sglang 0.4.2.post3__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 (88) hide show
  1. sglang/check_env.py +1 -0
  2. sglang/global_config.py +2 -0
  3. sglang/srt/constrained/outlines_backend.py +4 -1
  4. sglang/srt/entrypoints/engine.py +2 -2
  5. sglang/srt/layers/attention/flashinfer_backend.py +265 -147
  6. sglang/srt/layers/attention/triton_backend.py +358 -72
  7. sglang/srt/layers/attention/triton_ops/extend_attention.py +4 -4
  8. sglang/srt/layers/linear.py +12 -5
  9. 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
  10. 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
  11. 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
  12. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=AMD_Instinct_MI325X.json +200 -0
  13. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=AMD_Instinct_MI325X.json +200 -0
  14. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=AMD_Instinct_MI325X.json +200 -0
  15. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +178 -0
  16. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=AMD_Instinct_MI325X.json +200 -0
  17. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8.json +175 -0
  18. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +27 -5
  19. sglang/srt/layers/moe/fused_moe_triton/layer.py +2 -0
  20. sglang/srt/layers/moe/topk.py +1 -1
  21. sglang/srt/layers/quantization/__init__.py +51 -5
  22. sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  23. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  24. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +30 -30
  25. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  26. sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  27. sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  28. sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  29. sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  30. sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  31. sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +29 -29
  32. sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  33. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  34. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +33 -33
  35. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  36. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +31 -31
  37. sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  38. sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +27 -27
  39. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  40. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +31 -31
  41. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  42. sglang/srt/layers/quantization/configs/N=7168,K=1024,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  43. sglang/srt/layers/quantization/configs/N=7168,K=1152,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  44. sglang/srt/layers/quantization/configs/N=7168,K=128,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  45. sglang/srt/layers/quantization/configs/N=7168,K=16384,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  46. sglang/srt/layers/quantization/configs/N=7168,K=18432,device_name=NVIDIA_H20,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  47. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  48. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +24 -24
  49. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  50. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +30 -30
  51. sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=AMD_Instinct_MI325X,dtype=fp8_w8a8,block_shape=[128, 128].json +164 -0
  52. sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json +42 -42
  53. sglang/srt/layers/quantization/fp8_kernel.py +123 -17
  54. sglang/srt/layers/quantization/fp8_utils.py +33 -4
  55. sglang/srt/lora/backend/__init__.py +25 -5
  56. sglang/srt/lora/backend/base_backend.py +31 -9
  57. sglang/srt/lora/backend/flashinfer_backend.py +41 -4
  58. sglang/srt/lora/backend/triton_backend.py +34 -4
  59. sglang/srt/lora/layers.py +293 -0
  60. sglang/srt/lora/lora.py +101 -326
  61. sglang/srt/lora/lora_manager.py +101 -269
  62. sglang/srt/lora/mem_pool.py +174 -0
  63. sglang/srt/lora/triton_ops/__init__.py +7 -1
  64. sglang/srt/lora/triton_ops/gate_up_lora_b.py +170 -0
  65. sglang/srt/lora/triton_ops/qkv_lora_b.py +5 -5
  66. sglang/srt/lora/triton_ops/sgemm_lora_a.py +2 -2
  67. sglang/srt/lora/triton_ops/sgemm_lora_b.py +2 -2
  68. sglang/srt/lora/utils.py +141 -0
  69. sglang/srt/managers/detokenizer_manager.py +1 -0
  70. sglang/srt/managers/io_struct.py +4 -0
  71. sglang/srt/managers/schedule_batch.py +16 -3
  72. sglang/srt/managers/scheduler.py +29 -0
  73. sglang/srt/managers/tokenizer_manager.py +6 -0
  74. sglang/srt/managers/tp_worker_overlap_thread.py +4 -0
  75. sglang/srt/model_executor/cuda_graph_runner.py +16 -1
  76. sglang/srt/model_executor/model_runner.py +12 -2
  77. sglang/srt/models/deepseek_v2.py +17 -7
  78. sglang/srt/server_args.py +20 -1
  79. sglang/srt/speculative/eagle_draft_cuda_graph_runner.py +1 -0
  80. sglang/srt/speculative/eagle_utils.py +64 -21
  81. sglang/srt/speculative/eagle_worker.py +29 -8
  82. sglang/srt/utils.py +7 -0
  83. sglang/version.py +1 -1
  84. {sglang-0.4.2.post3.dist-info → sglang-0.4.3.dist-info}/METADATA +6 -5
  85. {sglang-0.4.2.post3.dist-info → sglang-0.4.3.dist-info}/RECORD +88 -55
  86. {sglang-0.4.2.post3.dist-info → sglang-0.4.3.dist-info}/LICENSE +0 -0
  87. {sglang-0.4.2.post3.dist-info → sglang-0.4.3.dist-info}/WHEEL +0 -0
  88. {sglang-0.4.2.post3.dist-info → sglang-0.4.3.dist-info}/top_level.txt +0 -0
sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=AMD_Radeon_Graphics,dtype=fp8_w8a8,block_shape=[128, 128].json CHANGED
@@ -1,51 +1,51 @@
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": 4,
6
+ "GROUP_SIZE_M": 8,
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": 8,
15
+ "GROUP_SIZE_M": 32,
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": 1,
24
+ "GROUP_SIZE_M": 16,
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": 16,
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
51
  "GROUP_SIZE_M": 32,
@@ -54,10 +54,10 @@
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
@@ -66,7 +66,7 @@
66
66
  "BLOCK_SIZE_M": 64,
67
67
  "BLOCK_SIZE_N": 16,
68
68
  "BLOCK_SIZE_K": 128,
69
- "GROUP_SIZE_M": 1,
69
+ "GROUP_SIZE_M": 16,
70
70
  "num_warps": 4,
71
71
  "num_stages": 2,
72
72
  "waves_per_eu": 0
@@ -75,7 +75,7 @@
75
75
  "BLOCK_SIZE_M": 64,
76
76
  "BLOCK_SIZE_N": 16,
77
77
  "BLOCK_SIZE_K": 128,
78
- "GROUP_SIZE_M": 32,
78
+ "GROUP_SIZE_M": 1,
79
79
  "num_warps": 4,
80
80
  "num_stages": 2,
81
81
  "waves_per_eu": 0
@@ -102,43 +102,43 @@
102
102
  "BLOCK_SIZE_M": 64,
103
103
  "BLOCK_SIZE_N": 16,
104
104
  "BLOCK_SIZE_K": 128,
105
- "GROUP_SIZE_M": 4,
105
+ "GROUP_SIZE_M": 1,
106
106
  "num_warps": 4,
107
107
  "num_stages": 2,
108
108
  "waves_per_eu": 0
109
109
  },
110
110
  "512": {
111
- "BLOCK_SIZE_M": 64,
112
- "BLOCK_SIZE_N": 16,
111
+ "BLOCK_SIZE_M": 32,
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
- "BLOCK_SIZE_M": 64,
121
- "BLOCK_SIZE_N": 16,
120
+ "BLOCK_SIZE_M": 32,
121
+ "BLOCK_SIZE_N": 128,
122
122
  "BLOCK_SIZE_K": 128,
123
- "GROUP_SIZE_M": 4,
123
+ "GROUP_SIZE_M": 1,
124
124
  "num_warps": 4,
125
125
  "num_stages": 2,
126
126
  "waves_per_eu": 0
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": 8,
132
+ "GROUP_SIZE_M": 4,
133
133
  "num_warps": 4,
134
134
  "num_stages": 2,
135
135
  "waves_per_eu": 0
136
136
  },
137
137
  "2048": {
138
- "BLOCK_SIZE_M": 128,
139
- "BLOCK_SIZE_N": 32,
138
+ "BLOCK_SIZE_M": 32,
139
+ "BLOCK_SIZE_N": 128,
140
140
  "BLOCK_SIZE_K": 128,
141
- "GROUP_SIZE_M": 8,
141
+ "GROUP_SIZE_M": 1,
142
142
  "num_warps": 4,
143
143
  "num_stages": 2,
144
144
  "waves_per_eu": 0
@@ -147,16 +147,16 @@
147
147
  "BLOCK_SIZE_M": 64,
148
148
  "BLOCK_SIZE_N": 128,
149
149
  "BLOCK_SIZE_K": 128,
150
- "GROUP_SIZE_M": 16,
150
+ "GROUP_SIZE_M": 4,
151
151
  "num_warps": 4,
152
152
  "num_stages": 2,
153
153
  "waves_per_eu": 0
154
154
  },
155
155
  "4096": {
156
156
  "BLOCK_SIZE_M": 64,
157
- "BLOCK_SIZE_N": 64,
157
+ "BLOCK_SIZE_N": 128,
158
158
  "BLOCK_SIZE_K": 128,
159
- "GROUP_SIZE_M": 16,
159
+ "GROUP_SIZE_M": 4,
160
160
  "num_warps": 4,
161
161
  "num_stages": 2,
162
162
  "waves_per_eu": 0
sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H20,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": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 4,
8
+ "num_stages": 5
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": 5
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 32,
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": 64,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 64,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 5
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 1,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 32,
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": 32,
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": 64,
68
+ "BLOCK_SIZE_N": 32,
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": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 64,
79
+ "num_warps": 4,
80
+ "num_stages": 5
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
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": 64,
92
+ "BLOCK_SIZE_N": 32,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 32,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 2
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 64,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 16,
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": 16,
119
+ "num_warps": 4,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 64,
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": 64,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 128,
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": 3
145
+ }
146
+ }
sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H20,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": 32,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 32,
7
+ "num_warps": 8,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 8,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 64,
23
+ "num_warps": 8,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 64,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 2
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 64,
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": 64,
53
+ "BLOCK_SIZE_K": 64,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 128,
60
+ "BLOCK_SIZE_N": 32,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 32,
63
+ "num_warps": 8,
64
+ "num_stages": 5
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 64,
71
+ "num_warps": 8,
72
+ "num_stages": 2
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 128,
76
+ "BLOCK_SIZE_N": 32,
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": 128,
84
+ "BLOCK_SIZE_N": 32,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 32,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 8,
96
+ "num_stages": 5
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
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": 64,
108
+ "BLOCK_SIZE_N": 64,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 32,
111
+ "num_warps": 4,
112
+ "num_stages": 2
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 64,
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": 64,
124
+ "BLOCK_SIZE_N": 64,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 2
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 64,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 2
145
+ }
146
+ }
sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H20,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": 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": 32,
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": 64,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 64,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 64,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 32,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 64,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 32,
69
+ "BLOCK_SIZE_K": 128,
70
+ "GROUP_SIZE_M": 32,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 64,
76
+ "BLOCK_SIZE_N": 64,
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": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 4,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 1,
95
+ "num_warps": 4,
96
+ "num_stages": 5
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": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 2
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": 64,
124
+ "BLOCK_SIZE_N": 64,
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": 128,
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": 64,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }