sglang 0.5.0rc1__py3-none-any.whl → 0.5.1__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 (203) hide show
  1. sglang/bench_one_batch.py +0 -7
  2. sglang/bench_one_batch_server.py +7 -2
  3. sglang/bench_serving.py +3 -3
  4. sglang/eval/llama3_eval.py +0 -1
  5. sglang/srt/configs/model_config.py +25 -9
  6. sglang/srt/configs/update_config.py +40 -5
  7. sglang/srt/constrained/xgrammar_backend.py +23 -11
  8. sglang/srt/conversation.py +2 -15
  9. sglang/srt/disaggregation/ascend/conn.py +1 -3
  10. sglang/srt/disaggregation/base/conn.py +1 -0
  11. sglang/srt/disaggregation/decode.py +1 -2
  12. sglang/srt/disaggregation/launch_lb.py +7 -1
  13. sglang/srt/disaggregation/mini_lb.py +11 -5
  14. sglang/srt/disaggregation/mooncake/conn.py +141 -47
  15. sglang/srt/disaggregation/prefill.py +261 -5
  16. sglang/srt/disaggregation/utils.py +2 -1
  17. sglang/srt/distributed/device_communicators/custom_all_reduce.py +1 -1
  18. sglang/srt/distributed/device_communicators/pynccl.py +68 -18
  19. sglang/srt/distributed/device_communicators/pynccl_wrapper.py +52 -0
  20. sglang/srt/distributed/naive_distributed.py +112 -0
  21. sglang/srt/distributed/parallel_state.py +90 -4
  22. sglang/srt/entrypoints/context.py +20 -1
  23. sglang/srt/entrypoints/engine.py +29 -4
  24. sglang/srt/entrypoints/http_server.py +76 -0
  25. sglang/srt/entrypoints/openai/protocol.py +4 -2
  26. sglang/srt/entrypoints/openai/serving_chat.py +23 -6
  27. sglang/srt/entrypoints/openai/serving_completions.py +10 -1
  28. sglang/srt/entrypoints/openai/serving_responses.py +2 -2
  29. sglang/srt/eplb/expert_distribution.py +2 -3
  30. sglang/srt/function_call/deepseekv3_detector.py +1 -1
  31. sglang/srt/hf_transformers_utils.py +24 -0
  32. sglang/srt/host_shared_memory.py +83 -0
  33. sglang/srt/layers/attention/ascend_backend.py +132 -22
  34. sglang/srt/layers/attention/flashattention_backend.py +24 -17
  35. sglang/srt/layers/attention/flashinfer_backend.py +14 -3
  36. sglang/srt/layers/attention/flashinfer_mla_backend.py +227 -76
  37. sglang/srt/layers/attention/triton_backend.py +109 -73
  38. sglang/srt/layers/attention/triton_ops/decode_attention.py +33 -2
  39. sglang/srt/layers/attention/triton_ops/extend_attention.py +32 -2
  40. sglang/srt/layers/attention/trtllm_mha_backend.py +398 -36
  41. sglang/srt/layers/attention/trtllm_mla_backend.py +49 -19
  42. sglang/srt/layers/attention/utils.py +94 -15
  43. sglang/srt/layers/attention/vision.py +40 -13
  44. sglang/srt/layers/attention/vision_utils.py +65 -0
  45. sglang/srt/layers/communicator.py +58 -10
  46. sglang/srt/layers/dp_attention.py +137 -27
  47. sglang/srt/layers/elementwise.py +94 -0
  48. sglang/srt/layers/flashinfer_comm_fusion.py +29 -1
  49. sglang/srt/layers/layernorm.py +8 -1
  50. sglang/srt/layers/linear.py +24 -0
  51. sglang/srt/layers/logits_processor.py +16 -18
  52. sglang/srt/layers/moe/__init__.py +31 -0
  53. sglang/srt/layers/moe/ep_moe/layer.py +37 -33
  54. sglang/srt/layers/moe/fused_moe_native.py +14 -25
  55. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=129,N=352,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  56. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_2_0/E=161,N=192,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  57. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_0/E=16,N=1024,device_name=NVIDIA_B200.json +146 -0
  58. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=128,N=384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  59. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=128,N=768,device_name=NVIDIA_H20.json +146 -0
  60. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=160,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  61. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=257,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  62. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=257,N=128,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  63. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=257,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  64. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=257,N=256,device_name=NVIDIA_H20-3e,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  65. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=257,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  66. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=128,N=384,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  67. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=129,N=352,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8.json +146 -0
  68. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=129,N=704,device_name=NVIDIA_B200,dtype=fp8_w8a8.json +146 -0
  69. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=161,N=384,device_name=NVIDIA_RTX_PRO_6000_Blackwell_Max-Q_Workstation_Edition,dtype=fp8_w8a8.json +146 -0
  70. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +69 -76
  71. sglang/srt/layers/moe/fused_moe_triton/layer.py +66 -123
  72. sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py +20 -18
  73. sglang/srt/layers/moe/moe_runner/__init__.py +3 -0
  74. sglang/srt/layers/moe/moe_runner/base.py +13 -0
  75. sglang/srt/layers/moe/rocm_moe_utils.py +141 -0
  76. sglang/srt/layers/moe/router.py +15 -9
  77. sglang/srt/layers/moe/token_dispatcher/__init__.py +6 -0
  78. sglang/srt/layers/moe/token_dispatcher/base_dispatcher.py +55 -14
  79. sglang/srt/layers/moe/token_dispatcher/deepep.py +11 -21
  80. sglang/srt/layers/moe/token_dispatcher/standard.py +1 -1
  81. sglang/srt/layers/moe/topk.py +167 -83
  82. sglang/srt/layers/moe/utils.py +159 -18
  83. sglang/srt/layers/multimodal.py +156 -40
  84. sglang/srt/layers/quantization/__init__.py +18 -46
  85. sglang/srt/layers/quantization/awq.py +22 -23
  86. sglang/srt/layers/quantization/base_config.py +2 -6
  87. sglang/srt/layers/quantization/blockwise_int8.py +4 -12
  88. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +72 -29
  89. sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py +2 -1
  90. sglang/srt/layers/quantization/fp8.py +127 -119
  91. sglang/srt/layers/quantization/fp8_kernel.py +195 -24
  92. sglang/srt/layers/quantization/fp8_utils.py +34 -9
  93. sglang/srt/layers/quantization/fpgemm_fp8.py +203 -0
  94. sglang/srt/layers/quantization/gptq.py +17 -21
  95. sglang/srt/layers/quantization/marlin_utils.py +26 -8
  96. sglang/srt/layers/quantization/marlin_utils_fp8.py +352 -0
  97. sglang/srt/layers/quantization/modelopt_quant.py +217 -98
  98. sglang/srt/layers/quantization/moe_wna16.py +10 -15
  99. sglang/srt/layers/quantization/mxfp4.py +222 -39
  100. sglang/srt/layers/quantization/quark/quark.py +390 -0
  101. sglang/srt/layers/quantization/quark/quark_moe.py +197 -0
  102. sglang/srt/layers/quantization/unquant.py +34 -70
  103. sglang/srt/layers/quantization/utils.py +77 -2
  104. sglang/srt/layers/quantization/w4afp8.py +7 -8
  105. sglang/srt/layers/quantization/w8a8_fp8.py +5 -13
  106. sglang/srt/layers/quantization/w8a8_int8.py +5 -13
  107. sglang/srt/layers/radix_attention.py +6 -0
  108. sglang/srt/layers/rotary_embedding.py +1 -0
  109. sglang/srt/layers/sampler.py +5 -2
  110. sglang/srt/lora/layers.py +6 -2
  111. sglang/srt/lora/lora_manager.py +21 -22
  112. sglang/srt/lora/lora_registry.py +3 -3
  113. sglang/srt/lora/mem_pool.py +26 -24
  114. sglang/srt/lora/utils.py +10 -12
  115. sglang/srt/managers/cache_controller.py +80 -19
  116. sglang/srt/managers/detokenizer_manager.py +10 -2
  117. sglang/srt/managers/io_struct.py +23 -0
  118. sglang/srt/managers/mm_utils.py +1 -1
  119. sglang/srt/managers/schedule_batch.py +22 -48
  120. sglang/srt/managers/scheduler.py +28 -20
  121. sglang/srt/managers/session_controller.py +1 -1
  122. sglang/srt/managers/template_manager.py +7 -5
  123. sglang/srt/managers/tokenizer_manager.py +88 -39
  124. sglang/srt/managers/tp_worker.py +1 -0
  125. sglang/srt/managers/utils.py +59 -1
  126. sglang/srt/mem_cache/allocator.py +10 -157
  127. sglang/srt/mem_cache/allocator_ascend.py +147 -0
  128. sglang/srt/mem_cache/chunk_cache.py +1 -1
  129. sglang/srt/mem_cache/hicache_storage.py +14 -4
  130. sglang/srt/mem_cache/memory_pool.py +3 -3
  131. sglang/srt/mem_cache/memory_pool_host.py +35 -2
  132. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +56 -12
  133. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +8 -4
  134. sglang/srt/mem_cache/storage/nixl/hicache_nixl.py +153 -59
  135. sglang/srt/mem_cache/storage/nixl/nixl_utils.py +19 -53
  136. sglang/srt/mem_cache/storage/nixl/test_hicache_nixl_storage.py +46 -7
  137. sglang/srt/model_executor/cuda_graph_runner.py +33 -33
  138. sglang/srt/model_executor/forward_batch_info.py +11 -10
  139. sglang/srt/model_executor/model_runner.py +93 -78
  140. sglang/srt/model_executor/npu_graph_runner.py +94 -0
  141. sglang/srt/model_loader/loader.py +24 -6
  142. sglang/srt/models/dbrx.py +12 -6
  143. sglang/srt/models/deepseek.py +2 -1
  144. sglang/srt/models/deepseek_nextn.py +5 -2
  145. sglang/srt/models/deepseek_v2.py +226 -223
  146. sglang/srt/models/ernie4.py +2 -2
  147. sglang/srt/models/glm4_moe.py +27 -65
  148. sglang/srt/models/glm4_moe_nextn.py +2 -1
  149. sglang/srt/models/glm4v.py +52 -1
  150. sglang/srt/models/glm4v_moe.py +8 -11
  151. sglang/srt/models/gpt_oss.py +41 -76
  152. sglang/srt/models/granitemoe.py +0 -1
  153. sglang/srt/models/grok.py +376 -48
  154. sglang/srt/models/interns1.py +12 -47
  155. sglang/srt/models/internvl.py +6 -51
  156. sglang/srt/models/llama.py +10 -2
  157. sglang/srt/models/llama4.py +18 -7
  158. sglang/srt/models/minicpm3.py +0 -1
  159. sglang/srt/models/mixtral.py +0 -2
  160. sglang/srt/models/nemotron_nas.py +435 -0
  161. sglang/srt/models/olmoe.py +0 -1
  162. sglang/srt/models/phi4mm.py +3 -21
  163. sglang/srt/models/qwen2.py +2 -2
  164. sglang/srt/models/qwen2_5_vl.py +2 -0
  165. sglang/srt/models/qwen2_moe.py +23 -23
  166. sglang/srt/models/qwen3.py +2 -2
  167. sglang/srt/models/qwen3_classification.py +84 -0
