sglang 0.5.0rc0__py3-none-any.whl → 0.5.0rc2__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 (170) hide show
  1. sglang/__init__.py +8 -3
  2. sglang/bench_one_batch.py +6 -1
  3. sglang/lang/chat_template.py +18 -0
  4. sglang/srt/bench_utils.py +137 -0
  5. sglang/srt/configs/model_config.py +8 -7
  6. sglang/srt/disaggregation/decode.py +8 -4
  7. sglang/srt/disaggregation/mooncake/conn.py +43 -25
  8. sglang/srt/disaggregation/mooncake/transfer_engine.py +29 -0
  9. sglang/srt/distributed/parallel_state.py +4 -2
  10. sglang/srt/entrypoints/context.py +3 -20
  11. sglang/srt/entrypoints/engine.py +13 -8
  12. sglang/srt/entrypoints/harmony_utils.py +2 -0
  13. sglang/srt/entrypoints/http_server.py +68 -5
  14. sglang/srt/entrypoints/openai/protocol.py +2 -9
  15. sglang/srt/entrypoints/openai/serving_chat.py +60 -265
  16. sglang/srt/entrypoints/openai/serving_completions.py +1 -0
  17. sglang/srt/entrypoints/openai/tool_server.py +4 -3
  18. sglang/srt/function_call/ebnf_composer.py +1 -0
  19. sglang/srt/function_call/function_call_parser.py +2 -0
  20. sglang/srt/function_call/glm4_moe_detector.py +1 -1
  21. sglang/srt/function_call/gpt_oss_detector.py +331 -0
  22. sglang/srt/function_call/kimik2_detector.py +3 -3
  23. sglang/srt/function_call/qwen3_coder_detector.py +219 -9
  24. sglang/srt/jinja_template_utils.py +6 -0
  25. sglang/srt/layers/attention/aiter_backend.py +370 -107
  26. sglang/srt/layers/attention/ascend_backend.py +3 -0
  27. sglang/srt/layers/attention/dual_chunk_flashattention_backend.py +1 -1
  28. sglang/srt/layers/attention/flashattention_backend.py +18 -0
  29. sglang/srt/layers/attention/flashinfer_backend.py +55 -13
  30. sglang/srt/layers/attention/flashinfer_mla_backend.py +1 -0
  31. sglang/srt/layers/attention/hybrid_attn_backend.py +1 -1
  32. sglang/srt/layers/attention/triton_backend.py +24 -27
  33. sglang/srt/layers/attention/trtllm_mha_backend.py +8 -6
  34. sglang/srt/layers/attention/trtllm_mla_backend.py +129 -25
  35. sglang/srt/layers/attention/vision.py +9 -1
  36. sglang/srt/layers/attention/wave_backend.py +627 -0
  37. sglang/srt/layers/attention/wave_ops/decode_attention.py +186 -0
  38. sglang/srt/layers/attention/wave_ops/extend_attention.py +149 -0
  39. sglang/srt/layers/attention/wave_ops/prefill_attention.py +79 -0
  40. sglang/srt/layers/communicator.py +11 -13
  41. sglang/srt/layers/dp_attention.py +118 -27
  42. sglang/srt/layers/flashinfer_comm_fusion.py +4 -4
  43. sglang/srt/layers/linear.py +1 -0
  44. sglang/srt/layers/logits_processor.py +12 -18
  45. sglang/srt/layers/moe/cutlass_moe.py +11 -16
  46. sglang/srt/layers/moe/cutlass_w4a8_moe.py +4 -5
  47. sglang/srt/layers/moe/ep_moe/kernels.py +43 -0
  48. sglang/srt/layers/moe/ep_moe/layer.py +60 -2
  49. 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
  50. 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
  51. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_0/E=16,N=1024,device_name=NVIDIA_B200.json +146 -0
  52. 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
  53. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_3_1/E=128,N=768,device_name=NVIDIA_H20.json +146 -0
  54. 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
  55. 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
  56. 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
  57. 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
  58. 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
  59. 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
  60. 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
  61. 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
  62. sglang/srt/layers/moe/fused_moe_triton/layer.py +7 -9
  63. sglang/srt/layers/moe/token_dispatcher/deepep.py +61 -24
  64. sglang/srt/layers/moe/topk.py +4 -1
  65. sglang/srt/layers/multimodal.py +156 -40
  66. sglang/srt/layers/quantization/__init__.py +10 -35
  67. sglang/srt/layers/quantization/awq.py +15 -16
  68. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +0 -1
  69. sglang/srt/layers/quantization/fp8_kernel.py +277 -0
  70. sglang/srt/layers/quantization/fp8_utils.py +22 -10
  71. sglang/srt/layers/quantization/gptq.py +12 -17
  72. sglang/srt/layers/quantization/marlin_utils.py +15 -5
  73. sglang/srt/layers/quantization/modelopt_quant.py +58 -41
