sglang 0.5.1.post2__py3-none-any.whl → 0.5.2__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 (256) hide show
  1. sglang/bench_one_batch.py +3 -0
  2. sglang/bench_one_batch_server.py +89 -54
  3. sglang/bench_serving.py +437 -40
  4. sglang/lang/interpreter.py +1 -1
  5. sglang/profiler.py +0 -1
  6. sglang/srt/configs/__init__.py +4 -0
  7. sglang/srt/configs/internvl.py +6 -0
  8. sglang/srt/configs/longcat_flash.py +104 -0
  9. sglang/srt/configs/model_config.py +37 -7
  10. sglang/srt/configs/qwen3_next.py +326 -0
  11. sglang/srt/connector/__init__.py +1 -1
  12. sglang/srt/connector/base_connector.py +1 -2
  13. sglang/srt/connector/redis.py +2 -2
  14. sglang/srt/connector/serde/__init__.py +1 -1
  15. sglang/srt/connector/serde/safe_serde.py +4 -3
  16. sglang/srt/custom_op.py +11 -1
  17. sglang/srt/debug_utils/dump_comparator.py +81 -44
  18. sglang/srt/debug_utils/dump_loader.py +97 -0
  19. sglang/srt/debug_utils/dumper.py +11 -3
  20. sglang/srt/debug_utils/text_comparator.py +73 -11
  21. sglang/srt/disaggregation/ascend/conn.py +75 -0
  22. sglang/srt/disaggregation/base/conn.py +1 -1
  23. sglang/srt/disaggregation/common/conn.py +15 -12
  24. sglang/srt/disaggregation/decode.py +6 -4
  25. sglang/srt/disaggregation/fake/conn.py +1 -1
  26. sglang/srt/disaggregation/mini_lb.py +6 -420
  27. sglang/srt/disaggregation/mooncake/conn.py +18 -10
  28. sglang/srt/disaggregation/nixl/conn.py +180 -16
  29. sglang/srt/disaggregation/prefill.py +6 -4
  30. sglang/srt/disaggregation/utils.py +5 -50
  31. sglang/srt/distributed/parallel_state.py +94 -58
  32. sglang/srt/entrypoints/engine.py +34 -14
  33. sglang/srt/entrypoints/http_server.py +172 -47
  34. sglang/srt/entrypoints/openai/protocol.py +90 -27
  35. sglang/srt/entrypoints/openai/serving_base.py +6 -2
  36. sglang/srt/entrypoints/openai/serving_chat.py +82 -26
  37. sglang/srt/entrypoints/openai/serving_completions.py +25 -4
  38. sglang/srt/entrypoints/openai/serving_embedding.py +8 -4
  39. sglang/srt/entrypoints/openai/serving_responses.py +7 -4
  40. sglang/srt/eplb/eplb_manager.py +28 -4
  41. sglang/srt/eplb/expert_distribution.py +55 -15
  42. sglang/srt/eplb/expert_location.py +8 -3
  43. sglang/srt/eplb/expert_location_updater.py +1 -1
  44. sglang/srt/function_call/deepseekv31_detector.py +222 -0
  45. sglang/srt/function_call/ebnf_composer.py +11 -9
  46. sglang/srt/function_call/function_call_parser.py +2 -0
  47. sglang/srt/function_call/glm4_moe_detector.py +1 -1
  48. sglang/srt/function_call/gpt_oss_detector.py +144 -256
  49. sglang/srt/function_call/qwen3_coder_detector.py +1 -1
  50. sglang/srt/hf_transformers_utils.py +28 -7
  51. sglang/srt/layers/activation.py +44 -9
  52. sglang/srt/layers/attention/aiter_backend.py +93 -68
  53. sglang/srt/layers/attention/ascend_backend.py +381 -136
  54. sglang/srt/layers/attention/fla/chunk.py +242 -0
  55. sglang/srt/layers/attention/fla/chunk_delta_h.py +314 -0
  56. sglang/srt/layers/attention/fla/chunk_o.py +178 -0
  57. sglang/srt/layers/attention/fla/chunk_scaled_dot_kkt.py +151 -0
  58. sglang/srt/layers/attention/fla/cumsum.py +300 -0
  59. sglang/srt/layers/attention/fla/fused_recurrent.py +640 -0
  60. sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py +232 -0
  61. sglang/srt/layers/attention/fla/index.py +37 -0
  62. sglang/srt/layers/attention/fla/l2norm.py +150 -0
  63. sglang/srt/layers/attention/fla/layernorm_gated.py +326 -0
  64. sglang/srt/layers/attention/fla/op.py +66 -0
  65. sglang/srt/layers/attention/fla/solve_tril.py +465 -0
  66. sglang/srt/layers/attention/fla/utils.py +331 -0
  67. sglang/srt/layers/attention/fla/wy_fast.py +158 -0
  68. sglang/srt/layers/attention/flashattention_backend.py +241 -7
  69. sglang/srt/layers/attention/flashinfer_backend.py +11 -6
  70. sglang/srt/layers/attention/flashinfer_mla_backend.py +21 -14
  71. sglang/srt/layers/attention/hybrid_attn_backend.py +47 -8
  72. sglang/srt/layers/attention/hybrid_linear_attn_backend.py +584 -0
  73. sglang/srt/layers/attention/intel_amx_backend.py +3 -0
  74. sglang/srt/layers/attention/mamba/causal_conv1d.py +128 -0
  75. sglang/srt/layers/attention/mamba/causal_conv1d_triton.py +1052 -0
  76. sglang/srt/layers/attention/mamba/mamba.py +64 -0
  77. sglang/srt/layers/attention/torch_native_backend.py +12 -6
  78. sglang/srt/layers/attention/trtllm_mla_backend.py +126 -36
  79. sglang/srt/layers/attention/wave_ops/decode_attention.py +2 -4
  80. sglang/srt/layers/attention/wave_ops/extend_attention.py +1 -3
  81. sglang/srt/layers/communicator.py +45 -8
  82. sglang/srt/layers/layernorm.py +54 -12
  83. sglang/srt/layers/logits_processor.py +10 -3
  84. sglang/srt/layers/moe/__init__.py +2 -1
  85. sglang/srt/layers/moe/cutlass_moe.py +0 -8
  86. sglang/srt/layers/moe/cutlass_w4a8_moe.py +4 -12
  87. sglang/srt/layers/moe/ep_moe/kernels.py +74 -0
  88. sglang/srt/layers/moe/ep_moe/layer.py +111 -56
  89. sglang/srt/layers/moe/fused_moe_native.py +5 -3
  90. sglang/srt/layers/moe/fused_moe_triton/__init__.py +5 -3
  91. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  92. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=129,N=352,device_name=NVIDIA_B200,dtype=fp8_w8a8.json +146 -0
  93. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/{E=128,N=768,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json → E=257,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json } +29 -29
  94. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=257,N=64,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  95. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  96. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H20-3e.json +146 -0
  97. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=256,device_name=NVIDIA_H20-3e.json +146 -0
  98. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=64,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  99. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +9 -1049
  100. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_config.py +212 -0
  101. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py +799 -0
  102. sglang/srt/layers/moe/fused_moe_triton/layer.py +56 -45
  103. sglang/srt/layers/moe/fused_moe_triton/moe_align_block_size.py +87 -0
  104. sglang/srt/layers/moe/moe_runner/__init__.py +2 -1
  105. sglang/srt/layers/moe/moe_runner/base.py +274 -1
  106. sglang/srt/layers/moe/moe_runner/runner.py +80 -0
  107. sglang/srt/layers/moe/moe_runner/triton.py +448 -0
  108. sglang/srt/layers/moe/token_dispatcher/__init__.py +16 -4
  109. sglang/srt/layers/moe/token_dispatcher/{base_dispatcher.py → base.py} +67 -17
  110. sglang/srt/layers/moe/token_dispatcher/deepep.py +41 -38
  111. sglang/srt/layers/moe/token_dispatcher/standard.py +44 -2
  112. sglang/srt/layers/moe/topk.py +43 -12
  113. sglang/srt/layers/moe/utils.py +6 -5
  114. sglang/srt/layers/quantization/awq.py +19 -7
  115. sglang/srt/layers/quantization/base_config.py +11 -6
  116. sglang/srt/layers/quantization/blockwise_int8.py +38 -27
  117. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +50 -30
  118. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +13 -1
  119. sglang/srt/layers/quantization/deep_gemm_wrapper/compile_utils.py +141 -235
