sglang 0.4.5__py3-none-any.whl → 0.4.5.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 (166) hide show
  1. sglang/__init__.py +2 -4
  2. sglang/bench_one_batch.py +23 -2
  3. sglang/bench_serving.py +6 -4
  4. sglang/lang/backend/anthropic.py +0 -4
  5. sglang/lang/backend/base_backend.py +1 -1
  6. sglang/lang/backend/openai.py +1 -1
  7. sglang/lang/backend/vertexai.py +0 -1
  8. sglang/lang/compiler.py +1 -7
  9. sglang/lang/tracer.py +3 -7
  10. sglang/srt/_custom_ops.py +0 -2
  11. sglang/srt/configs/model_config.py +37 -5
  12. sglang/srt/constrained/base_grammar_backend.py +26 -5
  13. sglang/srt/constrained/llguidance_backend.py +1 -0
  14. sglang/srt/constrained/outlines_backend.py +1 -0
  15. sglang/srt/constrained/outlines_jump_forward.py +14 -1
  16. sglang/srt/constrained/reasoner_grammar_backend.py +101 -0
  17. sglang/srt/constrained/triton_ops/bitmask_ops.py +141 -0
  18. sglang/srt/constrained/xgrammar_backend.py +27 -4
  19. sglang/srt/custom_op.py +0 -62
  20. sglang/srt/disaggregation/base/__init__.py +8 -0
  21. sglang/srt/disaggregation/base/conn.py +113 -0
  22. sglang/srt/disaggregation/decode.py +80 -11
  23. sglang/srt/disaggregation/mini_lb.py +58 -123
  24. sglang/srt/disaggregation/mooncake/__init__.py +6 -0
  25. sglang/srt/disaggregation/mooncake/conn.py +585 -0
  26. sglang/srt/disaggregation/mooncake/transfer_engine.py +77 -0
  27. sglang/srt/disaggregation/prefill.py +82 -22
  28. sglang/srt/disaggregation/utils.py +46 -0
  29. sglang/srt/entrypoints/EngineBase.py +53 -0
  30. sglang/srt/entrypoints/engine.py +36 -8
  31. sglang/srt/entrypoints/http_server.py +37 -8
  32. sglang/srt/entrypoints/http_server_engine.py +142 -0
  33. sglang/srt/entrypoints/verl_engine.py +42 -13
  34. sglang/srt/hf_transformers_utils.py +4 -0
  35. sglang/srt/layers/activation.py +6 -8
  36. sglang/srt/layers/attention/flashattention_backend.py +430 -257
  37. sglang/srt/layers/attention/flashinfer_backend.py +18 -9
  38. sglang/srt/layers/attention/torch_native_backend.py +6 -1
  39. sglang/srt/layers/attention/triton_backend.py +6 -0
  40. sglang/srt/layers/attention/triton_ops/extend_attention.py +13 -2
  41. sglang/srt/layers/attention/vision.py +1 -1
  42. sglang/srt/layers/dp_attention.py +2 -4
  43. sglang/srt/layers/elementwise.py +15 -2
  44. sglang/srt/layers/layernorm.py +1 -1
  45. sglang/srt/layers/linear.py +18 -3
  46. sglang/srt/layers/moe/ep_moe/layer.py +15 -29
  47. sglang/srt/layers/moe/ep_moe/token_dispatcher.py +145 -118
  48. sglang/srt/layers/moe/fused_moe_native.py +4 -0
  49. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  50. sglang/srt/layers/moe/fused_moe_triton/configs/E=264,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  51. sglang/srt/layers/moe/fused_moe_triton/configs/{E=257,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → E=264,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json } +34 -34
  52. sglang/srt/layers/moe/fused_moe_triton/configs/E=272,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  53. sglang/srt/layers/moe/fused_moe_triton/configs/E=272,N=64,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  54. sglang/srt/layers/moe/fused_moe_triton/configs/E=288,N=64,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  55. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +46 -34
  56. sglang/srt/layers/moe/fused_moe_triton/layer.py +7 -0
  57. sglang/srt/layers/moe/router.py +7 -1
  58. sglang/srt/layers/moe/topk.py +63 -45
  59. sglang/srt/layers/parameter.py +0 -2
  60. sglang/srt/layers/quantization/__init__.py +13 -5
  61. sglang/srt/layers/quantization/blockwise_int8.py +2 -0
  62. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py +12 -2
  63. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +72 -77
  64. sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py +4 -7
  65. sglang/srt/layers/quantization/fp8.py +131 -136
  66. sglang/srt/layers/quantization/fp8_kernel.py +328 -46
  67. sglang/srt/layers/quantization/fp8_utils.py +206 -253
  68. sglang/srt/layers/quantization/kv_cache.py +43 -52
  69. sglang/srt/layers/quantization/modelopt_quant.py +271 -4
  70. sglang/srt/layers/quantization/moe_wna16.py +2 -0
  71. sglang/srt/layers/quantization/utils.py +5 -11
  72. sglang/srt/layers/quantization/w8a8_fp8.py +156 -4
  73. sglang/srt/layers/quantization/w8a8_int8.py +8 -7
  74. sglang/srt/layers/radix_attention.py +28 -1
  75. sglang/srt/layers/rotary_embedding.py +15 -3
  76. sglang/srt/layers/sampler.py +5 -10
  77. sglang/srt/lora/backend/base_backend.py +18 -2
  78. sglang/srt/lora/backend/flashinfer_backend.py +1 -1
  79. sglang/srt/lora/backend/triton_backend.py +1 -1
  80. sglang/srt/lora/layers.py +1 -1
  81. sglang/srt/lora/lora.py +1 -1
  82. sglang/srt/lora/lora_manager.py +1 -1
  83. sglang/srt/managers/detokenizer_manager.py +0 -1
  84. sglang/srt/managers/io_struct.py +255 -97
  85. sglang/srt/managers/mm_utils.py +7 -5
  86. sglang/srt/managers/multimodal_processor.py +0 -2
  87. sglang/srt/managers/multimodal_processors/base_processor.py +117 -79
  88. sglang/srt/managers/multimodal_processors/janus_pro.py +3 -1
  89. sglang/srt/managers/multimodal_processors/mllama4.py +21 -36
  90. sglang/srt/managers/schedule_batch.py +64 -25
  91. sglang/srt/managers/scheduler.py +80 -82
  92. sglang/srt/managers/tokenizer_manager.py +18 -3
  93. sglang/srt/managers/tp_worker.py +1 -0
  94. sglang/srt/mem_cache/hiradix_cache.py +5 -1
  95. sglang/srt/mem_cache/memory_pool.py +21 -3
  96. sglang/srt/metrics/collector.py +9 -0
  97. sglang/srt/model_executor/cuda_graph_runner.py +9 -6
  98. sglang/srt/model_executor/forward_batch_info.py +234 -15
  99. sglang/srt/model_executor/model_runner.py +67 -35
  100. sglang/srt/model_loader/loader.py +31 -4
  101. sglang/srt/model_loader/weight_utils.py +4 -2
  102. sglang/srt/models/baichuan.py +2 -0
  103. sglang/srt/models/bert.py +398 -0
  104. sglang/srt/models/chatglm.py +1 -0
  105. sglang/srt/models/commandr.py +1 -0
  106. sglang/srt/models/dbrx.py +1 -0
  107. sglang/srt/models/deepseek.py +2 -1
  108. sglang/srt/models/deepseek_nextn.py +74 -70
  109. sglang/srt/models/deepseek_v2.py +494 -366
  110. sglang/srt/models/exaone.py +1 -0
  111. sglang/srt/models/gemma.py +1 -0
  112. sglang/srt/models/gemma2.py +1 -0
  113. sglang/srt/models/gemma3_causal.py +1 -0
  114. sglang/srt/models/gpt2.py +1 -0
  115. sglang/srt/models/gpt_bigcode.py +1 -0
  116. sglang/srt/models/granite.py +1 -0
  117. sglang/srt/models/grok.py +1 -0
  118. sglang/srt/models/internlm2.py +1 -0
  119. sglang/srt/models/llama.py +6 -5
  120. sglang/srt/models/llama4.py +101 -34
  121. sglang/srt/models/minicpm.py +1 -0
  122. sglang/srt/models/minicpm3.py +30 -200
  123. sglang/srt/models/mixtral.py +1 -0
  124. sglang/srt/models/mixtral_quant.py +1 -0
  125. sglang/srt/models/mllama.py +51 -8
  126. sglang/srt/models/mllama4.py +102 -29
  127. sglang/srt/models/olmo.py +1 -0
  128. sglang/srt/models/olmo2.py +1 -0
  129. sglang/srt/models/olmoe.py +1 -0
  130. sglang/srt/models/phi3_small.py +1 -0
  131. sglang/srt/models/qwen.py +1 -0
  132. sglang/srt/models/qwen2.py +5 -1
  133. sglang/srt/models/qwen2_5_vl.py +35 -70
  134. sglang/srt/models/qwen2_moe.py +15 -13
  135. sglang/srt/models/qwen2_vl.py +27 -25
  136. sglang/srt/models/qwen3.py +335 -0
  137. sglang/srt/models/qwen3_moe.py +423 -0
  138. sglang/srt/models/stablelm.py +1 -0
  139. sglang/srt/models/xverse.py +1 -0
  140. sglang/srt/models/xverse_moe.py +1 -0
  141. sglang/srt/openai_api/adapter.py +4 -1
  142. sglang/srt/patch_torch.py +11 -0
  143. sglang/srt/reasoning_parser.py +0 -1
  144. sglang/srt/sampling/sampling_batch_info.py +2 -3
  145. sglang/srt/server_args.py +55 -19
  146. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +4 -4
  147. sglang/srt/speculative/eagle_utils.py +1 -11
  148. sglang/srt/speculative/eagle_worker.py +10 -9
  149. sglang/srt/utils.py +136 -10
  150. sglang/test/attention/test_flashattn_backend.py +259 -221
  151. sglang/test/attention/test_flashattn_mla_backend.py +285 -0
  152. sglang/test/attention/test_prefix_chunk_info.py +224 -0
  153. sglang/test/runners.py +5 -1
  154. sglang/test/test_block_fp8.py +224 -0
  155. sglang/test/test_custom_ops.py +1 -1
  156. sglang/test/test_utils.py +19 -8
  157. sglang/version.py +1 -1
  158. {sglang-0.4.5.dist-info → sglang-0.4.5.post2.dist-info}/METADATA +15 -5
  159. {sglang-0.4.5.dist-info → sglang-0.4.5.post2.dist-info}/RECORD +162 -147
  160. {sglang-0.4.5.dist-info → sglang-0.4.5.post2.dist-info}/WHEEL +1 -1
  161. sglang/lang/__init__.py +0 -0
  162. sglang/srt/disaggregation/conn.py +0 -81
  163. sglang/srt/lora/backend/__init__.py +0 -25
  164. sglang/srt/server.py +0 -18
  165. {sglang-0.4.5.dist-info → sglang-0.4.5.post2.dist-info}/licenses/LICENSE +0 -0
  166. {sglang-0.4.5.dist-info → sglang-0.4.5.post2.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=264,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 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": 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": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 2
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": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 16,
71
+ "num_warps": 4,
72
+ "num_stages": 2
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": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 2
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": 2
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 64,
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": 32,
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": 32,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 64,
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=257,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json → E=264,N=256,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json } RENAMED
@@ -2,16 +2,16 @@
2
2
  "1": {
3
3
  "BLOCK_SIZE_M": 16,
4
4
  "BLOCK_SIZE_N": 128,
5
- "BLOCK_SIZE_K": 128,
6
- "GROUP_SIZE_M": 1,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 16,
7
7
  "num_warps": 4,
8
8
  "num_stages": 4
9
9
  },
10
10
  "2": {
11
11
  "BLOCK_SIZE_M": 16,
12
- "BLOCK_SIZE_N": 128,
12
+ "BLOCK_SIZE_N": 64,
13
13
  "BLOCK_SIZE_K": 128,
14
- "GROUP_SIZE_M": 16,
14
+ "GROUP_SIZE_M": 1,
15
15
  "num_warps": 4,
16
16
  "num_stages": 4
17
17
  },
@@ -19,23 +19,23 @@
19
19
  "BLOCK_SIZE_M": 16,
20
20
  "BLOCK_SIZE_N": 128,
21
21
  "BLOCK_SIZE_K": 128,
22
- "GROUP_SIZE_M": 32,
22
+ "GROUP_SIZE_M": 1,
23
23
  "num_warps": 4,
24
- "num_stages": 4
24
+ "num_stages": 3
25
25
  },
26
26
  "8": {
27
27
  "BLOCK_SIZE_M": 16,
28
28
  "BLOCK_SIZE_N": 128,
29
29
  "BLOCK_SIZE_K": 128,
30
- "GROUP_SIZE_M": 64,
30
+ "GROUP_SIZE_M": 1,
31
31
  "num_warps": 4,
32
- "num_stages": 4
32
+ "num_stages": 3
33
33
  },
34
34
  "16": {
35
35
  "BLOCK_SIZE_M": 16,
36
36
  "BLOCK_SIZE_N": 128,
37
37
  "BLOCK_SIZE_K": 128,
38
- "GROUP_SIZE_M": 1,
38
+ "GROUP_SIZE_M": 64,
39
39
  "num_warps": 4,
40
40
  "num_stages": 3
41
41
  },
@@ -43,17 +43,17 @@
43
43
  "BLOCK_SIZE_M": 16,
44
44
  "BLOCK_SIZE_N": 128,
45
45
  "BLOCK_SIZE_K": 128,
46
- "GROUP_SIZE_M": 1,
46
+ "GROUP_SIZE_M": 64,
47
47
  "num_warps": 4,
48
- "num_stages": 4
48
+ "num_stages": 3
49
49
  },
50
50
  "32": {
51
51
  "BLOCK_SIZE_M": 16,
52
52
  "BLOCK_SIZE_N": 128,
53
53
  "BLOCK_SIZE_K": 128,
54
- "GROUP_SIZE_M": 1,
54
+ "GROUP_SIZE_M": 32,
55
55
  "num_warps": 4,
56
- "num_stages": 5
56
+ "num_stages": 3
57
57
  },
58
58
  "48": {
59
59
  "BLOCK_SIZE_M": 16,
@@ -61,21 +61,21 @@
61
61
  "BLOCK_SIZE_K": 128,
62
62
  "GROUP_SIZE_M": 32,
63
63
  "num_warps": 4,
64
- "num_stages": 4
64
+ "num_stages": 3
65
65
  },
66
66
  "64": {
67
67
  "BLOCK_SIZE_M": 16,
68
68
  "BLOCK_SIZE_N": 128,
69
69
  "BLOCK_SIZE_K": 128,
70
- "GROUP_SIZE_M": 64,
70
+ "GROUP_SIZE_M": 1,
71
71
  "num_warps": 4,
72
- "num_stages": 4
72
+ "num_stages": 3
73
73
  },
74
74
  "96": {
75
75
  "BLOCK_SIZE_M": 16,
76
76
  "BLOCK_SIZE_N": 128,
77
77
  "BLOCK_SIZE_K": 128,
78
- "GROUP_SIZE_M": 16,
78
+ "GROUP_SIZE_M": 64,
79
79
  "num_warps": 4,
80
80
  "num_stages": 3
81
81
  },
@@ -83,7 +83,7 @@
83
83
  "BLOCK_SIZE_M": 16,
84
84
  "BLOCK_SIZE_N": 128,
85
85
  "BLOCK_SIZE_K": 128,
86
- "GROUP_SIZE_M": 32,
86
+ "GROUP_SIZE_M": 16,
87
87
  "num_warps": 4,
88
88
  "num_stages": 3
89
89
  },
@@ -91,45 +91,45 @@
91
91
  "BLOCK_SIZE_M": 16,
92
92
  "BLOCK_SIZE_N": 128,
93
93
  "BLOCK_SIZE_K": 128,
94
- "GROUP_SIZE_M": 16,
94
+ "GROUP_SIZE_M": 32,
95
95
  "num_warps": 4,
96
96
  "num_stages": 3
97
97
  },
98
98
  "512": {
99
99
  "BLOCK_SIZE_M": 64,
100
- "BLOCK_SIZE_N": 128,
100
+ "BLOCK_SIZE_N": 64,
101
101
  "BLOCK_SIZE_K": 128,
102
- "GROUP_SIZE_M": 32,
102
+ "GROUP_SIZE_M": 16,
103
103
  "num_warps": 4,
104
- "num_stages": 4
104
+ "num_stages": 3
105
105
  },
106
106
  "1024": {
107
107
  "BLOCK_SIZE_M": 64,
108
- "BLOCK_SIZE_N": 128,
108
+ "BLOCK_SIZE_N": 64,
109
109
  "BLOCK_SIZE_K": 128,
110
- "GROUP_SIZE_M": 16,
110
+ "GROUP_SIZE_M": 32,
111
111
  "num_warps": 4,
112
- "num_stages": 4
112
+ "num_stages": 3
113
113
  },
114
114
  "1536": {
115
115
  "BLOCK_SIZE_M": 64,
116
- "BLOCK_SIZE_N": 128,
116
+ "BLOCK_SIZE_N": 64,
117
117
  "BLOCK_SIZE_K": 128,
118
118
  "GROUP_SIZE_M": 32,
119
119
  "num_warps": 4,
120
- "num_stages": 4
120
+ "num_stages": 3
121
121
  },
122
122
  "2048": {
123
123
  "BLOCK_SIZE_M": 64,
124
- "BLOCK_SIZE_N": 128,
124
+ "BLOCK_SIZE_N": 64,
125
125
  "BLOCK_SIZE_K": 128,
126
126
  "GROUP_SIZE_M": 32,
127
127
  "num_warps": 4,
128
- "num_stages": 4
128
+ "num_stages": 3
129
129
  },
130
130
  "3072": {
131
- "BLOCK_SIZE_M": 64,
132
- "BLOCK_SIZE_N": 128,
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 64,
133
133
  "BLOCK_SIZE_K": 128,
134
134
  "GROUP_SIZE_M": 16,
135
135
  "num_warps": 4,
@@ -137,10 +137,10 @@
137
137
  },
138
138
  "4096": {
139
139
  "BLOCK_SIZE_M": 64,
140
- "BLOCK_SIZE_N": 128,
140
+ "BLOCK_SIZE_N": 64,
141
141
  "BLOCK_SIZE_K": 128,
142
- "GROUP_SIZE_M": 16,
142
+ "GROUP_SIZE_M": 32,
143
143
  "num_warps": 4,
144
- "num_stages": 4
144
+ "num_stages": 3
145
145
  }
146
146
  }
sglang/srt/layers/moe/fused_moe_triton/configs/E=272,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
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": 64,
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": 64,
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": 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": 64,
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": 64,
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": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 2
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
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": 2
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": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
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": 32,
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": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 64,
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": 32,
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": 32,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 64,
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=272,N=64,device_name=NVIDIA_A800-SXM4-80GB.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 64,
5
+ "BLOCK_SIZE_K": 64,
6
+ "GROUP_SIZE_M": 1,
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": 64,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 64,
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": 64,
29
+ "BLOCK_SIZE_K": 64,
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": 128,
37
+ "BLOCK_SIZE_K": 64,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 64,
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": 64,
53
+ "BLOCK_SIZE_K": 64,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 1,
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": 64,
70
+ "GROUP_SIZE_M": 32,
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": 64,
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": 64,
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": 64,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 2
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 2
145
+ }
146
+ }