sglang 0.4.2.post4__py3-none-any.whl → 0.4.3__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 (57) hide show
  1. sglang/global_config.py +2 -0
  2. sglang/srt/entrypoints/engine.py +2 -2
  3. sglang/srt/layers/attention/flashinfer_backend.py +235 -110
  4. sglang/srt/layers/attention/triton_backend.py +358 -72
  5. sglang/srt/layers/attention/triton_ops/extend_attention.py +4 -4
  6. sglang/srt/layers/linear.py +12 -5
  7. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8,block_shape=[128, 128].json +2 -2
  8. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  9. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +2 -2
  10. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json +200 -0
  11. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json +200 -0
  12. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json +200 -0
  13. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +178 -0
  14. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json +200 -0
  15. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +175 -0
  16. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +9 -2
  17. sglang/srt/layers/moe/fused_moe_triton/layer.py +2 -0
  18. sglang/srt/layers/moe/topk.py +1 -1
  19. sglang/srt/layers/quantization/__init__.py +51 -5
  20. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  21. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +30 -30
  22. sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  23. sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +29 -29
  24. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  25. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +33 -33
  26. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  27. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +31 -31
  28. sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  29. sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +27 -27
  30. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  31. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +31 -31
  32. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  33. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +24 -24
  34. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  35. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +30 -30
  36. sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  37. sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +42 -42
  38. sglang/srt/layers/quantization/fp8_kernel.py +123 -17
  39. sglang/srt/layers/quantization/fp8_utils.py +33 -4
  40. sglang/srt/managers/detokenizer_manager.py +1 -0
  41. sglang/srt/managers/io_struct.py +4 -0
  42. sglang/srt/managers/schedule_batch.py +16 -3
  43. sglang/srt/managers/scheduler.py +29 -0
  44. sglang/srt/managers/tokenizer_manager.py +6 -0
  45. sglang/srt/managers/tp_worker_overlap_thread.py +4 -0
  46. sglang/srt/model_executor/cuda_graph_runner.py +12 -1
  47. sglang/srt/model_executor/model_runner.py +12 -2
  48. sglang/srt/models/deepseek_v2.py +17 -7
  49. sglang/srt/server_args.py +20 -1
  50. sglang/srt/speculative/eagle_worker.py +28 -8
  51. sglang/srt/utils.py +7 -0
  52. sglang/version.py +1 -1
  53. {sglang-0.4.2.post4.dist-info → sglang-0.4.3.dist-info}/METADATA +4 -3
  54. {sglang-0.4.2.post4.dist-info → sglang-0.4.3.dist-info}/RECORD +57 -41
  55. {sglang-0.4.2.post4.dist-info → sglang-0.4.3.dist-info}/LICENSE +0 -0
  56. {sglang-0.4.2.post4.dist-info → sglang-0.4.3.dist-info}/WHEEL +0 -0
  57. {sglang-0.4.2.post4.dist-info → sglang-0.4.3.dist-info}/top_level.txt +0 -0
sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json ADDED
@@ -0,0 +1,164 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 16,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 2,
9
+ "waves_per_eu": 0
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 64,
13
+ "BLOCK_SIZE_N": 16,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 1,
16
+ "num_warps": 4,
17
+ "num_stages": 2,
18
+ "waves_per_eu": 0
19
+ },
20
+ "4": {
21
+ "BLOCK_SIZE_M": 64,
22
+ "BLOCK_SIZE_N": 16,
23
+ "BLOCK_SIZE_K": 128,
24
+ "GROUP_SIZE_M": 4,
25
+ "num_warps": 4,
26
+ "num_stages": 2,
27
+ "waves_per_eu": 0
28
+ },
29
+ "8": {
30
+ "BLOCK_SIZE_M": 64,
31
+ "BLOCK_SIZE_N": 16,
32
+ "BLOCK_SIZE_K": 128,
33
+ "GROUP_SIZE_M": 1,
34
+ "num_warps": 4,
35
+ "num_stages": 2,
36
+ "waves_per_eu": 0
37
+ },
38
+ "16": {
39
+ "BLOCK_SIZE_M": 64,
40
+ "BLOCK_SIZE_N": 16,
41
+ "BLOCK_SIZE_K": 128,
42
+ "GROUP_SIZE_M": 1,
43
+ "num_warps": 4,
44
+ "num_stages": 2,
45
+ "waves_per_eu": 0
46
+ },
47
+ "24": {
48
+ "BLOCK_SIZE_M": 64,
49
+ "BLOCK_SIZE_N": 16,
50
+ "BLOCK_SIZE_K": 128,
51
+ "GROUP_SIZE_M": 1,
52
+ "num_warps": 4,
53
+ "num_stages": 2,
54
+ "waves_per_eu": 0
55
+ },
56
+ "32": {
57
+ "BLOCK_SIZE_M": 64,
58
+ "BLOCK_SIZE_N": 16,
59
+ "BLOCK_SIZE_K": 128,
60
+ "GROUP_SIZE_M": 1,
61
+ "num_warps": 4,
62
+ "num_stages": 2,
63
+ "waves_per_eu": 0
64
+ },
65
+ "48": {
66
+ "BLOCK_SIZE_M": 64,
67
+ "BLOCK_SIZE_N": 16,
68
+ "BLOCK_SIZE_K": 128,
69
+ "GROUP_SIZE_M": 1,
70
+ "num_warps": 4,
71
+ "num_stages": 2,
72
+ "waves_per_eu": 0
73
+ },
74
+ "64": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 16,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2,
81
+ "waves_per_eu": 0
82
+ },
83
+ "96": {
84
+ "BLOCK_SIZE_M": 64,
85
+ "BLOCK_SIZE_N": 16,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 2,
90
+ "waves_per_eu": 0
91
+ },
92
+ "128": {
93
+ "BLOCK_SIZE_M": 64,
94
+ "BLOCK_SIZE_N": 16,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 1,
97
+ "num_warps": 4,
98
+ "num_stages": 2,
99
+ "waves_per_eu": 0
100
+ },
101
+ "256": {
102
+ "BLOCK_SIZE_M": 128,
103
+ "BLOCK_SIZE_N": 32,
104
+ "BLOCK_SIZE_K": 128,
105
+ "GROUP_SIZE_M": 1,
106
+ "num_warps": 4,
107
+ "num_stages": 2,
108
+ "waves_per_eu": 0
109
+ },
110
+ "512": {
111
+ "BLOCK_SIZE_M": 64,
112
+ "BLOCK_SIZE_N": 128,
113
+ "BLOCK_SIZE_K": 128,
114
+ "GROUP_SIZE_M": 4,
115
+ "num_warps": 4,
116
+ "num_stages": 2,
117
+ "waves_per_eu": 0
118
+ },
119
+ "1024": {
120
+ "BLOCK_SIZE_M": 64,
121
+ "BLOCK_SIZE_N": 64,
122
+ "BLOCK_SIZE_K": 128,
123
+ "GROUP_SIZE_M": 1,
124
+ "num_warps": 4,
125
+ "num_stages": 2,
126
+ "waves_per_eu": 0
127
+ },
128
+ "1536": {
129
+ "BLOCK_SIZE_M": 64,
130
+ "BLOCK_SIZE_N": 64,
131
+ "BLOCK_SIZE_K": 128,
132
+ "GROUP_SIZE_M": 1,
133
+ "num_warps": 4,
134
+ "num_stages": 2,
135
+ "waves_per_eu": 0
136
+ },
137
+ "2048": {
138
+ "BLOCK_SIZE_M": 64,
139
+ "BLOCK_SIZE_N": 128,
140
+ "BLOCK_SIZE_K": 128,
141
+ "GROUP_SIZE_M": 4,
142
+ "num_warps": 4,
143
+ "num_stages": 2,
144
+ "waves_per_eu": 0
145
+ },
146
+ "3072": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 128,
150
+ "GROUP_SIZE_M": 1,
151
+ "num_warps": 4,
152
+ "num_stages": 2,
153
+ "waves_per_eu": 0
154
+ },
155
+ "4096": {
156
+ "BLOCK_SIZE_M": 64,
157
+ "BLOCK_SIZE_N": 128,
158
+ "BLOCK_SIZE_K": 128,
159
+ "GROUP_SIZE_M": 1,
160
+ "num_warps": 4,
161
+ "num_stages": 2,
162
+ "waves_per_eu": 0
163
+ }
164
+ }
sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json CHANGED
@@ -1,70 +1,70 @@
1
1
  {
2
2
  "1": {
3
- "BLOCK_SIZE_M": 64,
3
+ "BLOCK_SIZE_M": 32,
4
4
  "BLOCK_SIZE_N": 16,
5
5
  "BLOCK_SIZE_K": 128,
6
- "GROUP_SIZE_M": 16,
6
+ "GROUP_SIZE_M": 1,
7
7
  "num_warps": 4,
8
8
  "num_stages": 2,
9
9
  "waves_per_eu": 0
10
10
  },
11
11
  "2": {
12
- "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_M": 32,
13
13
  "BLOCK_SIZE_N": 16,
14
14
  "BLOCK_SIZE_K": 128,
15
- "GROUP_SIZE_M": 1,
15
+ "GROUP_SIZE_M": 8,
16
16
  "num_warps": 4,
17
17
  "num_stages": 2,
18
18
  "waves_per_eu": 0
19
19
  },
20
20
  "4": {
21
- "BLOCK_SIZE_M": 64,
21
+ "BLOCK_SIZE_M": 32,
22
22
  "BLOCK_SIZE_N": 16,
23
23
  "BLOCK_SIZE_K": 128,
24
- "GROUP_SIZE_M": 4,
24
+ "GROUP_SIZE_M": 1,
25
25
  "num_warps": 4,
26
26
  "num_stages": 2,
27
27
  "waves_per_eu": 0
28
28
  },
29
29
  "8": {
30
- "BLOCK_SIZE_M": 64,
30
+ "BLOCK_SIZE_M": 32,
31
31
  "BLOCK_SIZE_N": 16,
32
32
  "BLOCK_SIZE_K": 128,
33
- "GROUP_SIZE_M": 1,
33
+ "GROUP_SIZE_M": 16,
34
34
  "num_warps": 4,
35
35
  "num_stages": 2,
36
36
  "waves_per_eu": 0
37
37
  },
38
38
  "16": {
39
- "BLOCK_SIZE_M": 64,
39
+ "BLOCK_SIZE_M": 32,
40
40
  "BLOCK_SIZE_N": 16,
41
41
  "BLOCK_SIZE_K": 128,
42
- "GROUP_SIZE_M": 1,
42
+ "GROUP_SIZE_M": 8,
43
43
  "num_warps": 4,
44
44
  "num_stages": 2,
45
45
  "waves_per_eu": 0
46
46
  },
47
47
  "24": {
48
- "BLOCK_SIZE_M": 64,
48
+ "BLOCK_SIZE_M": 32,
49
49
  "BLOCK_SIZE_N": 16,
50
50
  "BLOCK_SIZE_K": 128,
51
- "GROUP_SIZE_M": 1,
51
+ "GROUP_SIZE_M": 16,
52
52
  "num_warps": 4,
53
53
  "num_stages": 2,
54
54
  "waves_per_eu": 0
55
55
  },
56
56
  "32": {
57
- "BLOCK_SIZE_M": 64,
57
+ "BLOCK_SIZE_M": 32,
58
58
  "BLOCK_SIZE_N": 16,
59
59
  "BLOCK_SIZE_K": 128,
60
- "GROUP_SIZE_M": 1,
60
+ "GROUP_SIZE_M": 4,
61
61
  "num_warps": 4,
62
62
  "num_stages": 2,
63
63
  "waves_per_eu": 0
64
64
  },
65
65
  "48": {
66
- "BLOCK_SIZE_M": 64,
67
- "BLOCK_SIZE_N": 16,
66
+ "BLOCK_SIZE_M": 32,
67
+ "BLOCK_SIZE_N": 128,
68
68
  "BLOCK_SIZE_K": 128,
69
69
  "GROUP_SIZE_M": 1,
70
70
  "num_warps": 4,
@@ -72,8 +72,8 @@
72
72
  "waves_per_eu": 0
73
73
  },
74
74
  "64": {
75
- "BLOCK_SIZE_M": 64,
76
- "BLOCK_SIZE_N": 16,
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 128,
77
77
  "BLOCK_SIZE_K": 128,
78
78
  "GROUP_SIZE_M": 1,
79
79
  "num_warps": 4,
@@ -81,8 +81,8 @@
81
81
  "waves_per_eu": 0
82
82
  },
83
83
  "96": {
84
- "BLOCK_SIZE_M": 64,
85
- "BLOCK_SIZE_N": 16,
84
+ "BLOCK_SIZE_M": 32,
85
+ "BLOCK_SIZE_N": 128,
86
86
  "BLOCK_SIZE_K": 128,
87
87
  "GROUP_SIZE_M": 1,
88
88
  "num_warps": 4,
@@ -91,7 +91,7 @@
91
91
  },
92
92
  "128": {
93
93
  "BLOCK_SIZE_M": 64,
94
- "BLOCK_SIZE_N": 16,
94
+ "BLOCK_SIZE_N": 64,
95
95
  "BLOCK_SIZE_K": 128,
96
96
  "GROUP_SIZE_M": 1,
97
97
  "num_warps": 4,
@@ -99,8 +99,8 @@
99
99
  "waves_per_eu": 0
100
100
  },
101
101
  "256": {
102
- "BLOCK_SIZE_M": 128,
103
- "BLOCK_SIZE_N": 32,
102
+ "BLOCK_SIZE_M": 32,
103
+ "BLOCK_SIZE_N": 128,
104
104
  "BLOCK_SIZE_K": 128,
105
105
  "GROUP_SIZE_M": 1,
106
106
  "num_warps": 4,
@@ -111,14 +111,14 @@
111
111
  "BLOCK_SIZE_M": 64,
112
112
  "BLOCK_SIZE_N": 128,
113
113
  "BLOCK_SIZE_K": 128,
114
- "GROUP_SIZE_M": 4,
114
+ "GROUP_SIZE_M": 1,
115
115
  "num_warps": 4,
116
116
  "num_stages": 2,
117
117
  "waves_per_eu": 0
118
118
  },
119
119
  "1024": {
120
120
  "BLOCK_SIZE_M": 64,
121
- "BLOCK_SIZE_N": 64,
121
+ "BLOCK_SIZE_N": 128,
122
122
  "BLOCK_SIZE_K": 128,
123
123
  "GROUP_SIZE_M": 1,
124
124
  "num_warps": 4,
@@ -127,9 +127,9 @@
127
127
  },
128
128
  "1536": {
129
129
  "BLOCK_SIZE_M": 64,
130
- "BLOCK_SIZE_N": 64,
130
+ "BLOCK_SIZE_N": 128,
131
131
  "BLOCK_SIZE_K": 128,
132
- "GROUP_SIZE_M": 1,
132
+ "GROUP_SIZE_M": 32,
133
133
  "num_warps": 4,
134
134
  "num_stages": 2,
135
135
  "waves_per_eu": 0
@@ -138,7 +138,7 @@
138
138
  "BLOCK_SIZE_M": 64,
139
139
  "BLOCK_SIZE_N": 128,
140
140
  "BLOCK_SIZE_K": 128,
141
- "GROUP_SIZE_M": 4,
141
+ "GROUP_SIZE_M": 32,
142
142
  "num_warps": 4,
143
143
  "num_stages": 2,
144
144
  "waves_per_eu": 0
@@ -147,7 +147,7 @@
147
147
  "BLOCK_SIZE_M": 64,
148
148
  "BLOCK_SIZE_N": 128,
149
149
  "BLOCK_SIZE_K": 128,
150
- "GROUP_SIZE_M": 1,
150
+ "GROUP_SIZE_M": 32,
151
151
  "num_warps": 4,
152
152
  "num_stages": 2,
153
153
  "waves_per_eu": 0
@@ -156,7 +156,7 @@
156
156
  "BLOCK_SIZE_M": 64,
157
157
  "BLOCK_SIZE_N": 128,
158
158
  "BLOCK_SIZE_K": 128,
159
- "GROUP_SIZE_M": 1,
159
+ "GROUP_SIZE_M": 32,
160
160
  "num_warps": 4,
161
161
  "num_stages": 2,
162
162
  "waves_per_eu": 0
sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json ADDED
@@ -0,0 +1,164 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 64,
4
+ "BLOCK_SIZE_N": 16,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 4,
7
+ "num_warps": 4,
8
+ "num_stages": 2,
9
+ "waves_per_eu": 0
10
+ },
11
+ "2": {
12
+ "BLOCK_SIZE_M": 64,
13
+ "BLOCK_SIZE_N": 16,
14
+ "BLOCK_SIZE_K": 128,
15
+ "GROUP_SIZE_M": 32,
16
+ "num_warps": 4,
17
+ "num_stages": 2,
18
+ "waves_per_eu": 0
19
+ },
20
+ "4": {
21
+ "BLOCK_SIZE_M": 64,
22
+ "BLOCK_SIZE_N": 16,
23
+ "BLOCK_SIZE_K": 128,
24
+ "GROUP_SIZE_M": 1,
25
+ "num_warps": 4,
26
+ "num_stages": 2,
27
+ "waves_per_eu": 0
28
+ },
29
+ "8": {
30
+ "BLOCK_SIZE_M": 64,
31
+ "BLOCK_SIZE_N": 16,
32
+ "BLOCK_SIZE_K": 128,
33
+ "GROUP_SIZE_M": 4,
34
+ "num_warps": 4,
35
+ "num_stages": 2,
36
+ "waves_per_eu": 0
37
+ },
38
+ "16": {
39
+ "BLOCK_SIZE_M": 64,
40
+ "BLOCK_SIZE_N": 16,
41
+ "BLOCK_SIZE_K": 128,
42
+ "GROUP_SIZE_M": 16,
43
+ "num_warps": 4,
44
+ "num_stages": 2,
45
+ "waves_per_eu": 0
46
+ },
47
+ "24": {
48
+ "BLOCK_SIZE_M": 64,
49
+ "BLOCK_SIZE_N": 16,
50
+ "BLOCK_SIZE_K": 128,
51
+ "GROUP_SIZE_M": 16,
52
+ "num_warps": 4,
53
+ "num_stages": 2,
54
+ "waves_per_eu": 0
55
+ },
56
+ "32": {
57
+ "BLOCK_SIZE_M": 64,
58
+ "BLOCK_SIZE_N": 16,
59
+ "BLOCK_SIZE_K": 128,
60
+ "GROUP_SIZE_M": 16,
61
+ "num_warps": 4,
62
+ "num_stages": 2,
63
+ "waves_per_eu": 0
64
+ },
65
+ "48": {
66
+ "BLOCK_SIZE_M": 64,
67
+ "BLOCK_SIZE_N": 16,
68
+ "BLOCK_SIZE_K": 128,
69
+ "GROUP_SIZE_M": 16,
70
+ "num_warps": 4,
71
+ "num_stages": 2,
72
+ "waves_per_eu": 0
73
+ },
74
+ "64": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 16,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
79
+ "num_warps": 4,
80
+ "num_stages": 2,
81
+ "waves_per_eu": 0
82
+ },
83
+ "96": {
84
+ "BLOCK_SIZE_M": 64,
85
+ "BLOCK_SIZE_N": 16,
86
+ "BLOCK_SIZE_K": 128,
87
+ "GROUP_SIZE_M": 1,
88
+ "num_warps": 4,
89
+ "num_stages": 2,
90
+ "waves_per_eu": 0
91
+ },
92
+ "128": {
93
+ "BLOCK_SIZE_M": 64,
94
+ "BLOCK_SIZE_N": 32,
95
+ "BLOCK_SIZE_K": 128,
96
+ "GROUP_SIZE_M": 1,
97
+ "num_warps": 4,
98
+ "num_stages": 2,
99
+ "waves_per_eu": 0
100
+ },
101
+ "256": {
102
+ "BLOCK_SIZE_M": 64,
103
+ "BLOCK_SIZE_N": 32,
104
+ "BLOCK_SIZE_K": 128,
105
+ "GROUP_SIZE_M": 1,
106
+ "num_warps": 4,
107
+ "num_stages": 2,
108
+ "waves_per_eu": 0
109
+ },
110
+ "512": {
111
+ "BLOCK_SIZE_M": 128,
112
+ "BLOCK_SIZE_N": 32,
113
+ "BLOCK_SIZE_K": 128,
114
+ "GROUP_SIZE_M": 32,
115
+ "num_warps": 4,
116
+ "num_stages": 2,
117
+ "waves_per_eu": 0
118
+ },
119
+ "1024": {
120
+ "BLOCK_SIZE_M": 64,
121
+ "BLOCK_SIZE_N": 64,
122
+ "BLOCK_SIZE_K": 128,
123
+ "GROUP_SIZE_M": 4,
124
+ "num_warps": 4,
125
+ "num_stages": 2,
126
+ "waves_per_eu": 0
127
+ },
128
+ "1536": {
129
+ "BLOCK_SIZE_M": 64,
130
+ "BLOCK_SIZE_N": 64,
131
+ "BLOCK_SIZE_K": 128,
132
+ "GROUP_SIZE_M": 1,
133
+ "num_warps": 4,
134
+ "num_stages": 2,
135
+ "waves_per_eu": 0
136
+ },
137
+ "2048": {
138
+ "BLOCK_SIZE_M": 64,
139
+ "BLOCK_SIZE_N": 128,
140
+ "BLOCK_SIZE_K": 128,
141
+ "GROUP_SIZE_M": 1,
142
+ "num_warps": 4,
143
+ "num_stages": 2,
144
+ "waves_per_eu": 0
145
+ },
146
+ "3072": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 128,
149
+ "BLOCK_SIZE_K": 128,
150
+ "GROUP_SIZE_M": 4,
151
+ "num_warps": 4,
152
+ "num_stages": 2,
153
+ "waves_per_eu": 0
154
+ },
155
+ "4096": {
156
+ "BLOCK_SIZE_M": 64,
157
+ "BLOCK_SIZE_N": 128,
158
+ "BLOCK_SIZE_K": 128,
159
+ "GROUP_SIZE_M": 1,
160
+ "num_warps": 4,
161
+ "num_stages": 2,
162
+ "waves_per_eu": 0
163
+ }
164
+ }