sglang 0.5.2rc1__py3-none-any.whl → 0.5.3rc0__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 (265) hide show
  1. sglang/bench_one_batch_server.py +10 -1
  2. sglang/bench_serving.py +257 -29
  3. sglang/lang/interpreter.py +1 -1
  4. sglang/srt/configs/__init__.py +4 -0
  5. sglang/srt/configs/device_config.py +3 -1
  6. sglang/srt/configs/dots_vlm.py +139 -0
  7. sglang/srt/configs/internvl.py +6 -0
  8. sglang/srt/configs/load_config.py +1 -0
  9. sglang/srt/configs/model_config.py +50 -6
  10. sglang/srt/configs/qwen3_next.py +326 -0
  11. sglang/srt/connector/__init__.py +8 -1
  12. sglang/srt/connector/remote_instance.py +82 -0
  13. sglang/srt/constrained/base_grammar_backend.py +48 -12
  14. sglang/srt/constrained/llguidance_backend.py +0 -1
  15. sglang/srt/constrained/outlines_backend.py +0 -1
  16. sglang/srt/constrained/xgrammar_backend.py +28 -9
  17. sglang/srt/custom_op.py +11 -1
  18. sglang/srt/debug_utils/dump_comparator.py +81 -44
  19. sglang/srt/debug_utils/dump_loader.py +97 -0
  20. sglang/srt/debug_utils/dumper.py +11 -3
  21. sglang/srt/debug_utils/text_comparator.py +73 -11
  22. sglang/srt/disaggregation/base/conn.py +1 -1
  23. sglang/srt/disaggregation/common/conn.py +15 -12
  24. sglang/srt/disaggregation/decode.py +21 -10
  25. sglang/srt/disaggregation/decode_schedule_batch_mixin.py +4 -1
  26. sglang/srt/disaggregation/fake/conn.py +1 -1
  27. sglang/srt/disaggregation/mini_lb.py +6 -445
  28. sglang/srt/disaggregation/mooncake/conn.py +18 -10
  29. sglang/srt/disaggregation/nixl/conn.py +180 -16
  30. sglang/srt/disaggregation/prefill.py +5 -3
  31. sglang/srt/disaggregation/utils.py +5 -50
  32. sglang/srt/distributed/parallel_state.py +67 -43
  33. sglang/srt/entrypoints/engine.py +38 -17
  34. sglang/srt/entrypoints/grpc_request_manager.py +580 -0
  35. sglang/srt/entrypoints/grpc_server.py +680 -0
  36. sglang/srt/entrypoints/http_server.py +88 -53
  37. sglang/srt/entrypoints/openai/protocol.py +7 -4
  38. sglang/srt/entrypoints/openai/serving_base.py +46 -3
  39. sglang/srt/entrypoints/openai/serving_chat.py +39 -19
  40. sglang/srt/entrypoints/openai/serving_completions.py +15 -4
  41. sglang/srt/entrypoints/openai/serving_embedding.py +9 -4
  42. sglang/srt/entrypoints/openai/serving_rerank.py +3 -1
  43. sglang/srt/entrypoints/openai/serving_responses.py +7 -4
  44. sglang/srt/entrypoints/openai/serving_score.py +1 -0
  45. sglang/srt/eplb/eplb_manager.py +2 -2
  46. sglang/srt/eplb/expert_distribution.py +26 -13
  47. sglang/srt/eplb/expert_location.py +8 -3
  48. sglang/srt/eplb/expert_location_updater.py +1 -1
  49. sglang/srt/function_call/base_format_detector.py +3 -6
  50. sglang/srt/function_call/ebnf_composer.py +11 -9
  51. sglang/srt/function_call/function_call_parser.py +6 -0
  52. sglang/srt/function_call/glm4_moe_detector.py +1 -1
  53. sglang/srt/function_call/gpt_oss_detector.py +1 -1
  54. sglang/srt/function_call/qwen3_coder_detector.py +1 -1
  55. sglang/srt/grpc/__init__.py +1 -0
  56. sglang/srt/grpc/sglang_scheduler_pb2.py +106 -0
  57. sglang/srt/grpc/sglang_scheduler_pb2.pyi +427 -0
  58. sglang/srt/grpc/sglang_scheduler_pb2_grpc.py +236 -0
  59. sglang/srt/hf_transformers_utils.py +4 -0
  60. sglang/srt/layers/activation.py +142 -9
  61. sglang/srt/layers/attention/aiter_backend.py +93 -68
  62. sglang/srt/layers/attention/ascend_backend.py +11 -4
  63. sglang/srt/layers/attention/fla/chunk.py +242 -0
  64. sglang/srt/layers/attention/fla/chunk_delta_h.py +314 -0
  65. sglang/srt/layers/attention/fla/chunk_o.py +178 -0
  66. sglang/srt/layers/attention/fla/chunk_scaled_dot_kkt.py +151 -0
  67. sglang/srt/layers/attention/fla/cumsum.py +300 -0
  68. sglang/srt/layers/attention/fla/fused_recurrent.py +640 -0
  69. sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py +232 -0
  70. sglang/srt/layers/attention/fla/index.py +37 -0
  71. sglang/srt/layers/attention/fla/l2norm.py +150 -0
  72. sglang/srt/layers/attention/fla/layernorm_gated.py +326 -0
  73. sglang/srt/layers/attention/fla/op.py +66 -0
  74. sglang/srt/layers/attention/fla/solve_tril.py +465 -0
  75. sglang/srt/layers/attention/fla/utils.py +331 -0
  76. sglang/srt/layers/attention/fla/wy_fast.py +158 -0
  77. sglang/srt/layers/attention/flashinfer_backend.py +6 -4
  78. sglang/srt/layers/attention/flashinfer_mla_backend.py +16 -12
  79. sglang/srt/layers/attention/hybrid_attn_backend.py +57 -50
  80. sglang/srt/layers/attention/hybrid_linear_attn_backend.py +602 -0
  81. sglang/srt/layers/attention/intel_amx_backend.py +3 -0
  82. sglang/srt/layers/attention/mamba/causal_conv1d.py +128 -0
  83. sglang/srt/layers/attention/mamba/causal_conv1d_triton.py +1052 -0
  84. sglang/srt/layers/attention/mamba/mamba.py +64 -0
  85. sglang/srt/layers/attention/torch_native_backend.py +12 -6
  86. sglang/srt/layers/attention/triton_backend.py +18 -1
  87. sglang/srt/layers/attention/trtllm_mla_backend.py +124 -31
  88. sglang/srt/layers/attention/wave_ops/decode_attention.py +2 -4
  89. sglang/srt/layers/attention/wave_ops/extend_attention.py +1 -3
  90. sglang/srt/layers/communicator.py +45 -7
  91. sglang/srt/layers/dp_attention.py +30 -1
  92. sglang/srt/layers/layernorm.py +32 -15
  93. sglang/srt/layers/linear.py +34 -3
  94. sglang/srt/layers/logits_processor.py +29 -10
  95. sglang/srt/layers/moe/__init__.py +2 -1
  96. sglang/srt/layers/moe/cutlass_w4a8_moe.py +3 -3
  97. sglang/srt/layers/moe/ep_moe/kernels.py +1 -1
  98. sglang/srt/layers/moe/ep_moe/layer.py +182 -62
  99. sglang/srt/layers/moe/flashinfer_cutedsl_moe.py +156 -0
  100. sglang/srt/layers/moe/fused_moe_native.py +5 -3
  101. 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
  102. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=352,device_name=NVIDIA_RTX_5880_Ada_Generation,dtype=fp8_w8a8.json +146 -0
  103. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=512,device_name=NVIDIA_H20.json +146 -0
  104. 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
  105. 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
  106. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H20-3e.json +146 -0
  107. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=128,device_name=NVIDIA_H200.json +146 -0
  108. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=256,device_name=NVIDIA_H20-3e.json +146 -0
  109. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=256,device_name=NVIDIA_H200.json +146 -0
  110. 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
  111. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=512,N=64,device_name=NVIDIA_H200.json +146 -0
  112. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +5 -2
  113. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_config.py +1 -1
  114. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py +23 -20
  115. sglang/srt/layers/moe/fused_moe_triton/layer.py +61 -59
  116. sglang/srt/layers/moe/moe_runner/__init__.py +2 -1
  117. sglang/srt/layers/moe/moe_runner/base.py +274 -1
  118. sglang/srt/layers/moe/moe_runner/runner.py +80 -0
  119. sglang/srt/layers/moe/moe_runner/triton.py +448 -0
  120. sglang/srt/layers/moe/token_dispatcher/__init__.py +16 -4
  121. sglang/srt/layers/moe/token_dispatcher/{base_dispatcher.py → base.py} +67 -17
  122. sglang/srt/layers/moe/token_dispatcher/deepep.py +43 -39
  123. sglang/srt/layers/moe/token_dispatcher/standard.py +44 -2
  124. sglang/srt/layers/moe/topk.py +30 -9
  125. sglang/srt/layers/moe/utils.py +12 -7
  126. sglang/srt/layers/quantization/awq.py +19 -7
  127. sglang/srt/layers/quantization/base_config.py +11 -6
  128. sglang/srt/layers/quantization/blockwise_int8.py +38 -27
  129. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +50 -30
  130. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +13 -1
  131. sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py +27 -0
  132. sglang/srt/layers/quantization/fp8.py +76 -47
  133. sglang/srt/layers/quantization/fp8_utils.py +50 -31
  134. sglang/srt/layers/quantization/gptq.py +25 -17
  135. sglang/srt/layers/quantization/modelopt_quant.py +182 -49
  136. sglang/srt/layers/quantization/moe_wna16.py +21 -18
  137. sglang/srt/layers/quantization/mxfp4.py +68 -41
  138. sglang/srt/layers/quantization/quark/quark_moe.py +32 -27
  139. sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py +49 -30
  140. sglang/srt/layers/quantization/quark/utils.py +97 -0
  141. sglang/srt/layers/quantization/rocm_mxfp4_utils.py +13 -0
  142. sglang/srt/layers/quantization/unquant.py +135 -47
  143. sglang/srt/layers/quantization/w4afp8.py +30 -17
  144. sglang/srt/layers/quantization/w8a8_fp8.py +35 -20
  145. sglang/srt/layers/quantization/w8a8_int8.py +76 -38
  146. sglang/srt/layers/rocm_linear_utils.py +44 -0
  147. sglang/srt/layers/rotary_embedding.py +0 -18
  148. sglang/srt/layers/sampler.py +162 -18
  149. sglang/srt/lora/backend/base_backend.py +50 -8
  150. sglang/srt/lora/backend/triton_backend.py +90 -2
  151. sglang/srt/lora/layers.py +32 -0
  152. sglang/srt/lora/lora.py +4 -1
  153. sglang/srt/lora/lora_manager.py +35 -112
  154. sglang/srt/lora/mem_pool.py +24 -10
  155. sglang/srt/lora/utils.py +18 -9
  156. sglang/srt/managers/async_dynamic_batch_tokenizer.py +170 -0
  157. sglang/srt/managers/cache_controller.py +200 -199
  158. sglang/srt/managers/data_parallel_controller.py +105 -35