  74. sglang/srt/layers/quantization/mxfp4.py +20 -3
  75. sglang/srt/layers/quantization/utils.py +52 -2
  76. sglang/srt/layers/quantization/w4afp8.py +20 -11
  77. sglang/srt/layers/quantization/w8a8_int8.py +48 -34
  78. sglang/srt/layers/rotary_embedding.py +281 -2
  79. sglang/srt/layers/sampler.py +5 -2
  80. sglang/srt/lora/backend/base_backend.py +3 -23
  81. sglang/srt/lora/layers.py +66 -116
  82. sglang/srt/lora/lora.py +17 -62
  83. sglang/srt/lora/lora_manager.py +12 -48
  84. sglang/srt/lora/lora_registry.py +20 -9
  85. sglang/srt/lora/mem_pool.py +20 -63
  86. sglang/srt/lora/triton_ops/qkv_lora_b.py +1 -1
  87. sglang/srt/lora/utils.py +25 -58
  88. sglang/srt/managers/cache_controller.py +24 -29
  89. sglang/srt/managers/detokenizer_manager.py +1 -1
  90. sglang/srt/managers/io_struct.py +20 -6
  91. sglang/srt/managers/mm_utils.py +1 -2
  92. sglang/srt/managers/multimodal_processor.py +1 -1
  93. sglang/srt/managers/schedule_batch.py +43 -49
  94. sglang/srt/managers/schedule_policy.py +6 -6
  95. sglang/srt/managers/scheduler.py +18 -11
  96. sglang/srt/managers/scheduler_profiler_mixin.py +28 -8
  97. sglang/srt/managers/tokenizer_manager.py +53 -44
  98. sglang/srt/mem_cache/allocator.py +39 -214
  99. sglang/srt/mem_cache/allocator_ascend.py +158 -0
  100. sglang/srt/mem_cache/chunk_cache.py +1 -1
  101. sglang/srt/mem_cache/hicache_storage.py +1 -1
  102. sglang/srt/mem_cache/hiradix_cache.py +34 -24
  103. sglang/srt/mem_cache/lora_radix_cache.py +421 -0
  104. sglang/srt/mem_cache/memory_pool_host.py +33 -35
  105. sglang/srt/mem_cache/radix_cache.py +2 -5
  106. sglang/srt/mem_cache/storage/hf3fs/mini_3fs_metadata_server.py +443 -0
  107. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +139 -67
  108. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +6 -9
  109. sglang/srt/model_executor/cuda_graph_runner.py +29 -23
  110. sglang/srt/model_executor/forward_batch_info.py +33 -14
  111. sglang/srt/model_executor/model_runner.py +179 -81
  112. sglang/srt/model_loader/loader.py +18 -6
  113. sglang/srt/models/deepseek_nextn.py +2 -1
  114. sglang/srt/models/deepseek_v2.py +79 -38
  115. sglang/srt/models/gemma2.py +0 -34
  116. sglang/srt/models/gemma3n_mm.py +8 -9
  117. sglang/srt/models/glm4.py +6 -0
  118. sglang/srt/models/glm4_moe.py +11 -11
  119. sglang/srt/models/glm4_moe_nextn.py +2 -1
  120. sglang/srt/models/glm4v.py +589 -0
  121. sglang/srt/models/glm4v_moe.py +400 -0
  122. sglang/srt/models/gpt_oss.py +142 -20
  123. sglang/srt/models/granite.py +0 -25
  124. sglang/srt/models/llama.py +10 -27
  125. sglang/srt/models/llama4.py +19 -6
  126. sglang/srt/models/qwen2.py +2 -2
  127. sglang/srt/models/qwen2_5_vl.py +7 -3
  128. sglang/srt/models/qwen2_audio.py +10 -9
  129. sglang/srt/models/qwen2_moe.py +20 -5
  130. sglang/srt/models/qwen3.py +0 -24
  131. sglang/srt/models/qwen3_classification.py +78 -0
  132. sglang/srt/models/qwen3_moe.py +18 -5
  133. sglang/srt/models/registry.py +1 -1
  134. sglang/srt/models/step3_vl.py +6 -2
  135. sglang/srt/models/torch_native_llama.py +0 -24
  136. sglang/srt/multimodal/processors/base_processor.py +23 -13
  137. sglang/srt/multimodal/processors/glm4v.py +132 -0
  138. sglang/srt/multimodal/processors/qwen_audio.py +4 -2
  139. sglang/srt/operations.py +17 -2
  140. sglang/srt/reasoning_parser.py +316 -0
  141. sglang/srt/sampling/sampling_batch_info.py +7 -4
  142. sglang/srt/server_args.py +142 -140
  143. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +7 -21
  144. sglang/srt/speculative/eagle_draft_extend_cuda_graph_runner.py +7 -21
  145. sglang/srt/speculative/eagle_worker.py +16 -0
  146. sglang/srt/two_batch_overlap.py +16 -12
  147. sglang/srt/utils.py +3 -3
  148. sglang/srt/weight_sync/tensor_bucket.py +106 -0
  149. sglang/test/attention/test_trtllm_mla_backend.py +186 -36
  150. sglang/test/doc_patch.py +59 -0
  151. sglang/test/few_shot_gsm8k.py +1 -1
  152. sglang/test/few_shot_gsm8k_engine.py +1 -1
  153. sglang/test/run_eval.py +4 -1
  154. sglang/test/simple_eval_common.py +6 -0
  155. sglang/test/simple_eval_gpqa.py +2 -0
  156. sglang/test/test_fp4_moe.py +118 -36
  157. sglang/test/test_marlin_moe.py +1 -1
  158. sglang/test/test_marlin_utils.py +1 -1
  159. sglang/utils.py +1 -1
  160. sglang/version.py +1 -1
