sglang 0.5.4__py3-none-any.whl → 0.5.4.post2__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 (195) hide show
  1. sglang/bench_one_batch.py +149 -34
  2. sglang/bench_serving.py +73 -14
  3. sglang/compile_deep_gemm.py +13 -7
  4. sglang/launch_server.py +2 -0
  5. sglang/srt/batch_invariant_ops/__init__.py +2 -0
  6. sglang/srt/batch_invariant_ops/batch_invariant_ops.py +221 -4
  7. sglang/srt/checkpoint_engine/__init__.py +9 -0
  8. sglang/srt/checkpoint_engine/update.py +317 -0
  9. sglang/srt/compilation/backend.py +1 -1
  10. sglang/srt/configs/__init__.py +2 -0
  11. sglang/srt/configs/deepseek_ocr.py +542 -10
  12. sglang/srt/configs/deepseekvl2.py +95 -194
  13. sglang/srt/configs/kimi_linear.py +160 -0
  14. sglang/srt/configs/mamba_utils.py +66 -0
  15. sglang/srt/configs/model_config.py +30 -7
  16. sglang/srt/constants.py +7 -0
  17. sglang/srt/debug_utils/tensor_dump_forward_hook.py +149 -0
  18. sglang/srt/disaggregation/decode.py +34 -6
  19. sglang/srt/disaggregation/nixl/conn.py +2 -2
  20. sglang/srt/disaggregation/prefill.py +25 -3
  21. sglang/srt/distributed/device_communicators/custom_all_reduce.py +3 -1
  22. sglang/srt/distributed/parallel_state.py +9 -12
  23. sglang/srt/entrypoints/engine.py +31 -20
  24. sglang/srt/entrypoints/grpc_server.py +0 -1
  25. sglang/srt/entrypoints/http_server.py +94 -94
  26. sglang/srt/entrypoints/openai/protocol.py +7 -1
  27. sglang/srt/entrypoints/openai/serving_chat.py +42 -0
  28. sglang/srt/entrypoints/openai/serving_completions.py +10 -0
  29. sglang/srt/entrypoints/openai/serving_embedding.py +1 -0
  30. sglang/srt/environ.py +23 -2
  31. sglang/srt/eplb/expert_distribution.py +64 -1
  32. sglang/srt/eplb/expert_location.py +106 -36
  33. sglang/srt/function_call/function_call_parser.py +2 -0
  34. sglang/srt/function_call/minimax_m2.py +367 -0
  35. sglang/srt/grpc/compile_proto.py +3 -0
  36. sglang/srt/layers/activation.py +6 -0
  37. sglang/srt/layers/attention/ascend_backend.py +233 -5
  38. sglang/srt/layers/attention/attention_registry.py +3 -0
  39. sglang/srt/layers/attention/fla/chunk_delta_h.py +61 -32
  40. sglang/srt/layers/attention/fla/fused_recurrent.py +17 -4
  41. sglang/srt/layers/attention/fla/kda.py +1359 -0
  42. sglang/srt/layers/attention/fla/layernorm_gated.py +7 -1
  43. sglang/srt/layers/attention/flashattention_backend.py +19 -8
  44. sglang/srt/layers/attention/flashinfer_backend.py +10 -1
  45. sglang/srt/layers/attention/flashinfer_mla_backend.py +21 -11
  46. sglang/srt/layers/attention/flashmla_backend.py +1 -1
  47. sglang/srt/layers/attention/hybrid_linear_attn_backend.py +223 -0
  48. sglang/srt/layers/attention/mamba/mamba.py +20 -11
  49. sglang/srt/layers/attention/nsa/dequant_k_cache.py +138 -6
  50. sglang/srt/layers/attention/nsa/nsa_indexer.py +45 -22
  51. sglang/srt/layers/attention/nsa/quant_k_cache.py +44 -12
  52. sglang/srt/layers/attention/nsa/transform_index.py +1 -1
  53. sglang/srt/layers/attention/nsa_backend.py +157 -23
  54. sglang/srt/layers/attention/triton_backend.py +4 -1
  55. sglang/srt/layers/attention/trtllm_mha_backend.py +10 -4
  56. sglang/srt/layers/attention/trtllm_mla_backend.py +11 -15
  57. sglang/srt/layers/attention/utils.py +78 -0
  58. sglang/srt/layers/communicator.py +24 -1
  59. sglang/srt/layers/deep_gemm_wrapper/compile_utils.py +1 -1
  60. sglang/srt/layers/layernorm.py +35 -6
  61. sglang/srt/layers/logits_processor.py +9 -20
  62. sglang/srt/layers/moe/cutlass_w4a8_moe.py +138 -0
  63. sglang/srt/layers/moe/ep_moe/kernels.py +194 -0
  64. sglang/srt/layers/moe/ep_moe/layer.py +78 -289
  65. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=257,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  66. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=257,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128]_down.json +164 -0
  67. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +68 -22
  68. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_config.py +43 -3
  69. sglang/srt/layers/moe/fused_moe_triton/fused_moe_triton_kernels.py +106 -26
  70. sglang/srt/layers/moe/fused_moe_triton/layer.py +3 -3
  71. sglang/srt/layers/moe/fused_moe_triton/triton_kernels_moe.py +7 -4
  72. sglang/srt/layers/moe/moe_runner/deep_gemm.py +340 -55
  73. sglang/srt/layers/moe/moe_runner/runner.py +3 -0
  74. sglang/srt/layers/moe/moe_runner/triton_kernels.py +194 -0
  75. sglang/srt/layers/moe/token_dispatcher/__init__.py +4 -4
  76. sglang/srt/layers/moe/token_dispatcher/base.py +11 -5
  77. sglang/srt/layers/moe/token_dispatcher/deepep.py +25 -18
  78. sglang/srt/layers/moe/token_dispatcher/standard.py +1 -1
  79. sglang/srt/layers/moe/topk.py +35 -10
  80. sglang/srt/layers/moe/utils.py +3 -4
  81. sglang/srt/layers/pooler.py +21 -2
  82. sglang/srt/layers/quantization/__init__.py +13 -84
  83. sglang/srt/layers/quantization/auto_round.py +394 -0
  84. sglang/srt/layers/quantization/awq.py +0 -3
  85. sglang/srt/layers/quantization/base_config.py +7 -0
  86. sglang/srt/layers/quantization/fp8.py +68 -63
  87. sglang/srt/layers/quantization/fp8_kernel.py +1 -1
  88. sglang/srt/layers/quantization/fp8_utils.py +2 -2
  89. sglang/srt/layers/quantization/gguf.py +566 -0
  90. sglang/srt/layers/quantization/modelopt_quant.py +168 -11
  91. sglang/srt/layers/quantization/mxfp4.py +30 -38
  92. sglang/srt/layers/quantization/unquant.py +23 -45
  93. sglang/srt/layers/quantization/w4afp8.py +38 -2
  94. sglang/srt/layers/radix_attention.py +5 -2
  95. sglang/srt/layers/rotary_embedding.py +130 -46
  96. sglang/srt/layers/sampler.py +12 -1
  97. sglang/srt/lora/lora_registry.py +9 -0
  98. sglang/srt/managers/async_mm_data_processor.py +122 -0
  99. sglang/srt/managers/data_parallel_controller.py +30 -3
  100. sglang/srt/managers/detokenizer_manager.py +3 -0
  101. sglang/srt/managers/io_struct.py +29 -4
  102. sglang/srt/managers/multi_tokenizer_mixin.py +22 -1
  103. sglang/srt/managers/schedule_batch.py +74 -15
  104. sglang/srt/managers/scheduler.py +185 -144
  105. sglang/srt/managers/scheduler_metrics_mixin.py +22 -14
  106. sglang/srt/managers/scheduler_output_processor_mixin.py +40 -3
  107. sglang/srt/managers/scheduler_pp_mixin.py +7 -2
  108. sglang/srt/managers/scheduler_profiler_mixin.py +3 -4
  109. sglang/srt/managers/scheduler_runtime_checker_mixin.py +45 -0
  110. sglang/srt/managers/scheduler_update_weights_mixin.py +18 -3
  111. sglang/srt/managers/session_controller.py +6 -5
  112. sglang/srt/managers/tokenizer_manager.py +165 -78
  113. sglang/srt/managers/tp_worker.py +24 -1
  114. sglang/srt/mem_cache/base_prefix_cache.py +23 -4
  115. sglang/srt/mem_cache/common.py +1 -0
  116. sglang/srt/mem_cache/hicache_storage.py +7 -1
  117. sglang/srt/mem_cache/memory_pool.py +253 -57
  118. sglang/srt/mem_cache/memory_pool_host.py +12 -5
  119. sglang/srt/mem_cache/radix_cache.py +4 -0
  120. sglang/srt/mem_cache/storage/hf3fs/mini_3fs_metadata_server.py +3 -2
  121. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +1 -1
  122. sglang/srt/metrics/collector.py +46 -3
  123. sglang/srt/model_executor/cuda_graph_runner.py +15 -3
  124. sglang/srt/model_executor/forward_batch_info.py +55 -14
  125. sglang/srt/model_executor/model_runner.py +77 -170
  126. sglang/srt/model_executor/npu_graph_runner.py +7 -3
  127. sglang/srt/model_executor/piecewise_cuda_graph_runner.py +22 -12
  128. sglang/srt/model_loader/weight_utils.py +1 -1
  129. sglang/srt/models/bailing_moe.py +9 -2
  130. sglang/srt/models/deepseek_nextn.py +11 -2
  131. sglang/srt/models/deepseek_v2.py +296 -78
  132. sglang/srt/models/glm4.py +391 -77
  133. sglang/srt/models/glm4_moe.py +322 -354
  134. sglang/srt/models/glm4_moe_nextn.py +4 -14
  135. sglang/srt/models/glm4v.py +196 -55
  136. sglang/srt/models/glm4v_moe.py +29 -197
  137. sglang/srt/models/gpt_oss.py +1 -10
  138. sglang/srt/models/kimi_linear.py +678 -0
  139. sglang/srt/models/llama4.py +1 -1
  140. sglang/srt/models/llama_eagle3.py +11 -1
  141. sglang/srt/models/longcat_flash.py +2 -2
  142. sglang/srt/models/minimax_m2.py +922 -0
  143. sglang/srt/models/nvila.py +355 -0
  144. sglang/srt/models/nvila_lite.py +184 -0
  145. sglang/srt/models/qwen2.py +23 -2
  146. sglang/srt/models/qwen2_moe.py +30 -15
  147. sglang/srt/models/qwen3.py +35 -5
  148. sglang/srt/models/qwen3_moe.py +18 -12
  149. sglang/srt/models/qwen3_next.py +7 -0
  150. sglang/srt/multimodal/customized_mm_processor_utils.py +35 -0
  151. sglang/srt/multimodal/processors/base_processor.py +1 -0
  152. sglang/srt/multimodal/processors/glm4v.py +1 -1
  153. sglang/srt/multimodal/processors/{vila.py → nvila.py} +32 -24
  154. sglang/srt/multimodal/processors/points_v15_chat.py +2 -2
  155. sglang/srt/multiplex/multiplexing_mixin.py +209 -0
  156. sglang/srt/multiplex/pdmux_context.py +164 -0
  157. sglang/srt/parser/conversation.py +7 -1
  158. sglang/srt/parser/reasoning_parser.py +28 -1
  159. sglang/srt/sampling/custom_logit_processor.py +67 -1
  160. sglang/srt/sampling/penaltylib/frequency_penalty.py +6 -8
  161. sglang/srt/sampling/penaltylib/min_new_tokens.py +7 -8
  162. sglang/srt/sampling/penaltylib/orchestrator.py +43 -3
  163. sglang/srt/sampling/penaltylib/presence_penalty.py +6 -8
  164. sglang/srt/server_args.py +459 -199
  165. sglang/srt/single_batch_overlap.py +2 -4
  166. sglang/srt/speculative/draft_utils.py +16 -0
  167. sglang/srt/speculative/eagle_info.py +42 -36
  168. sglang/srt/speculative/eagle_info_v2.py +68 -25
  169. sglang/srt/speculative/eagle_utils.py +261 -16
  170. sglang/srt/speculative/eagle_worker.py +11 -3
  171. sglang/srt/speculative/eagle_worker_v2.py +15 -9
  172. sglang/srt/speculative/spec_info.py +305 -31
  173. sglang/srt/speculative/spec_utils.py +44 -8
  174. sglang/srt/tracing/trace.py +121 -12
  175. sglang/srt/utils/common.py +142 -74
  176. sglang/srt/utils/hf_transformers_utils.py +38 -12
  177. sglang/srt/utils/torch_memory_saver_adapter.py +20 -0
  178. sglang/test/kits/radix_cache_server_kit.py +50 -0
  179. sglang/test/runners.py +31 -7
  180. sglang/test/simple_eval_common.py +5 -3
  181. sglang/test/simple_eval_humaneval.py +1 -0
  182. sglang/test/simple_eval_math.py +1 -0
  183. sglang/test/simple_eval_mmlu.py +1 -0
  184. sglang/test/simple_eval_mmmu_vlm.py +1 -0
  185. sglang/test/test_deterministic.py +235 -12
  186. sglang/test/test_deterministic_utils.py +2 -1
  187. sglang/test/test_utils.py +7 -1
  188. sglang/version.py +1 -1
  189. {sglang-0.5.4.dist-info → sglang-0.5.4.post2.dist-info}/METADATA +15 -28
  190. {sglang-0.5.4.dist-info → sglang-0.5.4.post2.dist-info}/RECORD +194 -175
  191. sglang/srt/models/vila.py +0 -306
  192. /sglang/test/{kit_matched_stop.py → kits/matched_stop_kit.py} +0 -0
  193. {sglang-0.5.4.dist-info → sglang-0.5.4.post2.dist-info}/WHEEL +0 -0
  194. {sglang-0.5.4.dist-info → sglang-0.5.4.post2.dist-info}/licenses/LICENSE +0 -0
  195. {sglang-0.5.4.dist-info → sglang-0.5.4.post2.dist-info}/top_level.txt +0 -0
sglang/srt/models/llama_eagle3.py CHANGED
@@ -19,6 +19,7 @@ from sglang.srt.utils import add_prefix
19
19
  # https://github.com/SafeAILab/EAGLE/blob/main/eagle/model/cnets.py
20
20
  """Inference-only LLaMA-EAGLE model compatible with HuggingFace weights."""
21
21
 
22
+ import copy
22
23
  from typing import Iterable, Optional, Tuple
23
24
 
24
25
  import torch
@@ -161,6 +162,10 @@ class LlamaModel(nn.Module):
161
162
  if hidden_states.shape[-1] != embeds.shape[-1]:
162
163
  hidden_states = self.fc(hidden_states)
163
164
 
165
+ # idle batch
166
+ if hidden_states.shape[0] == 0:
167
+ return hidden_states, [hidden_states]
168
+
164
169
  residual = None
165
170
  hidden_states, residual = self.midlayer(
166
171
  positions,
@@ -212,7 +217,12 @@ class LlamaForCausalLMEagle3(LlamaForCausalLM):
212
217
  prefix=add_prefix("lm_head", prefix),
213
218
  )
214
219
 
215
- self.logits_processor = LogitsProcessor(config)
220
+ config_ = copy.deepcopy(config)
221
+ config_.vocab_size = (
222
+ config_.draft_vocab_size
223
+ ) # draft logits processor has it's own vocab size
224
+ self.logits_processor = LogitsProcessor(config_)
225
+
216
226
  self.capture_aux_hidden_states = True
217
227
  self.hot_token_id = None
218
228
 
sglang/srt/models/longcat_flash.py CHANGED
@@ -821,8 +821,8 @@ class LongcatFlashForCausalLM(nn.Module):
821
821
  experts = layer.mlp.experts
822
822
  if isinstance(experts, DeepEPMoE):
823
823
  for w in [
824
- experts.w13_weight_fp8,
825
- experts.w2_weight_fp8,
824
+ (experts.w13_weight, experts.w13_weight_scale_inv),
825
+ (experts.w2_weight, experts.w2_weight_scale_inv),
826
826
  ]:
827
827
  requant_weight_ue8m0_inplace(w[0], w[1], weight_block_size)
828
828