=B4uEwuftvaj9gHGdoDBnVhxbNRmzUtzu4LH0u-O7voA,3804
952
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=B4uEwuftvaj9gHGdoDBnVhxbNRmzUtzu4LH0u-O7voA,3804
953
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=ZTPWtJA3JBL2jhy7C60RdsntKCN8oQ-DDIL17ok7OB4,3257
954
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=mokCWoXdKi8p4mLYqgljjwDRJWK5I2oF6_MJuObi5sU,3254
955
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=kLviGvVngpgOuelfKtvv9Is7MWQ89rGxlomMRP6t0Ic,3250
956
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=bIVRtaaHThozH54VIte0Nk0sOGV67K4s2YZUE6QWx2s,3252
957
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=_YXzQ6N3QpF3Ou1Fy-51YyL-J3i5gOBVCgSM42vOT9I,549
958
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=csaz7AaVDTvCuzaptN-e8K1PNuIwZm9OwnPSJydHI90,3803
959
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=scfO3_ncCtyrqcYSnIoAZTMfvBzjB4o_0_bdiiVSNh4,3803
960
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=scfO3_ncCtyrqcYSnIoAZTMfvBzjB4o_0_bdiiVSNh4,3803
961
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=CE1wRLyFONo4_icKO8fcTTX-5giKNJ9_1F-2mr-lGQU,3257
962
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=JdUaLiMmf8oEbwuhPHMIncvWzXS2SxOEgfM80ZjM7l0,3259
963
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=vlys0Zi_CaaU41OHGbWSBtbVglFi98bgqEySBMc9Sdg,3258
964
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=-sqiMkGGMzhrs1WdhEfwiNZd2r-NmhEsfvJxczLZJ-g,3258
965
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=GY9VBPi21K6vJlF1NOEzCyqMS7LX3xq5dRxrK0jvIHk,3244
966
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=8LWF55ZPjrOY_sEdRGqf1eLcTNySgUiiWNWsN4EGxLY,3247
967
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=xEN9qWPikXf4njuYyKQJVW0SM5cDReje-euoWbr64WE,3258
968
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=_Cc0EqUzl6d93OxWJRWYbYpEaTIp0glJhdfV-GSAi5M,552
969
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=16384,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=ZSHvdnC2vOXI2HPW1iNI9HdihoLcNYlRLMF85pqjWZE,551
970
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=SkyMLsoxGoHdO4kgTerihone7eEi0nmHlrvZUI1I_V4,3804
971
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=6Jo2hw2gQpyiNoCRZpGItu4MBkYytzdW-VggWUC4fPE,3804
972
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=6Jo2hw2gQpyiNoCRZpGItu4MBkYytzdW-VggWUC4fPE,3804
973
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=xbDfUYLphVtZWJojZWODlxGMCoiIgxn4LsnD9ge3r9A,3257
974
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128,128].json",sha256=hqh8TQw3t5hPM9u7rmHPuaMjwgxmQ-Zt35fSTgOS0HQ,3261
975
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=Ggy4hejkcWjiw5Bi-wGzSP5JLVuvOjip_rbjXFBJZbs,3257
976
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=RzkrUzR_nzxnD4L2jF0_8aDX-nidn6AjhVXlJK50VyY,3259
977
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=qKG9hmaxN_7tCB_06L1dh0csxs3TGeya9B-X6W-tNhg,3245
978
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=jb7vGi1RJefImkT3BZU_9iOkiCulcd5oDjxpVSt7big,3246
979
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=at0_Rcc0-lSzI9MFj-TjnjyDt0HckCZYAZ19q-7p5zI,3257
980
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=_Cc0EqUzl6d93OxWJRWYbYpEaTIp0glJhdfV-GSAi5M,552
981
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=18432,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=TWpzs48j0QwApAsBWt3iIlu6cqR46Meslyp96MOANcc,551
982
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=i5b52A1Oe8kCdPrPLBGud7OMHm8779JD0rBocYO_lo4,3797
983
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=U20Q4JwG63kU-6cc241VHGdpettCWbBXRJ9EZ-fbkqA,3803
984
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=U20Q4JwG63kU-6cc241VHGdpettCWbBXRJ9EZ-fbkqA,3803
985
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4uWiQMh3cZY_EtLA0a3PU8Z1VCunF2PpolTPYeP9Rjo,3256
986
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=OgXOVvRKqsgzlJwjHNxNCsJ_o3POBFigwCqafbh_aKc,3258
987
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=toHzCprq0KetQI0-9IrLYCIm1bQ0nSeP1gXArU0GogI,3245
988
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=_0V6CEfYgBsaUnF5DwNWWseo8N1Ph_R0la_XN8HzcuM,3259
989
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2048,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=pGZZj_gZms1T9Zgjs4tbIm90LhbEy1UUkkgrto9jPts,551
990
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=fqnjZCn0gbY7fO9JwZOHMYJJHe8gceWhWCZOFPRUlYM,3802
991
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=OTZt3ell0OZ7Cg5L17K2NPU4UwayAkTihV5HjUmUiAw,3810
992
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=OTZt3ell0OZ7Cg5L17K2NPU4UwayAkTihV5HjUmUiAw,3810
993
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=LdtOyXsA9r18GiFkmDOkiRinsDSZBZ8NYapL59EZ4iM,3264
994
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=07GarBHmiiYkyqn-qxEtrAcgCETuUbqm6HqlbH9yJi8,3252
995
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=xMNxtLL_8tyg4TWSt_llz_IJ2qlxc2NEwhUzhV1VsG8,3252
996
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=kEuvCsW3YNByF-DALYqPZpW3TL8ZbtQ5gUNq7-8YvZ4,3252
997
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=2304,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=4uNqB71a6ctZ-c4tF3r66vOsHFrqcR28g_UWy0N8iBo,550
998
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=QkrfZ69jxW_mweigtHL5R0Sv_WcSBp7wjFX75G9kbHw,3805
999
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=xMGmoN2ZTjKQBZS-k75mFTPpAEbPR3kyMwqZVtgbEiM,3802
1000
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=xMGmoN2ZTjKQBZS-k75mFTPpAEbPR3kyMwqZVtgbEiM,3802
1001
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=PD4AJYCkHfy2ivv9baMouFXzBTy0eKMumbAfxfm91HI,3256
1002
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128,128].json",sha256=iu8M35YR-RDpKWbjXSRzk02sW9nr_dtbhalfLSNtxNs,3251
1003
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=FFBjSWlpKXMxfAUUYUqXbOK_Hd7qBeBsfbcaa9uB4qY,3249
1004
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=256,device_name=NVIDIA_L20,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=41m0bvskFUzVtlr_yppBr4PZ0cVkqHvy9Hrc5pUCUyY,552
1005
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=2VxMGfWtxTzXcF0bP3d5s7rc1cKb5TNBAn-WiCKAngw,3804
1006
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=VtQGO3fEiyhbKG4sl07cuVc6id2EtKeV05ozLmN_ENQ,3807
1007
+ "vllm/model_executor/layers/quantization/utils/configs/N=7168,K=8192,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=VtQGO3fEiyhbKG4sl07cuVc6id2EtKeV05ozLmN_ENQ,3807
1008
+ "vllm/model_executor/layers/quantization/utils/configs/N=8192,K=1536,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=W3cYFteFIZLu5c1K41cOh4_-WZzFU6-jGnZocDzmKaA,3796
1009
+ "vllm/model_executor/layers/quantization/utils/configs/N=8192,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=HIoWSUgAOcNaK2kj2YwDjDa23PzQVTT2C2ePW985Ovw,3805
1010
+ "vllm/model_executor/layers/quantization/utils/configs/N=8192,K=1536,device_name=AMD_Instinct_MI325_OAM,dtype=fp8_w8a8,block_shape=[128,128].json",sha256=HIoWSUgAOcNaK2kj2YwDjDa23PzQVTT2C2ePW985Ovw,3805
1011
+ vllm/model_executor/layers/quantization/utils/configs/README.md,sha256=kfjjurECwd-xH4EDjuueS0Xezi86c_pYu2yELgiw8Ew,102
1012
+ vllm/model_executor/layers/rotary_embedding/__init__.py,sha256=eNRXBBkE_8lcMx2H4vAOa3IaRKKGG-ToDvpGdr0Y_tI,9778
1013
+ vllm/model_executor/layers/rotary_embedding/base.py,sha256=v3S_jYgWIgFFYhb9IfAA2LxPPwgqt0SXFEN1QWQJe6g,8288
1014
+ vllm/model_executor/layers/rotary_embedding/common.py,sha256=ky0KV5EGCjzjfu196978-vOfBAyx3OmFJQ1sqOVvPgU,5297
1015
+ vllm/model_executor/layers/rotary_embedding/deepseek_scaling_rope.py,sha256=gz1nzkviimur19z78ehADynbP261_3ltjoFE7mHJ04o,5551
1016
+ vllm/model_executor/layers/rotary_embedding/dual_chunk_rope.py,sha256=BJc52pbq6NlkweRK7DEsyrJxIPhWzRKLfBIXXCQA4r0,8368
1017
+ vllm/model_executor/layers/rotary_embedding/dynamic_ntk_alpha_rope.py,sha256=gkD_OmIqduxIUvBgqEqFu22dVDm4zzOmtJ5SgRQJOAc,1289
1018
+ vllm/model_executor/layers/rotary_embedding/dynamic_ntk_scaling_rope.py,sha256=6lZQAj1v-nvRbZkLIR9_4a9nzAczr09YByOF7IXDr_U,2681
1019
+ vllm/model_executor/layers/rotary_embedding/ernie45_vl_rope.py,sha256=-kC66gUp0ONA7QbSBZtCvzImb3yBovIyPk8hcfCLKck,3029
1020
+ vllm/model_executor/layers/rotary_embedding/linear_scaling_rope.py,sha256=5a1OVeRvwUeU35x7xjUbhpc_UNtWd2MRWlZ1QftnxkI,4643
1021
+ vllm/model_executor/layers/rotary_embedding/llama3_rope.py,sha256=EMeHz5tB4hKDXr6ce4GFfFpt8x03HKYlmedwXUgDcHw,1785
1022
+ vllm/model_executor/layers/rotary_embedding/llama4_vision_rope.py,sha256=_f4tKV35lmZ4a3HvmTMzbqVNEGk6qwq4qoIugKubzhk,3197
1023
+ vllm/model_executor/layers/rotary_embedding/mrope.py,sha256=XEzGoRwsyDoHgpWNxCezfqMdcpJEQ1XlY2PUv2hKkDw,14180
1024
+ vllm/model_executor/layers/rotary_embedding/ntk_scaling_rope.py,sha256=KzRfhq8fROqVm5Eu6ETmQZyvc8LZjIKzde0ZGcr8Pkc,1462
1025
+ vllm/model_executor/layers/rotary_embedding/phi3_long_rope_scaled_rope.py,sha256=AT89hmBGCmGUtX1hoy4agKZfYdHd98eJnE1fGlzkQVA,5699
1026
+ vllm/model_executor/layers/rotary_embedding/yarn_scaling_rope.py,sha256=5INGA5Su2LI_DU4WR2Ob0ZSlcWBSzi7FIJSYKd9befU,2772
1027
+ vllm/model_executor/model_loader/__init__.py,sha256=MgfCw4mANJzhITt5XANh9LMMb3E26G2PsBx61NJBV2s,4861
1028
+ vllm/model_executor/model_loader/base_loader.py,sha256=BnF4wUJhm-eWW7tzYMO0UbE-Q9ORX90S2lVww7MNjcg,2075
1029
+ vllm/model_executor/model_loader/bitsandbytes_loader.py,sha256=O3PbLUnXn90YEzHKPMhs7tRFfkzM_xGS38ePV8gq3K0,34791
1030
+ vllm/model_executor/model_loader/default_loader.py,sha256=RP416MSmYfa1PFHDTGNf0fpfMiM-Vu4SNMYORrziWlI,12749
1031
+ vllm/model_executor/model_loader/dummy_loader.py,sha256=QStJ9SXJTYqzTWI6ytieoh0KNsQWlfL_sFz7bcHqtEY,1115
1032
+ vllm/model_executor/model_loader/gguf_loader.py,sha256=0dW2KsH6Iwngcg803H6QacPJWzgh9ITb0B2rJ4UQv5Q,7648
1033
+ vllm/model_executor/model_loader/online_quantization.py,sha256=8pAFlVZ-qCZFyl4LwvnJ1RpIwgM6WCS9llmphe95w9o,8882
1034
+ vllm/model_executor/model_loader/runai_streamer_loader.py,sha256=7YpbtdAaeMdq0wpI6GvkrvcUgt48EkxRRhgAYAqfdio,4394
1035
+ vllm/model_executor/model_loader/sharded_state_loader.py,sha256=2EA6hKlh5_qeY25KKpBPAH7kpvZHUk1noIoQZfu04UY,7917
1036
+ vllm/model_executor/model_loader/tensorizer.py,sha256=dhi_PV99hi-gyEqEWtwVmH8EY5XbluVRLgcxRMrENOE,29807
1037
+ vllm/model_executor/model_loader/tensorizer_loader.py,sha256=brQv1ywScnNSpYMAt9l0zbLz1hUbEbStsEGhVcwpuJw,5826
1038
+ vllm/model_executor/model_loader/tpu.py,sha256=Yk2kqIhQxistx25MQriqsuUu8-PyFdu_lPpZCiOzOkU,4798
1039
+ vllm/model_executor/model_loader/utils.py,sha256=ybqIuK2aaUEb1aa1vhE-L0I16dAs_c6-To2DpcqRRkg,11072
1040
+ vllm/model_executor/model_loader/weight_utils.py,sha256=hntWq1O-76Z0G-WaTQDCqcY6bvaew8OAgqaKq-R-7WU,40193
1041
+ vllm/model_executor/models/__init__.py,sha256=JC8KFZavvBlB4woyRgOBd_vcuoiA31nmBOR5dGcJSjo,997
1042
+ vllm/model_executor/models/adapters.py,sha256=aWg8ybmgRTxbhSyen-1fcy20aCJTVkvXS1te95WPK0o,19339
1043
+ vllm/model_executor/models/afmoe.py,sha256=IayWZsTpQytjqd7xPNmbLH_UKOGcMU4jiRI-quV8SZs,26759
1044
+ vllm/model_executor/models/aimv2.py,sha256=w4Mun4d93Fmp3d2tqjCaqhdPfX6fYzaTltHfyhkcge8,8242
1045
+ vllm/model_executor/models/apertus.py,sha256=fHHKTASupvKB4D9bwcxOTVkVxZ_p7KszRf5WV5vdf5M,21734
1046
+ vllm/model_executor/models/arcee.py,sha256=_FIAt9vwuVkPe1UQQXk1rLt0QtNbpcBJjeUYeYB06S0,16594
1047
+ vllm/model_executor/models/arctic.py,sha256=rjJ36pd3fTEeOUbsQJy-FpN7jnqnlIjjwPZC9285wo4,23830
1048
+ vllm/model_executor/models/aria.py,sha256=93GaVD47WLug13-5fNJYoYLzq_KK9U6Pcf6c41W4jXs,23523
1049
+ vllm/model_executor/models/aya_vision.py,sha256=eTnjMz3pZUYvH54zG04-eMpQhruR2ZnuawyIauv1l8w,16568
1050
+ vllm/model_executor/models/baichuan.py,sha256=mTC3QXOHVsuOX2cg7bS7Gbf8CUbdU5xhrBvlon6YjK4,17900
1051
+ vllm/model_executor/models/bailing_moe.py,sha256=xD1phEbU7ehVg1Gw5gzYWqCYy69Jjd77hm0nip9Fxec,23182
1052
+ vllm/model_executor/models/bamba.py,sha256=o5EqDsG7S8DBEI4cG-0ZTunZYFj2i4-21tzed4DGm9M,18103
1053
+ vllm/model_executor/models/bee.py,sha256=BXEICXixtwrG50Roqwk1cujdmQ4pljJGfxQgUt7ZqtA,5450
1054
+ vllm/model_executor/models/bert.py,sha256=fY5WOuhUKrcoUPDbc6dKKbCg6n91MU98eWhlMmnICXg,31345
1055
+ vllm/model_executor/models/bert_with_rope.py,sha256=CbohS09BZMV8Y7WHVbMGvsz5-KT6M_Zo0LwAplbuGs4,25587
1056
+ vllm/model_executor/models/blip.py,sha256=eUrhdTb8Sfvy49Kdb9vj8dVEgbHzHbnfh9QfvDGvcvQ,11852
1057
+ vllm/model_executor/models/blip2.py,sha256=xfWKeShF_mWqqTaT3hmbS1G62No4jUJrmn0S3o6XNzI,23302
1058
+ vllm/model_executor/models/bloom.py,sha256=6M5wXg2MJSPZhsoS-vzqzIa75LYv9YzZfIhvfvrI5Tw,13961
1059
+ vllm/model_executor/models/chameleon.py,sha256=FJ8Ik_Sl4QOkyyAKTz7O-3jTESx4rYp9vMXHXNXvJac,41415
1060
+ vllm/model_executor/models/chatglm.py,sha256=_CuJF3KAbYF9Q1mid_pgDdm35Er3BnjlSKcpskPhEwo,17463
1061
+ vllm/model_executor/models/clip.py,sha256=XzcRRflXnB-QLOM0zsIDGnqTHl3I-3G9UNjgbwyscnA,31768
1062
+ vllm/model_executor/models/cohere2_vision.py,sha256=5SxbFLm4Kda3rdaeXVjDcySAp59aEImLgrP71pGQTO0,17069
1063
+ vllm/model_executor/models/commandr.py,sha256=x906MvAeohu2NFmSuB2gzdB0NnnnH8m11OBcLhZpxoQ,17564
1064
+ vllm/model_executor/models/config.py,sha256=S_wWVBEAroo77ldsMPMomyUTWkFhXQenmJAaF32RZkc,21097
1065
+ vllm/model_executor/models/dbrx.py,sha256=9hpx2vCyc7D2bY2Fh4kyQEG7U8_BxWmA7oFWJNiUoI0,17375
1066
+ vllm/model_executor/models/deepencoder.py,sha256=9NJVyE12W6YV63lat-i9a71d77KNxDQmPSubPNWuYVw,23686
1067
+ vllm/model_executor/models/deepseek_eagle.py,sha256=02mvbqtH51KWwMo98_LJj5B8gulhEhgw-RZ6NfegBqA,9647
1068
+ vllm/model_executor/models/deepseek_mtp.py,sha256=I9NTBWFICBT3C7qoLMxY-q7YgWUSazYE9Ow8QZ359f4,13536
1069
+ vllm/model_executor/models/deepseek_ocr.py,sha256=rEiFC4YN26-WAA2arDbaaVijMMrn3g7Q2G6dJ1n3b9A,20413
1070
+ vllm/model_executor/models/deepseek_v2.py,sha256=49Voz6biughQA3F1JawuN7oGoEoT3eCS8k-AAqq3KPc,62896
1071
+ vllm/model_executor/models/deepseek_vl2.py,sha256=WUe0Z-ybIy-EQZoZHz9U-JY52X2cydZktGE60Jkhdnw,23427
1072
+ vllm/model_executor/models/dots1.py,sha256=MGt5TbW0U6VU4H9vakbvCbU1Ev7uJB5m8BIeMpvvMqE,21150
1073
+ vllm/model_executor/models/dots_ocr.py,sha256=KDqDeVNe4AC-AJD_gC1zO9LKK1L2_Ob_juQp31wPixM,31255
1074
+ vllm/model_executor/models/ernie45.py,sha256=gxkle1tmFbYsDkQx8AjoYZ-2_e93ALUtHMMdVnPA_fY,2244
1075
+ vllm/model_executor/models/ernie45_moe.py,sha256=dWtzn8XTxiXKZ17BCapFVoPcxEagjgqjVhxaDlX1Hjo,28246
1076
+ vllm/model_executor/models/ernie45_vl.py,sha256=QsHhqgkiNL20dUV-vTCM5Ecs3DwV6yHD9fxYNTUDWoM,62336
1077
+ vllm/model_executor/models/ernie45_vl_moe.py,sha256=Opo6FY7hr_oPpoCzRH3sKudZdSDtmmU8q8vwhlbRK8M,30470
1078
+ vllm/model_executor/models/ernie_mtp.py,sha256=2Lh7jpKYkisUeuVVc4spslBo_eYCYWEhHlO32pHddZA,10568
1079
+ vllm/model_executor/models/exaone.py,sha256=VwajqefS7tZRmvDODg6D5Gg9u8tb3gnU8Vcoc2Ec5lg,19658
1080
+ vllm/model_executor/models/exaone4.py,sha256=_TiubIDwQauMAtBHT5y1SI_oceH_USLtYwFI91FgCwY,19214
1081
+ vllm/model_executor/models/fairseq2_llama.py,sha256=dIadCgG1u8yASjN4qmOTvkli4xxu-Ygl3t0iL2Czsik,6375
1082
+ vllm/model_executor/models/falcon.py,sha256=ELhut8ADTWeUIA-faX-KoEa1n7048ShVjE_TWTw5X9s,20542
1083
+ vllm/model_executor/models/falcon_h1.py,sha256=-pMErltxdkZAuxEJwR8jWS27G4PjYJJchbdxJVRCgww,24470
1084
+ vllm/model_executor/models/flex_olmo.py,sha256=eWFSijc0m35t7xMZKdsb0cXZv1OdI3vbAw8nAujahMI,5624
1085
+ vllm/model_executor/models/fuyu.py,sha256=9Dfy6RAbw7BgHkx91KV89SUFnUTdTS-DpRFXf7t65nQ,12952
1086
+ vllm/model_executor/models/gemma.py,sha256=Q6n0OQV6hvyxOkvnQkZliH7LGIWXY5WzH9n5UGEjJyU,15553
1087
+ vllm/model_executor/models/gemma2.py,sha256=Zl7xD8oXoqtAzKHsL59inuzfArf1vC7n0BtEibLujJQ,16482
1088
+ vllm/model_executor/models/gemma3.py,sha256=4_4TldiJaOOca65DJy8xQBPGdZTc5P2ZoZXu5pQJMsc,22057
1089
+ vllm/model_executor/models/gemma3_mm.py,sha256=TRdT-to2Jav-DzFshE-untgGSchHzOz5uoY15Y-NVcY,24891
1090
+ vllm/model_executor/models/gemma3n.py,sha256=lQvv9DTakkz4CBMWylc4HP85c6Dmi_WzwfMkJAIrwLk,44047
1091
+ vllm/model_executor/models/gemma3n_mm.py,sha256=B2apzMJ_dWnEwMlw12y942qkHeyipjdNu6zEixvpfBY,29732
1092
+ vllm/model_executor/models/glm.py,sha256=FhBl3j5GNEFJFqiPe76ccfMH2BW8z7FzaQ6Rs-0rB9c,1059
1093
+ vllm/model_executor/models/glm4.py,sha256=F4BEsUgGGKmEe03PM9uG14dfXHu7G1nHcZ06QmMZB1Q,11003
1094
+ vllm/model_executor/models/glm4_1v.py,sha256=4TE4dDEKnG1tjZnWlB2vVU1VjdMu4hY6qjfYafB2DrQ,66326
1095
+ vllm/model_executor/models/glm4_moe.py,sha256=mvpQ1YYfadkLxmT5SRYpa7Mz5RYp4MGoD2dekgEy2IU,28516
1096
+ vllm/model_executor/models/glm4_moe_mtp.py,sha256=ILXawxkleeK_jaTrKBIhSGnMSPnCX6h-YbNGWqyBdXU,14106
1097
+ vllm/model_executor/models/glm4v.py,sha256=Hy0gNqX40iKdJGvYYnvN3JYnQNLV4d3gk3xEYGFGlR8,25720
1098
+ vllm/model_executor/models/gpt2.py,sha256=VDrK4N1_IxXn1YRMze2u3ZDpo5t-jGkdx7wZIohP2pw,14310
1099
+ vllm/model_executor/models/gpt_bigcode.py,sha256=UxvOCdzvDIcUBV40t2qjUvrW9buiLtiBszHfMcZIjaw,12129
1100
+ vllm/model_executor/models/gpt_j.py,sha256=t6x3vWA0Oya2vBrRlb2Z6Snf-7RRSTWgJ5-IZslXjVI,12685
1101
+ vllm/model_executor/models/gpt_neox.py,sha256=d_CdlRdt3rfGxQN2lBx1vAvijqH3LmsNDVRfRWXJafw,12839
1102
+ vllm/model_executor/models/gpt_oss.py,sha256=kO8X33Dsu7x-iglE1JmLKxvWkYZw1DrRE6AXQn2VO1o,28494
1103
+ vllm/model_executor/models/granite.py,sha256=3fbaGtlQhrBJWWUDH2bb-CVKbjGPCUzJw51ZMLyemZo,19277
1104
+ vllm/model_executor/models/granite_speech.py,sha256=C1pHZFSNl8mfH4qkPmXm7qAVuiuU4jNZrPZnzlWQ8Dw,34655
1105
+ vllm/model_executor/models/granitemoe.py,sha256=SSjpWh4392j9d3SjMrt0X1OLA5Mt0axPsix2VklJXcY,21460
1106
+ vllm/model_executor/models/granitemoehybrid.py,sha256=26h3vdW7uCjKWLAn3ctCqMrX5WyczWRJXLQ2_19hf0s,26035
1107
+ vllm/model_executor/models/granitemoeshared.py,sha256=n8LFN8wWqZ9W9BBHTeh42c5ftUPxq9zFUHP7_9W-tXE,12255
1108
+ vllm/model_executor/models/gritlm.py,sha256=dgvZPU2beOzObEdf1iBpTcpDYDyJvUymfWEPRPuAJfw,8284
1109
+ vllm/model_executor/models/grok1.py,sha256=TLurRBB0JczpFKCLOdNFpOFHHYPYXOpcUpm1tDKt2Nk,20729
1110
+ vllm/model_executor/models/h2ovl.py,sha256=q2Aq8oC1n-XhFdPpu-rL1o5UlBagoz4DISQcQZk6otk,17846
1111
+ vllm/model_executor/models/hunyuan_v1.py,sha256=5Iwn8RQSyT9VqsPzBKfrBLZd_3CT2LpLKWbHF4HzEFs,40026
1112
+ vllm/model_executor/models/hyperclovax_vision.py,sha256=4gkUTlc-VUM_eIODLuW7VGvSOu3wER6HwztLNATezjU,40447
1113
+ vllm/model_executor/models/idefics2_vision_model.py,sha256=WfD9atpKbpIcZJYPw3Y-TyGNu02l069CqwpBvnwnu-g,15481
1114
+ vllm/model_executor/models/idefics3.py,sha256=TfUyfJcgKmx2btXFtMiGL7-mWhqjXnxSNJvb0GVZDHE,24612
1115
+ vllm/model_executor/models/interfaces.py,sha256=Owwhn5zZUuUFsvEQ1cMdeawS6RcoqTn6pNgPVUZLR9o,33564
1116
+ vllm/model_executor/models/interfaces_base.py,sha256=YiA2ajHjfsfzh2ox1ZqfAO9OSYviodSx0TcgRlmnLG4,6265
1117
+ vllm/model_executor/models/intern_vit.py,sha256=zAXli_HJc4hHdE5Q_3HxH3EJabMTUTWt-YdfASbhurQ,15041
1118
+ vllm/model_executor/models/internlm2.py,sha256=I2b08eL2_exZYuAP2NVIM5UYmMd9CZkPXf58_o37bbs,16382
1119
+ vllm/model_executor/models/internlm2_ve.py,sha256=cX8yOH9ir1mT7pMNTgcAljJ8CQvWKQK87rhBx9Ox0Qo,5491
1120
+ vllm/model_executor/models/interns1.py,sha256=Udo0PZf1Wq6jZYzYefkXION0e6dcofdi7eINBJ5gMfE,30635
1121
+ vllm/model_executor/models/interns1_vit.py,sha256=iAhX4TadBjvFQzpqHJzlOHQnTRqeihM6bmbFUNHJXUA,14851
1122
+ vllm/model_executor/models/internvl.py,sha256=N0wXOprX0W8Oo7hjkWeMK52ieuyaNaVrT_w1HNlyTzc,49913
1123
+ vllm/model_executor/models/jais.py,sha256=iNJ9_77JlqHr74z0i_5WgGWLMcuKXpphCbnW739Yh8s,14192
1124
+ vllm/model_executor/models/jamba.py,sha256=vvYRH1kP0c8xfpvQRTFKdzcqb76HfpLvVubZNs2w5B8,21644
1125
+ vllm/model_executor/models/jina_vl.py,sha256=2Q7RF3RrsecoSEUwyK2bAl7ZTf1zwExUZiRdCjXH3WM,5244
1126
+ vllm/model_executor/models/keye.py,sha256=m_OdOQGDx1PE3dONOh-55GbRh8I2mW5Q4YF12CjjyA0,61357
1127
+ vllm/model_executor/models/keye_vl1_5.py,sha256=KJEMR34TguKFFGuC1tnTf-T4akAiKDP6XGsMblIZWO4,24902
1128
+ vllm/model_executor/models/kimi_linear.py,sha256=GhoVtuXRbm1LI47zjFjPrq4VczRaP_pDEZmHE2o7xwE,23993
1129
+ vllm/model_executor/models/kimi_vl.py,sha256=qt3VKGQKcFL-44Vew2_Mx5ByexP0WLiSGmOCqPoNghA,22169
1130
+ vllm/model_executor/models/lfm2.py,sha256=wSkM28pZBx1xoxjsSFALwBsFceN4cy1unobYXizJAsc,18915
1131
+ vllm/model_executor/models/lfm2_moe.py,sha256=4rSatjTQQAsndeWvqxNEqpEjnxX6Vr073AY3yU404Vc,27463
1132
+ vllm/model_executor/models/lightonocr.py,sha256=8WKI50EwvphQp4rn2NKKWYW94NhgdVlflL1Wl6aK4rs,7062
1133
+ vllm/model_executor/models/llama.py,sha256=7MIYZPnssHbqeNpnY_yXwkPiSmcTsipbuoK5bPzLsxw,27676
1134
+ vllm/model_executor/models/llama4.py,sha256=BcdkU9NtsZvkfYzhoL9bfKP3-7QxenJJGi_U62eXZRE,35361
1135
+ vllm/model_executor/models/llama4_eagle.py,sha256=nebxeHwtL7ktNf98oF5uLJm1b5Qf1XxSZssOKvqpBnE,9153
1136
+ vllm/model_executor/models/llama_eagle.py,sha256=xtpep0xbNNDlJJqtxc3T-FyzeHMAKJ7FTcvFXix6BE8,8176
1137
+ vllm/model_executor/models/llama_eagle3.py,sha256=hwRjjoKjbL5qvuBbGd_Fxj0nwTtZfuIOHzfePPkorsw,13704
1138
+ vllm/model_executor/models/llava.py,sha256=s71aQeButQnRZwRE6s2w7a1O84lje-v5MBUm7ncjI_Q,29040
1139
+ vllm/model_executor/models/llava_next.py,sha256=sIONContY3Fs8UaxJV6V648JWIPZ50FoAlpvk_vBl8o,22042
1140
+ vllm/model_executor/models/llava_next_video.py,sha256=_uMNJ5LfCwqOwZhXbVU5UWHsfpLJmsnyqGJOAR-UlTo,16144
1141
+ vllm/model_executor/models/llava_onevision.py,sha256=UBXJkxS_mquBX_VninkmJgGdTvmPkfeiWoYT9KsxIsM,32819
1142
+ vllm/model_executor/models/longcat_flash.py,sha256=syprwLVItbGfq8BNE-V1g2YnDU3z3k9-wVwbTVoUyp4,27904
1143
+ vllm/model_executor/models/longcat_flash_mtp.py,sha256=tN7zbhFVQj7XvokoXYjXezg6LQhmVILKigcm2FknOLY,15551
1144
+ vllm/model_executor/models/mamba.py,sha256=su7lsV3dqkgTgsc4TTBca8QVuhg2hl4e5OXvqFMNW7w,9787
1145
+ vllm/model_executor/models/mamba2.py,sha256=xgGkDtsg6IJWwFomY9-PjvGuH3WDPx3pGbfwgEd7fqE,10048
1146
+ vllm/model_executor/models/medusa.py,sha256=ljZQYNwpS3jkemq1r2ghTc2Z1LW8ncpdfh6qzY20TKg,6738
1147
+ vllm/model_executor/models/midashenglm.py,sha256=DphN5qkKL4PH6f-uhh0XXB7oD3Ct5bdTacgjtGU8wRg,27759
1148
+ vllm/model_executor/models/mimo.py,sha256=4b4Ezgz1uHqVOyhjxm8_FOd8IthV0WXoGSqTs_zq2Qs,7332
1149
+ vllm/model_executor/models/mimo_mtp.py,sha256=LZGDXqszX94QsLS32RRGuC8d-cFlM2O02mjwgC9oRc4,11061
1150
+ vllm/model_executor/models/minicpm.py,sha256=CyjHU8p9gkKeSNcEz1pq7g-g2Acbi_5J6GQp92Rro10,24045
1151
+ vllm/model_executor/models/minicpm3.py,sha256=ROK3IRlIDw1Z5US2qAmMSLTuQ9LaNg06ulVXPrPIc2Y,8857
1152
+ vllm/model_executor/models/minicpm_eagle.py,sha256=L8T30an7NZ08dPVEaDcb2cFT9iXYxgO7vCahMPxsoE8,14590
1153
+ vllm/model_executor/models/minicpmo.py,sha256=mD4uCUlkpmdeW7kPJoU-BITGtex3uuvvH2RZN22Iy9g,26859
1154
+ vllm/model_executor/models/minicpmv.py,sha256=-CXKHVREtVU4K2Op5XmWaSNfxJMJzmHWR270GjDLDl4,59187
1155
+ vllm/model_executor/models/minimax_m2.py,sha256=rybMPshccwNvMCGvvlAqeIu4XdLjXDYSTTvinhCPOgs,20712
1156
+ vllm/model_executor/models/minimax_text_01.py,sha256=8Tf5UrbWkbLFhkuzfu5ICvNM0fJWl0f50DlhBsb6opc,38050
1157
+ vllm/model_executor/models/minimax_vl_01.py,sha256=O_vkO-GV_dZ0xt4kr-XuU8IXCqOdq7rtdtR9VL7lGY8,14493
1158
+ vllm/model_executor/models/mistral3.py,sha256=YKm4atzzYypLrmPQVGvTBYpyERkxME6utu7bwLgkXl0,22082
1159
+ vllm/model_executor/models/mixtral.py,sha256=OeZFod2PCUoO50iBUnNTiXZ7HqScD3s-NjPab5FhNYA,23588
1160
+ vllm/model_executor/models/mllama4.py,sha256=mhwBssNoktKbBR-HOMXC5tkNUR7GJ3KG0yyzTgRTIio,41858
1161
+ vllm/model_executor/models/mlp_speculator.py,sha256=HDvlP-QB88Z6facl3fDKmIck7dtSWjpifteV4i1E23c,8457
1162
+ vllm/model_executor/models/modernbert.py,sha256=5h5k0_Vf_g112kDJ2UKBJAkY_CxYq2Btn91f6VjyKM8,16018
1163
+ vllm/model_executor/models/module_mapping.py,sha256=YdEB0K1EaW-BbLfkr-IvlO-E9pv7G5qrYUn_vQwnLis,1688
1164
+ vllm/model_executor/models/molmo.py,sha256=6go4WUSUPp3Z2ByITkQBoJ9TWmBlzLHy6nUcDK2vDa8,51981
1165
+ vllm/model_executor/models/moonvit.py,sha256=F7lXCIVGtWrA0j0uIL_-KcwmGq1fdEFiDbfp23p7i-w,25043
1166
+ vllm/model_executor/models/mpt.py,sha256=DIgijo0b9h3y-UqGDrNM0XzEjEOyCbNM0BavsN24kO4,12113
1167
+ vllm/model_executor/models/nano_nemotron_vl.py,sha256=xe82WUAS_tlPBMx6caLOh8Kl4o31pXMYgmcql31orns,64730
1168
+ vllm/model_executor/models/nemotron.py,sha256=slXHckrHiRQ2OXKHM39vWe-Bu1Y6kim4QqRKM0P7yVo,18745
1169
+ vllm/model_executor/models/nemotron_h.py,sha256=ECbHsy524D2b8OmfPsmNSw7ihZ6jFhOj1ulUrT3JlIg,30536
1170
+ vllm/model_executor/models/nemotron_nas.py,sha256=ksIXP72Sk-aNicKh2eO9Fr6ZND5aWaN_07YwGT0-5F4,17602
1171
+ vllm/model_executor/models/nemotron_vl.py,sha256=nA69LMmQ7wM6ahMBDwoCKfxhAZ65mswUmPdTWzAUeNA,22522
1172
+ vllm/model_executor/models/nvlm_d.py,sha256=MqXkN1AQ-SJv9UCAbpdTvOe4-R1yVcG3wEOv94z45nY,7636
1173
+ vllm/model_executor/models/olmo.py,sha256=A7fG3BTpsDU9qLaN8fxCUnowCRC6aoo7Zw1YzbGUKHo,14385
1174
+ vllm/model_executor/models/olmo2.py,sha256=QkkGsy5qcHaXGZC00WHImE6dmDb38TqzQjvKHjRRUvo,16562
1175
+ vllm/model_executor/models/olmoe.py,sha256=qSFpWxdCPAAVwZSTqocns7fMEMdt7QpYXYJZ92xzWvg,18573
1176
+ vllm/model_executor/models/openpangu.py,sha256=fueP2O-5EPP6q49oJInqMsC_O1qNzkbKCbuNB-mIcf0,39633
1177
+ vllm/model_executor/models/openpangu_mtp.py,sha256=1MEp3hje6mdaayhvzH7AeSVxkF29eEpch16tVM6rJow,10369
1178
+ vllm/model_executor/models/opt.py,sha256=7i2HYQQAMXduVzayx9nACyQmKCAXjiIBla9jLB98Ri0,15675
1179
+ vllm/model_executor/models/orion.py,sha256=lNQno1vaPtbi4PuV-oP92nbtzy6istFzC9yeLmLARKo,13477
1180
+ vllm/model_executor/models/ouro.py,sha256=616UEBvdC-U25qT1rh1eSfQArXLIxk_0n6alEnWK7h8,19143
1181
+ vllm/model_executor/models/ovis.py,sha256=4m2NXluwFSlLrJS72Un_6avItp1YYiNirf-mYVcLzgk,20342
1182
+ vllm/model_executor/models/ovis2_5.py,sha256=mnDNcLHbd1MoZYT2IwW99vwkZDdZJSRC-GmcJG83bzk,25087
1183
+ vllm/model_executor/models/paddleocr_vl.py,sha256=un_shrr0BJwQjXtWBCgb3-sX306Hcu1zcSTpn30jk0k,50179
1184
+ vllm/model_executor/models/paligemma.py,sha256=g3C3aG3yKSAaVAdrgw7chWdWA_Y7gqWtOqcWJGQGtCI,13595
1185
+ vllm/model_executor/models/persimmon.py,sha256=mqQMaTrMo3zv52XfVIZhGwmgFlpeXTea1pUx3W7efko,13776
1186
+ vllm/model_executor/models/phi.py,sha256=srs6KU2ZiKJ34TwmXNbBid6SZduYV49QAHfqHSzGFyg,13404
1187
+ vllm/model_executor/models/phi3.py,sha256=N380ApMP7_3VL-rlWeVD27SUjC-kLXbldzYNWgG8ApM,456
1188
+ vllm/model_executor/models/phi3v.py,sha256=OYg-00ZSvzanq1eq24Yj7vCkATEITb15HNiOVvcNnhQ,26075
1189
+ vllm/model_executor/models/phi4_multimodal.py,sha256=DrtKdfv1axdlEHcb5AfCryb64DO5x6lkPGTtKThx4xI,52221
1190
+ vllm/model_executor/models/phi4mm.py,sha256=d_yQk1aIZcHNmmCvMy63s5shz5fqOcsc-1tCectHI-A,45294
1191
+ vllm/model_executor/models/phi4mm_audio.py,sha256=EML6r9O_PY-LQZQwCbC45nvf0QY2j0pevPYKPuZNsE0,50483
1192
+ vllm/model_executor/models/phi4mm_utils.py,sha256=un2oSpNyKdnDN9wLMnBn2CzxWO1xNeRwAPyV4pI6Gj0,67179
1193
+ vllm/model_executor/models/phimoe.py,sha256=dui6VS8XPmnF_Os91wl-jibUydAILWwqyUrEORplt3c,23629
1194
+ vllm/model_executor/models/pixtral.py,sha256=C6pNRJv2iYEYDSX4ZiMlMCNrrt-RuJ-9b5cI_X-p2Tw,47382
1195
+ vllm/model_executor/models/plamo2.py,sha256=7i-jTBbzxDLZhzPL_v5OJrqCDwQEcVAg85hOKWjk2Yk,37541
1196
+ vllm/model_executor/models/qwen.py,sha256=olqSTQYvQVYu_qkf8Np5GlTzWCALD8R_2zNy9HcQVxc,13107
1197
+ vllm/model_executor/models/qwen2.py,sha256=IHRGQTusGeDLRLrP8QutHdD7kluHjUuPQizNTgxTuqc,20102
1198
+ vllm/model_executor/models/qwen2_5_omni_thinker.py,sha256=25Sa8LHxotOl-kwxHIWvl3zYhCaUVVEYn1smu6BZEWs,48152
1199
+ vllm/model_executor/models/qwen2_5_vl.py,sha256=YJ7a9v59MTT9YbYYika0ANapkg4oOV6ieW-FpTatmYg,58171
1200
+ vllm/model_executor/models/qwen2_audio.py,sha256=otDiUl9mrisB1AFvbi_fOpiYEHp0vPW7Qz2Ql0roaP8,16912
1201
+ vllm/model_executor/models/qwen2_moe.py,sha256=tAbQLs9W5dwAnlastPpcBEeVnbQXSno5hpMgGfke9I0,22960
1202
+ vllm/model_executor/models/qwen2_rm.py,sha256=K3lpU9hDakDURPIClLkzT81ZwOAUAsBPtvRpevJbsB0,4131
1203
+ vllm/model_executor/models/qwen2_vl.py,sha256=uWqfefV76y2rgJSV3OSblsj-j1pZ8eGB_VOk1trh5oQ,58174
1204
+ vllm/model_executor/models/qwen3.py,sha256=gpXVO4HSNTNAfJvIHq70CckXlkbDKyWAtz32v5n3kC8,12512
1205
+ vllm/model_executor/models/qwen3_moe.py,sha256=MTC1ASNBU2TU003rvsSQe7K0yQJjs4Zj2uZBWH4jsmY,29268
1206
+ vllm/model_executor/models/qwen3_next.py,sha256=SS1VjmVBnYZvHPzcCSJK1LcWSXfOr9ptjqYyJNQWdwQ,51108
1207
+ vllm/model_executor/models/qwen3_next_mtp.py,sha256=beovwMRrBMh5AIqjwydzTVLbR9XJ2TsIylrfA8HxL-M,10691
1208
+ vllm/model_executor/models/qwen3_omni_moe_thinker.py,sha256=aFVgPC3Yo6utJOauJuH7xG5QG2eiDHtE-EO1LHAO8JE,67973
1209
+ vllm/model_executor/models/qwen3_vl.py,sha256=7AMkx1v5LWSgyeAcplZA_qcbiR5Z1HEtQi-N7eNUaRw,64317
1210
+ vllm/model_executor/models/qwen3_vl_moe.py,sha256=duC1778HFfYpdAesqA6ZUeN_4_8lGy9JAM46oy5-LxI,17302
1211
+ vllm/model_executor/models/qwen_vl.py,sha256=oepn8QFaGvE6S8xrb84-lN3dCC2FjiKyQs5O97VUrdw,25182
1212
+ vllm/model_executor/models/radio.py,sha256=wZd79PQaK33rgnguwVacOop8-ab7zn4vSbeKFcAacAM,18240
1213
+ vllm/model_executor/models/registry.py,sha256=Er0DgkpmrrnolXh3lt-M5UT27ml7AXUR0Mv6my4EM30,45475
1214
+ vllm/model_executor/models/roberta.py,sha256=k9DeYeEXRgMRUjWM5YpZvXlRRGFFiWi64UhEIgzUiOk,9630
1215
+ vllm/model_executor/models/rvl.py,sha256=-XiyV1RZrjlWPfJhnL0K8ulP02dHOLbt0DiLuWdvhLk,3567
1216
+ vllm/model_executor/models/seed_oss.py,sha256=V22A979tLGcKlSZNnPBK7ueWee6p26gymwsw_pjy-2A,18034
1217
+ vllm/model_executor/models/siglip.py,sha256=LOD_bJiXthyc5i4iISn2ZaLTRNSDK0anQgL25B2SDBA,40188
1218
+ vllm/model_executor/models/siglip2navit.py,sha256=OlRJqBks1S8Pn0F7YLXQ7UPNemRHwQCq6NO_jwsCMS8,27112
1219
+ vllm/model_executor/models/skyworkr1v.py,sha256=71_9fQW6nTyXeu5oiPYIcIXT_k67D7S1cVBOWEdj-ZU,31839
1220
+ vllm/model_executor/models/smolvlm.py,sha256=ZsY0x3CaHHUuTBggSWCwGv-HfUAiwks8jZRNoMd7hXs,1450
1221
+ vllm/model_executor/models/solar.py,sha256=Ie2iykC4NuzgVXcw7KA_SQLt_Q3AYPJ1mv_Rv_AL2bE,18452
1222
+ vllm/model_executor/models/stablelm.py,sha256=S9ysUCdcmfk6K--6aOuYsrZQ13oFKI-Z8OlgmQ4OpOM,13774
1223
+ vllm/model_executor/models/starcoder2.py,sha256=cM3Y_KcdpXj7HXEDmKB5DK47HgD8XNjhm0W4c-RIqx4,13434
1224
+ vllm/model_executor/models/step3_text.py,sha256=74YHwA4kFE9ux8BMuY6aoWYUpPR99oiOXuqUDCRGVw8,20024
1225
+ vllm/model_executor/models/step3_vl.py,sha256=ODzDhYpoljoDuJZIb3Gi0ZGc7YuOaQACBfdZ5XsU740,40314
1226
+ vllm/model_executor/models/swin.py,sha256=6HMXPVyWBQm59cAC4lunRkzSFCcHk28pi_eCKQkNWpI,16978
1227
+ vllm/model_executor/models/tarsier.py,sha256=EylnPwlRQkKEZgC_nhqQUewRe3-WaMiCL3SJwxHihgM,22850
1228
+ vllm/model_executor/models/telechat2.py,sha256=wOEquSezPeLg2qqUNZ7ncsIzpfxHxRCrFoproADyARE,6245
1229
+ vllm/model_executor/models/teleflm.py,sha256=ssXMikvm-C8nI7vFXnPCDaDQfQt97XQVDcxw1U152T8,3050
1230
+ vllm/model_executor/models/terratorch.py,sha256=cgMC4sd93oI-T-EHlz93_-0rXuLPkN3bBA0jYXnxRLU,11246
1231
+ vllm/model_executor/models/ultravox.py,sha256=qEs2_cGYwJAl6lwX4Eqwem3cOLFhXFYTZ35vC2ELoXg,25502
1232
+ vllm/model_executor/models/utils.py,sha256=CteTUotEpkkolPVFZGIlfvupBqBvVkJqxmt7Ct1QdhM,28954
1233
+ vllm/model_executor/models/vision.py,sha256=5E7o7uN5yOb8tR4kdl_v7cByz5AUMOaLu5Pc_D-bcNo,19421
1234
+ vllm/model_executor/models/voxtral.py,sha256=vM8bPKblSS_oiQ9yRkpeGvQSeKO4X_e-K_2FKz64KYs,30727
1235
+ vllm/model_executor/models/whisper.py,sha256=4cJa-L4mUg3ufXZJ52DNQf1K9xH1TAIrXq2Dzpjzngo,33017
1236
+ vllm/model_executor/models/zamba2.py,sha256=8MAKCEqir7JT9Ry9BubmgcFTmAhYjmAnAsJjLxla3Mk,35500
1237
+ vllm/model_executor/models/transformers/__init__.py,sha256=KR0hAoCBU4oqFfOpZ1kT9u80zQ_s-Z6HWpTfkvra-iM,4264
1238
+ vllm/model_executor/models/transformers/base.py,sha256=tIvuOzBCQXQwp3G-zvojKBXAuBYimcg8T1bet7Xvibw,18553
1239
+ vllm/model_executor/models/transformers/causal.py,sha256=WkomK1hlsPamiJU7wi-Y-kXtPKo9j8e5mEb7rUn2_hM,2609
1240
+ vllm/model_executor/models/transformers/legacy.py,sha256=NY96kEaStY66o4Sa3tOFCqLhhwgiIf3eO2dnViXowIQ,3229
1241
+ vllm/model_executor/models/transformers/moe.py,sha256=A1rpf3HxIxzS-PAgkkWaDXNTeH3dOowkfcTzAXeUxoE,13706
1242
+ vllm/model_executor/models/transformers/multimodal.py,sha256=TiuayxGZs0CRu2ULpfV0Ug81nl9SvIvr8NLV1X1NSwk,16398
1243
+ vllm/model_executor/models/transformers/pooling.py,sha256=b7wvQD_BcNhKrRTfatvYapeHZq4TbP7gnaB5SiGcjbU,4606
1244
+ vllm/model_executor/models/transformers/utils.py,sha256=jnTmC7PGngKepgcH2bVvno9Pc97IHpAsdnclYzSS19k,7345
1245
+ vllm/model_executor/warmup/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1246
+ vllm/model_executor/warmup/deep_gemm_warmup.py,sha256=HP_WkAMOTu9t2viVV7jtK9xA4Cb_ueNUpWZ_pRZ-QZs,10885
1247
+ vllm/model_executor/warmup/kernel_warmup.py,sha256=mPsA-O_KblSFAvqMfIiYfz5lO2d4k4qTnuvvs1Hf00k,3574
1248
+ vllm/multimodal/__init__.py,sha256=12IM-iHLb9Uko5kVWANuRPcW6-fxcPP7VjJ7yoQSjpY,1031
1249
+ vllm/multimodal/audio.py,sha256=KxTk3rbnBIyKxphd119G6hnvzIo9yYmPes9KFZ-WrPc,3384
1250
+ vllm/multimodal/base.py,sha256=lfe3t_e1h6x7kjZQkuwsdOWNbzb2ulAOSW7qvM_lfjo,701
1251
+ vllm/multimodal/cache.py,sha256=5oUbZwMoMmGdg3Jz0x-qfHR3i-GCIgB4Us7bJSFuE6s,23113
1252
+ vllm/multimodal/evs.py,sha256=jesJ5ZpXWjL1BcSDVx5x409viFLIHQQnTaAlwrcn1Qo,10887
1253
+ vllm/multimodal/hasher.py,sha256=27tiR_7YEdENFbYhjCK-wUF3ezT43AaAOBuBzRlq7Jc,3710
1254
+ vllm/multimodal/image.py,sha256=FglAc-z9-wLqIEVEDBjDDCkDoIOsRDvZnqXRAKV4jnE,4375
1255
+ vllm/multimodal/inputs.py,sha256=IQsqiXCZQa6Ay7Rhf6kk-SNV5UBN-6KF4mu-gUdjiK8,31855
1256
+ vllm/multimodal/parse.py,sha256=SzDRitBZ6P9QNTOoOocy9nQ8GTw71Llf6VdkZ7_7P2M,16362
1257
+ vllm/multimodal/processing.py,sha256=dij1vmL-70sXfmfPzgmpq61IkcVXvrgnYWMv8HzBm44,71565
1258
+ vllm/multimodal/profiling.py,sha256=_Aj_-CR_4Apta8G14vZ8UQpHQ6SjqYFDMlAeFZfXcyw,12156
1259
+ vllm/multimodal/registry.py,sha256=A9npb5qvipZFRvdB8wLKeY9IW2GKlV0Vs203Bz4tohM,12221
1260
+ vllm/multimodal/utils.py,sha256=nQgtdBdKmG05nu5bzFHh2TAFu653EYML0D3PF7Pc3w8,16505
1261
+ vllm/multimodal/video.py,sha256=QDSRmTi9H4s31JwCMhBsqo2PGVeDCkNTVV3wN3k_vH4,10192
1262
+ vllm/platforms/__init__.py,sha256=9MJt87rVxIFAeoDiKAQWoMsJOhD8noEyUZ1SUakJWvo,9896
1263
+ vllm/platforms/cpu.py,sha256=DicpyAZrWqn8Xg-7GChHw1I4pGxrDJl_tTKndasi94A,15363
1264
+ vllm/platforms/cuda.py,sha256=IUzz015G5rBDzGC4oscX-vagjDIP2WY-RiyXIra0_Eo,23556
1265
+ vllm/platforms/interface.py,sha256=FXAIKQQ8tqooJFnPqcliWx3p1WcYFrEmuSXJL-t2kYY,20382
1266
+ vllm/platforms/rocm.py,sha256=9dqbQyZfoKMJ4CEHjdcVkQvqXWjm1sPxfXYI-DEge44,16332
1267
+ vllm/platforms/tpu.py,sha256=MMy0w0pGLTIGsTFV_SCbH73Bz9joN-GuQBX3bdBZqxs,9189
1268
+ vllm/platforms/xpu.py,sha256=6IQsLWGxKY8sHsI6aHoRlnPVhNQzKAWb3ocPY94qTnE,9909
1269
+ vllm/plugins/__init__.py,sha256=26-U9UX1KQT395H1mq43lZIZjDHM0qlK5GZokF_5bQU,2683
1270
+ vllm/plugins/io_processors/__init__.py,sha256=UhJEF--e0B816MCdrtyBxon4RQgsuM6vxLTBu6V6Y4A,2517
1271
+ vllm/plugins/io_processors/interface.py,sha256=mcuEoIDY4SgvWT5Lp5QLRo29OuNkGA--PsLArS193DU,2574
1272
+ vllm/plugins/lora_resolvers/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1273
+ vllm/plugins/lora_resolvers/filesystem_resolver.py,sha256=Frj6--zxNA52jqW6UhIEjH5P9DoTaup3N2WeMHyF4V0,1957
1274
+ vllm/profiler/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1275
+ vllm/profiler/gpu_profiler.py,sha256=fUNPYYf5m3F7oKuDogPR0BC-c9w7R0gdK86oCSrL3Zk,1210
1276
+ vllm/profiler/layerwise_profile.py,sha256=p_EhWfWFVE-Nkhiz6VlwR5S5Y7OUEx12mvJp3TE7lsQ,13424
1277
+ vllm/profiler/utils.py,sha256=0rGR3LYSg6uZitAxN0bKgruTtVwtvytsNfbFZz4QLjk,4659
1278
+ vllm/ray/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1279
+ vllm/ray/lazy_utils.py,sha256=3fL417TkcJ3fEBJxln9nVO-m7Ynwm_B2W8SfaGtrRts,555
1280
+ vllm/ray/ray_env.py,sha256=hi3BEix6Bwyaaunm-2KvvhCuqV4AD3ooK6SQQDowizA,2562
1281
+ vllm/reasoning/__init__.py,sha256=iE1WyvEZBZhFeW8kNREYw_aiMwvs98Irr9obyTbvBXw,2346
1282
+ vllm/reasoning/abs_reasoning_parsers.py,sha256=NT8sV1niBEZwvebX9wH3GhkWeolOkr5bTvUR_9uHU1I,9578
1283
+ vllm/reasoning/basic_parsers.py,sha256=j0Wl8__LD3slLuEbrxtz_300dyD3spQ-wqJAG9l58vw,6466
1284
+ vllm/reasoning/deepseek_r1_reasoning_parser.py,sha256=aCkmFQVv4d-4EqYMbNtdHnozRs_5tLrOJWfXXHT6U-E,2351
1285
+ vllm/reasoning/deepseek_v3_reasoning_parser.py,sha256=xoymChTclMZWe1MLIysKFen4NrRwoUGmWr1Gd1MaL_0,2176
1286
+ vllm/reasoning/ernie45_reasoning_parser.py,sha256=FWSQqSHiUxh9VeR0jP7fMbWVjeiwnKs8iqy9Eudy-oI,6828
1287
+ vllm/reasoning/glm4_moe_reasoning_parser.py,sha256=9iogIA3Ss8_GgqAScUZnbpGygakEf1OBqMWJ5ERc5_s,6893
1288
+ vllm/reasoning/gptoss_reasoning_parser.py,sha256=e1_-mx24_2pLk0OWu9AjF3x-jzArD_G0YCvF8l5mR_4,6796
1289
+ vllm/reasoning/granite_reasoning_parser.py,sha256=d6BikPcKaIBd5WkZaIm2dG1H8HUrCcjPR4gEePz15js,15196
1290
+ vllm/reasoning/hunyuan_a13b_reasoning_parser.py,sha256=sTm9xUNnoJ73zevO_Z0hkx_O4SriOKNe6e7THs_NCYU,9571
1291
+ vllm/reasoning/identity_reasoning_parser.py,sha256=o1Wt5C74ROikHK-IiAoSPR7Jwq78IIH--D3VS9ZrFgA,1983
1292
+ vllm/reasoning/minimax_m2_reasoning_parser.py,sha256=ytHKX-ZUidvk3QxwLpXCeukD2GXKnh49IjKqajZtZFo,2098
1293
+ vllm/reasoning/mistral_reasoning_parser.py,sha256=5Qikid7ihtaYP6pCSc6Bqw-D5kxLJiaFb1QiovmlniY,2020
1294
+ vllm/reasoning/olmo3_reasoning_parser.py,sha256=SjzElXOcMXM6wr2jdVMzDjGr2MO1Zv7pthMAqXkgKYU,11007
1295
+ vllm/reasoning/qwen3_reasoning_parser.py,sha256=6yJYZVXxa-QL3I2ZHqmFdhZOpmoJh4pWex-6YWegSrU,2487
1296
+ vllm/reasoning/seedoss_reasoning_parser.py,sha256=Oz4zSgr6seYtssSyDKsFi7J-b9kTy-IcuC01HNJkv_c,832
1297
+ vllm/reasoning/step3_reasoning_parser.py,sha256=gl1-8lvpbv2H4rXscf-bVRFZUgJYbX2GbmhOueuWo0U,4023
1298
+ vllm/third_party/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1299
+ vllm/third_party/pynvml.py,sha256=HRQEbE5ZB-AgMIySG4j24hjlEgSGLrNvHJUq7UawCfA,234653
1300
+ vllm/transformers_utils/__init__.py,sha256=BTdT_ym1ftRVhgxBUkdJQqmO2LiKDX2MZZEEG8U_05E,932
1301
+ vllm/transformers_utils/config.py,sha256=C6oXxTAtF0W451ntS68aVsdGgAd1kC3c4fRj1W5ElmM,40964
1302
+ vllm/transformers_utils/config_parser_base.py,sha256=qSyL1AYBvvnQTII-lS9_N602OJ_Huf5BkUZYBhQEFtU,523
1303
+ vllm/transformers_utils/detokenizer_utils.py,sha256=7UA9nCScq2mHFjzJJghpB2vg0gLbnTtfqP4-zS7ogA0,7591
1304
+ vllm/transformers_utils/dynamic_module.py,sha256=K1nKutDfT9VCxbU58gw_o3H49BBM7pu-1jFh2UmIsXU,1781
1305
+ vllm/transformers_utils/processor.py,sha256=jDYf2gxIwVK0x-FWGVPAPXQnwk1PjrXPxUkbQNB5NLo,14255
1306
+ vllm/transformers_utils/runai_utils.py,sha256=xJ0HgkaBGZrOZ5G6RaYBbBgyAI4lhgEdLEN8j0XfrqA,3337
1307
+ vllm/transformers_utils/s3_utils.py,sha256=S8pMrLcoUOXiexsioyzBjuYz4wst_YZC-yqTpC0QCsI,2783
1308
+ vllm/transformers_utils/tokenizer.py,sha256=xsDWDQGgV2NCpC-9atJfp9wb4heoZiqqFJV9NFGDd2g,10049
1309
+ vllm/transformers_utils/tokenizer_base.py,sha256=FdVRLQrtI0eLNxfTVJYKLODHlAvmtTuWN1fU_-maOD0,4000
1310
+ vllm/transformers_utils/utils.py,sha256=wgpz-HriKQ5V2ehKvz0KxggCBReaLMXYJTGJlm1q-nY,3523
1311
+ vllm/transformers_utils/chat_templates/__init__.py,sha256=U1sUyX9swSjxaULlg0B6rSroU5H8upeyInuHsF74SEE,208
1312
+ vllm/transformers_utils/chat_templates/registry.py,sha256=f5ULuuY1mJUrpwFwJXY_QqDskleozjJKMomEhr-jUBc,2563
1313
+ vllm/transformers_utils/chat_templates/template_basic.jinja,sha256=DMH0156UMA7eoJelXKUMEDzB-SigjbyCOBxIu9OyFJE,78
1314
+ vllm/transformers_utils/chat_templates/template_blip2.jinja,sha256=ltMbjFdK7T4HUcN_OQaX4hj2r0PGlS1EJ9zhSlnTz1c,332
1315
+ vllm/transformers_utils/chat_templates/template_chatml.jinja,sha256=CKxCWf_KemM_DntV70Hf03WNkDvxznolyW-03SJJw54,370
1316
+ vllm/transformers_utils/chat_templates/template_deepseek_ocr.jinja,sha256=sN0qqjXqEmWsTBafWtrvXbYnSaeCOTgfDjzmwvk4Njk,499
1317
+ vllm/transformers_utils/chat_templates/template_deepseek_vl2.jinja,sha256=WX32uOZ7h8_xqrWvmsI5R-6Ns8ZcXVn74CKB7FJOifA,785
1318
+ vllm/transformers_utils/chat_templates/template_fuyu.jinja,sha256=hzdsPgeUMaZnd5L23QPiz2oC6_wMBy5WgZkXMVs3Dgo,85
1319
+ vllm/transformers_utils/chat_templates/template_minicpmv45.jinja,sha256=tNvJgf7S0je1GqFxPUQ-q5nSGwXoINdqW3-VOFNekh8,4231
1320
+ vllm/transformers_utils/configs/__init__.py,sha256=nAMevARyiyYCMgWG7uY8eIh7YnFsZEBRKAzO5vxjkGA,2804
1321
+ vllm/transformers_utils/configs/afmoe.py,sha256=Q4tVnYfF3An5FH4XYPZcx0_d6MCBaQKkffkG86KqspM,3022
1322
+ vllm/transformers_utils/configs/arctic.py,sha256=hYIgzKHaE5jGkBC6VPJvB-bX7cGYtotXRa4kpGW1vcw,9036
1323
+ vllm/transformers_utils/configs/chatglm.py,sha256=0Lw_vCbBpbsWOSiLQG6q9LpgeWmgbTAIJJwlHYbMWPk,2731
1324
+ vllm/transformers_utils/configs/deepseek_vl2.py,sha256=0X1rctqdTJmbNu1uFybVeYSpDxAr299Ly0qZjl0SQLU,3881
1325
+ vllm/transformers_utils/configs/dotsocr.py,sha256=_mxuA6xDpO6N2-FD4R5XVqLDP79D8i0OAZcPvqUCSdU,2445
1326
+ vllm/transformers_utils/configs/eagle.py,sha256=zJojMzcPx40UjBLocIKbr54zNl0l8NyAPatOhQvtWaQ,2758
1327
+ vllm/transformers_utils/configs/falcon.py,sha256=XNrg2x08UDZngZqtv5T0CiEFZq7qhoJPWdEYnI64e_Q,2937
1328
+ vllm/transformers_utils/configs/flex_olmo.py,sha256=k40reY_SOJlUjVOj0kZL--r7pUAr7gxXbR4mLwE2g8s,2771
1329
+ vllm/transformers_utils/configs/jais.py,sha256=0mHMFC7GztEyIRf8IJGF6ELhzyVlCqPnJpOEFBUpIRY,10441
1330
+ vllm/transformers_utils/configs/kimi_linear.py,sha256=g130hve0CYnQXmfeTuoKfMB8IxcrWgTnAOi_Y7QFr-8,5051
1331
+ vllm/transformers_utils/configs/kimi_vl.py,sha256=wGV24Na51emzxtMP6EbQ2pereXMCjkIQPo1A2pgcSZs,1362
1332
+ vllm/transformers_utils/configs/lfm2_moe.py,sha256=lS-DKxLmLXfbsmb5SLqguWKkzRS7_t9g6HDp_vc8jqM,7311
1333
+ vllm/transformers_utils/configs/medusa.py,sha256=ECZ7guyPJJob-viGpZh1LZSdi0VPQg1w84pym4C6Q3o,1930
1334
+ vllm/transformers_utils/configs/midashenglm.py,sha256=UZ_nZXvFS511bU5BVNlVYxK4AR-MDWJK5y5LcCM51oM,3628
1335
+ vllm/transformers_utils/configs/mistral.py,sha256=AUVBdefZViZLc79PEYcyjPIY8Pagqv48kHDqVd0MgJE,6224
1336
+ vllm/transformers_utils/configs/mlp_speculator.py,sha256=aTNL7dpJxF6piYE6EysUiB11sBAOFw_xlmlDOFEoEa8,2403
1337
+ vllm/transformers_utils/configs/moonvit.py,sha256=wENIYbaocro7cWlpcc2yrIaWSyzkLpABVTEJ-JQ7Aak,1232
1338
+ vllm/transformers_utils/configs/nemotron.py,sha256=qpJrnD42CpsmsAMm0BsOO_-w-VNWzi_sMvFRw-stCuU,9114
1339
+ vllm/transformers_utils/configs/nemotron_h.py,sha256=eFZ7oDbzNKQI2KmyNyu8QILtVrq8VYBegKFQEXecmDk,13099
1340
+ vllm/transformers_utils/configs/olmo3.py,sha256=RgFgVcGGKstm97xN6iX67SDUUPx2uh3l-bYqLHV2CwU,2700
1341
+ vllm/transformers_utils/configs/ovis.py,sha256=zjfzz8-btPauEhKwQm9dRqmsZ8jakSaVCqqZ2rJZv2g,7535
1342
+ vllm/transformers_utils/configs/qwen3_next.py,sha256=iEZyeB8Vp5N_3WvqXJs0eQWbeO6YuNY5BHQeBmQ7rvQ,14618
1343
+ vllm/transformers_utils/configs/radio.py,sha256=bde0vTUtEfMNZEAefJoyMFC7uDFR35zOPeAiwl9vP2I,3575
1344
+ vllm/transformers_utils/configs/step3_vl.py,sha256=rrE5Ryr8onP2z4SnfYnQMwXWCgUBVQIptVa5_JezmB8,4919
1345
+ vllm/transformers_utils/configs/ultravox.py,sha256=i6GhU7jGY8e5g89o9CB4L758FF4qERwEQ0w_teYmw2U,4848
1346
+ vllm/transformers_utils/configs/speculators/__init__.py,sha256=48Vcuw16i9R99vM3iDVkRkX107yJtFeTtdWgLQE-XFg,107
1347
+ vllm/transformers_utils/configs/speculators/algos.py,sha256=ZUKraC3NROj6rIWXDUqWJBELGFbNI3P0-4Or4W39enQ,1538
1348
+ vllm/transformers_utils/configs/speculators/base.py,sha256=6ltGW5jguo0u1Sznw2I1pXfIAe0ojteBKsm9DhaMOsQ,4382
1349
+ vllm/transformers_utils/processors/__init__.py,sha256=4wL98m-Rn6Kk0z97_EKXY-vbplB-uUDvVJ5i8YNGRmk,637
1350
+ vllm/transformers_utils/processors/deepseek_ocr.py,sha256=qhDq_pr2PZjisj5Mj0SBY6mjX0ivSSGmp3z_V0i8K3U,15088
1351
+ vllm/transformers_utils/processors/deepseek_vl2.py,sha256=YWmA16RtLAOE_eMHTJwHJSFoRFlBU3wD7cR3ZXP9JyQ,15121
1352
+ vllm/transformers_utils/processors/ovis.py,sha256=wzvMI1WEBK408Yb4PxgepUwJ96-LQV2n-lxyUI9mcjM,19426
1353
+ vllm/transformers_utils/processors/ovis2_5.py,sha256=AGj8iHo5IrDmPlRmljIYnh9T2DED9Ior1HFSJWwXsGo,19505
1354
+ vllm/transformers_utils/tokenizers/__init__.py,sha256=v23gaQAc8oJb8DSlzX9Ngu-AHR0kvsDHyKqs1QL_880,378
1355
+ vllm/transformers_utils/tokenizers/mistral.py,sha256=IphhCqDE4kszzoFADJtaeXq0H720BP8fwzkJ25Nkx4o,19285
1356
+ vllm/triton_utils/__init__.py,sha256=iNVjNTtxuTNOarMYSaEySaTuQEiCJX-rlZYKPoIE3vI,517
1357
+ vllm/triton_utils/importing.py,sha256=CcaJ7kpP7bKP3yfuR_1w_vNRBxiZk8qNdcatjsvoAYI,3589
1358
+ vllm/usage/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1359
+ vllm/usage/usage_lib.py,sha256=zNXchtEzMFhPnwC5Iu5Y0dCaYTrLzXfqme2Gig2qK94,9594
1360
+ vllm/utils/__init__.py,sha256=94NTO3qmkISsyvgWRKzfL47WNKE_W2jpURJTkYdp2WY,2764
1361
+ vllm/utils/argparse_utils.py,sha256=e0Prgr0slR0c8_LHbvl_mk-U_OuTML-5-DkmVrhhrDE,18980
1362
+ vllm/utils/async_utils.py,sha256=mFUKypqMLu9HDa3yz0qIIvTTUcGEmkhYuqVm-sbEqms,11061
1363
+ vllm/utils/cache.py,sha256=p6Cvrul9e1fOuw4h9-n74qmY_lHMQyS0BqhHXQqN-_Y,6206
1364
+ vllm/utils/collection_utils.py,sha256=sgBSATvwIGj2eT-uMqilxbKrteFSHNNP51wA_Ie0Eas,3789
1365
+ vllm/utils/counter.py,sha256=6Xa0WjHey2LwyPJkGfYa37mgOsQfcHJF9kRypIcaHDQ,1165
1366
+ vllm/utils/deep_gemm.py,sha256=e0wAu0aEBVLTbpkheTqAXYuKv0ixXp7yrLsrS4gf-i4,13320
1367
+ vllm/utils/flashinfer.py,sha256=3YFLqTPqo_icRDM8UaGjHzYNwYZdOCRukHyhE9UNdJ8,15524
1368
+ vllm/utils/func_utils.py,sha256=pLXKoub-DlGrxZRwliCGmuMy2eiEHrglZPEQSbvPjKg,7706
1369
+ vllm/utils/gc_utils.py,sha256=OErpKbAVLFLRM8yn9ZTLA5lO7IppKdtK7VoEMHUYGZA,4816
1370
+ vllm/utils/hashing.py,sha256=UVfZVZMT7aF6LeIsdlrFPdCkf-IQ7SwD5uOXqA1Fwq0,1908
1371
+ vllm/utils/import_utils.py,sha256=I8gCRdfNOA0Fh_Z6CWiqYmhQBzzm8IhWLKd2TCIzjhY,12687
1372
+ vllm/utils/jsontree.py,sha256=VNxFH00KfcdtZyMGPUvzVWTZ7YXF55Che-gXp5otTu4,3840
1373
+ vllm/utils/math_utils.py,sha256=MskLJpEF6qy1In1mRyRLzmuyDRBlJgtZHfyu2Okq1c4,774
1374
+ vllm/utils/mem_constants.py,sha256=DerNf9-iEKE2kx8c6ufmLY23mn7vV3K1vskiAw3SWJU,390
1375
+ vllm/utils/mem_utils.py,sha256=2ks5KiJh8Fv_44DIZ0MP0nA3sdGNKm4rlGom554jMHE,8834
1376
+ vllm/utils/nccl.py,sha256=IdZhFS4VG6StR6ORf8SwJIaI0ICxYEf17AAfESlCvkA,1927
1377
+ vllm/utils/network_utils.py,sha256=eEN9MnjoOAJMqNWTKhMYQcHexksJrUBIJsCUk8-LU18,10058
1378
+ vllm/utils/platform_utils.py,sha256=V5bnsv9S_IGUJtlyqhIG0Nn-BldsEwrqDAvRTgSp_QY,1873
1379
+ vllm/utils/profiling.py,sha256=UWku3rWFiMP4UclIokbOxF1fh5xVuwiOc4vjsq2Fdo4,1469
1380
+ vllm/utils/registry.py,sha256=81hvRTITtaDAClpSEQiOo8IH_rVnpwu0wblZW0LIg-8,1504
1381
+ vllm/utils/serial_utils.py,sha256=HDxai84PXIsCfL_8GD3BsUNHsFC6zFZdLoPhEtaGUms,4874
1382
+ vllm/utils/system_utils.py,sha256=Ryn0HQUm3l0e1k3UuGa15VSYs289T00f3zngjeDka8I,6798
1383
+ vllm/utils/tensor_schema.py,sha256=Q2GzQ-CzlvQjA9qvZBcP8f3layD454s7c5nRt_FeZWU,9140
1384
+ vllm/utils/torch_utils.py,sha256=KV2zul2DZsuWWBwTaCCBcaBvMZVPOpmAa23PeOPb8Nc,21574
1385
+ vllm/v1/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1386
+ vllm/v1/cudagraph_dispatcher.py,sha256=Gw72XuFhouir0qsnRq7P50FQV9s_A9KCaOjPDmwPhHI,6388
1387
+ vllm/v1/kv_cache_interface.py,sha256=4AtFx9VYBJm23MmoPtKftVuP62voiSYI9QpUroLcAMw,14377
1388
+ vllm/v1/outputs.py,sha256=xOGepdlWzMwtRIAOklfoWwpIRO7I_df_FbmtNEb385w,8388
1389
+ vllm/v1/request.py,sha256=-cLeUSKamIJgJg21NBkWPS0B023Q-hDzXT46hal9tMc,9351
1390
+ vllm/v1/serial_utils.py,sha256=NUgH_AOsd8Tidf-DSJf7FiDpVp6GN_V3MnThwXhyNfI,20445
1391
+ vllm/v1/utils.py,sha256=bN8fkzu6oInYP_vOjUMxjmKiuSZLLXVfml6Cl3XdIBk,13923
1392
+ vllm/v1/attention/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1393
+ vllm/v1/attention/backends/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1394
+ vllm/v1/attention/backends/cpu_attn.py,sha256=OpZk7MQwCJfJ5JVE4d1MpDRq0ys6brXwya_byMKwJlQ,17932
1395
+ vllm/v1/attention/backends/flash_attn.py,sha256=Jk5acYhmCg2JP9DLehfGyXXdqDILzSgBGIi2KhCarY8,40340
1396
+ vllm/v1/attention/backends/flashinfer.py,sha256=A18QAnXgZ-lFcKMIxJN0xrG6JEt_umn0x4e8FXVIcSk,63414
1397
+ vllm/v1/attention/backends/flex_attention.py,sha256=DZA45y61xRZNyg0jqMObbN-Jycrz1d1E3o6nKhXbWPI,35450
1398
+ vllm/v1/attention/backends/gdn_attn.py,sha256=GcXAzEytL7-5lx8oF5p4hcxgqoyklFjcahxzUeg3Cl0,15593
1399
+ vllm/v1/attention/backends/linear_attn.py,sha256=YaiAPm7YM2uvhfPkd6p7UOCgE5cfyEhSMmS7I2vzVhE,2326
1400
+ vllm/v1/attention/backends/mamba1_attn.py,sha256=P_gggtrN3dCOKIpu8tfaPA0OxHMillVnw7-Fq15SrlE,6566
1401
+ vllm/v1/attention/backends/mamba2_attn.py,sha256=I-QqkgHIi16wTrbmRdaOYYbXGjdERvJeMsp_tVP4e4Y,14536
1402
+ vllm/v1/attention/backends/mamba_attn.py,sha256=Z43eFX-wNSVASzAGk8g3MQloYYPXLGsa_KilMUH2W7E,4008
1403
+ vllm/v1/attention/backends/pallas.py,sha256=vPmofZaWK01Ofrrefk-j80q-Zb5Y0iIgVLsnCwWwkwg,15525
1404
+ vllm/v1/attention/backends/rocm_aiter_fa.py,sha256=ZPfP5L0Woor0fAHb_XRvTmBDoVr9N-m7AEeGud8Yxb0,31314
1405
+ vllm/v1/attention/backends/rocm_aiter_unified_attn.py,sha256=DREArjCjHWwI_UP_4hL8rjxxEvdVFMYv0cIBh1cZxT0,6716
1406
+ vllm/v1/attention/backends/rocm_attn.py,sha256=eGKNMm1G4MRHm998JuhpRqLGK6nCxqVPgFL_oCgZjkU,13224
1407
+ vllm/v1/attention/backends/short_conv_attn.py,sha256=4H_h50v2RK1zT6ty_hUw0D2oscm91Y0qeE8R7VYUm0k,3642
1408
+ vllm/v1/attention/backends/tree_attn.py,sha256=PVik7ukvzab34j4eeL9Lm7kumn4mdsvcxSg8KYIc1wU,15654
1409
+ vllm/v1/attention/backends/triton_attn.py,sha256=2iAyYumjs2UZ-E3QHzd8lhYnOZ3tiJv3jE6kZ5c2AKw,13282
1410
+ vllm/v1/attention/backends/utils.py,sha256=wZkchoiCF_VxEAJyJWcH6fUM2-3CsZq17yhsSiO3hHU,41991
1411
+ vllm/v1/attention/backends/xformers.py,sha256=3AP_Da9hTFiAkZkwF7Z382MjzRV0ttWivcTDbeLuZjk,14647
1412
+ vllm/v1/attention/backends/mla/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1413
+ vllm/v1/attention/backends/mla/common.py,sha256=EIRu_L9yWMtDAtDnZWA9y0If36fdCsUzYzonkV1r0IA,79707
1414
+ vllm/v1/attention/backends/mla/cutlass_mla.py,sha256=cwlNCp_QvHkkxmGEUW9ROAHNaVgGOJxKqzrDNVu8dD4,8816
1415
+ vllm/v1/attention/backends/mla/flashattn_mla.py,sha256=nlgR2aWiBSpcxglJcdhat21bXjS9_iZQjIG60sPYSG0,11951
1416
+ vllm/v1/attention/backends/mla/flashinfer_mla.py,sha256=RZ6pud3_GHBG0SzKjH9U6JUNexBXKcRxZPD6CJ165mw,5611
1417
+ vllm/v1/attention/backends/mla/flashmla.py,sha256=yyHnT1aw3mYuIS1vwkFOT-6P-ESMK61EBSeIU86OEp8,11254
1418
+ vllm/v1/attention/backends/mla/flashmla_sparse.py,sha256=PDU1TbtDdAcP0Vp5WsR8lZOUKk6C88R6LgGskLe4Mws,19461
1419
+ vllm/v1/attention/backends/mla/indexer.py,sha256=Nx-KYQgDwyR947CPTYIgtkjLttTriYynpruGycnTgAQ,12925
1420
+ vllm/v1/attention/backends/mla/rocm_aiter_mla.py,sha256=8O3FhVVv1sp5-9oHfjKQTlktPbFQ28MjN3YXERSE10Y,10076
1421
+ vllm/v1/attention/backends/mla/triton_mla.py,sha256=ahsT84ViQyAuFsP38zMfrEkG5mxUtt_Lpz_k-8ZFhk4,5135
1422
+ vllm/v1/core/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1423
+ vllm/v1/core/block_pool.py,sha256=hRu98ZEedyb7Fir70IH7__SNxYJ5RJkkyZMFrhmp6nk,16101
1424
+ vllm/v1/core/encoder_cache_manager.py,sha256=IMT_S9-qVSz_BpWp5Ln98BCdKFgbAxczEb2kvsXY2Bs,13588
1425
+ vllm/v1/core/kv_cache_coordinator.py,sha256=ocMoTpKIGJG6Yd5_2cTRaWqXQOjuE_oAgaWxQJd_1B0,17834
1426
+ vllm/v1/core/kv_cache_manager.py,sha256=Oe9O0JUlHvqzRMhNe4EicSesKfKmhCVd0I3FIzGRQs8,17058
1427
+ vllm/v1/core/kv_cache_utils.py,sha256=ohR5CqCCdu5Oew8X7IvytfvD0cY8Tl8rA0sVL5nyPc8,52504
1428
+ vllm/v1/core/single_type_kv_cache_manager.py,sha256=4NjGppMWbWv6CZ0BgspIbJ8TkLf7OnxIYTXnBEfO-vY,29494
1429
+ vllm/v1/core/sched/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1430
+ vllm/v1/core/sched/async_scheduler.py,sha256=TcNLl6fzesE9tqnyj9QE-VkB2kMpVcUAY2Psz045HDk,2510
1431
+ vllm/v1/core/sched/interface.py,sha256=p6JwVQ-NsthS0BpnZiKyDZzlas3JGXrq4MSyo4UWOaQ,6888
1432
+ vllm/v1/core/sched/output.py,sha256=OgGvZnaAvLjtwhx8czw2G0vnoh6tY1L2NNrw07m6gw4,7524
1433
+ vllm/v1/core/sched/request_queue.py,sha256=k-ATvNUkkLcgOCAtcYtlxVq7TdOLlLWm5N2n-6KwPYs,7437
1434
+ vllm/v1/core/sched/scheduler.py,sha256=gtnLvHHhR7o7DjYj9ykx84Pgeucss6qfwutLTEJzUOg,73996
1435
+ vllm/v1/core/sched/utils.py,sha256=1QRUNDCcMmg1VWbm07iYgkrSETwWGyM2uSn6gpnMneY,2374
1436
+ vllm/v1/engine/__init__.py,sha256=_6hqIqU44-2hwRGLxiGauxgmH-_Y6ymX-MgUnIK1nqc,5937
1437
+ vllm/v1/engine/async_llm.py,sha256=9Q43P6WObVlQwEa9wPN7MiNwZ7h7CWpMztX4GBvOyNM,31228
1438
+ vllm/v1/engine/coordinator.py,sha256=U1oOQQmip60PDIPfSibkkiKtXb-3ZLU1T_ehwQBrpqM,16464
1439
+ vllm/v1/engine/core.py,sha256=eoAIMtfg8P3Q3idFhof3RuGYSdrE-FKkqhWPG60DDww,56686
1440
+ vllm/v1/engine/core_client.py,sha256=SIYe49uxQvkfQdTkarnp2A4wDkq8K2uYtuP2esJyvFo,53495
1441
+ vllm/v1/engine/detokenizer.py,sha256=l8qurzKWZtZlKWS4qEvvvmcHsYZCH1C0EbXuG1YftM4,13170
1442
+ vllm/v1/engine/exceptions.py,sha256=Uqu6bUKEvMYpeVJYxlqYeveRcd9WnsFUENKt0JLMins,732
1443
+ vllm/v1/engine/llm_engine.py,sha256=_yIqeKLu861RFp4bdzZIT2t-TydQndaWe4mDuMurQQE,15569
1444
+ vllm/v1/engine/logprobs.py,sha256=imRkakbwzCOhwZu3vxkC_RQvHsv3vCmKSQmW4bQ_WRY,6099
1445
+ vllm/v1/engine/output_processor.py,sha256=JvneDhV1YvmToxzYD2MTq0VGiMEnWqZWf_8kJ1vyty4,24627
1446
+ vllm/v1/engine/parallel_sampling.py,sha256=cWNTjIH6qXgQ0lzN6Jbe-dGCR5p0iruyaHNd2qitQqA,5313
1447
+ vllm/v1/engine/processor.py,sha256=aBCS1bp4yHKjdErIomO50KRs2zRLgxn5xJzO-AUhDYU,26056
1448
+ vllm/v1/engine/utils.py,sha256=Xq6iYVxsC2Kv3SQaTF4fPVETc4uGDcyKZXMwZM843Vg,41578
1449
+ vllm/v1/executor/__init__.py,sha256=VwU0pRwKjvcCcXII_7N2FIYT2oNj1fquLjvUS-qqE2U,227
1450
+ vllm/v1/executor/abstract.py,sha256=DtnQ83Vk3xlVD9zUaHka42L_xj9D-MuM5s54Sdq2VDs,13247
1451
+ vllm/v1/executor/multiproc_executor.py,sha256=y3af5nUYHhJHIQBeQ0f4yFre2GXB4VP54oRs7yJ0CQY,34416
1452
+ vllm/v1/executor/ray_distributed_executor.py,sha256=tl8FrTK1jqQk2SKF-Rt6RtXacGD6ND-hXIjTrquAbkc,289
1453
+ vllm/v1/executor/ray_executor.py,sha256=Jx658pDLy-P0GqWtBHXv2RJXynr2ujaTqFUqPB5UVB4,25084
1454
+ vllm/v1/executor/ray_utils.py,sha256=MzthlsqlXOeYkeYrkKmLNYC5HcNgezzTgLTjfvZXVEI,19165
1455
+ vllm/v1/executor/uniproc_executor.py,sha256=jBmZS7HLUnjmFwxobfLfgPCw7KkHixZNpGiCL4ZI4hs,7254
1456
+ vllm/v1/kv_offload/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1457
+ vllm/v1/kv_offload/abstract.py,sha256=rEo4jQBZZ3AuYMDWNprePENX8wDFc086h40wFj4gmyg,5262
1458
+ vllm/v1/kv_offload/arc_manager.py,sha256=vUgKdsb11DwRqCcMkzW-lTf2zlMvGN2AaEHGy24G4LU,9545
1459
+ vllm/v1/kv_offload/backend.py,sha256=2ewriYjByJ2UwiWX-YJYhGkriJB0lChRnAbgtN0_8rY,2898
1460
+ vllm/v1/kv_offload/cpu.py,sha256=QHWAveJb4DUxczQD4I1eALTv_tTkQc8wV7SlqXxkuVA,3704
1461
+ vllm/v1/kv_offload/factory.py,sha256=iHoQ7Nz6tsCfg1UrrC0Ik9oI6APUQp_2y-oeGtSZ8Fk,1995
1462
+ vllm/v1/kv_offload/lru_manager.py,sha256=QO85HtD0uV-Et_UkhZdtgXk38-SZqHQtIi0M-qnWULY,4964
1463
+ vllm/v1/kv_offload/mediums.py,sha256=8i5qik1oKldR3Gyf7S5veHM5h1prPZVR4xN05HSOUKQ,880
1464
+ vllm/v1/kv_offload/spec.py,sha256=RjRUWjVpaWe-TSEbT0JHkZKc6CXLaOIcys8kr5Wuawo,1997
1465
+ vllm/v1/kv_offload/backends/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1466
+ vllm/v1/kv_offload/backends/cpu.py,sha256=McAbCUX_6yYAipJ8bd418EOKxT_I4eIbzr0fiqeR46Y,2228
1467
+ vllm/v1/kv_offload/worker/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1468
+ vllm/v1/kv_offload/worker/cpu_gpu.py,sha256=-QQqQ8w2GskgNT2N1rXM41lZxOxdXzVeTJrtPr1VkqE,6824
1469
+ vllm/v1/kv_offload/worker/worker.py,sha256=Dl9fhiT_ARmjZ_e_NLGYtu1jx9uHZ7IaiGbYy8AI6Gs,4662
1470
+ vllm/v1/metrics/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1471
+ vllm/v1/metrics/loggers.py,sha256=Cxfksi4GsetqXKpRbYZ578L4B9s73vbfTOy_M6w_hc0,45920
1472
+ vllm/v1/metrics/prometheus.py,sha256=dhI7DQbT7X3gFlcQOXmoa7PY5JPyw3fvFm0azVEP-ok,2777
1473
+ vllm/v1/metrics/ray_wrappers.py,sha256=BKVIWOw2MRZxMYYbXEBLK2Mqt4EotvyQJVB_SY4Qqmc,5590
1474
+ vllm/v1/metrics/reader.py,sha256=drwtEKJn4gOnaCFvlaYEjibmQx7cQHIsDiQii4B9Uk4,8694
1475
+ vllm/v1/metrics/stats.py,sha256=xH4C8eBqW5-4ki9l03zQ6Kshgg-bpujG0rOXDOBVgbk,14358
1476
+ vllm/v1/pool/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1477
+ vllm/v1/pool/metadata.py,sha256=jKgnOkKfQfYbesiTGOXuNNnEyqwota5SlZ0yW4Wmbb4,2735
1478
+ vllm/v1/sample/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1479
+ vllm/v1/sample/metadata.py,sha256=jLeOuEJuWBJhgcy9TBL2gfjFj2DaJLWk-zPXARUEbsI,1142
1480
+ vllm/v1/sample/rejection_sampler.py,sha256=LhJB3h1sVw0olJJsFWt8l6nAWbAkFrFH5AU9HbnMaZs,29454
1481
+ vllm/v1/sample/sampler.py,sha256=FFaBav-YKXICwJoICbAQOHE2mxImkc-GIcl68bWoL_k,12447
1482
+ vllm/v1/sample/logits_processor/__init__.py,sha256=NJ5q8cgaprlm5kvfpCr1Qly1BpBT-9THiA8qm3_RMGE,12025
1483
+ vllm/v1/sample/logits_processor/builtin.py,sha256=0ByV310FwL442pkn2WCN48eoCjCCxzkD23CWKLLHqKk,10173
1484
+ vllm/v1/sample/logits_processor/interface.py,sha256=p6dGdoKdK94lIUv6RktdOjSHP8adIPNbL4lHxXDfC14,3218
1485
+ vllm/v1/sample/logits_processor/state.py,sha256=Fk_xGySqz7y2MiAi8hL0Q95DNXWelyu5E4XB-pD0d4E,5666
1486
+ vllm/v1/sample/ops/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1487
+ vllm/v1/sample/ops/bad_words.py,sha256=z573nbjiCHu1wCQvnUSFaHS7pvgjnc5xnzDJpzNv3xU,1659
1488
+ vllm/v1/sample/ops/logprobs.py,sha256=krdqPu-zZQChtxbssdvIQtm49l4EckUTljJNVSxZfjk,942
1489
+ vllm/v1/sample/ops/penalties.py,sha256=OOrkMB8o3sFEbYjnCq8fvTbT0EEdL7Zn2U1Euf44CjM,1929
1490
+ vllm/v1/sample/ops/topk_topp_sampler.py,sha256=QhBVYQKbT2mgSNkpt8Q0cvFwyG1fg8bG-QUNxYok7oI,10751
1491
+ vllm/v1/sample/tpu/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1492
+ vllm/v1/sample/tpu/metadata.py,sha256=-bfn1jLwhVOguzItVzMr44WzBKIvbhozwBZOLTKK8gQ,4624
1493
+ vllm/v1/sample/tpu/sampler.py,sha256=_Xe3UmjT308GuyK2u4K3tRSBF4pC3l_vzvsCOLI_-jk,7724
1494
+ vllm/v1/spec_decode/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1495
+ vllm/v1/spec_decode/eagle.py,sha256=WqM9jDfYqKK7kboXN8sgASntw3UD6aY02UT0XiIyOnY,53230
1496
+ vllm/v1/spec_decode/medusa.py,sha256=feMJOjJZaFvLQj_YYEvn_8iOvjfDSPpA8NES4BP1oPA,2510
1497
+ vllm/v1/spec_decode/metadata.py,sha256=pmlkh7PIgjwRn2Ad6BUENBve_v6Vuv5nPB9XCrICjag,2353
1498
+ vllm/v1/spec_decode/metrics.py,sha256=OEmiOz6V2BoHxaT_bXj_mef1_9z0WNfLnUZKJlGxIXA,7856
1499
+ vllm/v1/spec_decode/ngram_proposer.py,sha256=I52ne46-vgDtkqpKcquPRHMsgY3LeeO9xOGRhRlvaGw,11166
1500
+ vllm/v1/spec_decode/suffix_decoding.py,sha256=YyStrQ42YKDfGe50wkm8Qlm8t1p0MFUAMkJq_rI9rCI,4542
1501
+ vllm/v1/spec_decode/utils.py,sha256=rAfBMyGOY1Y3P1Uj595UWhXVxjV6JW3TsKWzPVvG9eI,589
1502
+ vllm/v1/structured_output/__init__.py,sha256=fpxi5LLBbJ1C5Zx-KY9_bN-ALmcHhizzIEIV5GQiuEQ,14146
1503
+ vllm/v1/structured_output/backend_guidance.py,sha256=oMIwUIobGba_daERjvGua059g-ouDPv-2oa8m4fP3ZY,9000
1504
+ vllm/v1/structured_output/backend_lm_format_enforcer.py,sha256=Srnwxs9RmrtdxO2JL1PkwDOAczjrfZsA5eHh-uuAEH4,6307
1505
+ vllm/v1/structured_output/backend_outlines.py,sha256=e4MIkyIQk6oiKL5cdCJQsK7ZmM0QzKWgp31qLM5W8aA,11748
1506
+ vllm/v1/structured_output/backend_types.py,sha256=0dr2J70DYkeZuE0wPf2kDQe89RPD6LFjnlVT7GbxTaM,3824
1507
+ vllm/v1/structured_output/backend_xgrammar.py,sha256=KpPS5L7MpjeRAuqpo8OX2iEsddUQZ2ac-mapG2KFNpw,12669
1508
+ vllm/v1/structured_output/request.py,sha256=PeAz0cG86T-VwR4y1p75TIM9aSKD09Ey7luq0g4h990,3313
1509
+ vllm/v1/structured_output/utils.py,sha256=asVkojQENVyfdRzIAHaJkqfhyVMtA5uNv77VMpc5LyI,16914
1510
+ vllm/v1/worker/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1511
+ vllm/v1/worker/block_table.py,sha256=xqwUEZb3MkElxoiIluWOp2tDj1vr6nidb4awfQZuonM,12488
1512
+ vllm/v1/worker/cpu_model_runner.py,sha256=DryZJ9f7JeiH8JuzCqDuEA1JwhyA-2yTIp1SM0A7BfI,4210
1513
+ vllm/v1/worker/cpu_worker.py,sha256=lotTVOV7KIRPbbWo6jIeZHIwGFWuTFKjZedNouObCzw,8187
1514
+ vllm/v1/worker/dp_utils.py,sha256=YqyWjLM0hxw9V5wjW0Sr23l1SGHlfdOQlgYVk9lXOCY,8402
1515
+ vllm/v1/worker/ec_connector_model_runner_mixin.py,sha256=ngXHt8jcjaBCpa0q4yPh_t4irf5xNTEf8qaChEsIY1s,2966
1516
+ vllm/v1/worker/gpu_input_batch.py,sha256=GReRxD-s8zLv7_D3owobbPa_5X0PbIv2dpHAaIP2wlE,40769
1517
+ vllm/v1/worker/gpu_model_runner.py,sha256=uwCGPLx2iODmXIqfUOup0oBMpLdVl6w0KmIuSeRsIik,223702
1518
+ vllm/v1/worker/gpu_ubatch_wrapper.py,sha256=IOReZAvnIHXvpUtfuFkmwWEB9yLIApX3MjvyUoABrqk,17907
1519
+ vllm/v1/worker/gpu_worker.py,sha256=AkM_lCv1BRt_3q8ulGamH1yNqGanwKl64c164eU35-o,36743
1520
+ vllm/v1/worker/kv_connector_model_runner_mixin.py,sha256=3ZJn6cNspFQI7CvDikabrgUAtd6dUmLAV_JN9TRoEjQ,5391
1521
+ vllm/v1/worker/lora_model_runner_mixin.py,sha256=OxPRmZz4a0EinqP73ZO4BmSBfUzZQn2nMuKNWB358nE,7492
1522
+ vllm/v1/worker/tpu_input_batch.py,sha256=3kO6wZAX-pWgl7KGyg7ErO012TAITOzOmWypzgBlEiM,25033
1523
+ vllm/v1/worker/tpu_model_runner.py,sha256=Jo_yXNY0ILlOrO83IxCSVm2LQLQ0X8ElijUpXfmHaGY,93561
1524
+ vllm/v1/worker/tpu_worker.py,sha256=ckTU91MlDcioAZooc-AvRRpqeLVrI4BhK5hi7VPzYx8,14690
1525
+ vllm/v1/worker/ubatch_utils.py,sha256=1_zwPq9ZjBDEv671mvjmlVxMT7KmG41Zr99J-MYL0gY,2416
1526
+ vllm/v1/worker/ubatching.py,sha256=hr4vZHjqczxIHjtJcSXANDednk9pDOHpM-NUnz4Jd_U,7994
1527
+ vllm/v1/worker/utils.py,sha256=9z1WoLZCw9F0Aa6vIiP9orWdlLe_IG7QoWUWFVjmpKA,13439
1528
+ vllm/v1/worker/worker_base.py,sha256=pcHpi3rHHRL__4cEpOm4qUIbWdh1yduISKgws8kpYeo,14447
1529
+ vllm/v1/worker/xpu_model_runner.py,sha256=rF7RfKGey-XYPzdJcnYlqgLRGsTfFuCaOxZnSRwirqE,1559
1530
+ vllm/v1/worker/xpu_worker.py,sha256=VfkCRbzZJGuKSbmzFLpVa5JZqzBh0J1MS10Cd5OB_NM,7976
1531
+ vllm/vllm_flash_attn/.gitkeep,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
1532
+ vllm_cpu_amxbf16-0.11.2.post2.dist-info/METADATA,sha256=M9GGIlKK3ep8NMjXt5p1lwnvCUhz7j5g-xp0P4KGbuU,16393
1533
+ vllm_cpu_amxbf16-0.11.2.post2.dist-info/WHEEL,sha256=o8YcOpH5mC_n9ScWnBte7z1kkjfSI6p87bunUssSWm8,113
1534
+ vllm_cpu_amxbf16-0.11.2.post2.dist-info/entry_points.txt,sha256=ErfiCUEEMrGDD3jBwf8c54AolBCFv7qrc8Ix9iqzzfs,184
1535
+ vllm_cpu_amxbf16-0.11.2.post2.dist-info/top_level.txt,sha256=fAgb8Pt4zQoKTUA3ZnKEIgcjh0L97_dwEjYDTL5MEEo,5
1536
+ vllm_cpu_amxbf16-0.11.2.post2.dist-info/RECORD,,