sglang 0.4.10.post2__py3-none-any.whl → 0.5.0rc1__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 (175) hide show
  1. sglang/__init__.py +8 -3
  2. sglang/bench_one_batch.py +119 -17
  3. sglang/lang/chat_template.py +18 -0
  4. sglang/srt/bench_utils.py +137 -0
  5. sglang/srt/configs/model_config.py +42 -7
  6. sglang/srt/conversation.py +9 -5
  7. sglang/srt/disaggregation/base/conn.py +5 -2
  8. sglang/srt/disaggregation/decode.py +14 -4
  9. sglang/srt/disaggregation/decode_schedule_batch_mixin.py +3 -0
  10. sglang/srt/disaggregation/mooncake/conn.py +286 -160
  11. sglang/srt/disaggregation/mooncake/transfer_engine.py +29 -0
  12. sglang/srt/disaggregation/prefill.py +2 -0
  13. sglang/srt/distributed/parallel_state.py +15 -11
  14. sglang/srt/entrypoints/context.py +227 -0
  15. sglang/srt/entrypoints/engine.py +15 -9
  16. sglang/srt/entrypoints/harmony_utils.py +372 -0
  17. sglang/srt/entrypoints/http_server.py +74 -4
  18. sglang/srt/entrypoints/openai/protocol.py +218 -1
  19. sglang/srt/entrypoints/openai/serving_chat.py +41 -11
  20. sglang/srt/entrypoints/openai/serving_responses.py +1273 -0
  21. sglang/srt/entrypoints/openai/tool_server.py +175 -0
  22. sglang/srt/entrypoints/tool.py +87 -0
  23. sglang/srt/eplb/expert_location.py +5 -1
  24. sglang/srt/function_call/ebnf_composer.py +1 -0
  25. sglang/srt/function_call/function_call_parser.py +2 -0
  26. sglang/srt/function_call/glm4_moe_detector.py +1 -1
  27. sglang/srt/function_call/gpt_oss_detector.py +331 -0
  28. sglang/srt/function_call/kimik2_detector.py +3 -3
  29. sglang/srt/function_call/qwen3_coder_detector.py +219 -9
  30. sglang/srt/hf_transformers_utils.py +30 -3
  31. sglang/srt/jinja_template_utils.py +14 -1
  32. sglang/srt/layers/attention/aiter_backend.py +375 -115
  33. sglang/srt/layers/attention/ascend_backend.py +3 -0
  34. sglang/srt/layers/attention/dual_chunk_flashattention_backend.py +1700 -0
  35. sglang/srt/layers/attention/flashattention_backend.py +18 -0
  36. sglang/srt/layers/attention/flashinfer_backend.py +52 -13
  37. sglang/srt/layers/attention/hybrid_attn_backend.py +1 -1
  38. sglang/srt/layers/attention/triton_backend.py +85 -14
  39. sglang/srt/layers/attention/triton_ops/decode_attention.py +17 -0
  40. sglang/srt/layers/attention/triton_ops/extend_attention.py +143 -98
  41. sglang/srt/layers/attention/trtllm_mha_backend.py +332 -0
  42. sglang/srt/layers/attention/trtllm_mla_backend.py +119 -22
  43. sglang/srt/layers/attention/vision.py +22 -6
  44. sglang/srt/layers/attention/wave_backend.py +627 -0
  45. sglang/srt/layers/attention/wave_ops/decode_attention.py +186 -0
  46. sglang/srt/layers/attention/wave_ops/extend_attention.py +149 -0
  47. sglang/srt/layers/attention/wave_ops/prefill_attention.py +79 -0
  48. sglang/srt/layers/communicator.py +29 -14
  49. sglang/srt/layers/dp_attention.py +12 -0
  50. sglang/srt/layers/flashinfer_comm_fusion.py +4 -4
  51. sglang/srt/layers/linear.py +3 -7
  52. sglang/srt/layers/moe/cutlass_moe.py +12 -3
  53. sglang/srt/layers/moe/cutlass_w4a8_moe.py +4 -5
  54. sglang/srt/layers/moe/ep_moe/kernels.py +43 -0
  55. sglang/srt/layers/moe/ep_moe/layer.py +135 -73
  56. 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 +146 -0
  57. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=384,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  58. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +101 -12
  59. sglang/srt/layers/moe/fused_moe_triton/layer.py +412 -33
  60. sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py +188 -3
  61. sglang/srt/layers/moe/token_dispatcher/deepep.py +61 -24
  62. sglang/srt/layers/moe/topk.py +16 -4
  63. sglang/srt/layers/moe/utils.py +16 -0
  64. sglang/srt/layers/quantization/__init__.py +27 -3
  65. sglang/srt/layers/quantization/fp4.py +557 -0
  66. sglang/srt/layers/quantization/fp8.py +3 -6
  67. sglang/srt/layers/quantization/fp8_kernel.py +277 -0
  68. sglang/srt/layers/quantization/fp8_utils.py +51 -10
  69. sglang/srt/layers/quantization/modelopt_quant.py +258 -68
  70. sglang/srt/layers/quantization/mxfp4.py +654 -0
  71. sglang/srt/layers/quantization/mxfp4_tensor.py +133 -0
  72. sglang/srt/layers/quantization/quark/schemes/__init__.py +6 -0
  73. sglang/srt/layers/quantization/quark/schemes/quark_scheme.py +55 -0
  74. sglang/srt/layers/quantization/quark/schemes/quark_w4a4_mxfp4.py +118 -0
  75. sglang/srt/layers/quantization/quark/utils.py +107 -0
  76. sglang/srt/layers/quantization/unquant.py +60 -6
  77. sglang/srt/layers/quantization/w4afp8.py +21 -12
  78. sglang/srt/layers/quantization/w8a8_int8.py +48 -34
  79. sglang/srt/layers/rotary_embedding.py +506 -3
  80. sglang/srt/layers/utils.py +9 -0
  81. sglang/srt/layers/vocab_parallel_embedding.py +8 -3
  82. sglang/srt/lora/backend/base_backend.py +3 -23
  83. sglang/srt/lora/layers.py +60 -114
  84. sglang/srt/lora/lora.py +17 -62
  85. sglang/srt/lora/lora_manager.py +82 -62
  86. sglang/srt/lora/lora_registry.py +23 -11
  87. sglang/srt/lora/mem_pool.py +63 -68
  88. sglang/srt/lora/triton_ops/qkv_lora_b.py +1 -1
  89. sglang/srt/lora/utils.py +25 -58
  90. sglang/srt/managers/cache_controller.py +75 -58
  91. sglang/srt/managers/detokenizer_manager.py +1 -1
  92. sglang/srt/managers/io_struct.py +20 -8
  93. sglang/srt/managers/mm_utils.py +6 -13
  94. sglang/srt/managers/multimodal_processor.py +1 -1
  95. sglang/srt/managers/schedule_batch.py +61 -25
  96. sglang/srt/managers/schedule_policy.py +6 -6
  97. sglang/srt/managers/scheduler.py +41 -19
  98. sglang/srt/managers/scheduler_output_processor_mixin.py +1 -2
  99. sglang/srt/managers/scheduler_profiler_mixin.py +28 -8
  100. sglang/srt/managers/scheduler_recv_skipper.py +37 -0
  101. sglang/srt/managers/scheduler_update_weights_mixin.py +6 -0
  102. sglang/srt/managers/template_manager.py +35 -1
  103. sglang/srt/managers/tokenizer_manager.py +47 -30
  104. sglang/srt/managers/tp_worker.py +3 -0
  105. sglang/srt/managers/tp_worker_overlap_thread.py +3 -0
  106. sglang/srt/mem_cache/allocator.py +61 -87
  107. sglang/srt/mem_cache/hicache_storage.py +1 -1
  108. sglang/srt/mem_cache/hiradix_cache.py +80 -22
  109. sglang/srt/mem_cache/lora_radix_cache.py +421 -0
  110. sglang/srt/mem_cache/memory_pool_host.py +34 -36
  111. sglang/srt/mem_cache/multimodal_cache.py +33 -13
  112. sglang/srt/mem_cache/radix_cache.py +2 -5
  113. sglang/srt/mem_cache/storage/hf3fs/client_hf3fs.py +2 -2
  114. sglang/srt/mem_cache/storage/hf3fs/mini_3fs_metadata_server.py +443 -0
  115. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +139 -67
  116. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +6 -9
  117. sglang/srt/model_executor/cuda_graph_runner.py +29 -9
  118. sglang/srt/model_executor/forward_batch_info.py +61 -19
  119. sglang/srt/model_executor/model_runner.py +148 -37
  120. sglang/srt/model_loader/loader.py +18 -6
  121. sglang/srt/model_loader/weight_utils.py +10 -0
  122. sglang/srt/models/bailing_moe.py +425 -0
  123. sglang/srt/models/deepseek_v2.py +137 -59
  124. sglang/srt/models/ernie4.py +426 -0
  125. sglang/srt/models/ernie4_eagle.py +203 -0
  126. sglang/srt/models/gemma2.py +0 -34
  127. sglang/srt/models/gemma3n_mm.py +38 -0
  128. sglang/srt/models/glm4.py +6 -0
  129. sglang/srt/models/glm4_moe.py +28 -16
  130. sglang/srt/models/glm4v.py +589 -0
  131. sglang/srt/models/glm4v_moe.py +400 -0
  132. sglang/srt/models/gpt_oss.py +1251 -0
  133. sglang/srt/models/granite.py +0 -25
  134. sglang/srt/models/llama.py +0 -25
  135. sglang/srt/models/llama4.py +1 -1
  136. sglang/srt/models/qwen2.py +6 -0
  137. sglang/srt/models/qwen2_5_vl.py +7 -3
  138. sglang/srt/models/qwen2_audio.py +10 -9
  139. sglang/srt/models/qwen2_moe.py +6 -0
  140. sglang/srt/models/qwen3.py +0 -24
  141. sglang/srt/models/qwen3_moe.py +32 -6
  142. sglang/srt/models/registry.py +1 -1
  143. sglang/srt/models/step3_vl.py +9 -0
  144. sglang/srt/models/torch_native_llama.py +0 -24
  145. sglang/srt/models/transformers.py +2 -5
  146. sglang/srt/multimodal/processors/base_processor.py +23 -13
  147. sglang/srt/multimodal/processors/glm4v.py +132 -0
  148. sglang/srt/multimodal/processors/qwen_audio.py +4 -2
  149. sglang/srt/multimodal/processors/step3_vl.py +3 -1
  150. sglang/srt/reasoning_parser.py +332 -37
  151. sglang/srt/server_args.py +186 -75
  152. sglang/srt/speculative/eagle_worker.py +16 -0
  153. sglang/srt/two_batch_overlap.py +169 -9
  154. sglang/srt/utils.py +41 -5
  155. sglang/srt/weight_sync/tensor_bucket.py +106 -0
  156. sglang/test/attention/test_trtllm_mla_backend.py +186 -36
  157. sglang/test/doc_patch.py +59 -0
  158. sglang/test/few_shot_gsm8k.py +1 -1
  159. sglang/test/few_shot_gsm8k_engine.py +1 -1
  160. sglang/test/run_eval.py +4 -1
  161. sglang/test/runners.py +2 -2
  162. sglang/test/simple_eval_common.py +6 -0
  163. sglang/test/simple_eval_gpqa.py +2 -0
  164. sglang/test/test_fp4_moe.py +118 -36
  165. sglang/test/test_utils.py +1 -1
  166. sglang/utils.py +1 -1
  167. sglang/version.py +1 -1
  168. {sglang-0.4.10.post2.dist-info → sglang-0.5.0rc1.dist-info}/METADATA +36 -38
  169. {sglang-0.4.10.post2.dist-info → sglang-0.5.0rc1.dist-info}/RECORD +174 -141
  170. sglang/srt/lora/backend/flashinfer_backend.py +0 -131
  171. /sglang/{api.py → lang/api.py} +0 -0
  172. /sglang/{lang/backend → srt/layers/quantization/quark}/__init__.py +0 -0
  173. {sglang-0.4.10.post2.dist-info → sglang-0.5.0rc1.dist-info}/WHEEL +0 -0
  174. {sglang-0.4.10.post2.dist-info → sglang-0.5.0rc1.dist-info}/licenses/LICENSE +0 -0
  175. {sglang-0.4.10.post2.dist-info → sglang-0.5.0rc1.dist-info}/top_level.txt +0 -0
sglang/srt/layers/attention/ascend_backend.py CHANGED
@@ -75,6 +75,9 @@ class AscendAttnBackend(AttentionBackend):
75
75
  )
76
76
  self.forward_metadata.seq_lens_cpu_int = forward_batch.seq_lens_cpu.int()
77
77
 
78
+ def get_cuda_graph_seq_len_fill_value(self):
79
+ return 1
80
+
78
81
  def forward_extend(
79
82
  self,
80
83
  q,