  120. sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py +5 -10
  121. sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py +31 -22
  122. sglang/srt/layers/quantization/fp8.py +78 -48
  123. sglang/srt/layers/quantization/fp8_kernel.py +2 -2
  124. sglang/srt/layers/quantization/fp8_utils.py +45 -31
  125. sglang/srt/layers/quantization/gptq.py +25 -17
  126. sglang/srt/layers/quantization/modelopt_quant.py +107 -40
  127. sglang/srt/layers/quantization/moe_wna16.py +21 -18
  128. sglang/srt/layers/quantization/mxfp4.py +93 -68
  129. sglang/srt/layers/quantization/mxfp4_tensor.py +3 -1
  130. sglang/srt/layers/quantization/quark/quark_moe.py +32 -27
  131. sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py +49 -30
  132. sglang/srt/layers/quantization/quark/utils.py +97 -0
  133. sglang/srt/layers/quantization/rocm_mxfp4_utils.py +13 -0
  134. sglang/srt/layers/quantization/unquant.py +135 -47
  135. sglang/srt/layers/quantization/utils.py +13 -0
  136. sglang/srt/layers/quantization/w4afp8.py +60 -42
  137. sglang/srt/layers/quantization/w8a8_fp8.py +35 -20
  138. sglang/srt/layers/quantization/w8a8_int8.py +83 -41
  139. sglang/srt/layers/rocm_linear_utils.py +44 -0
  140. sglang/srt/layers/rotary_embedding.py +28 -19
  141. sglang/srt/layers/sampler.py +29 -5
  142. sglang/srt/layers/utils.py +0 -14
  143. sglang/srt/lora/backend/base_backend.py +50 -8
  144. sglang/srt/lora/backend/triton_backend.py +90 -2
  145. sglang/srt/lora/layers.py +32 -0
  146. sglang/srt/lora/lora.py +4 -1
  147. sglang/srt/lora/lora_manager.py +35 -112
  148. sglang/srt/lora/mem_pool.py +24 -10
  149. sglang/srt/lora/utils.py +18 -9
  150. sglang/srt/managers/cache_controller.py +396 -365
  151. sglang/srt/managers/data_parallel_controller.py +30 -15
  152. sglang/srt/managers/detokenizer_manager.py +18 -2
  153. sglang/srt/managers/disagg_service.py +46 -0
  154. sglang/srt/managers/io_struct.py +190 -11
  155. sglang/srt/managers/mm_utils.py +6 -1
  156. sglang/srt/managers/multi_tokenizer_mixin.py +579 -0
  157. sglang/srt/managers/schedule_batch.py +27 -44
  158. sglang/srt/managers/schedule_policy.py +4 -3
  159. sglang/srt/managers/scheduler.py +148 -122
  160. sglang/srt/managers/scheduler_metrics_mixin.py +114 -8
  161. sglang/srt/managers/scheduler_output_processor_mixin.py +29 -19
  162. sglang/srt/managers/scheduler_profiler_mixin.py +1 -1
  163. sglang/srt/managers/scheduler_update_weights_mixin.py +8 -1
  164. sglang/srt/managers/template_manager.py +3 -3
  165. sglang/srt/managers/tokenizer_communicator_mixin.py +491 -0
  166. sglang/srt/managers/tokenizer_manager.py +77 -480
  167. sglang/srt/managers/tp_worker.py +16 -4
  168. sglang/srt/managers/tp_worker_overlap_thread.py +8 -10
  169. sglang/srt/mem_cache/allocator.py +1 -1
  170. sglang/srt/mem_cache/chunk_cache.py +1 -1
  171. sglang/srt/mem_cache/hicache_storage.py +53 -40
  172. sglang/srt/mem_cache/hiradix_cache.py +196 -104
  173. sglang/srt/mem_cache/lora_radix_cache.py +1 -1
  174. sglang/srt/mem_cache/memory_pool.py +395 -53
  175. sglang/srt/mem_cache/memory_pool_host.py +27 -19
  176. sglang/srt/mem_cache/radix_cache.py +6 -6
  177. sglang/srt/mem_cache/radix_cache_cpp.py +1 -1
  178. sglang/srt/mem_cache/storage/hf3fs/hf3fs_client.py +164 -0
  179. sglang/srt/mem_cache/storage/hf3fs/{client_hf3fs.py → hf3fs_usrbio_client.py} +5 -1
  180. sglang/srt/mem_cache/storage/hf3fs/mini_3fs_metadata_server.py +61 -34
  181. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +152 -23
  182. sglang/srt/mem_cache/storage/lmcache/lmc_radix_cache.py +280 -0
  183. sglang/srt/mem_cache/storage/lmcache/unit_test.py +121 -0
  184. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +154 -95
  185. sglang/srt/mem_cache/storage/mooncake_store/test_mooncake_store.py +161 -0
  186. sglang/srt/mem_cache/swa_radix_cache.py +1 -3
  187. sglang/srt/metrics/collector.py +484 -63
  188. sglang/srt/metrics/startup_func_log_and_timer.py +150 -0
  189. sglang/srt/metrics/utils.py +48 -0
  190. sglang/srt/model_executor/cpu_graph_runner.py +640 -0
  191. sglang/srt/model_executor/cuda_graph_runner.py +13 -5
  192. sglang/srt/model_executor/forward_batch_info.py +72 -18
  193. sglang/srt/model_executor/model_runner.py +190 -32
  194. sglang/srt/model_loader/__init__.py +9 -3
  195. sglang/srt/model_loader/loader.py +33 -28
  196. sglang/srt/model_loader/utils.py +12 -0
  197. sglang/srt/model_loader/weight_utils.py +2 -1
  198. sglang/srt/models/deepseek_v2.py +323 -53
  199. sglang/srt/models/gemma3n_mm.py +1 -1
  200. sglang/srt/models/glm4_moe.py +10 -1
  201. sglang/srt/models/glm4v.py +4 -2
  202. sglang/srt/models/gpt_oss.py +7 -19
  203. sglang/srt/models/internvl.py +28 -0
  204. sglang/srt/models/llama4.py +9 -0
  205. sglang/srt/models/llama_eagle3.py +17 -0
  206. sglang/srt/models/longcat_flash.py +1026 -0
  207. sglang/srt/models/longcat_flash_nextn.py +699 -0
  208. sglang/srt/models/minicpmv.py +165 -3
  209. sglang/srt/models/mllama4.py +25 -0
  210. sglang/srt/models/opt.py +637 -0
  211. sglang/srt/models/qwen2.py +33 -3
  212. sglang/srt/models/qwen2_5_vl.py +91 -42
  213. sglang/srt/models/qwen2_moe.py +79 -14
  214. sglang/srt/models/qwen3.py +8 -2
  215. sglang/srt/models/qwen3_moe.py +39 -8
  216. sglang/srt/models/qwen3_next.py +1039 -0
  217. sglang/srt/models/qwen3_next_mtp.py +109 -0
  218. sglang/srt/models/torch_native_llama.py +1 -1
  219. sglang/srt/models/transformers.py +1 -1
  220. sglang/srt/multimodal/processors/base_processor.py +4 -2
  221. sglang/srt/multimodal/processors/glm4v.py +9 -9
  222. sglang/srt/multimodal/processors/internvl.py +141 -129
  223. sglang/srt/{conversation.py → parser/conversation.py} +38 -5
  224. sglang/srt/parser/harmony_parser.py +588 -0
  225. sglang/srt/parser/reasoning_parser.py +309 -0
  226. sglang/srt/sampling/penaltylib/orchestrator.py +14 -2
  227. sglang/srt/sampling/sampling_batch_info.py +18 -15
  228. sglang/srt/server_args.py +307 -80
  229. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +5 -0
  230. sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +10 -1
  231. sglang/srt/speculative/eagle_worker.py +216 -120
  232. sglang/srt/speculative/spec_info.py +5 -0
  233. sglang/srt/speculative/standalone_worker.py +109 -0
  234. sglang/srt/tokenizer/tiktoken_tokenizer.py +6 -1
  235. sglang/srt/utils.py +96 -7
  236. sglang/srt/weight_sync/utils.py +1 -1
  237. sglang/test/attention/test_trtllm_mla_backend.py +181 -8
  238. sglang/test/few_shot_gsm8k.py +1 -0
  239. sglang/test/runners.py +4 -0
  240. sglang/test/test_cutlass_moe.py +24 -6
  241. sglang/test/test_cutlass_w4a8_moe.py +24 -9
  242. sglang/test/test_disaggregation_utils.py +66 -0
  243. sglang/test/test_utils.py +25 -1
  244. sglang/utils.py +5 -0
  245. sglang/version.py +1 -1
  246. {sglang-0.5.1.post2.dist-info → sglang-0.5.2.dist-info}/METADATA +13 -10
  247. {sglang-0.5.1.post2.dist-info → sglang-0.5.2.dist-info}/RECORD +253 -201
  248. sglang/srt/disaggregation/launch_lb.py +0 -131
  249. sglang/srt/mem_cache/storage/mooncake_store/unit_test.py +0 -40
  250. sglang/srt/reasoning_parser.py +0 -553
  251. /sglang/srt/{model_parallel.py → layers/model_parallel.py} +0 -0
  252. /sglang/srt/{code_completion_parser.py → parser/code_completion_parser.py} +0 -0
  253. /sglang/srt/{jinja_template_utils.py → parser/jinja_template_utils.py} +0 -0
  254. {sglang-0.5.1.post2.dist-info → sglang-0.5.2.dist-info}/WHEEL +0 -0
  255. {sglang-0.5.1.post2.dist-info → sglang-0.5.2.dist-info}/licenses/LICENSE +0 -0
  256. {sglang-0.5.1.post2.dist-info → sglang-0.5.2.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H20-3e.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": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 64,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
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": 128,
45
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 64,
61
+ "BLOCK_SIZE_K": 64,
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": 64,
69
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 64,
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": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 16,
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": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 64,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 4,
120
+ "num_stages": 2
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 64,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 8,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 64,
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=512,N=256,device_name=NVIDIA_H20-3e.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 3
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": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 64,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
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": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 64,
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": 64,
53
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 64,
69
+ "BLOCK_SIZE_K": 64,
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": 64,
78
+ "GROUP_SIZE_M": 64,
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": 64,
86
+ "GROUP_SIZE_M": 32,
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": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 16,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
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": 64,
118
+ "GROUP_SIZE_M": 64,
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": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
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": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=64,device_name=NVIDIA_H100_80GB_HBM3.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": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 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": 128,
45
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 64,
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": 64,
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": 64,
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": 64,
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": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 16,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 2
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 1,
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": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 8,
120
+ "num_stages": 2
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 8,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
135
+ "num_warps": 4,
136
+ "num_stages": 2
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 32,
143
+ "num_warps": 8,
144
+ "num_stages": 3
145
+ }
146
+ }