sglang 0.4.4__py3-none-any.whl → 0.4.4.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 (176) hide show
  1. sglang/__init__.py +2 -0
  2. sglang/api.py +6 -0
  3. sglang/bench_one_batch.py +1 -1
  4. sglang/bench_one_batch_server.py +1 -1
  5. sglang/bench_serving.py +3 -1
  6. sglang/check_env.py +3 -4
  7. sglang/lang/backend/openai.py +18 -5
  8. sglang/lang/chat_template.py +28 -7
  9. sglang/lang/interpreter.py +7 -3
  10. sglang/lang/ir.py +10 -0
  11. sglang/srt/_custom_ops.py +1 -1
  12. sglang/srt/code_completion_parser.py +174 -0
  13. sglang/srt/configs/__init__.py +2 -6
  14. sglang/srt/configs/deepseekvl2.py +667 -0
  15. sglang/srt/configs/janus_pro.py +3 -4
  16. sglang/srt/configs/load_config.py +1 -0
  17. sglang/srt/configs/model_config.py +63 -11
  18. sglang/srt/configs/utils.py +25 -0
  19. sglang/srt/connector/__init__.py +51 -0
  20. sglang/srt/connector/base_connector.py +112 -0
  21. sglang/srt/connector/redis.py +85 -0
  22. sglang/srt/connector/s3.py +122 -0
  23. sglang/srt/connector/serde/__init__.py +31 -0
  24. sglang/srt/connector/serde/safe_serde.py +29 -0
  25. sglang/srt/connector/serde/serde.py +43 -0
  26. sglang/srt/connector/utils.py +35 -0
  27. sglang/srt/conversation.py +88 -0
  28. sglang/srt/disaggregation/conn.py +81 -0
  29. sglang/srt/disaggregation/decode.py +495 -0
  30. sglang/srt/disaggregation/mini_lb.py +285 -0
  31. sglang/srt/disaggregation/prefill.py +249 -0
  32. sglang/srt/disaggregation/utils.py +44 -0
  33. sglang/srt/distributed/parallel_state.py +10 -3
  34. sglang/srt/entrypoints/engine.py +55 -5
  35. sglang/srt/entrypoints/http_server.py +71 -12
  36. sglang/srt/function_call_parser.py +164 -54
  37. sglang/srt/hf_transformers_utils.py +28 -3
  38. sglang/srt/layers/activation.py +4 -2
  39. sglang/srt/layers/attention/base_attn_backend.py +1 -1
  40. sglang/srt/layers/attention/flashattention_backend.py +295 -0
  41. sglang/srt/layers/attention/flashinfer_backend.py +1 -1
  42. sglang/srt/layers/attention/flashmla_backend.py +284 -0
  43. sglang/srt/layers/attention/triton_backend.py +171 -38
  44. sglang/srt/layers/attention/triton_ops/decode_attention.py +94 -31
  45. sglang/srt/layers/attention/triton_ops/extend_attention.py +14 -5
  46. sglang/srt/layers/attention/utils.py +53 -0
  47. sglang/srt/layers/attention/vision.py +9 -28
  48. sglang/srt/layers/dp_attention.py +62 -23
  49. sglang/srt/layers/elementwise.py +411 -0
  50. sglang/srt/layers/layernorm.py +24 -2
  51. sglang/srt/layers/linear.py +17 -5
  52. sglang/srt/layers/logits_processor.py +26 -7
  53. sglang/srt/layers/moe/ep_moe/kernels.py +110 -11
  54. sglang/srt/layers/moe/ep_moe/layer.py +273 -1
  55. sglang/srt/layers/moe/ep_moe/token_dispatcher.py +416 -0
  56. sglang/srt/layers/moe/fused_moe_native.py +2 -1
  57. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=64,device_name=NVIDIA_L20,dtype=int8_w8a8.json +146 -0
  58. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=64,device_name=NVIDIA_L40S,dtype=int8_w8a8.json +146 -0
  59. 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 +146 -0
  60. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  61. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +23 -32
  62. sglang/srt/layers/moe/fused_moe_triton/layer.py +1 -2
  63. sglang/srt/layers/moe/router.py +342 -0
  64. sglang/srt/layers/moe/topk.py +31 -18
  65. sglang/srt/layers/parameter.py +1 -1
  66. sglang/srt/layers/quantization/__init__.py +184 -126
  67. sglang/srt/layers/quantization/base_config.py +5 -0
  68. sglang/srt/layers/quantization/blockwise_int8.py +1 -1
  69. sglang/srt/layers/quantization/compressed_tensors/__init__.py +0 -0
  70. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py +652 -0
  71. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +658 -0
  72. sglang/srt/layers/quantization/compressed_tensors/schemes/__init__.py +9 -0
  73. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_scheme.py +56 -0
  74. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +162 -0
  75. sglang/srt/layers/quantization/compressed_tensors/utils.py +218 -0
  76. sglang/srt/layers/quantization/fp8.py +76 -34
  77. sglang/srt/layers/quantization/fp8_kernel.py +24 -8
  78. sglang/srt/layers/quantization/fp8_utils.py +284 -28
  79. sglang/srt/layers/quantization/gptq.py +36 -9
  80. sglang/srt/layers/quantization/kv_cache.py +98 -0
  81. sglang/srt/layers/quantization/modelopt_quant.py +9 -7
  82. sglang/srt/layers/quantization/utils.py +153 -0
  83. sglang/srt/layers/quantization/w8a8_fp8.py +70 -19
  84. sglang/srt/layers/rotary_embedding.py +66 -87
  85. sglang/srt/layers/sampler.py +1 -1
  86. sglang/srt/lora/layers.py +68 -0
  87. sglang/srt/lora/lora.py +2 -22
  88. sglang/srt/lora/lora_manager.py +47 -23
  89. sglang/srt/lora/mem_pool.py +110 -51
  90. sglang/srt/lora/utils.py +12 -1
  91. sglang/srt/managers/cache_controller.py +4 -5
  92. sglang/srt/managers/data_parallel_controller.py +31 -9
  93. sglang/srt/managers/expert_distribution.py +81 -0
  94. sglang/srt/managers/io_struct.py +39 -3
  95. sglang/srt/managers/mm_utils.py +373 -0
  96. sglang/srt/managers/multimodal_processor.py +68 -0
  97. sglang/srt/managers/multimodal_processors/base_processor.py +275 -0
  98. sglang/srt/managers/multimodal_processors/deepseek_vl_v2.py +119 -0
  99. sglang/srt/managers/multimodal_processors/gemma3.py +83 -0
  100. sglang/srt/managers/{image_processors → multimodal_processors}/janus_pro.py +20 -15
  101. sglang/srt/managers/{image_processors → multimodal_processors}/llava.py +10 -15
  102. sglang/srt/managers/multimodal_processors/minicpm.py +167 -0
  103. sglang/srt/managers/{image_processors → multimodal_processors}/mlama.py +7 -8
  104. sglang/srt/managers/{image_processors → multimodal_processors}/qwen_vl.py +28 -22
  105. sglang/srt/managers/schedule_batch.py +134 -31
  106. sglang/srt/managers/scheduler.py +325 -38
  107. sglang/srt/managers/scheduler_output_processor_mixin.py +4 -1
  108. sglang/srt/managers/session_controller.py +1 -1
  109. sglang/srt/managers/tokenizer_manager.py +59 -23
  110. sglang/srt/managers/tp_worker.py +1 -1
  111. sglang/srt/managers/tp_worker_overlap_thread.py +3 -3
  112. sglang/srt/managers/utils.py +6 -1
  113. sglang/srt/mem_cache/hiradix_cache.py +27 -8
  114. sglang/srt/mem_cache/memory_pool.py +258 -98
  115. sglang/srt/mem_cache/paged_allocator.py +2 -2
  116. sglang/srt/mem_cache/radix_cache.py +4 -4
  117. sglang/srt/model_executor/cuda_graph_runner.py +85 -28
  118. sglang/srt/model_executor/forward_batch_info.py +81 -15
  119. sglang/srt/model_executor/model_runner.py +70 -6
  120. sglang/srt/model_loader/loader.py +160 -2
  121. sglang/srt/model_loader/weight_utils.py +45 -0
  122. sglang/srt/models/deepseek_janus_pro.py +29 -86
  123. sglang/srt/models/deepseek_nextn.py +22 -10
  124. sglang/srt/models/deepseek_v2.py +326 -192
  125. sglang/srt/models/deepseek_vl2.py +358 -0
  126. sglang/srt/models/gemma3_causal.py +684 -0
  127. sglang/srt/models/gemma3_mm.py +462 -0
  128. sglang/srt/models/grok.py +374 -119
  129. sglang/srt/models/llama.py +47 -7
  130. sglang/srt/models/llama_eagle.py +1 -0
  131. sglang/srt/models/llama_eagle3.py +196 -0
  132. sglang/srt/models/llava.py +3 -3
  133. sglang/srt/models/llavavid.py +3 -3
  134. sglang/srt/models/minicpmo.py +1995 -0
  135. sglang/srt/models/minicpmv.py +62 -137
  136. sglang/srt/models/mllama.py +4 -4
  137. sglang/srt/models/phi3_small.py +1 -1
  138. sglang/srt/models/qwen2.py +3 -0
  139. sglang/srt/models/qwen2_5_vl.py +68 -146
  140. sglang/srt/models/qwen2_classification.py +75 -0
  141. sglang/srt/models/qwen2_moe.py +9 -1
  142. sglang/srt/models/qwen2_vl.py +25 -63
  143. sglang/srt/openai_api/adapter.py +145 -47
  144. sglang/srt/openai_api/protocol.py +23 -2
  145. sglang/srt/sampling/sampling_batch_info.py +1 -1
  146. sglang/srt/sampling/sampling_params.py +6 -6
  147. sglang/srt/server_args.py +104 -14
  148. sglang/srt/speculative/build_eagle_tree.py +7 -347
  149. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +41 -5
  150. sglang/srt/speculative/eagle_utils.py +208 -252
  151. sglang/srt/speculative/eagle_worker.py +139 -53
  152. sglang/srt/speculative/spec_info.py +6 -1
  153. sglang/srt/torch_memory_saver_adapter.py +22 -0
  154. sglang/srt/utils.py +182 -21
  155. sglang/test/__init__.py +0 -0
  156. sglang/test/attention/__init__.py +0 -0
  157. sglang/test/attention/test_flashattn_backend.py +312 -0
  158. sglang/test/runners.py +2 -0
  159. sglang/test/test_activation.py +2 -1
  160. sglang/test/test_block_fp8.py +5 -4
  161. sglang/test/test_block_fp8_ep.py +2 -1
  162. sglang/test/test_dynamic_grad_mode.py +58 -0
  163. sglang/test/test_layernorm.py +3 -2
  164. sglang/test/test_utils.py +55 -4
  165. sglang/utils.py +31 -0
  166. sglang/version.py +1 -1
  167. {sglang-0.4.4.dist-info → sglang-0.4.4.post2.dist-info}/METADATA +12 -8
  168. {sglang-0.4.4.dist-info → sglang-0.4.4.post2.dist-info}/RECORD +171 -125
  169. {sglang-0.4.4.dist-info → sglang-0.4.4.post2.dist-info}/WHEEL +1 -1
  170. sglang/srt/configs/qwen2_5_vl_config.py +0 -1006
  171. sglang/srt/managers/image_processor.py +0 -55
  172. sglang/srt/managers/image_processors/base_image_processor.py +0 -219
  173. sglang/srt/managers/image_processors/minicpmv.py +0 -86
  174. sglang/srt/managers/multi_modality_padding.py +0 -134
  175. {sglang-0.4.4.dist-info → sglang-0.4.4.post2.dist-info/licenses}/LICENSE +0 -0
  176. {sglang-0.4.4.dist-info → sglang-0.4.4.post2.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=64,device_name=NVIDIA_L40S,dtype=int8_w8a8.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 64,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 64,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 256,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 4,
40
+ "num_stages": 2
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 64,
47
+ "num_warps": 8,
48
+ "num_stages": 2
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 32,
55
+ "num_warps": 8,
56
+ "num_stages": 2
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 256,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 256,
70
+ "GROUP_SIZE_M": 64,
71
+ "num_warps": 8,
72
+ "num_stages": 2
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 256,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 8,
80
+ "num_stages": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 256,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 8,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 256,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 2
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 256,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 2
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 256,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 4,
112
+ "num_stages": 2
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 32,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 16,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 256,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 8,
136
+ "num_stages": 2
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
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 ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 32,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 64,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 32,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 16,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 64,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 64,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 32,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 64,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 64,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 64,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 32,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 64,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 32,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 64,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 32,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/fused_moe.py CHANGED
@@ -11,7 +11,6 @@ from typing import Any, Callable, Dict, List, Optional, Tuple
11
11
  import torch
12
12
  import triton
13
13
  import triton.language as tl
14
- from vllm import _custom_ops as vllm_ops
15
14
 
16
15
  from sglang.srt.layers.moe.topk import select_experts
17
16
  from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
@@ -46,6 +45,8 @@ if _is_cuda:
46
45
  from sglang.srt.layers.quantization.fp8_kernel import (
47
46
  sglang_per_token_group_quant_fp8,
48
47
  )
48
+ else:
49
+ from vllm import _custom_ops as vllm_ops
49
50
 
50
51
  if _is_cuda or _is_hip:
51
52
  from sgl_kernel import moe_align_block_size as sgl_moe_align_block_size
@@ -456,44 +457,34 @@ def moe_align_block_size(
456
457
  (max_num_m_blocks,), dtype=torch.int32, device=topk_ids.device
457
458
  )
458
459
  num_tokens_post_pad = torch.empty((1), dtype=torch.int32, device=topk_ids.device)
459
- if num_experts >= 224:
460
- if enable_moe_align_block_size_triton:
461
- moe_align_block_size_triton(
462
- topk_ids,
463
- num_experts,
464
- block_size,
465
- sorted_ids,
466
- expert_ids,
467
- num_tokens_post_pad,
468
- )
469
- else:
470
- token_cnts_buffer = torch.zeros(
471
- (num_experts + 1) * num_experts,
472
- dtype=torch.int32,
473
- device=topk_ids.device,
474
- )
475
- cumsum_buffer = torch.zeros(
476
- num_experts + 1, dtype=torch.int32, device=topk_ids.device
477
- )
478
-
479
- sgl_moe_align_block_size(
480
- topk_ids,
481
- num_experts,
482
- block_size,
483
- sorted_ids,
484
- expert_ids,
485
- num_tokens_post_pad,
486
- token_cnts_buffer,
487
- cumsum_buffer,
488
- )
460
+ if enable_moe_align_block_size_triton:
461
+ moe_align_block_size_triton(
462
+ topk_ids,
463
+ num_experts,
464
+ block_size,
465
+ sorted_ids,
466
+ expert_ids,
467
+ num_tokens_post_pad,
468
+ )
489
469
  else:
490
- vllm_ops.moe_align_block_size(
470
+ token_cnts_buffer = torch.zeros(
471
+ (num_experts + 1) * num_experts,
472
+ dtype=torch.int32,
473
+ device=topk_ids.device,
474
+ )
475
+ cumsum_buffer = torch.zeros(
476
+ num_experts + 1, dtype=torch.int32, device=topk_ids.device
477
+ )
478
+
479
+ sgl_moe_align_block_size(
491
480
  topk_ids,
492
481
  num_experts,
493
482
  block_size,
494
483
  sorted_ids,
495
484
  expert_ids,
496
485
  num_tokens_post_pad,
486
+ token_cnts_buffer,
487
+ cumsum_buffer,
497
488
  )
498
489
  return sorted_ids, expert_ids, num_tokens_post_pad
499
490
 
sglang/srt/layers/moe/fused_moe_triton/layer.py CHANGED
@@ -305,6 +305,7 @@ class FusedMoE(torch.nn.Module):
305
305
  self.use_presharded_weights = use_presharded_weights
306
306
  self.inplace = inplace
307
307
  self.no_combine = no_combine
308
+ self.local_num_experts = num_experts
308
309
 
309
310
  if quant_config is None:
310
311
  self.quant_method: Optional[QuantizeMethodBase] = (
@@ -629,8 +630,6 @@ class FusedMoE(torch.nn.Module):
629
630
  custom_routing_function=self.custom_routing_function,
630
631
  correction_bias=self.correction_bias,
631
632
  activation=self.activation,
632
- inplace=self.inplace,
633
- no_combine=self.no_combine,
634
633
  )
635
634
 
636
635
  if self.reduce_results and self.tp_size > 1: