sglang 0.4.6.post5__py3-none-any.whl → 0.4.7__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 (318) hide show
  1. sglang/bench_offline_throughput.py +10 -4
  2. sglang/bench_one_batch_server.py +67 -11
  3. sglang/bench_serving.py +85 -74
  4. sglang/lang/backend/runtime_endpoint.py +24 -1
  5. sglang/profiler.py +167 -0
  6. sglang/srt/_custom_ops.py +34 -0
  7. sglang/srt/configs/internvl.py +8 -12
  8. sglang/srt/configs/model_config.py +27 -1
  9. sglang/srt/constrained/base_grammar_backend.py +5 -2
  10. sglang/srt/constrained/llguidance_backend.py +9 -8
  11. sglang/srt/constrained/outlines_backend.py +5 -4
  12. sglang/srt/constrained/xgrammar_backend.py +18 -18
  13. sglang/srt/conversation.py +46 -8
  14. sglang/srt/custom_op.py +38 -3
  15. sglang/srt/debug_utils.py +74 -0
  16. sglang/srt/disaggregation/common/__init__.py +1 -0
  17. sglang/srt/disaggregation/common/conn.py +407 -0
  18. sglang/srt/disaggregation/decode.py +67 -3
  19. sglang/srt/disaggregation/fake/conn.py +1 -0
  20. sglang/srt/disaggregation/kv_events.py +60 -5
  21. sglang/srt/disaggregation/launch_lb.py +140 -0
  22. sglang/srt/disaggregation/mini_lb.py +29 -48
  23. sglang/srt/disaggregation/mooncake/conn.py +432 -140
  24. sglang/srt/disaggregation/mooncake/transfer_engine.py +32 -16
  25. sglang/srt/disaggregation/nixl/conn.py +124 -432
  26. sglang/srt/disaggregation/prefill.py +2 -0
  27. sglang/srt/disaggregation/utils.py +38 -1
  28. sglang/srt/distributed/device_communicators/pymscclpp.py +315 -0
  29. sglang/srt/distributed/parallel_state.py +52 -5
  30. sglang/srt/entrypoints/EngineBase.py +6 -0
  31. sglang/srt/entrypoints/engine.py +102 -5
  32. sglang/srt/entrypoints/http_server.py +15 -2
  33. sglang/srt/function_call/base_format_detector.py +138 -86
  34. sglang/srt/function_call/deepseekv3_detector.py +54 -6
  35. sglang/srt/function_call/ebnf_composer.py +33 -19
  36. sglang/srt/function_call/function_call_parser.py +27 -0
  37. sglang/srt/function_call/llama32_detector.py +33 -14
  38. sglang/srt/function_call/mistral_detector.py +73 -26
  39. sglang/srt/function_call/pythonic_detector.py +86 -20
  40. sglang/srt/function_call/qwen25_detector.py +64 -10
  41. sglang/srt/function_call/utils.py +17 -0
  42. sglang/srt/hf_transformers_utils.py +4 -0
  43. sglang/srt/layers/attention/aiter_backend.py +488 -123
  44. sglang/srt/layers/attention/base_attn_backend.py +4 -0
  45. sglang/srt/layers/attention/cutlass_mla_backend.py +2 -19
  46. sglang/srt/layers/attention/flashattention_backend.py +103 -18
  47. sglang/srt/layers/attention/flashinfer_backend.py +45 -1
  48. sglang/srt/layers/attention/flashinfer_mla_backend.py +37 -1
  49. sglang/srt/layers/attention/intel_amx_backend.py +128 -0
  50. sglang/srt/layers/attention/tbo_backend.py +232 -0
  51. sglang/srt/layers/attention/torch_native_backend.py +3 -0
  52. sglang/srt/layers/attention/triton_backend.py +244 -5
  53. sglang/srt/layers/attention/triton_ops/extend_attention.py +12 -4
  54. sglang/srt/layers/communicator.py +260 -194
  55. sglang/srt/layers/dp_attention.py +6 -5
  56. sglang/srt/layers/layernorm.py +30 -19
  57. sglang/srt/layers/moe/cutlass_moe.py +170 -7
  58. sglang/srt/layers/moe/cutlass_moe_params.py +169 -0
  59. sglang/srt/layers/moe/ep_moe/kernels.py +27 -6
  60. sglang/srt/layers/moe/ep_moe/layer.py +94 -40
  61. sglang/srt/layers/moe/ep_moe/token_dispatcher.py +13 -8
  62. sglang/srt/layers/moe/fused_moe_native.py +4 -0
  63. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=257,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  64. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=257,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  65. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=257,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  66. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=257,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  67. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=257,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  68. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  69. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=257,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  70. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  71. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +220 -25
  72. sglang/srt/layers/moe/fused_moe_triton/layer.py +34 -4
  73. sglang/srt/layers/moe/topk.py +44 -18
  74. sglang/srt/layers/multimodal.py +3 -3
  75. sglang/srt/layers/quantization/__init__.py +3 -2
  76. sglang/srt/layers/quantization/blockwise_int8.py +3 -0
  77. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +5 -0
  78. sglang/srt/layers/quantization/deep_gemm.py +55 -56
  79. sglang/srt/layers/quantization/fp8.py +28 -23
  80. sglang/srt/layers/quantization/fp8_kernel.py +118 -66
  81. sglang/srt/layers/quantization/fp8_utils.py +165 -49
  82. sglang/srt/layers/quantization/modelopt_quant.py +334 -7
  83. sglang/srt/layers/quantization/moe_wna16.py +3 -0
  84. sglang/srt/layers/quantization/w8a8_fp8.py +3 -0
  85. sglang/srt/layers/quantization/w8a8_int8.py +3 -0
  86. sglang/srt/layers/rotary_embedding.py +6 -12
  87. sglang/srt/layers/sampler.py +80 -79
  88. sglang/srt/layers/utils.py +6 -0
  89. sglang/srt/lora/layers.py +12 -15
  90. sglang/srt/lora/lora.py +49 -5
  91. sglang/srt/lora/lora_manager.py +19 -5
  92. sglang/srt/lora/mem_pool.py +24 -16
  93. sglang/srt/lora/utils.py +17 -13
  94. sglang/srt/managers/data_parallel_controller.py +13 -5
  95. sglang/srt/managers/eplb_algorithms/__init__.py +63 -0
  96. sglang/srt/managers/eplb_algorithms/deepseek.py +223 -0
  97. sglang/srt/managers/{deepseek_eplb.py → eplb_algorithms/deepseek_vec.py} +5 -7
  98. sglang/srt/managers/eplb_manager.py +55 -14
  99. sglang/srt/managers/expert_distribution.py +220 -46
  100. sglang/srt/managers/expert_location.py +110 -56
  101. sglang/srt/managers/expert_location_dispatch.py +23 -6
  102. sglang/srt/managers/io_struct.py +15 -4
  103. sglang/srt/managers/mm_utils.py +88 -38
  104. sglang/srt/managers/multimodal_processors/base_processor.py +188 -16
  105. sglang/srt/managers/multimodal_processors/gemma3.py +4 -31
  106. sglang/srt/managers/multimodal_processors/internvl.py +4 -0
  107. sglang/srt/managers/multimodal_processors/kimi_vl.py +15 -34
  108. sglang/srt/managers/multimodal_processors/minicpm.py +2 -1
  109. sglang/srt/managers/multimodal_processors/phi4mm.py +87 -0
  110. sglang/srt/managers/multimodal_processors/qwen_vl.py +22 -64
  111. sglang/srt/managers/schedule_batch.py +140 -38
  112. sglang/srt/managers/scheduler.py +305 -112
  113. sglang/srt/managers/tokenizer_manager.py +134 -17
  114. sglang/srt/managers/utils.py +0 -4
  115. sglang/srt/metrics/collector.py +9 -0
  116. sglang/srt/model_executor/cuda_graph_runner.py +72 -61
  117. sglang/srt/model_executor/expert_location_updater.py +157 -22
  118. sglang/srt/model_executor/forward_batch_info.py +38 -17
  119. sglang/srt/model_executor/model_runner.py +96 -56
  120. sglang/srt/model_loader/utils.py +67 -1
  121. sglang/srt/models/deepseek_nextn.py +1 -1
  122. sglang/srt/models/deepseek_v2.py +609 -234
  123. sglang/srt/models/gemma3_causal.py +7 -0
  124. sglang/srt/models/gemma3_mm.py +19 -14
  125. sglang/srt/models/idefics2.py +342 -0
  126. sglang/srt/models/kimi_vl.py +4 -4
  127. sglang/srt/models/llama.py +1 -1
  128. sglang/srt/models/minicpmo.py +2 -5
  129. sglang/srt/models/minicpmv.py +3 -295
  130. sglang/srt/models/phi4mm.py +512 -0
  131. sglang/srt/models/qwen2.py +38 -9
  132. sglang/srt/models/qwen2_5_vl.py +3 -9
  133. sglang/srt/models/qwen2_eagle.py +4 -1
  134. sglang/srt/models/qwen2_moe.py +58 -191
  135. sglang/srt/models/qwen2_vl.py +3 -9
  136. sglang/srt/models/qwen3.py +41 -10
  137. sglang/srt/models/qwen3_moe.py +230 -191
  138. sglang/srt/models/registry.py +9 -1
  139. sglang/srt/models/transformers.py +291 -0
  140. sglang/srt/openai_api/adapter.py +86 -24
  141. sglang/srt/openai_api/protocol.py +31 -2
  142. sglang/srt/openai_api/utils.py +172 -0
  143. sglang/srt/operations.py +37 -2
  144. sglang/srt/operations_strategy.py +200 -24
  145. sglang/srt/sampling/sampling_batch_info.py +13 -1
  146. sglang/srt/sampling/sampling_params.py +2 -1
  147. sglang/srt/server_args.py +114 -27
  148. sglang/srt/speculative/build_eagle_tree.py +8 -8
  149. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +8 -11
  150. sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +253 -0
  151. sglang/srt/speculative/eagle_utils.py +51 -91
  152. sglang/srt/speculative/eagle_worker.py +101 -21
  153. sglang/srt/two_batch_overlap.py +635 -0
  154. sglang/srt/utils.py +129 -7
  155. sglang/test/runners.py +16 -7
  156. sglang/test/send_one.py +4 -0
  157. sglang/test/test_cutlass_moe.py +3 -3
  158. sglang/test/test_fp4_moe.py +248 -0
  159. sglang/test/test_utils.py +79 -6
  160. sglang/version.py +1 -1
  161. {sglang-0.4.6.post5.dist-info → sglang-0.4.7.dist-info}/METADATA +14 -11
  162. {sglang-0.4.6.post5.dist-info → sglang-0.4.7.dist-info}/RECORD +318 -291
  163. {sglang-0.4.6.post5.dist-info → sglang-0.4.7.dist-info}/WHEEL +1 -1
  164. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  165. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  166. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  167. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  168. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  169. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json → triton_3_1_0/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json} +0 -0
  170. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  171. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  172. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  173. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  174. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  175. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=144,N=512,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=144,N=512,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  176. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=16,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  177. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1024,device_name=NVIDIA_H200.json → triton_3_1_0/E=16,N=1024,device_name=NVIDIA_H200.json} +0 -0
  178. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json → triton_3_1_0/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json} +0 -0