  168. sglang/srt/models/qwen3_moe.py +27 -43
  169. sglang/srt/models/step3_vl.py +8 -3
  170. sglang/srt/models/xverse_moe.py +11 -5
  171. sglang/srt/multimodal/processors/base_processor.py +3 -3
  172. sglang/srt/multimodal/processors/internvl.py +7 -2
  173. sglang/srt/multimodal/processors/llava.py +11 -7
  174. sglang/srt/offloader.py +433 -0
  175. sglang/srt/operations.py +22 -2
  176. sglang/srt/reasoning_parser.py +4 -3
  177. sglang/srt/sampling/sampling_batch_info.py +7 -4
  178. sglang/srt/server_args.py +264 -105
  179. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +8 -21
  180. sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +7 -21
  181. sglang/srt/speculative/eagle_utils.py +36 -13
  182. sglang/srt/speculative/eagle_worker.py +56 -3
  183. sglang/srt/tokenizer/tiktoken_tokenizer.py +161 -0
  184. sglang/srt/two_batch_overlap.py +20 -19
  185. sglang/srt/utils.py +68 -70
  186. sglang/test/runners.py +8 -5
  187. sglang/test/test_block_fp8.py +5 -6
  188. sglang/test/test_block_fp8_ep.py +13 -19
  189. sglang/test/test_cutlass_moe.py +4 -6
  190. sglang/test/test_cutlass_w4a8_moe.py +4 -3
  191. sglang/test/test_fp4_moe.py +4 -3
  192. sglang/test/test_marlin_moe.py +1 -1
  193. sglang/test/test_marlin_utils.py +1 -1
  194. sglang/test/test_utils.py +7 -0
  195. sglang/utils.py +0 -1
  196. sglang/version.py +1 -1
  197. {sglang-0.5.0rc1.dist-info → sglang-0.5.1.dist-info}/METADATA +11 -11
  198. {sglang-0.5.0rc1.dist-info → sglang-0.5.1.dist-info}/RECORD +201 -171
  199. sglang/srt/layers/quantization/fp4.py +0 -557
  200. sglang/srt/layers/quantization/scalar_type.py +0 -352
  201. {sglang-0.5.0rc1.dist-info → sglang-0.5.1.dist-info}/WHEEL +0 -0
  202. {sglang-0.5.0rc1.dist-info → sglang-0.5.1.dist-info}/licenses/LICENSE +0 -0
  203. {sglang-0.5.0rc1.dist-info → sglang-0.5.1.dist-info}/top_level.txt +0 -0
sglang/srt/models/deepseek_nextn.py CHANGED
@@ -20,8 +20,9 @@ import torch
20
20
  from torch import nn
21
21
  from transformers import PretrainedConfig
22
22
 
23
- from sglang.srt.distributed import get_tensor_model_parallel_world_size
23
+ from sglang.srt.distributed import get_pp_group, get_tensor_model_parallel_world_size
24
24
  from sglang.srt.eplb.expert_distribution import get_global_expert_distribution_recorder
25
+ from sglang.srt.layers.dp_attention import is_dp_attention_enabled
25
26
  from sglang.srt.layers.layernorm import RMSNorm
26
27
  from sglang.srt.layers.logits_processor import LogitsProcessor
27
28
  from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -56,7 +57,7 @@ class DeepseekModelNextN(nn.Module):
56
57
  self.embed_tokens = VocabParallelEmbedding(
57
58
  config.vocab_size,
58
59
  config.hidden_size,
59
- enable_tp=not global_server_args_dict["enable_dp_attention"],
60
+ enable_tp=not is_dp_attention_enabled(),
60
61
  prefix=add_prefix("embed_tokens", prefix),
61
62
  )
62
63
 
@@ -134,6 +135,8 @@ class DeepseekV3ForCausalLMNextN(DeepseekV3ForCausalLM):
134
135
  self.config = config
135
136
  self.tp_size = get_tensor_model_parallel_world_size()
136
137
  self.quant_config = quant_config
138
+ # if not set, model load will be broken in DeepseekV3ForCausalLM load_weights()
139
+ self.pp_group = get_pp_group()
137
140
  self.determine_num_fused_shared_experts("DeepseekV3ForCausalLMNextN")
138
141
 
139
142
  self.model = DeepseekModelNextN(