  161. {sglang-0.5.0rc0.dist-info → sglang-0.5.0rc2.dist-info}/METADATA +27 -31
  162. {sglang-0.5.0rc0.dist-info → sglang-0.5.0rc2.dist-info}/RECORD +166 -142
  163. sglang/lang/backend/__init__.py +0 -0
  164. sglang/srt/function_call/harmony_tool_parser.py +0 -130
  165. sglang/srt/layers/quantization/scalar_type.py +0 -352
  166. sglang/srt/lora/backend/flashinfer_backend.py +0 -131
  167. /sglang/{api.py → lang/api.py} +0 -0
  168. {sglang-0.5.0rc0.dist-info → sglang-0.5.0rc2.dist-info}/WHEEL +0 -0
  169. {sglang-0.5.0rc0.dist-info → sglang-0.5.0rc2.dist-info}/licenses/LICENSE +0 -0
  170. {sglang-0.5.0rc0.dist-info → sglang-0.5.0rc2.dist-info}/top_level.txt +0 -0
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 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": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 64,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 32,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 64,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
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": 16,
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": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 16,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
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": 64,
116
+ "BLOCK_SIZE_N": 128,
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": 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": 32,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
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 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": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
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": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 32,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 64,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
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": 16,
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": 16,
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": 16,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
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": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 3
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": 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": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
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 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": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 4
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": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
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": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 64,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 4,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
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": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 16,
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": 32,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 32,
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": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
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": 4
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/triton_3_4_0/E=128,N=768,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].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": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 256,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 256,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 4
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 256,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
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": 4,
80
+ "num_stages": 4
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": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
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": 256,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
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": 256,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
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": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 256,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ }
146
+ }