  179. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  180. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  181. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  182. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  183. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  184. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  185. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=16,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  186. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  187. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  188. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  189. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json} +0 -0
  190. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  191. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  192. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  193. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  194. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json} +0 -0
  195. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  196. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json → triton_3_1_0/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json} +0 -0
  197. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  198. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_1_0/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  199. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=20,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=20,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  200. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=24,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=24,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  201. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json} +0 -0
  202. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json} +0 -0
  203. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json} +0 -0
  204. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json} +0 -0
  205. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  206. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_H20,block_shape=[128, 128].json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_H20,block_shape=[128, 128].json} +0 -0
  207. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  208. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=128,device_name=NVIDIA_L20Y,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  209. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  210. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  211. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  212. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  213. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json} +0 -0
  214. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  215. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_1_0/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  216. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=64,device_name=NVIDIA_L20,dtype=int8_w8a8.json → triton_3_1_0/E=256,N=64,device_name=NVIDIA_L20,dtype=int8_w8a8.json} +0 -0
  217. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=256,N=64,device_name=NVIDIA_L40S,dtype=int8_w8a8.json → triton_3_1_0/E=256,N=64,device_name=NVIDIA_L40S,dtype=int8_w8a8.json} +0 -0
  218. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=64,N=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  219. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  220. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_1_0/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  221. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  222. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  223. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  224. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=1280,device_name=NVIDIA_H200.json → triton_3_1_0/E=64,N=1280,device_name=NVIDIA_H200.json} +0 -0
  225. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  226. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  227. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=2560,device_name=NVIDIA_H200.json → triton_3_1_0/E=64,N=2560,device_name=NVIDIA_H200.json} +0 -0
  228. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  229. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  230. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  231. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=320,device_name=NVIDIA_H200.json → triton_3_1_0/E=64,N=320,device_name=NVIDIA_H200.json} +0 -0
  232. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_1_0/E=64,N=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  233. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  234. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  235. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json} +0 -0
  236. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  237. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  238. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  239. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=64,N=640,device_name=NVIDIA_H200.json → triton_3_1_0/E=64,N=640,device_name=NVIDIA_H200.json} +0 -0
  240. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=14336,device_name=AMD_Instinct_MI300X.json → triton_3_1_0/E=8,N=14336,device_name=AMD_Instinct_MI300X.json} +0 -0
  241. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=14336,device_name=AMD_Instinct_MI325X.json → triton_3_1_0/E=8,N=14336,device_name=AMD_Instinct_MI325X.json} +0 -0
  242. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=14336,device_name=AMD_Radeon_Graphics.json → triton_3_1_0/E=8,N=14336,device_name=AMD_Radeon_Graphics.json} +0 -0
  243. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  244. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  245. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=14336,device_name=NVIDIA_H200.json → triton_3_1_0/E=8,N=14336,device_name=NVIDIA_H200.json} +0 -0
  246. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=AMD_Instinct_MI300X.json → triton_3_1_0/E=8,N=1792,device_name=AMD_Instinct_MI300X.json} +0 -0
  247. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=AMD_Instinct_MI325X.json → triton_3_1_0/E=8,N=1792,device_name=AMD_Instinct_MI325X.json} +0 -0
  248. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=AMD_Radeon_Graphics.json → triton_3_1_0/E=8,N=1792,device_name=AMD_Radeon_Graphics.json} +0 -0
  249. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json → triton_3_1_0/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json} +0 -0
  250. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  251. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  252. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  253. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=1792,device_name=NVIDIA_H200.json → triton_3_1_0/E=8,N=1792,device_name=NVIDIA_H200.json} +0 -0
  254. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  255. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  256. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  257. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  258. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=2048,device_name=NVIDIA_H200.json → triton_3_1_0/E=8,N=2048,device_name=NVIDIA_H200.json} +0 -0
  259. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=AMD_Instinct_MI300X.json → triton_3_1_0/E=8,N=3584,device_name=AMD_Instinct_MI300X.json} +0 -0
  260. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=AMD_Instinct_MI325X.json → triton_3_1_0/E=8,N=3584,device_name=AMD_Instinct_MI325X.json} +0 -0
  261. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=AMD_Radeon_Graphics.json → triton_3_1_0/E=8,N=3584,device_name=AMD_Radeon_Graphics.json} +0 -0
  262. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json} +0 -0
  263. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  264. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json} +0 -0
  265. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  266. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  267. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  268. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_H200.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_H200.json} +0 -0
  269. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=3584,device_name=NVIDIA_L40S.json → triton_3_1_0/E=8,N=3584,device_name=NVIDIA_L40S.json} +0 -0
  270. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json} +0 -0
  271. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json} +0 -0
  272. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=4096,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8.json} +0 -0
  273. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  274. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  275. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  276. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  277. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=4096,device_name=NVIDIA_H200.json → triton_3_1_0/E=8,N=4096,device_name=NVIDIA_H200.json} +0 -0
  278. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=AMD_Instinct_MI300X.json → triton_3_1_0/E=8,N=7168,device_name=AMD_Instinct_MI300X.json} +0 -0
  279. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=AMD_Instinct_MI325X.json → triton_3_1_0/E=8,N=7168,device_name=AMD_Instinct_MI325X.json} +0 -0
  280. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=AMD_Radeon_Graphics.json → triton_3_1_0/E=8,N=7168,device_name=AMD_Radeon_Graphics.json} +0 -0
  281. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json → triton_3_1_0/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json} +0 -0
  282. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  283. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_1_0/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  284. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  285. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=7168,device_name=NVIDIA_H200.json → triton_3_1_0/E=8,N=7168,device_name=NVIDIA_H200.json} +0 -0
  286. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json} +0 -0
  287. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json} +0 -0
  288. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=8192,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=8192,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8.json} +0 -0
  289. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json} +0 -0
  290. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json → triton_3_1_0/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json} +0 -0
  291. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=192,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_2_0/E=128,N=192,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  292. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_2_0/E=128,N=192,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  293. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=192,device_name=NVIDIA_H20.json → triton_3_2_0/E=128,N=192,device_name=NVIDIA_H20.json} +0 -0
  294. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=192,device_name=NVIDIA_H200.json → triton_3_2_0/E=128,N=192,device_name=NVIDIA_H200.json} +0 -0
  295. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=384,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_2_0/E=128,N=384,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  296. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  297. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=384,device_name=NVIDIA_H20.json → triton_3_2_0/E=128,N=384,device_name=NVIDIA_H20.json} +0 -0
  298. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=128,N=384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  299. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=384,device_name=NVIDIA_H200.json → triton_3_2_0/E=128,N=384,device_name=NVIDIA_H200.json} +0 -0
  300. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_2_0/E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  301. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=768,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_2_0/E=128,N=768,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  302. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=768,device_name=NVIDIA_H100_80GB_HBM3.json → triton_3_2_0/E=128,N=768,device_name=NVIDIA_H100_80GB_HBM3.json} +0 -0
  303. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=128,N=768,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  304. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=768,device_name=NVIDIA_H20.json → triton_3_2_0/E=128,N=768,device_name=NVIDIA_H20.json} +0 -0
  305. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=768,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=128,N=768,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  306. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=768,device_name=NVIDIA_H200.json → triton_3_2_0/E=128,N=768,device_name=NVIDIA_H200.json} +0 -0
  307. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=128,N=96,device_name=NVIDIA_H20.json → triton_3_2_0/E=128,N=96,device_name=NVIDIA_H20.json} +0 -0
  308. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=264,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json → triton_3_2_0/E=264,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json} +0 -0
  309. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=264,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=264,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  310. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=264,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=264,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  311. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=264,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=264,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  312. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=272,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json → triton_3_2_0/E=272,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json} +0 -0
  313. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=272,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=272,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  314. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=272,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json → triton_3_2_0/E=272,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json} +0 -0
  315. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=272,N=64,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_2_0/E=272,N=64,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  316. /sglang/srt/layers/moe/fused_moe_triton/configs/{E=288,N=64,device_name=NVIDIA_A800-SXM4-80GB.json → triton_3_2_0/E=288,N=64,device_name=NVIDIA_A800-SXM4-80GB.json} +0 -0
  317. {sglang-0.4.6.post5.dist-info → sglang-0.4.7.dist-info}/licenses/LICENSE +0 -0
  318. {sglang-0.4.6.post5.dist-info → sglang-0.4.7.dist-info}/top_level.txt +0 -0
sglang/srt/hf_transformers_utils.py CHANGED
@@ -203,6 +203,10 @@ def get_tokenizer(
203
203
  raise ValueError("Cannot use the fast tokenizer in slow tokenizer mode.")
204
204
  kwargs["use_fast"] = False
205
205
 
206
+ # TODO(Xinyuan): Remove this once we have a proper tokenizer for Devstral
207
+ if tokenizer_name == "mistralai/Devstral-Small-2505":
208
+ tokenizer_name = "mistralai/Mistral-Small-3.1-24B-Instruct-2503"
209
+
206
210
  is_gguf = check_gguf_file(tokenizer_name)
207
211
  if is_gguf:
208
212
  kwargs["gguf_file"] = tokenizer_name