sglang 0.4.1.post3__py3-none-any.whl → 0.4.1.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 (63) hide show
  1. sglang/bench_one_batch.py +2 -0
  2. sglang/srt/layers/attention/__init__.py +14 -5
  3. sglang/srt/layers/attention/double_sparsity_backend.py +0 -52
  4. sglang/srt/layers/attention/flashinfer_backend.py +211 -81
  5. sglang/srt/layers/attention/torch_native_backend.py +1 -38
  6. sglang/srt/layers/attention/triton_backend.py +20 -11
  7. sglang/srt/layers/attention/triton_ops/decode_attention.py +4 -0
  8. sglang/srt/layers/logits_processor.py +167 -212
  9. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  10. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H200.json +146 -0
  11. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  12. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H200.json +146 -0
  13. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  14. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H200.json +146 -0
  15. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  16. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H200.json +146 -0
  17. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  18. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=NVIDIA_H200.json +146 -0
  19. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  20. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_H200.json +146 -0
  21. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  22. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_H200.json +146 -0
  23. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  24. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_H200.json +146 -0
  25. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  26. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H200.json +146 -0
  27. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  28. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H200.json +146 -0
  29. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=NVIDIA_H200,dtype=fp8_w8a8.json +146 -0
  30. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +187 -29
  31. sglang/srt/layers/moe/fused_moe_triton/layer.py +14 -6
  32. sglang/srt/layers/quantization/fp8.py +2 -2
  33. sglang/srt/layers/sampler.py +57 -21
  34. sglang/srt/layers/torchao_utils.py +17 -3
  35. sglang/srt/managers/io_struct.py +1 -2
  36. sglang/srt/managers/schedule_batch.py +26 -2
  37. sglang/srt/managers/schedule_policy.py +159 -90
  38. sglang/srt/managers/scheduler.py +62 -26
  39. sglang/srt/managers/tokenizer_manager.py +22 -20
  40. sglang/srt/managers/tp_worker.py +16 -4
  41. sglang/srt/managers/tp_worker_overlap_thread.py +3 -4
  42. sglang/srt/model_executor/cuda_graph_runner.py +118 -73
  43. sglang/srt/model_executor/forward_batch_info.py +33 -8
  44. sglang/srt/model_executor/model_runner.py +63 -61
  45. sglang/srt/models/deepseek_v2.py +34 -7
  46. sglang/srt/models/grok.py +97 -26
  47. sglang/srt/openai_api/adapter.py +0 -17
  48. sglang/srt/openai_api/protocol.py +3 -3
  49. sglang/srt/sampling/sampling_batch_info.py +21 -0
  50. sglang/srt/sampling/sampling_params.py +9 -1
  51. sglang/srt/server.py +9 -5
  52. sglang/srt/server_args.py +108 -57
  53. sglang/srt/speculative/build_eagle_tree.py +347 -0
  54. sglang/srt/speculative/eagle_utils.py +618 -0
  55. sglang/srt/speculative/eagle_worker.py +170 -0
  56. sglang/srt/speculative/spec_info.py +5 -0
  57. sglang/srt/utils.py +15 -2
  58. sglang/version.py +1 -1
  59. {sglang-0.4.1.post3.dist-info → sglang-0.4.1.post4.dist-info}/METADATA +9 -8
  60. {sglang-0.4.1.post3.dist-info → sglang-0.4.1.post4.dist-info}/RECORD +63 -39
  61. {sglang-0.4.1.post3.dist-info → sglang-0.4.1.post4.dist-info}/WHEEL +1 -1
  62. {sglang-0.4.1.post3.dist-info → sglang-0.4.1.post4.dist-info}/LICENSE +0 -0
  63. {sglang-0.4.1.post3.dist-info → sglang-0.4.1.post4.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H200,dtype=fp8_w8a8.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": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 64,
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": 256,
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": 256,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 256,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 8,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
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": 64,
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": 64,
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": 64,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 32,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 4
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": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 256,
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": 128,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 16,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 32,
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": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 8,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 32,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H200.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": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 32,
12
+ "BLOCK_SIZE_N": 256,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 32,
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": 16,
23
+ "num_warps": 8,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 32,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 256,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 8,
40
+ "num_stages": 2
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": 4,
48
+ "num_stages": 2
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": 32,
60
+ "BLOCK_SIZE_N": 256,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 64,
63
+ "num_warps": 4,
64
+ "num_stages": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 32,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 8,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 8,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 256,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 8,
120
+ "num_stages": 4
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": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
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": 32,
143
+ "num_warps": 8,
144
+ "num_stages": 4
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8.json ADDED
@@ -0,0 +1,146 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 128,
5
+ "BLOCK_SIZE_K": 256,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 256,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 4,
16
+ "num_stages": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 256,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 8,
24
+ "num_stages": 5
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": 5
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
39
+ "num_warps": 4,
40
+ "num_stages": 5
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": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 4
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": 4
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": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 4,
80
+ "num_stages": 4
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": 4
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": 128,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 128,
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": 128,
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": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H200.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": 5
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 16,
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": 32,
23
+ "num_warps": 8,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 32,
28
+ "BLOCK_SIZE_N": 128,
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": 128,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 16,
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": 256,
46
+ "GROUP_SIZE_M": 16,
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": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 32,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 5
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 32,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 8,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 8,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 256,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 8,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 8,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 8,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 8,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 16,
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": 64,
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": 16,
143
+ "num_warps": 8,
144
+ "num_stages": 3
145
+ }
146
+ }