sglang 0.4.4.post4__py3-none-any.whl → 0.4.5.post1__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 (134) hide show
  1. sglang/bench_one_batch.py +21 -0
  2. sglang/bench_serving.py +10 -4
  3. sglang/lang/chat_template.py +24 -0
  4. sglang/srt/configs/model_config.py +40 -4
  5. sglang/srt/constrained/base_grammar_backend.py +26 -5
  6. sglang/srt/constrained/llguidance_backend.py +1 -0
  7. sglang/srt/constrained/outlines_backend.py +1 -0
  8. sglang/srt/constrained/reasoner_grammar_backend.py +101 -0
  9. sglang/srt/constrained/xgrammar_backend.py +1 -0
  10. sglang/srt/conversation.py +29 -4
  11. sglang/srt/disaggregation/base/__init__.py +8 -0
  12. sglang/srt/disaggregation/base/conn.py +113 -0
  13. sglang/srt/disaggregation/decode.py +18 -5
  14. sglang/srt/disaggregation/mini_lb.py +53 -122
  15. sglang/srt/disaggregation/mooncake/__init__.py +6 -0
  16. sglang/srt/disaggregation/mooncake/conn.py +615 -0
  17. sglang/srt/disaggregation/mooncake/transfer_engine.py +108 -0
  18. sglang/srt/disaggregation/prefill.py +43 -19
  19. sglang/srt/disaggregation/utils.py +31 -0
  20. sglang/srt/entrypoints/EngineBase.py +53 -0
  21. sglang/srt/entrypoints/engine.py +36 -8
  22. sglang/srt/entrypoints/http_server.py +37 -8
  23. sglang/srt/entrypoints/http_server_engine.py +142 -0
  24. sglang/srt/entrypoints/verl_engine.py +37 -10
  25. sglang/srt/hf_transformers_utils.py +4 -0
  26. sglang/srt/layers/attention/flashattention_backend.py +609 -202
  27. sglang/srt/layers/attention/flashinfer_backend.py +13 -7
  28. sglang/srt/layers/attention/vision.py +1 -1
  29. sglang/srt/layers/dp_attention.py +2 -4
  30. sglang/srt/layers/elementwise.py +15 -2
  31. sglang/srt/layers/linear.py +1 -0
  32. sglang/srt/layers/moe/ep_moe/token_dispatcher.py +145 -118
  33. sglang/srt/layers/moe/fused_moe_native.py +5 -0
  34. sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  35. sglang/srt/layers/moe/fused_moe_triton/configs/E=144,N=512,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  36. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  37. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1024,device_name=NVIDIA_H200.json +146 -0
  38. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  39. sglang/srt/layers/moe/fused_moe_triton/configs/E=20,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  40. sglang/srt/layers/moe/fused_moe_triton/configs/E=24,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  41. 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
  42. sglang/srt/layers/moe/fused_moe_triton/configs/E=264,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  43. 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
  44. sglang/srt/layers/moe/fused_moe_triton/configs/E=272,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  45. sglang/srt/layers/moe/fused_moe_triton/configs/E=272,N=64,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  46. sglang/srt/layers/moe/fused_moe_triton/configs/E=288,N=64,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  47. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +51 -24
  48. sglang/srt/layers/moe/fused_moe_triton/layer.py +7 -0
  49. sglang/srt/layers/moe/router.py +7 -1
  50. sglang/srt/layers/moe/topk.py +37 -16
  51. sglang/srt/layers/quantization/__init__.py +13 -5
  52. sglang/srt/layers/quantization/blockwise_int8.py +2 -0
  53. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py +4 -0
  54. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +68 -45
  55. sglang/srt/layers/quantization/fp8.py +28 -14
  56. sglang/srt/layers/quantization/fp8_kernel.py +130 -4
  57. sglang/srt/layers/quantization/fp8_utils.py +34 -6
  58. sglang/srt/layers/quantization/kv_cache.py +43 -52
  59. sglang/srt/layers/quantization/modelopt_quant.py +271 -4
  60. sglang/srt/layers/quantization/moe_wna16.py +2 -0
  61. sglang/srt/layers/quantization/w8a8_fp8.py +154 -4
  62. sglang/srt/layers/quantization/w8a8_int8.py +3 -0
  63. sglang/srt/layers/radix_attention.py +14 -0
  64. sglang/srt/layers/rotary_embedding.py +75 -1
  65. sglang/srt/managers/io_struct.py +254 -97
  66. sglang/srt/managers/mm_utils.py +3 -2
  67. sglang/srt/managers/multimodal_processors/base_processor.py +114 -77
  68. sglang/srt/managers/multimodal_processors/janus_pro.py +3 -1
  69. sglang/srt/managers/multimodal_processors/mllama4.py +146 -0
  70. sglang/srt/managers/schedule_batch.py +62 -21
  71. sglang/srt/managers/scheduler.py +71 -14
  72. sglang/srt/managers/tokenizer_manager.py +17 -3
  73. sglang/srt/managers/tp_worker.py +1 -0
  74. sglang/srt/mem_cache/memory_pool.py +14 -1
  75. sglang/srt/metrics/collector.py +9 -0
  76. sglang/srt/model_executor/cuda_graph_runner.py +7 -4
  77. sglang/srt/model_executor/forward_batch_info.py +234 -15
  78. sglang/srt/model_executor/model_runner.py +49 -9
  79. sglang/srt/model_loader/loader.py +31 -4
  80. sglang/srt/model_loader/weight_utils.py +4 -2
  81. sglang/srt/models/baichuan.py +2 -0
  82. sglang/srt/models/chatglm.py +1 -0
  83. sglang/srt/models/commandr.py +1 -0
  84. sglang/srt/models/dbrx.py +1 -0
  85. sglang/srt/models/deepseek.py +1 -0
  86. sglang/srt/models/deepseek_v2.py +248 -61
  87. sglang/srt/models/exaone.py +1 -0
  88. sglang/srt/models/gemma.py +1 -0
  89. sglang/srt/models/gemma2.py +1 -0
  90. sglang/srt/models/gemma3_causal.py +1 -0
  91. sglang/srt/models/gpt2.py +1 -0
  92. sglang/srt/models/gpt_bigcode.py +1 -0
  93. sglang/srt/models/granite.py +1 -0
  94. sglang/srt/models/grok.py +1 -0
  95. sglang/srt/models/internlm2.py +1 -0
  96. sglang/srt/models/llama.py +13 -4
  97. sglang/srt/models/llama4.py +487 -0
  98. sglang/srt/models/minicpm.py +1 -0
  99. sglang/srt/models/minicpm3.py +2 -0
  100. sglang/srt/models/mixtral.py +1 -0
  101. sglang/srt/models/mixtral_quant.py +1 -0
  102. sglang/srt/models/mllama.py +51 -8
  103. sglang/srt/models/mllama4.py +227 -0
  104. sglang/srt/models/olmo.py +1 -0
  105. sglang/srt/models/olmo2.py +1 -0
  106. sglang/srt/models/olmoe.py +1 -0
  107. sglang/srt/models/phi3_small.py +1 -0
  108. sglang/srt/models/qwen.py +1 -0
  109. sglang/srt/models/qwen2.py +1 -0
  110. sglang/srt/models/qwen2_5_vl.py +35 -70
  111. sglang/srt/models/qwen2_moe.py +1 -0
  112. sglang/srt/models/qwen2_vl.py +27 -25
  113. sglang/srt/models/stablelm.py +1 -0
  114. sglang/srt/models/xverse.py +1 -0
  115. sglang/srt/models/xverse_moe.py +1 -0
  116. sglang/srt/openai_api/adapter.py +4 -1
  117. sglang/srt/patch_torch.py +11 -0
  118. sglang/srt/server_args.py +34 -0
  119. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +4 -4
  120. sglang/srt/speculative/eagle_utils.py +1 -11
  121. sglang/srt/speculative/eagle_worker.py +6 -2
  122. sglang/srt/utils.py +120 -9
  123. sglang/test/attention/test_flashattn_backend.py +259 -221
  124. sglang/test/attention/test_flashattn_mla_backend.py +285 -0
  125. sglang/test/attention/test_prefix_chunk_info.py +224 -0
  126. sglang/test/test_block_fp8.py +57 -0
  127. sglang/test/test_utils.py +19 -8
  128. sglang/version.py +1 -1
  129. {sglang-0.4.4.post4.dist-info → sglang-0.4.5.post1.dist-info}/METADATA +14 -4
  130. {sglang-0.4.4.post4.dist-info → sglang-0.4.5.post1.dist-info}/RECORD +133 -109
  131. sglang/srt/disaggregation/conn.py +0 -81
  132. {sglang-0.4.4.post4.dist-info → sglang-0.4.5.post1.dist-info}/WHEEL +0 -0
  133. {sglang-0.4.4.post4.dist-info → sglang-0.4.5.post1.dist-info}/licenses/LICENSE +0 -0
  134. {sglang-0.4.4.post4.dist-info → sglang-0.4.5.post1.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=128,N=512,device_name=NVIDIA_H100_80GB_HBM3.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": 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": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 5
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": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 1,
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": 16,
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": 4
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": 32,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 16,
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": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 2
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 8,
80
+ "num_stages": 3
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 32,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 16,
87
+ "num_warps": 4,
88
+ "num_stages": 5
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": 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": 8,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 16,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 256,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 2
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 32,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 1,
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": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 64,
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": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 2
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=144,N=512,device_name=NVIDIA_H100_80GB_HBM3.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": 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": 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": 128,
22
+ "GROUP_SIZE_M": 32,
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": 64,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 128,
37
+ "BLOCK_SIZE_K": 64,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 5
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": 128,
53
+ "BLOCK_SIZE_K": 64,
54
+ "GROUP_SIZE_M": 32,
55
+ "num_warps": 8,
56
+ "num_stages": 3
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": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 64,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 16,
76
+ "BLOCK_SIZE_N": 32,
77
+ "BLOCK_SIZE_K": 64,
78
+ "GROUP_SIZE_M": 32,
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": 1,
87
+ "num_warps": 8,
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": 8,
96
+ "num_stages": 4
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": 8,
104
+ "num_stages": 2
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 2
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 32,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 4,
120
+ "num_stages": 2
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 32,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 2
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1024,device_name=NVIDIA_H100_80GB_HBM3.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 32,
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": 64,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 256,
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": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 16,
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": 8,
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": 64,
47
+ "num_warps": 8,
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": 32,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 256,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 16,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
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": 3
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": 8,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 64,
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": 64,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
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": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 8,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1024,device_name=NVIDIA_H200.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 5
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": 5
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": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 16,
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": 64,
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": 64,
47
+ "num_warps": 8,
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": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 16,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
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": 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": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 32,
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": 64,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 64,
127
+ "num_warps": 8,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 8,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }