sglang 0.4.4.post2__py3-none-any.whl → 0.4.4.post4__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 (108) hide show
  1. sglang/bench_serving.py +72 -10
  2. sglang/srt/_custom_ops.py +59 -92
  3. sglang/srt/configs/deepseekvl2.py +10 -1
  4. sglang/srt/configs/model_config.py +6 -16
  5. sglang/srt/constrained/base_grammar_backend.py +5 -1
  6. sglang/srt/custom_op.py +5 -0
  7. sglang/srt/distributed/device_communicators/custom_all_reduce.py +28 -80
  8. sglang/srt/distributed/device_communicators/custom_all_reduce_utils.py +2 -2
  9. sglang/srt/distributed/parallel_state.py +32 -5
  10. sglang/srt/entrypoints/engine.py +0 -5
  11. sglang/srt/entrypoints/http_server.py +7 -1
  12. sglang/srt/entrypoints/verl_engine.py +2 -0
  13. sglang/srt/function_call_parser.py +0 -1
  14. sglang/srt/layers/attention/flashattention_backend.py +582 -125
  15. sglang/srt/layers/attention/flashinfer_backend.py +5 -7
  16. sglang/srt/layers/attention/flashinfer_mla_backend.py +1 -3
  17. sglang/srt/layers/attention/flashmla_backend.py +1 -1
  18. sglang/srt/layers/dp_attention.py +12 -1
  19. sglang/srt/layers/moe/ep_moe/kernels.py +142 -0
  20. sglang/srt/layers/moe/ep_moe/layer.py +79 -80
  21. sglang/srt/layers/moe/ep_moe/token_dispatcher.py +382 -199
  22. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_H20,block_shape=[128, 128].json +146 -0
  23. sglang/srt/layers/moe/fused_moe_triton/configs/E=257,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  24. sglang/srt/layers/moe/fused_moe_triton/configs/E=264,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  25. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +403 -47
  26. sglang/srt/layers/moe/topk.py +79 -6
  27. sglang/srt/layers/quantization/__init__.py +137 -165
  28. sglang/srt/layers/quantization/awq.py +200 -0
  29. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors.py +2 -1
  30. sglang/srt/layers/quantization/compressed_tensors/compressed_tensors_moe.py +34 -10
  31. sglang/srt/layers/quantization/fp8_kernel.py +2 -1
  32. sglang/srt/layers/quantization/fp8_utils.py +1 -4
  33. sglang/srt/layers/quantization/gptq.py +30 -40
  34. sglang/srt/layers/quantization/moe_wna16.py +501 -0
  35. sglang/srt/layers/quantization/utils.py +1 -1
  36. sglang/srt/layers/quantization/w8a8_fp8.py +1 -1
  37. sglang/srt/lora/backend/base_backend.py +4 -4
  38. sglang/srt/lora/backend/flashinfer_backend.py +12 -9
  39. sglang/srt/lora/backend/triton_backend.py +5 -8
  40. sglang/srt/lora/layers.py +19 -33
  41. sglang/srt/lora/lora_manager.py +20 -7
  42. sglang/srt/lora/mem_pool.py +12 -6
  43. sglang/srt/lora/triton_ops/gate_up_lora_b.py +10 -4
  44. sglang/srt/lora/triton_ops/qkv_lora_b.py +8 -3
  45. sglang/srt/lora/triton_ops/sgemm_lora_a.py +16 -5
  46. sglang/srt/lora/triton_ops/sgemm_lora_b.py +11 -6
  47. sglang/srt/lora/utils.py +6 -0
  48. sglang/srt/managers/cache_controller.py +34 -11
  49. sglang/srt/managers/io_struct.py +4 -2
  50. sglang/srt/managers/mm_utils.py +202 -156
  51. sglang/srt/managers/multimodal_processor.py +0 -2
  52. sglang/srt/managers/multimodal_processors/base_processor.py +45 -77
  53. sglang/srt/managers/multimodal_processors/clip.py +44 -0
  54. sglang/srt/managers/multimodal_processors/deepseek_vl_v2.py +17 -58
  55. sglang/srt/managers/multimodal_processors/gemma3.py +12 -27
  56. sglang/srt/managers/multimodal_processors/janus_pro.py +21 -47
  57. sglang/srt/managers/multimodal_processors/llava.py +34 -14
  58. sglang/srt/managers/multimodal_processors/minicpm.py +35 -38
  59. sglang/srt/managers/multimodal_processors/mlama.py +10 -23
  60. sglang/srt/managers/multimodal_processors/qwen_vl.py +22 -45
  61. sglang/srt/managers/schedule_batch.py +185 -127
  62. sglang/srt/managers/scheduler.py +29 -23
  63. sglang/srt/managers/tokenizer_manager.py +1 -2
  64. sglang/srt/managers/tp_worker.py +3 -0
  65. sglang/srt/managers/utils.py +1 -6
  66. sglang/srt/mem_cache/hiradix_cache.py +62 -52
  67. sglang/srt/mem_cache/memory_pool.py +72 -6
  68. sglang/srt/mem_cache/paged_allocator.py +39 -0
  69. sglang/srt/metrics/collector.py +23 -53
  70. sglang/srt/model_executor/cuda_graph_runner.py +16 -13
  71. sglang/srt/model_executor/forward_batch_info.py +10 -10
  72. sglang/srt/model_executor/model_runner.py +64 -59
  73. sglang/srt/model_loader/loader.py +19 -1
  74. sglang/srt/model_loader/weight_utils.py +6 -3
  75. sglang/srt/models/clip.py +568 -0
  76. sglang/srt/models/deepseek_janus_pro.py +12 -17
  77. sglang/srt/models/deepseek_v2.py +339 -123
  78. sglang/srt/models/deepseek_vl2.py +105 -104
  79. sglang/srt/models/gemma3_causal.py +12 -2
  80. sglang/srt/models/gemma3_mm.py +20 -80
  81. sglang/srt/models/llama.py +4 -1
  82. sglang/srt/models/llava.py +31 -19
  83. sglang/srt/models/llavavid.py +16 -7
  84. sglang/srt/models/minicpmo.py +63 -147
  85. sglang/srt/models/minicpmv.py +17 -27
  86. sglang/srt/models/mllama.py +29 -14
  87. sglang/srt/models/qwen2.py +9 -6
  88. sglang/srt/models/qwen2_5_vl.py +21 -31
  89. sglang/srt/models/qwen2_vl.py +20 -21
  90. sglang/srt/openai_api/adapter.py +106 -93
  91. sglang/srt/openai_api/protocol.py +10 -5
  92. sglang/srt/patch_torch.py +71 -0
  93. sglang/srt/platforms/interface.py +371 -0
  94. sglang/srt/server_args.py +120 -25
  95. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +5 -5
  96. sglang/srt/speculative/eagle_utils.py +140 -28
  97. sglang/srt/speculative/eagle_worker.py +94 -25
  98. sglang/srt/utils.py +137 -51
  99. sglang/test/runners.py +27 -2
  100. sglang/test/test_custom_ops.py +55 -0
  101. sglang/test/test_utils.py +14 -27
  102. sglang/utils.py +2 -2
  103. sglang/version.py +1 -1
  104. {sglang-0.4.4.post2.dist-info → sglang-0.4.4.post4.dist-info}/METADATA +10 -5
  105. {sglang-0.4.4.post2.dist-info → sglang-0.4.4.post4.dist-info}/RECORD +108 -99
  106. {sglang-0.4.4.post2.dist-info → sglang-0.4.4.post4.dist-info}/WHEEL +0 -0
  107. {sglang-0.4.4.post2.dist-info → sglang-0.4.4.post4.dist-info}/licenses/LICENSE +0 -0
  108. {sglang-0.4.4.post2.dist-info → sglang-0.4.4.post4.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_H20,block_shape=[128, 128].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": 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": 64,
14
+ "GROUP_SIZE_M": 32,
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": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 64,
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": 64,
38
+ "GROUP_SIZE_M": 32,
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": 64,
46
+ "GROUP_SIZE_M": 32,
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": 64,
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": 64,
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": 256,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 16,
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": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 64,
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": 64,
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": 256,
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": 64,
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": 64,
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": 64,
126
+ "GROUP_SIZE_M": 32,
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": 64,
134
+ "GROUP_SIZE_M": 16,
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": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 2
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 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": 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": 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": 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": 4
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": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 32,
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": 64,
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": 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": 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": 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": 16,
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": 16,
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": 4
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=264,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": 64,
4
+ "BLOCK_SIZE_N": 64,
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": 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": 1,
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": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
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": 64,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 16,
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": 32,
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": 64,
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": 16,
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": 32,
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": 32,
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": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 64,
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": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }