sglang 0.5.3rc0__py3-none-any.whl → 0.5.4__py3-none-any.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 (482) hide show
  1. sglang/bench_one_batch.py +54 -37
  2. sglang/bench_one_batch_server.py +340 -34
  3. sglang/bench_serving.py +340 -159
  4. sglang/check_env.py +1 -1
  5. sglang/compile_deep_gemm.py +6 -2
  6. sglang/global_config.py +1 -25
  7. sglang/lang/api.py +6 -0
  8. sglang/lang/backend/runtime_endpoint.py +1 -1
  9. sglang/lang/interpreter.py +1 -0
  10. sglang/lang/ir.py +13 -0
  11. sglang/launch_server.py +9 -2
  12. sglang/profiler.py +20 -3
  13. sglang/srt/_custom_ops.py +1 -1
  14. sglang/srt/batch_invariant_ops/__init__.py +27 -0
  15. sglang/srt/batch_invariant_ops/batch_invariant_ops.py +547 -0
  16. sglang/srt/checkpoint_engine/checkpoint_engine_worker.py +142 -0
  17. sglang/srt/compilation/backend.py +437 -0
  18. sglang/srt/compilation/compilation_config.py +20 -0
  19. sglang/srt/compilation/compilation_counter.py +47 -0
  20. sglang/srt/compilation/compile.py +210 -0
  21. sglang/srt/compilation/compiler_interface.py +503 -0
  22. sglang/srt/compilation/cuda_piecewise_backend.py +228 -0
  23. sglang/srt/compilation/fix_functionalization.py +134 -0
  24. sglang/srt/compilation/fx_utils.py +83 -0
  25. sglang/srt/compilation/inductor_pass.py +140 -0
  26. sglang/srt/compilation/pass_manager.py +66 -0
  27. sglang/srt/compilation/piecewise_context_manager.py +40 -0
  28. sglang/srt/compilation/weak_ref_tensor_jit.py +16 -0
  29. sglang/srt/configs/__init__.py +8 -0
  30. sglang/srt/configs/deepseek_ocr.py +262 -0
  31. sglang/srt/configs/deepseekvl2.py +194 -96
  32. sglang/srt/configs/dots_ocr.py +64 -0
  33. sglang/srt/configs/dots_vlm.py +2 -7
  34. sglang/srt/configs/falcon_h1.py +309 -0
  35. sglang/srt/configs/load_config.py +33 -2
  36. sglang/srt/configs/mamba_utils.py +117 -0
  37. sglang/srt/configs/model_config.py +284 -118
  38. sglang/srt/configs/modelopt_config.py +30 -0
  39. sglang/srt/configs/nemotron_h.py +286 -0
  40. sglang/srt/configs/olmo3.py +105 -0
  41. sglang/srt/configs/points_v15_chat.py +29 -0
  42. sglang/srt/configs/qwen3_next.py +11 -47
  43. sglang/srt/configs/qwen3_omni.py +613 -0
  44. sglang/srt/configs/qwen3_vl.py +576 -0
  45. sglang/srt/connector/remote_instance.py +1 -1
  46. sglang/srt/constrained/base_grammar_backend.py +6 -1
  47. sglang/srt/constrained/llguidance_backend.py +5 -0
  48. sglang/srt/constrained/outlines_backend.py +1 -1
  49. sglang/srt/constrained/outlines_jump_forward.py +1 -1
  50. sglang/srt/constrained/reasoner_grammar_backend.py +9 -6
  51. sglang/srt/constrained/utils.py +12 -0
  52. sglang/srt/constrained/xgrammar_backend.py +26 -15
  53. sglang/srt/debug_utils/dumper.py +10 -3
  54. sglang/srt/disaggregation/ascend/conn.py +2 -2
  55. sglang/srt/disaggregation/ascend/transfer_engine.py +48 -10
  56. sglang/srt/disaggregation/base/conn.py +17 -4
  57. sglang/srt/disaggregation/common/conn.py +268 -98
  58. sglang/srt/disaggregation/decode.py +172 -39
  59. sglang/srt/disaggregation/decode_kvcache_offload_manager.py +185 -0
  60. sglang/srt/disaggregation/decode_schedule_batch_mixin.py +25 -16
  61. sglang/srt/disaggregation/fake/conn.py +11 -3
  62. sglang/srt/disaggregation/mooncake/conn.py +203 -555
  63. sglang/srt/disaggregation/nixl/conn.py +217 -63
  64. sglang/srt/disaggregation/prefill.py +113 -270
  65. sglang/srt/disaggregation/utils.py +36 -5
  66. sglang/srt/distributed/device_communicators/all_reduce_utils.py +16 -0
  67. sglang/srt/distributed/device_communicators/custom_all_reduce.py +6 -6
  68. sglang/srt/distributed/device_communicators/pymscclpp.py +2 -2
  69. sglang/srt/distributed/device_communicators/pynccl.py +24 -12
  70. sglang/srt/distributed/device_communicators/pynccl_allocator.py +2 -2
  71. sglang/srt/distributed/device_communicators/shm_broadcast.py +4 -2
  72. sglang/srt/distributed/device_communicators/symm_mem.py +164 -0
  73. sglang/srt/distributed/naive_distributed.py +5 -4
  74. sglang/srt/distributed/parallel_state.py +203 -97
  75. sglang/srt/elastic_ep/elastic_ep.py +74 -0
  76. sglang/srt/entrypoints/context.py +3 -2
  77. sglang/srt/entrypoints/engine.py +85 -65
  78. sglang/srt/entrypoints/grpc_server.py +632 -305
  79. sglang/srt/entrypoints/harmony_utils.py +2 -2
  80. sglang/srt/entrypoints/http_server.py +169 -17
  81. sglang/srt/entrypoints/http_server_engine.py +1 -7
  82. sglang/srt/entrypoints/openai/protocol.py +327 -34
  83. sglang/srt/entrypoints/openai/serving_base.py +74 -8
  84. sglang/srt/entrypoints/openai/serving_chat.py +202 -118
  85. sglang/srt/entrypoints/openai/serving_classify.py +204 -0
  86. sglang/srt/entrypoints/openai/serving_completions.py +20 -4
  87. sglang/srt/entrypoints/openai/serving_embedding.py +1 -0
  88. sglang/srt/entrypoints/openai/serving_responses.py +47 -2
  89. sglang/srt/entrypoints/openai/serving_tokenize.py +144 -0
  90. sglang/srt/environ.py +323 -0
  91. sglang/srt/eplb/eplb_algorithms/__init__.py +18 -1
  92. sglang/srt/eplb/eplb_algorithms/deepseek.py +0 -2
  93. sglang/srt/eplb/eplb_algorithms/elasticity_aware.py +87 -0
  94. sglang/srt/eplb/expert_distribution.py +3 -4
  95. sglang/srt/eplb/expert_location.py +30 -5
  96. sglang/srt/eplb/expert_location_dispatch.py +2 -2
  97. sglang/srt/eplb/expert_location_updater.py +2 -2
  98. sglang/srt/function_call/base_format_detector.py +17 -18
  99. sglang/srt/function_call/function_call_parser.py +21 -16
  100. sglang/srt/function_call/glm4_moe_detector.py +4 -8
  101. sglang/srt/function_call/gpt_oss_detector.py +24 -1
  102. sglang/srt/function_call/json_array_parser.py +61 -0
  103. sglang/srt/function_call/kimik2_detector.py +17 -4
  104. sglang/srt/function_call/utils.py +98 -7
  105. sglang/srt/grpc/compile_proto.py +245 -0
  106. sglang/srt/grpc/grpc_request_manager.py +915 -0
  107. sglang/srt/grpc/health_servicer.py +189 -0
  108. sglang/srt/grpc/scheduler_launcher.py +181 -0
  109. sglang/srt/grpc/sglang_scheduler_pb2.py +81 -68
  110. sglang/srt/grpc/sglang_scheduler_pb2.pyi +124 -61
  111. sglang/srt/grpc/sglang_scheduler_pb2_grpc.py +92 -1
  112. sglang/srt/layers/activation.py +11 -7
  113. sglang/srt/layers/attention/aiter_backend.py +17 -18
  114. sglang/srt/layers/attention/ascend_backend.py +125 -10
  115. sglang/srt/layers/attention/attention_registry.py +226 -0
  116. sglang/srt/layers/attention/base_attn_backend.py +32 -4
  117. sglang/srt/layers/attention/cutlass_mla_backend.py +3 -3
  118. sglang/srt/layers/attention/double_sparsity_backend.py +2 -2
  119. sglang/srt/layers/attention/dual_chunk_flashattention_backend.py +1 -1
  120. sglang/srt/layers/attention/fla/chunk.py +0 -1
  121. sglang/srt/layers/attention/fla/chunk_o.py +1 -1
  122. sglang/srt/layers/attention/fla/chunk_scaled_dot_kkt.py +2 -2
  123. sglang/srt/layers/attention/fla/fused_recurrent.py +4 -4
  124. sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py +2 -2
  125. sglang/srt/layers/attention/fla/index.py +0 -2
  126. sglang/srt/layers/attention/fla/layernorm_gated.py +50 -32
  127. sglang/srt/layers/attention/fla/utils.py +0 -3
  128. sglang/srt/layers/attention/fla/wy_fast.py +0 -2
  129. sglang/srt/layers/attention/flashattention_backend.py +52 -15
  130. sglang/srt/layers/attention/flashinfer_backend.py +357 -212
  131. sglang/srt/layers/attention/flashinfer_mla_backend.py +31 -33
  132. sglang/srt/layers/attention/flashmla_backend.py +9 -7
  133. sglang/srt/layers/attention/hybrid_attn_backend.py +12 -4
  134. sglang/srt/layers/attention/hybrid_linear_attn_backend.py +236 -133
  135. sglang/srt/layers/attention/intel_amx_backend.py +1 -1
  136. sglang/srt/layers/attention/mamba/causal_conv1d.py +2 -1
  137. sglang/srt/layers/attention/mamba/causal_conv1d_triton.py +24 -103
  138. sglang/srt/layers/attention/mamba/mamba.py +514 -1
  139. sglang/srt/layers/attention/mamba/mamba2_metadata.py +211 -0
  140. sglang/srt/layers/attention/mamba/mixer2_rms_norm_gated.py +120 -0
  141. sglang/srt/layers/attention/mamba/ops/__init__.py +2 -0
  142. sglang/srt/layers/attention/mamba/ops/layernorm_gated.py +172 -0
  143. sglang/srt/layers/attention/mamba/ops/mamba_ssm.py +442 -0
  144. sglang/srt/layers/attention/mamba/ops/ssd_bmm.py +214 -0
  145. sglang/srt/layers/attention/mamba/ops/ssd_chunk_scan.py +562 -0
  146. sglang/srt/layers/attention/mamba/ops/ssd_chunk_state.py +646 -0
  147. sglang/srt/layers/attention/mamba/ops/ssd_combined.py +261 -0
  148. sglang/srt/layers/attention/mamba/ops/ssd_state_passing.py +264 -0
  149. sglang/srt/layers/attention/npu_ops/mla_preprocess.py +393 -0
  150. sglang/srt/layers/attention/nsa/dequant_k_cache.py +163 -0
  151. sglang/srt/layers/attention/nsa/index_buf_accessor.py +354 -0
  152. sglang/srt/layers/attention/nsa/nsa_indexer.py +718 -0
  153. sglang/srt/layers/attention/nsa/quant_k_cache.py +255 -0
  154. sglang/srt/layers/attention/nsa/tilelang_kernel.py +785 -0
  155. sglang/srt/layers/attention/nsa/transform_index.py +144 -0
  156. sglang/srt/layers/attention/nsa/triton_kernel.py +136 -0
  157. sglang/srt/layers/attention/nsa/utils.py +23 -0
  158. sglang/srt/layers/attention/nsa_backend.py +1201 -0
  159. sglang/srt/layers/attention/tbo_backend.py +6 -6
  160. sglang/srt/layers/attention/torch_flex_backend.py +325 -0
  161. sglang/srt/layers/attention/triton_backend.py +249 -42
  162. sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py +2 -2
  163. sglang/srt/layers/attention/triton_ops/extend_attention.py +539 -44
  164. sglang/srt/layers/attention/trtllm_mha_backend.py +7 -9
  165. sglang/srt/layers/attention/trtllm_mla_backend.py +523 -48
  166. sglang/srt/layers/attention/utils.py +11 -7
  167. sglang/srt/layers/attention/vision.py +61 -3
  168. sglang/srt/layers/attention/wave_backend.py +4 -4
  169. sglang/srt/layers/attention/xpu_backend.py +1028 -0
  170. sglang/srt/layers/communicator.py +19 -7
  171. sglang/srt/layers/{quantization/deep_gemm_wrapper → deep_gemm_wrapper}/compile_utils.py +4 -8
  172. sglang/srt/layers/deep_gemm_wrapper/configurer.py +25 -0
  173. sglang/srt/layers/{quantization/deep_gemm_wrapper → deep_gemm_wrapper}/entrypoint.py +3 -3
  174. sglang/srt/layers/dp_attention.py +28 -1
  175. sglang/srt/layers/elementwise.py +3 -1
  176. sglang/srt/layers/layernorm.py +47 -15
  177. sglang/srt/layers/linear.py +30 -5
  178. sglang/srt/layers/logits_processor.py +161 -18
  179. sglang/srt/layers/modelopt_utils.py +11 -0
  180. sglang/srt/layers/moe/cutlass_moe.py +0 -2
  181. sglang/srt/layers/moe/cutlass_w4a8_moe.py +213 -21
  182. sglang/srt/layers/moe/ep_moe/kernels.py +36 -458
  183. sglang/srt/layers/moe/ep_moe/layer.py +243 -448
  184. sglang/srt/layers/moe/flashinfer_cutedsl_moe.py +52 -25
  185. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  186. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=256,device_name=NVIDIA_B200.json +146 -0
  187. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=256,device_name=NVIDIA_H800,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  188. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H800,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  189. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=256,device_name=NVIDIA_B200.json +146 -0
  190. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_config.py +17 -5
  191. sglang/srt/layers/moe/fused_moe_triton/layer.py +86 -81
  192. sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py +18 -42
  193. sglang/srt/layers/moe/moe_runner/deep_gemm.py +304 -0
  194. sglang/srt/layers/moe/moe_runner/runner.py +3 -0
  195. sglang/srt/layers/moe/moe_runner/triton.py +3 -1
  196. sglang/srt/layers/moe/rocm_moe_utils.py +0 -1
  197. sglang/srt/layers/moe/router.py +51 -15
  198. sglang/srt/layers/moe/token_dispatcher/__init__.py +10 -0
  199. sglang/srt/layers/moe/token_dispatcher/base.py +1 -1
  200. sglang/srt/layers/moe/token_dispatcher/deepep.py +177 -106
  201. sglang/srt/layers/moe/token_dispatcher/mooncake.py +386 -0
  202. sglang/srt/layers/moe/token_dispatcher/standard.py +46 -0
  203. sglang/srt/layers/moe/topk.py +3 -2
  204. sglang/srt/layers/moe/utils.py +27 -1
  205. sglang/srt/layers/parameter.py +23 -6
  206. sglang/srt/layers/quantization/__init__.py +2 -53
  207. sglang/srt/layers/quantization/awq.py +183 -6
  208. sglang/srt/layers/quantization/awq_triton.py +29 -0
  209. sglang/srt/layers/quantization/base_config.py +20 -1
  210. sglang/srt/layers/quantization/compressed_tensors/__init__.py +7 -0
  211. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py +21 -49
  212. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +421 -70
  213. sglang/srt/layers/quantization/compressed_tensors/schemes/__init__.py +5 -0
  214. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a16_fp8.py +4 -22
  215. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_int8.py +173 -0
  216. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16.py +339 -0
  217. sglang/srt/layers/quantization/fp8.py +86 -20
  218. sglang/srt/layers/quantization/fp8_kernel.py +55 -10
  219. sglang/srt/layers/quantization/fp8_utils.py +43 -15
  220. sglang/srt/layers/quantization/fpgemm_fp8.py +2 -3
  221. sglang/srt/layers/quantization/gptq.py +0 -1
  222. sglang/srt/layers/quantization/int8_kernel.py +18 -2
  223. sglang/srt/layers/quantization/marlin_utils.py +12 -0
  224. sglang/srt/layers/quantization/modelopt_quant.py +141 -81
  225. sglang/srt/layers/quantization/mxfp4.py +17 -34
  226. sglang/srt/layers/quantization/petit.py +1 -1
  227. sglang/srt/layers/quantization/quark/quark.py +3 -1
  228. sglang/srt/layers/quantization/quark/quark_moe.py +18 -5
  229. sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py +0 -7
  230. sglang/srt/layers/quantization/unquant.py +1 -4
  231. sglang/srt/layers/quantization/utils.py +0 -1
  232. sglang/srt/layers/quantization/w4afp8.py +51 -24
  233. sglang/srt/layers/quantization/w8a8_int8.py +45 -27
  234. sglang/srt/layers/radix_attention.py +59 -9
  235. sglang/srt/layers/rotary_embedding.py +750 -46
  236. sglang/srt/layers/sampler.py +84 -16
  237. sglang/srt/layers/sparse_pooler.py +98 -0
  238. sglang/srt/layers/utils.py +23 -1
  239. sglang/srt/layers/vocab_parallel_embedding.py +4 -1
  240. sglang/srt/lora/backend/base_backend.py +3 -3
  241. sglang/srt/lora/backend/chunked_backend.py +348 -0
  242. sglang/srt/lora/backend/triton_backend.py +9 -4
  243. sglang/srt/lora/eviction_policy.py +139 -0
  244. sglang/srt/lora/lora.py +7 -5
  245. sglang/srt/lora/lora_manager.py +33 -7
  246. sglang/srt/lora/lora_registry.py +1 -1
  247. sglang/srt/lora/mem_pool.py +41 -17
  248. sglang/srt/lora/triton_ops/__init__.py +4 -0
  249. sglang/srt/lora/triton_ops/chunked_sgmv_expand.py +214 -0
  250. sglang/srt/lora/triton_ops/chunked_sgmv_shrink.py +176 -0
  251. sglang/srt/lora/utils.py +7 -5
  252. sglang/srt/managers/cache_controller.py +83 -152
  253. sglang/srt/managers/data_parallel_controller.py +156 -87
  254. sglang/srt/managers/detokenizer_manager.py +51 -24
  255. sglang/srt/managers/io_struct.py +223 -129
  256. sglang/srt/managers/mm_utils.py +49 -10
  257. sglang/srt/managers/multi_tokenizer_mixin.py +83 -98
  258. sglang/srt/managers/multimodal_processor.py +1 -2
  259. sglang/srt/managers/overlap_utils.py +130 -0
  260. sglang/srt/managers/schedule_batch.py +340 -529
  261. sglang/srt/managers/schedule_policy.py +158 -18
  262. sglang/srt/managers/scheduler.py +665 -620
  263. sglang/srt/managers/scheduler_input_blocker.py +1 -1
  264. sglang/srt/managers/scheduler_metrics_mixin.py +150 -131
  265. sglang/srt/managers/scheduler_output_processor_mixin.py +337 -122
  266. sglang/srt/managers/scheduler_pp_mixin.py +341 -0
  267. sglang/srt/managers/scheduler_profiler_mixin.py +62 -15
  268. sglang/srt/managers/scheduler_runtime_checker_mixin.py +217 -0
  269. sglang/srt/managers/scheduler_update_weights_mixin.py +40 -14
  270. sglang/srt/managers/tokenizer_communicator_mixin.py +141 -19
  271. sglang/srt/managers/tokenizer_manager.py +462 -226
  272. sglang/srt/managers/tp_worker.py +217 -156
  273. sglang/srt/managers/utils.py +79 -47
  274. sglang/srt/mem_cache/allocator.py +21 -22
  275. sglang/srt/mem_cache/allocator_ascend.py +42 -28
  276. sglang/srt/mem_cache/base_prefix_cache.py +3 -3
  277. sglang/srt/mem_cache/chunk_cache.py +20 -2
  278. sglang/srt/mem_cache/common.py +480 -0
  279. sglang/srt/mem_cache/evict_policy.py +38 -0
  280. sglang/srt/mem_cache/hicache_storage.py +44 -2
  281. sglang/srt/mem_cache/hiradix_cache.py +134 -34
  282. sglang/srt/mem_cache/mamba_radix_cache.py +993 -0
  283. sglang/srt/mem_cache/memory_pool.py +602 -208
  284. sglang/srt/mem_cache/memory_pool_host.py +134 -183
  285. sglang/srt/mem_cache/multimodal_cache.py +0 -1
  286. sglang/srt/mem_cache/radix_cache.py +263 -78
  287. sglang/srt/mem_cache/radix_cache_cpp.py +29 -21
  288. sglang/srt/mem_cache/storage/__init__.py +10 -0
  289. sglang/srt/mem_cache/storage/aibrix_kvcache/aibrix_kvcache_storage.py +157 -0
  290. sglang/srt/mem_cache/storage/aibrix_kvcache/unit_test.py +97 -0
  291. sglang/srt/mem_cache/storage/backend_factory.py +223 -0
  292. sglang/srt/mem_cache/storage/eic/eic_storage.py +777 -0
  293. sglang/srt/mem_cache/storage/eic/test_unit.py +115 -0
  294. sglang/srt/mem_cache/storage/hf3fs/hf3fs_client.py +0 -1
  295. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +180 -59
  296. sglang/srt/mem_cache/storage/lmcache/lmc_radix_cache.py +15 -9
  297. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +217 -26
  298. sglang/srt/mem_cache/storage/nixl/hicache_nixl.py +38 -9
  299. sglang/srt/mem_cache/storage/nixl/nixl_utils.py +1 -1
  300. sglang/srt/mem_cache/storage/nixl/test_hicache_nixl_storage.py +17 -2
  301. sglang/srt/mem_cache/swa_radix_cache.py +115 -58
  302. sglang/srt/metrics/collector.py +113 -120
  303. sglang/srt/metrics/func_timer.py +3 -8
  304. sglang/srt/metrics/utils.py +8 -1
  305. sglang/srt/model_executor/cpu_graph_runner.py +2 -2
  306. sglang/srt/model_executor/cuda_graph_runner.py +81 -36
  307. sglang/srt/model_executor/forward_batch_info.py +40 -50
  308. sglang/srt/model_executor/model_runner.py +507 -319
  309. sglang/srt/model_executor/npu_graph_runner.py +11 -5
  310. sglang/srt/model_executor/piecewise_cuda_graph_runner.py +539 -0
  311. sglang/srt/model_loader/__init__.py +1 -1
  312. sglang/srt/model_loader/loader.py +438 -37
  313. sglang/srt/model_loader/utils.py +0 -1
  314. sglang/srt/model_loader/weight_utils.py +200 -27
  315. sglang/srt/models/apertus.py +2 -3
  316. sglang/srt/models/arcee.py +2 -2
  317. sglang/srt/models/bailing_moe.py +40 -56
  318. sglang/srt/models/bailing_moe_nextn.py +3 -4
  319. sglang/srt/models/bert.py +1 -1
  320. sglang/srt/models/deepseek_nextn.py +25 -4
  321. sglang/srt/models/deepseek_ocr.py +1516 -0
  322. sglang/srt/models/deepseek_v2.py +793 -235
  323. sglang/srt/models/dots_ocr.py +171 -0
  324. sglang/srt/models/dots_vlm.py +0 -1
  325. sglang/srt/models/dots_vlm_vit.py +1 -1
  326. sglang/srt/models/falcon_h1.py +570 -0
  327. sglang/srt/models/gemma3_causal.py +0 -2
  328. sglang/srt/models/gemma3_mm.py +17 -1
  329. sglang/srt/models/gemma3n_mm.py +2 -3
  330. sglang/srt/models/glm4_moe.py +17 -40
  331. sglang/srt/models/glm4_moe_nextn.py +4 -4
  332. sglang/srt/models/glm4v.py +3 -2
  333. sglang/srt/models/glm4v_moe.py +6 -6
  334. sglang/srt/models/gpt_oss.py +12 -35
  335. sglang/srt/models/grok.py +10 -23
  336. sglang/srt/models/hunyuan.py +2 -7
  337. sglang/srt/models/interns1.py +0 -1
  338. sglang/srt/models/kimi_vl.py +1 -7
  339. sglang/srt/models/kimi_vl_moonvit.py +4 -2
  340. sglang/srt/models/llama.py +6 -2
  341. sglang/srt/models/llama_eagle3.py +1 -1
  342. sglang/srt/models/longcat_flash.py +6 -23
  343. sglang/srt/models/longcat_flash_nextn.py +4 -15
  344. sglang/srt/models/mimo.py +2 -13
  345. sglang/srt/models/mimo_mtp.py +1 -2
  346. sglang/srt/models/minicpmo.py +7 -5
  347. sglang/srt/models/mixtral.py +1 -4
  348. sglang/srt/models/mllama.py +1 -1
  349. sglang/srt/models/mllama4.py +27 -6
  350. sglang/srt/models/nemotron_h.py +511 -0
  351. sglang/srt/models/olmo2.py +31 -4
  352. sglang/srt/models/opt.py +5 -5
  353. sglang/srt/models/phi.py +1 -1
  354. sglang/srt/models/phi4mm.py +1 -1
  355. sglang/srt/models/phimoe.py +0 -1
  356. sglang/srt/models/pixtral.py +0 -3
  357. sglang/srt/models/points_v15_chat.py +186 -0
  358. sglang/srt/models/qwen.py +0 -1
  359. sglang/srt/models/qwen2.py +0 -7
  360. sglang/srt/models/qwen2_5_vl.py +5 -5
  361. sglang/srt/models/qwen2_audio.py +2 -15
  362. sglang/srt/models/qwen2_moe.py +70 -4
  363. sglang/srt/models/qwen2_vl.py +6 -3
  364. sglang/srt/models/qwen3.py +18 -3
  365. sglang/srt/models/qwen3_moe.py +50 -38
  366. sglang/srt/models/qwen3_next.py +43 -21
  367. sglang/srt/models/qwen3_next_mtp.py +3 -4
  368. sglang/srt/models/qwen3_omni_moe.py +661 -0
  369. sglang/srt/models/qwen3_vl.py +791 -0
  370. sglang/srt/models/qwen3_vl_moe.py +343 -0
  371. sglang/srt/models/registry.py +15 -3
  372. sglang/srt/models/roberta.py +55 -3
  373. sglang/srt/models/sarashina2_vision.py +268 -0
  374. sglang/srt/models/solar.py +505 -0
  375. sglang/srt/models/starcoder2.py +357 -0
  376. sglang/srt/models/step3_vl.py +3 -5
  377. sglang/srt/models/torch_native_llama.py +9 -2
  378. sglang/srt/models/utils.py +61 -0
  379. sglang/srt/multimodal/processors/base_processor.py +21 -9
  380. sglang/srt/multimodal/processors/deepseek_ocr.py +37 -0
  381. sglang/srt/multimodal/processors/deepseek_vl_v2.py +0 -3
  382. sglang/srt/multimodal/processors/dots_vlm.py +2 -4
  383. sglang/srt/multimodal/processors/glm4v.py +1 -5
  384. sglang/srt/multimodal/processors/internvl.py +20 -10
  385. sglang/srt/multimodal/processors/janus_pro.py +0 -1
  386. sglang/srt/multimodal/processors/mllama4.py +0 -8
  387. sglang/srt/multimodal/processors/phi4mm.py +0 -1
  388. sglang/srt/multimodal/processors/points_v15_chat.py +52 -0
  389. sglang/srt/multimodal/processors/qwen_vl.py +83 -17
  390. sglang/srt/multimodal/processors/sarashina2_vision.py +81 -0
  391. sglang/srt/multimodal/processors/step3_vl.py +1 -1
  392. sglang/srt/parser/conversation.py +41 -0
  393. sglang/srt/parser/jinja_template_utils.py +6 -0
  394. sglang/srt/parser/reasoning_parser.py +0 -1
  395. sglang/srt/sampling/custom_logit_processor.py +77 -2
  396. sglang/srt/sampling/sampling_batch_info.py +36 -23
  397. sglang/srt/sampling/sampling_params.py +75 -0
  398. sglang/srt/server_args.py +1300 -338
  399. sglang/srt/server_args_config_parser.py +146 -0
  400. sglang/srt/single_batch_overlap.py +161 -0
  401. sglang/srt/speculative/base_spec_worker.py +34 -0
  402. sglang/srt/speculative/cpp_ngram/ngram.cpp +374 -0
  403. sglang/srt/speculative/cpp_ngram/ngram.h +110 -0
  404. sglang/srt/speculative/cpp_ngram/ngram_cache.py +138 -0
  405. sglang/srt/speculative/cpp_ngram/ngram_cache_binding.cpp +43 -0
  406. sglang/srt/speculative/cpp_ngram/param.h +125 -0
  407. sglang/srt/speculative/cpp_ngram/queue.h +71 -0
  408. sglang/srt/speculative/draft_utils.py +226 -0
  409. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +26 -8
  410. sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +26 -3
  411. sglang/srt/speculative/eagle_info.py +786 -0
  412. sglang/srt/speculative/eagle_info_v2.py +458 -0
  413. sglang/srt/speculative/eagle_utils.py +113 -1270
  414. sglang/srt/speculative/eagle_worker.py +120 -285
  415. sglang/srt/speculative/eagle_worker_v2.py +702 -0
  416. sglang/srt/speculative/ngram_info.py +433 -0
  417. sglang/srt/speculative/ngram_worker.py +246 -0
  418. sglang/srt/speculative/spec_info.py +49 -0
  419. sglang/srt/speculative/spec_utils.py +641 -0
  420. sglang/srt/speculative/standalone_worker.py +4 -14
  421. sglang/srt/tokenizer/tiktoken_tokenizer.py +2 -2
  422. sglang/srt/tracing/trace.py +32 -6
  423. sglang/srt/two_batch_overlap.py +35 -18
  424. sglang/srt/utils/__init__.py +2 -0
  425. sglang/srt/{bench_utils.py → utils/bench_utils.py} +4 -2
  426. sglang/srt/{utils.py → utils/common.py} +583 -113
  427. sglang/srt/{hf_transformers_utils.py → utils/hf_transformers_utils.py} +86 -19
  428. sglang/srt/{host_shared_memory.py → utils/host_shared_memory.py} +0 -1
  429. sglang/srt/{offloader.py → utils/offloader.py} +4 -4
  430. sglang/srt/{patch_torch.py → utils/patch_torch.py} +8 -0
  431. sglang/srt/utils/profile_merger.py +199 -0
  432. sglang/srt/utils/rpd_utils.py +452 -0
  433. sglang/srt/utils/slow_rank_detector.py +71 -0
  434. sglang/srt/{torch_memory_saver_adapter.py → utils/torch_memory_saver_adapter.py} +5 -7
  435. sglang/srt/warmup.py +8 -4
  436. sglang/srt/weight_sync/utils.py +1 -1
  437. sglang/test/attention/test_flashattn_backend.py +1 -1
  438. sglang/test/attention/test_flashattn_mla_backend.py +0 -1
  439. sglang/test/attention/test_prefix_chunk_info.py +0 -2
  440. sglang/test/attention/test_trtllm_mla_backend.py +221 -53
  441. sglang/test/few_shot_gsm8k_engine.py +2 -4
  442. sglang/test/get_logits_ut.py +57 -0
  443. sglang/test/kit_matched_stop.py +157 -0
  444. sglang/test/longbench_v2/__init__.py +1 -0
  445. sglang/test/longbench_v2/test_longbench_v2_eval.py +238 -0
  446. sglang/test/longbench_v2/validate_longbench_v2.py +337 -0
  447. sglang/test/longbench_v2/validate_longbench_v2_standalone.py +306 -0
  448. sglang/test/run_eval.py +120 -11
  449. sglang/test/runners.py +3 -1
  450. sglang/test/send_one.py +42 -7
  451. sglang/test/simple_eval_common.py +8 -2
  452. sglang/test/simple_eval_gpqa.py +0 -1
  453. sglang/test/simple_eval_humaneval.py +0 -3
  454. sglang/test/simple_eval_longbench_v2.py +344 -0
  455. sglang/test/simple_eval_mmmu_vlm.py +441 -0
  456. sglang/test/test_block_fp8.py +3 -4
  457. sglang/test/test_block_fp8_deep_gemm_blackwell.py +0 -1
  458. sglang/test/test_cutlass_moe.py +1 -2
  459. sglang/test/test_cutlass_w4a8_moe.py +10 -20
  460. sglang/test/test_deterministic.py +430 -0
  461. sglang/test/test_deterministic_utils.py +73 -0
  462. sglang/test/test_disaggregation_utils.py +93 -1
  463. sglang/test/test_marlin_moe.py +0 -1
  464. sglang/test/test_programs.py +1 -1
  465. sglang/test/test_utils.py +432 -16
  466. sglang/utils.py +10 -1
  467. sglang/version.py +1 -1
  468. {sglang-0.5.3rc0.dist-info → sglang-0.5.4.dist-info}/METADATA +64 -43
  469. {sglang-0.5.3rc0.dist-info → sglang-0.5.4.dist-info}/RECORD +476 -346
  470. sglang/srt/entrypoints/grpc_request_manager.py +0 -580
  471. sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py +0 -32
  472. sglang/srt/managers/tp_worker_overlap_thread.py +0 -319
  473. sglang/srt/mem_cache/lora_radix_cache.py +0 -421
  474. sglang/srt/speculative/build_eagle_tree.py +0 -427
  475. sglang/test/test_block_fp8_ep.py +0 -358
  476. /sglang/srt/layers/{quantization/deep_gemm_wrapper → deep_gemm_wrapper}/__init__.py +0 -0
  477. /sglang/srt/{remote_instance_weight_loader_utils.py → model_loader/remote_instance_weight_loader_utils.py} +0 -0
  478. /sglang/srt/{aio_rwlock.py → utils/aio_rwlock.py} +0 -0
  479. /sglang/srt/{poll_based_barrier.py → utils/poll_based_barrier.py} +0 -0
  480. {sglang-0.5.3rc0.dist-info → sglang-0.5.4.dist-info}/WHEEL +0 -0
  481. {sglang-0.5.3rc0.dist-info → sglang-0.5.4.dist-info}/licenses/LICENSE +0 -0
  482. {sglang-0.5.3rc0.dist-info → sglang-0.5.4.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/flashinfer_cutedsl_moe.py CHANGED
@@ -1,4 +1,4 @@
1
- from typing import Any, Dict, Optional
1
+ from typing import Optional, Union
2
2
 
3
3
  import torch
4
4
  from flashinfer.cute_dsl.blockscaled_gemm import grouped_gemm_nt_masked
@@ -20,7 +20,7 @@ def get_cute_dtype(input: torch.Tensor) -> str:
20
20
 
21
21
 
22
22
  def flashinfer_cutedsl_moe_masked(
23
- hidden_states: torch.Tensor,
23
+ hidden_states: Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]],
24
24
  input_global_scale: torch.Tensor,
25
25
  w1: torch.Tensor,
26
26
  w1_blockscale: torch.Tensor,
@@ -30,13 +30,18 @@ def flashinfer_cutedsl_moe_masked(
30
30
  w2_blockscale: torch.Tensor,
31
31
  w2_alpha,
32
32
  masked_m: torch.Tensor,
33
+ down_sm_count: Optional[int] = None,
34
+ down_signals: Optional[torch.Tensor] = None,
35
+ down_start_event: Optional[torch.cuda.Event] = None,
33
36
  ):
34
37
  """
35
38
  Perform masked Mixture-of-Experts computation with FlashInfer's CuteDSL
36
39
  kernels.
37
40
 
38
41
  Args:
39
- hidden_states (torch.Tensor): [num_experts, m, k], bf16
42
+ hidden_states: Either of the following case
43
+ * torch.Tensor: [num_experts, m, k], bf16
44
+ * tuple[torch.Tensor, torch.Tensor]: [num_experts, m, k // 2], uint8, [num_experts, m, k // 16], float8_e4m3fn
40
45
  input_global_scale (torch.Tensor): (l,)
41
46
  w1 (torch.Tensor): fp4 weights, [l, 2 * n, k // 2], uint8
42
47
  w1_blockscale (torch.Tensor): blockscale factors, e4m3,
@@ -48,13 +53,10 @@ def flashinfer_cutedsl_moe_masked(
48
53
  masked_m (torch.Tensor): Masked dimension indices
49
54
 
50
55
  Notes:
51
- - Assumes max(masked_m) <= m.
56
+ - Assumes max(masked_m) == m.