  159. sglang/srt/managers/detokenizer_manager.py +8 -4
  160. sglang/srt/managers/disagg_service.py +46 -0
  161. sglang/srt/managers/io_struct.py +199 -12
  162. sglang/srt/managers/mm_utils.py +1 -0
  163. sglang/srt/managers/multi_tokenizer_mixin.py +351 -397
  164. sglang/srt/managers/schedule_batch.py +77 -56
  165. sglang/srt/managers/schedule_policy.py +4 -3
  166. sglang/srt/managers/scheduler.py +191 -139
  167. sglang/srt/managers/scheduler_metrics_mixin.py +116 -9
  168. sglang/srt/managers/scheduler_output_processor_mixin.py +55 -11
  169. sglang/srt/managers/scheduler_profiler_mixin.py +1 -1
  170. sglang/srt/managers/template_manager.py +3 -3
  171. sglang/srt/managers/tokenizer_communicator_mixin.py +569 -0
  172. sglang/srt/managers/tokenizer_manager.py +260 -519
  173. sglang/srt/managers/tp_worker.py +53 -4
  174. sglang/srt/managers/tp_worker_overlap_thread.py +42 -19
  175. sglang/srt/mem_cache/allocator.py +1 -1
  176. sglang/srt/mem_cache/hicache_storage.py +18 -33
  177. sglang/srt/mem_cache/hiradix_cache.py +108 -48
  178. sglang/srt/mem_cache/memory_pool.py +347 -48
  179. sglang/srt/mem_cache/memory_pool_host.py +121 -57
  180. sglang/srt/mem_cache/radix_cache.py +0 -2
  181. sglang/srt/mem_cache/storage/hf3fs/hf3fs_client.py +164 -0
  182. sglang/srt/mem_cache/storage/hf3fs/{client_hf3fs.py → hf3fs_usrbio_client.py} +5 -1
  183. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +95 -5
  184. sglang/srt/mem_cache/storage/lmcache/lmc_radix_cache.py +280 -0
  185. sglang/srt/mem_cache/storage/lmcache/unit_test.py +121 -0
  186. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +81 -20
  187. sglang/srt/mem_cache/storage/mooncake_store/test_mooncake_store.py +161 -0
  188. sglang/srt/mem_cache/swa_radix_cache.py +0 -2
  189. sglang/srt/metrics/collector.py +502 -77
  190. sglang/srt/metrics/startup_func_log_and_timer.py +150 -0
  191. sglang/srt/metrics/utils.py +48 -0
  192. sglang/srt/model_executor/cpu_graph_runner.py +640 -0
  193. sglang/srt/model_executor/cuda_graph_runner.py +13 -5
  194. sglang/srt/model_executor/forward_batch_info.py +75 -19
  195. sglang/srt/model_executor/model_runner.py +357 -30
  196. sglang/srt/model_loader/__init__.py +9 -3
  197. sglang/srt/model_loader/loader.py +128 -4
  198. sglang/srt/model_loader/weight_utils.py +2 -1
  199. sglang/srt/models/apertus.py +686 -0
  200. sglang/srt/models/bailing_moe.py +798 -218
  201. sglang/srt/models/bailing_moe_nextn.py +168 -0
  202. sglang/srt/models/deepseek_v2.py +346 -48
  203. sglang/srt/models/dots_vlm.py +174 -0
  204. sglang/srt/models/dots_vlm_vit.py +337 -0
  205. sglang/srt/models/ernie4.py +1 -1
  206. sglang/srt/models/gemma3n_mm.py +1 -1
  207. sglang/srt/models/glm4_moe.py +11 -2
  208. sglang/srt/models/glm4v.py +4 -2
  209. sglang/srt/models/glm4v_moe.py +3 -0
  210. sglang/srt/models/gpt_oss.py +1 -1
  211. sglang/srt/models/internvl.py +28 -0
  212. sglang/srt/models/llama4.py +9 -0
  213. sglang/srt/models/llama_eagle3.py +13 -0
  214. sglang/srt/models/longcat_flash.py +2 -2
  215. sglang/srt/models/minicpmv.py +165 -3
  216. sglang/srt/models/mllama4.py +25 -0
  217. sglang/srt/models/opt.py +637 -0
  218. sglang/srt/models/qwen2.py +7 -0
  219. sglang/srt/models/qwen2_5_vl.py +27 -3
  220. sglang/srt/models/qwen2_moe.py +60 -13
  221. sglang/srt/models/qwen3.py +8 -2
  222. sglang/srt/models/qwen3_moe.py +40 -9
  223. sglang/srt/models/qwen3_next.py +1042 -0
  224. sglang/srt/models/qwen3_next_mtp.py +112 -0
  225. sglang/srt/models/step3_vl.py +1 -1
  226. sglang/srt/models/torch_native_llama.py +1 -1
  227. sglang/srt/multimodal/processors/dots_vlm.py +99 -0
  228. sglang/srt/multimodal/processors/glm4v.py +9 -9
  229. sglang/srt/multimodal/processors/internvl.py +141 -129
  230. sglang/srt/multimodal/processors/qwen_vl.py +15 -5
  231. sglang/srt/offloader.py +27 -3
  232. sglang/srt/{reasoning_parser.py → parser/reasoning_parser.py} +1 -1
  233. sglang/srt/remote_instance_weight_loader_utils.py +69 -0
  234. sglang/srt/sampling/sampling_batch_info.py +18 -15
  235. sglang/srt/server_args.py +355 -37
  236. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +5 -0
  237. sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +10 -1
  238. sglang/srt/speculative/eagle_utils.py +0 -2
  239. sglang/srt/speculative/eagle_worker.py +197 -112
  240. sglang/srt/speculative/spec_info.py +5 -0
  241. sglang/srt/speculative/standalone_worker.py +109 -0
  242. sglang/srt/tracing/trace.py +552 -0
  243. sglang/srt/utils.py +46 -3
  244. sglang/srt/weight_sync/utils.py +1 -1
  245. sglang/test/attention/test_trtllm_mla_backend.py +169 -5
  246. sglang/test/few_shot_gsm8k.py +1 -0
  247. sglang/test/runners.py +4 -0
  248. sglang/test/test_cutlass_moe.py +24 -6
  249. sglang/test/test_disaggregation_utils.py +66 -0
  250. sglang/test/test_fp4_moe.py +370 -1
  251. sglang/test/test_utils.py +28 -1
  252. sglang/utils.py +12 -0
  253. sglang/version.py +1 -1
  254. {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/METADATA +59 -123
  255. {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/RECORD +263 -200
  256. sglang/srt/disaggregation/launch_lb.py +0 -118
  257. sglang/srt/mem_cache/storage/mooncake_store/unit_test.py +0 -40
  258. /sglang/srt/{model_parallel.py → layers/model_parallel.py} +0 -0
  259. /sglang/srt/{code_completion_parser.py → parser/code_completion_parser.py} +0 -0
  260. /sglang/srt/{conversation.py → parser/conversation.py} +0 -0
  261. /sglang/srt/{harmony_parser.py → parser/harmony_parser.py} +0 -0
  262. /sglang/srt/{jinja_template_utils.py → parser/jinja_template_utils.py} +0 -0
  263. {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/WHEEL +0 -0
  264. {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/licenses/LICENSE +0 -0
  265. {sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/top_level.txt +0 -0
{sglang-0.5.2rc1.dist-info → sglang-0.5.3rc0.dist-info}/METADATA RENAMED
@@ -1,7 +1,7 @@
1
1
  Metadata-Version: 2.4
2
2
  Name: sglang
3
- Version: 0.5.2rc1
4
- Summary: SGLang is yet another fast serving framework for large language models and vision language models.
3
+ Version: 0.5.3rc0
4
+ Summary: SGLang is a fast serving framework for large language models and vision language models.
5
5
  License: Apache License
6
6
  Version 2.0, January 2004
7
7
  http://www.apache.org/licenses/
@@ -217,84 +217,54 @@ Requires-Dist: tqdm
217
217
  Requires-Dist: numpy
218
218
  Requires-Dist: IPython
219
219
  Requires-Dist: setproctitle
220
- Provides-Extra: runtime-common
221
- Requires-Dist: blobfile==3.0.0; extra == "runtime-common"
222
- Requires-Dist: build; extra == "runtime-common"
223
- Requires-Dist: compressed-tensors; extra == "runtime-common"
224
- Requires-Dist: datasets; extra == "runtime-common"
225
- Requires-Dist: einops; extra == "runtime-common"
226
- Requires-Dist: fastapi; extra == "runtime-common"
227
- Requires-Dist: hf_transfer; extra == "runtime-common"
228
- Requires-Dist: huggingface_hub; extra == "runtime-common"
229
- Requires-Dist: interegular; extra == "runtime-common"
230
- Requires-Dist: llguidance<0.8.0,>=0.7.11; extra == "runtime-common"
231
- Requires-Dist: modelscope; extra == "runtime-common"
232
- Requires-Dist: msgspec; extra == "runtime-common"
233
- Requires-Dist: ninja; extra == "runtime-common"
234
- Requires-Dist: openai==1.99.1; extra == "runtime-common"
235
- Requires-Dist: openai-harmony==0.0.4; extra == "runtime-common"
236
- Requires-Dist: orjson; extra == "runtime-common"
237
- Requires-Dist: outlines==0.1.11; extra == "runtime-common"
238
- Requires-Dist: packaging; extra == "runtime-common"
239
- Requires-Dist: partial_json_parser; extra == "runtime-common"
240
- Requires-Dist: pillow; extra == "runtime-common"
241
- Requires-Dist: prometheus-client>=0.20.0; extra == "runtime-common"
242
- Requires-Dist: psutil; extra == "runtime-common"
243
- Requires-Dist: pybase64; extra == "runtime-common"
244
- Requires-Dist: pydantic; extra == "runtime-common"
245
- Requires-Dist: pynvml; extra == "runtime-common"
246
- Requires-Dist: python-multipart; extra == "runtime-common"
247
- Requires-Dist: pyzmq>=25.1.2; extra == "runtime-common"
248
- Requires-Dist: sentencepiece; extra == "runtime-common"
249
- Requires-Dist: soundfile==0.13.1; extra == "runtime-common"
250
- Requires-Dist: scipy; extra == "runtime-common"
251
- Requires-Dist: timm==1.0.16; extra == "runtime-common"
252
- Requires-Dist: tiktoken; extra == "runtime-common"
253
- Requires-Dist: torchao==0.9.0; extra == "runtime-common"
254
- Requires-Dist: transformers==4.56.0; extra == "runtime-common"
255
- Requires-Dist: uvicorn; extra == "runtime-common"
256
- Requires-Dist: uvloop; extra == "runtime-common"
257
- Requires-Dist: xgrammar==0.1.23; extra == "runtime-common"
258
- Provides-Extra: srt
259
- Requires-Dist: sglang[runtime_common]; extra == "srt"
260
- Requires-Dist: sgl-kernel==0.3.8; extra == "srt"
261
- Requires-Dist: torch==2.8.0; extra == "srt"
262
- Requires-Dist: torchaudio==2.8.0; extra == "srt"
263
- Requires-Dist: torchvision; extra == "srt"
264
- Requires-Dist: cuda-python; extra == "srt"
265
- Requires-Dist: flashinfer_python==0.3.0; extra == "srt"
266
- Provides-Extra: blackwell
267
- Requires-Dist: sglang[runtime_common]; extra == "blackwell"
268
- Requires-Dist: sgl-kernel; extra == "blackwell"
269
- Requires-Dist: torch==2.8.0; extra == "blackwell"
270
- Requires-Dist: torchaudio==2.8.0; extra == "blackwell"
271
- Requires-Dist: torchvision; extra == "blackwell"
272
- Requires-Dist: cuda-python; extra == "blackwell"
273
- Requires-Dist: flashinfer_python==0.3.0; extra == "blackwell"
274
- Provides-Extra: srt-hip
275
- Requires-Dist: sglang[runtime_common]; extra == "srt-hip"
276
- Requires-Dist: torch; extra == "srt-hip"
277
- Requires-Dist: petit_kernel==0.0.2; extra == "srt-hip"
278
- Requires-Dist: wave-lang==1.0.1; extra == "srt-hip"
279
- Provides-Extra: srt-cpu
280
- Requires-Dist: sglang[runtime_common]; extra == "srt-cpu"
281
- Provides-Extra: srt-npu
282
- Requires-Dist: sglang[runtime_common]; extra == "srt-npu"
283
- Provides-Extra: srt-xpu
284
- Requires-Dist: sglang[runtime_common]; extra == "srt-xpu"
285
- Provides-Extra: srt-hpu
286
- Requires-Dist: sglang[runtime_common]; extra == "srt-hpu"
287
- Provides-Extra: openai
288
- Requires-Dist: openai==1.99.1; extra == "openai"
289
- Requires-Dist: tiktoken; extra == "openai"
290
- Provides-Extra: anthropic
291
- Requires-Dist: anthropic>=0.20.0; extra == "anthropic"
292
- Provides-Extra: litellm
293
- Requires-Dist: litellm>=1.0.0; extra == "litellm"
294
- Provides-Extra: torch-memory-saver
295
- Requires-Dist: torch_memory_saver==0.0.8; extra == "torch-memory-saver"
296
- Provides-Extra: decord
297
- Requires-Dist: decord; extra == "decord"
220
+ Requires-Dist: blobfile==3.0.0
221
+ Requires-Dist: build
222
+ Requires-Dist: compressed-tensors
223
+ Requires-Dist: datasets
224
+ Requires-Dist: einops
225
+ Requires-Dist: fastapi
226
+ Requires-Dist: hf_transfer
227
+ Requires-Dist: huggingface_hub
228
+ Requires-Dist: interegular
229
+ Requires-Dist: llguidance<0.8.0,>=0.7.11
230
+ Requires-Dist: modelscope
231
+ Requires-Dist: msgspec
232
+ Requires-Dist: ninja
233
+ Requires-Dist: openai==1.99.1
234
+ Requires-Dist: openai-harmony==0.0.4
235
+ Requires-Dist: orjson
236
+ Requires-Dist: outlines==0.1.11
237
+ Requires-Dist: packaging
238
+ Requires-Dist: partial_json_parser
239
+ Requires-Dist: pillow
240
+ Requires-Dist: prometheus-client>=0.20.0
241
+ Requires-Dist: psutil
242
+ Requires-Dist: pybase64
243
+ Requires-Dist: pydantic
244
+ Requires-Dist: pynvml
245
+ Requires-Dist: python-multipart
246
+ Requires-Dist: pyzmq>=25.1.2
247
+ Requires-Dist: scipy
248
+ Requires-Dist: sentencepiece
249
+ Requires-Dist: soundfile==0.13.1
250
+ Requires-Dist: timm==1.0.16
251
+ Requires-Dist: tiktoken
252
+ Requires-Dist: torchao==0.9.0
253
+ Requires-Dist: transformers==4.56.1
254
+ Requires-Dist: uvicorn
255
+ Requires-Dist: uvloop
256
+ Requires-Dist: xgrammar==0.1.24
257
+ Requires-Dist: sgl-kernel==0.3.9.post2
258
+ Requires-Dist: torch==2.8.0
259
+ Requires-Dist: torchaudio==2.8.0
260
+ Requires-Dist: torchvision
261
+ Requires-Dist: cuda-python
262
+ Requires-Dist: flashinfer_python==0.3.1
263
+ Requires-Dist: openai==1.99.1
264
+ Requires-Dist: tiktoken
265
+ Requires-Dist: anthropic>=0.20.0
266
+ Requires-Dist: torch_memory_saver==0.0.8
267
+ Requires-Dist: decord
298
268
  Provides-Extra: test
299
269
  Requires-Dist: accelerate; extra == "test"
300
270
  Requires-Dist: expecttest; extra == "test"
@@ -305,52 +275,18 @@ Requires-Dist: peft; extra == "test"
305
275
  Requires-Dist: sentence_transformers; extra == "test"
306
276
  Requires-Dist: pytest; extra == "test"
307
277
  Requires-Dist: tabulate; extra == "test"
278
+ Provides-Extra: tracing
279
+ Requires-Dist: opentelemetry-sdk; extra == "tracing"
280
+ Requires-Dist: opentelemetry-api; extra == "tracing"
281
+ Requires-Dist: opentelemetry-exporter-otlp; extra == "tracing"
282
+ Requires-Dist: opentelemetry-exporter-otlp-proto-grpc; extra == "tracing"
308
283
  Provides-Extra: all
309
- Requires-Dist: sglang[srt]; extra == "all"
310
- Requires-Dist: sglang[openai]; extra == "all"
311
- Requires-Dist: sglang[anthropic]; extra == "all"
312
- Requires-Dist: sglang[torch_memory_saver]; extra == "all"
313
- Requires-Dist: sglang[decord]; extra == "all"
314
- Provides-Extra: all-hip
315
- Requires-Dist: sglang[srt_hip]; extra == "all-hip"
316
- Requires-Dist: sglang[openai]; extra == "all-hip"
317
- Requires-Dist: sglang[anthropic]; extra == "all-hip"
318
- Requires-Dist: sglang[decord]; extra == "all-hip"
319
- Provides-Extra: all-xpu
320
- Requires-Dist: sglang[srt_xpu]; extra == "all-xpu"
321
- Requires-Dist: sglang[openai]; extra == "all-xpu"
322
- Requires-Dist: sglang[anthropic]; extra == "all-xpu"
323
- Requires-Dist: sglang[decord]; extra == "all-xpu"
324
- Provides-Extra: all-hpu
325
- Requires-Dist: sglang[srt_hpu]; extra == "all-hpu"
326
- Requires-Dist: sglang[openai]; extra == "all-hpu"
327
- Requires-Dist: sglang[anthropic]; extra == "all-hpu"
328
- Requires-Dist: sglang[decord]; extra == "all-hpu"
329
- Provides-Extra: all-cpu
330
- Requires-Dist: sglang[srt_cpu]; extra == "all-cpu"
331
- Requires-Dist: sglang[openai]; extra == "all-cpu"
332
- Requires-Dist: sglang[anthropic]; extra == "all-cpu"
333
- Requires-Dist: sglang[decord]; extra == "all-cpu"
334
- Provides-Extra: all-npu
335
- Requires-Dist: sglang[srt_npu]; extra == "all-npu"
336
- Requires-Dist: sglang[openai]; extra == "all-npu"
337
- Requires-Dist: sglang[anthropic]; extra == "all-npu"
338
- Requires-Dist: sglang[decord]; extra == "all-npu"
284
+ Requires-Dist: sglang[test]; extra == "all"
285
+ Provides-Extra: blackwell
286
+ Requires-Dist: nvidia-cutlass-dsl==4.1.0; extra == "blackwell"
287
+ Requires-Dist: sglang[test]; extra == "blackwell"
339
288
  Provides-Extra: dev
340
- Requires-Dist: sglang[all]; extra == "dev"
341
289
  Requires-Dist: sglang[test]; extra == "dev"
342
- Provides-Extra: dev-hip
343
- Requires-Dist: sglang[all_hip]; extra == "dev-hip"
344
- Requires-Dist: sglang[test]; extra == "dev-hip"
345
- Provides-Extra: dev-xpu
346
- Requires-Dist: sglang[all_xpu]; extra == "dev-xpu"
347
- Requires-Dist: sglang[test]; extra == "dev-xpu"
348
- Provides-Extra: dev-hpu
349
- Requires-Dist: sglang[all_hpu]; extra == "dev-hpu"
350
- Requires-Dist: sglang[test]; extra == "dev-hpu"
351
- Provides-Extra: dev-cpu
352
- Requires-Dist: sglang[all_cpu]; extra == "dev-cpu"
353
- Requires-Dist: sglang[test]; extra == "dev-cpu"
354
290
  Dynamic: license-file
355
291
 
356
292
  <div align="center" id="sglangtop">