52
57
  """
53
58
 
54
59
  # === Assertions on dtypes ===
55
- assert (
56
- input_global_scale.dtype == torch.float32
57
- ), f"input_global_scale must be float32, got {input_global_scale.dtype}"
58
60
  assert w1.dtype == torch.uint8, f"w1 must be uint8 (fp4 packed), got {w1.dtype}"
59
61
  assert (
60
62
  w1_blockscale.dtype == torch.float8_e4m3fn
@@ -75,7 +77,31 @@ def flashinfer_cutedsl_moe_masked(
75
77
 
76
78
  # === Assertions on shapes ===
77
79
  n = w2.shape[-1] * 2 # intermediate dimension
78
- num_experts, m, k = hidden_states.shape
80
+
81
+ if isinstance(hidden_states, tuple):
82
+ assert (
83
+ input_global_scale is None
84
+ ), "input_global_scale is needed when input needs quant"
85
+
86
+ a_q = hidden_states[0].view(torch.uint8)
87
+ a_q_sf = hidden_states[1].view(torch.float8_e4m3fn)
88
+ m, k_by_2, num_experts = a_q.shape
89
+ k = k_by_2 * 2
90
+ else:
91
+ num_experts, m, k = hidden_states.shape
92
+
93
+ assert (
94
+ input_global_scale.dtype == torch.float32
95
+ ), f"input_global_scale must be float32, got {input_global_scale.dtype}"
96
+ assert input_global_scale.shape == (
97
+ num_experts,
98
+ ), f"input_global_scale must be (l,), got {input_global_scale.shape}"
99
+
100
+ a_q, a_q_sf = scaled_fp4_grouped_quant(
101
+ hidden_states,
102
+ input_global_scale,
103
+ masked_m,
104
+ )
79
105
 
80
106
  assert w1.shape[-2] == 2 * n, f"w1 last-2 dim must be 2*n, got {w1.shape}"
81
107
  assert (
@@ -85,10 +111,6 @@ def flashinfer_cutedsl_moe_masked(
85
111
  k,
86
112
  n // 2,
87
113
  ), f"w2 shape mismatch, got {w2.shape[-2:]}, expected {(k, n//2)}"
88
-
89
- assert input_global_scale.shape == (
90
- num_experts,
91
- ), f"input_global_scale must be (l,), got {input_global_scale.shape}"
92
114
  assert w1_alpha.shape == (
93
115
  num_experts,
94
116
  ), f"w1_alpha must be (l,), got {w1_alpha.shape}"
@@ -99,27 +121,21 @@ def flashinfer_cutedsl_moe_masked(
99
121
  num_experts,
100
122
  ), f"w2_alpha must be (l,), got {w2_alpha.shape}"
101
123
 
102
- aq, aq_sf = scaled_fp4_grouped_quant(
103
- hidden_states,
104
- input_global_scale,
105
- masked_m,
106
- )
124
+ # TODO(kaixih@nvidia): dtype should be based on inputs.
107
125
  gateup_output = torch.empty(
108
- (num_experts, m, n * 2), dtype=hidden_states.dtype, device=aq.device
126
+ (num_experts, m, n * 2), dtype=torch.bfloat16, device=a_q.device
109
127
  )
110
128
  gateup_output = gateup_output.permute(1, 2, 0) # requirement of kernel
111
129
  sf_vec_size = 16
112
- assert aq_sf.dtype == torch.float8_e4m3fn
113
- assert aq.dtype == torch.uint8
130
+ assert a_q_sf.dtype == torch.float8_e4m3fn
131
+ assert a_q.dtype == torch.uint8
114
132
  ab_dtype = "float4_e2m1fn"
115
133
  sf_dtype = "float8_e4m3fn"
116
-
117
- c_dtype = get_cute_dtype(hidden_states)
134
+ c_dtype = "bfloat16"
118
135
 
119
136
  # Gemm1
120
-
121
137
  grouped_gemm_nt_masked(
122
- (aq, aq_sf),
138
+ (a_q, a_q_sf),
123
139
  (w1.permute(1, 2, 0), w1_blockscale),
124
140
  gateup_output,
125
141
  masked_m,
@@ -138,8 +154,11 @@ def flashinfer_cutedsl_moe_masked(
138
154
  masked_m,
139
155
  )
140
156
 
157
+ if down_start_event is not None:
158
+ down_start_event.record()
159
+
141
160
  # Gemm2
142
- out = torch.empty_like(hidden_states)
161
+ out = torch.empty((num_experts, m, k), dtype=torch.bfloat16, device=a_q.device)
143
162
  out = out.permute(1, 2, 0) # requirement of kernel
144
163
  grouped_gemm_nt_masked(
145
164
  (diq, diq_sf),
@@ -152,5 +171,13 @@ def flashinfer_cutedsl_moe_masked(
152
171
  sf_vec_size=sf_vec_size,
153
172
  alpha=w2_alpha.view(1, 1, num_experts),
154
173
  alpha_dtype=get_cute_dtype(w2_alpha),
174
+ **(
175
+ dict(
176
+ sm_count=down_sm_count,
177
+ dst_signals=down_signals,
178
+ )
179
+ if down_sm_count is not None or down_signals is not None
180
+ else {}
181
+ ),
155
182
  ) # in logical [m, k, l]
156
183
  return out.permute(2, 0, 1)
sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=256,device_name=NVIDIA_B200.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 8,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 64,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 256,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 8,
48
+ "num_stages": 2
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 8,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 8,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 8,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 2
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 8,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 2
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 2
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=256,device_name=NVIDIA_H800,dtype=fp8_w8a8,block_shape=[128, 128].json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 64,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 64,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 16,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 32,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }