sglang 0.4.3.post4__py3-none-any.whl → 0.4.4.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 (131) hide show
  1. sglang/bench_serving.py +1 -1
  2. sglang/lang/chat_template.py +29 -0
  3. sglang/srt/_custom_ops.py +19 -17
  4. sglang/srt/configs/__init__.py +2 -0
  5. sglang/srt/configs/janus_pro.py +629 -0
  6. sglang/srt/configs/model_config.py +24 -14
  7. sglang/srt/conversation.py +80 -2
  8. sglang/srt/custom_op.py +64 -3
  9. sglang/srt/distributed/device_communicators/custom_all_reduce.py +18 -17
  10. sglang/srt/distributed/parallel_state.py +10 -1
  11. sglang/srt/entrypoints/engine.py +5 -3
  12. sglang/srt/entrypoints/http_server.py +1 -1
  13. sglang/srt/function_call_parser.py +33 -2
  14. sglang/srt/hf_transformers_utils.py +16 -1
  15. sglang/srt/layers/attention/flashinfer_backend.py +1 -1
  16. sglang/srt/layers/attention/flashinfer_mla_backend.py +317 -57
  17. sglang/srt/layers/attention/triton_backend.py +1 -3
  18. sglang/srt/layers/attention/triton_ops/decode_attention.py +6 -6
  19. sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py +3 -3
  20. sglang/srt/layers/attention/triton_ops/extend_attention.py +4 -4
  21. sglang/srt/layers/attention/triton_ops/rocm_mla_decode_rope.py +3 -3
  22. sglang/srt/layers/attention/vision.py +43 -62
  23. sglang/srt/layers/dp_attention.py +30 -2
  24. sglang/srt/layers/elementwise.py +411 -0
  25. sglang/srt/layers/linear.py +1 -1
  26. sglang/srt/layers/logits_processor.py +1 -0
  27. sglang/srt/layers/moe/ep_moe/kernels.py +2 -1
  28. sglang/srt/layers/moe/ep_moe/layer.py +25 -9
  29. sglang/srt/layers/moe/fused_moe_triton/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  30. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  31. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  32. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.json +146 -0
  33. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  34. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=64,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  35. sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +63 -23
  36. sglang/srt/layers/moe/fused_moe_triton/layer.py +16 -4
  37. sglang/srt/layers/moe/router.py +342 -0
  38. sglang/srt/layers/parameter.py +10 -0
  39. sglang/srt/layers/quantization/__init__.py +90 -68
  40. sglang/srt/layers/quantization/blockwise_int8.py +1 -2
  41. sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  42. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  43. sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  44. sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  45. sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  46. sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  47. sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  48. sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  49. sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  50. sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  51. sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  52. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  53. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  54. sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  55. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  56. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  57. sglang/srt/layers/quantization/configs/N=7168,K=1024,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  58. sglang/srt/layers/quantization/configs/N=7168,K=1152,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  59. sglang/srt/layers/quantization/configs/N=7168,K=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  60. sglang/srt/layers/quantization/configs/N=7168,K=16384,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  61. sglang/srt/layers/quantization/configs/N=7168,K=16384,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  62. sglang/srt/layers/quantization/configs/N=7168,K=18432,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  63. sglang/srt/layers/quantization/configs/N=7168,K=18432,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  64. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  65. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  66. sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=NVIDIA_H20,dtype=int8_w8a8,block_shape=[128, 128].json +146 -0
  67. sglang/srt/layers/quantization/fp8.py +174 -106
  68. sglang/srt/layers/quantization/fp8_kernel.py +210 -38
  69. sglang/srt/layers/quantization/fp8_utils.py +156 -15
  70. sglang/srt/layers/quantization/modelopt_quant.py +5 -1
  71. sglang/srt/layers/quantization/w8a8_fp8.py +128 -0
  72. sglang/srt/layers/quantization/w8a8_int8.py +152 -3
  73. sglang/srt/layers/rotary_embedding.py +5 -3
  74. sglang/srt/layers/sampler.py +29 -35
  75. sglang/srt/layers/vocab_parallel_embedding.py +0 -1
  76. sglang/srt/lora/backend/__init__.py +9 -12
  77. sglang/srt/managers/cache_controller.py +74 -8
  78. sglang/srt/managers/data_parallel_controller.py +1 -1
  79. sglang/srt/managers/image_processor.py +37 -631
  80. sglang/srt/managers/image_processors/base_image_processor.py +219 -0
  81. sglang/srt/managers/image_processors/janus_pro.py +79 -0
  82. sglang/srt/managers/image_processors/llava.py +152 -0
  83. sglang/srt/managers/image_processors/minicpmv.py +86 -0
  84. sglang/srt/managers/image_processors/mlama.py +60 -0
  85. sglang/srt/managers/image_processors/qwen_vl.py +161 -0
  86. sglang/srt/managers/io_struct.py +32 -15
  87. sglang/srt/managers/multi_modality_padding.py +134 -0
  88. sglang/srt/managers/schedule_batch.py +213 -118
  89. sglang/srt/managers/schedule_policy.py +40 -8
  90. sglang/srt/managers/scheduler.py +176 -683
  91. sglang/srt/managers/scheduler_output_processor_mixin.py +614 -0
  92. sglang/srt/managers/tokenizer_manager.py +6 -6
  93. sglang/srt/managers/tp_worker_overlap_thread.py +4 -1
  94. sglang/srt/mem_cache/base_prefix_cache.py +6 -8
  95. sglang/srt/mem_cache/chunk_cache.py +12 -44
  96. sglang/srt/mem_cache/hiradix_cache.py +71 -34
  97. sglang/srt/mem_cache/memory_pool.py +81 -17
  98. sglang/srt/mem_cache/paged_allocator.py +283 -0
  99. sglang/srt/mem_cache/radix_cache.py +117 -36
  100. sglang/srt/model_executor/cuda_graph_runner.py +68 -20
  101. sglang/srt/model_executor/forward_batch_info.py +23 -10
  102. sglang/srt/model_executor/model_runner.py +63 -63
  103. sglang/srt/model_loader/loader.py +2 -1
  104. sglang/srt/model_loader/weight_utils.py +1 -1
  105. sglang/srt/models/deepseek_janus_pro.py +2127 -0
  106. sglang/srt/models/deepseek_nextn.py +23 -3
  107. sglang/srt/models/deepseek_v2.py +200 -191
  108. sglang/srt/models/grok.py +374 -119
  109. sglang/srt/models/minicpmv.py +28 -89
  110. sglang/srt/models/mllama.py +1 -1
  111. sglang/srt/models/qwen2.py +0 -1
  112. sglang/srt/models/qwen2_5_vl.py +25 -50
  113. sglang/srt/models/qwen2_vl.py +33 -49
  114. sglang/srt/openai_api/adapter.py +59 -35
  115. sglang/srt/openai_api/protocol.py +8 -1
  116. sglang/srt/sampling/penaltylib/frequency_penalty.py +0 -1
  117. sglang/srt/sampling/penaltylib/presence_penalty.py +0 -1
  118. sglang/srt/server_args.py +24 -16
  119. sglang/srt/speculative/eagle_worker.py +75 -39
  120. sglang/srt/utils.py +104 -9
  121. sglang/test/runners.py +104 -10
  122. sglang/test/test_block_fp8.py +106 -16
  123. sglang/test/test_custom_ops.py +88 -0
  124. sglang/test/test_utils.py +20 -4
  125. sglang/utils.py +0 -4
  126. sglang/version.py +1 -1
  127. {sglang-0.4.3.post4.dist-info → sglang-0.4.4.post1.dist-info}/METADATA +9 -10
  128. {sglang-0.4.3.post4.dist-info → sglang-0.4.4.post1.dist-info}/RECORD +131 -84
  129. {sglang-0.4.3.post4.dist-info → sglang-0.4.4.post1.dist-info}/WHEEL +1 -1
  130. {sglang-0.4.3.post4.dist-info → sglang-0.4.4.post1.dist-info}/LICENSE +0 -0
  131. {sglang-0.4.3.post4.dist-info → sglang-0.4.4.post1.dist-info}/top_level.txt +0 -0
sglang/srt/layers/logits_processor.py CHANGED
@@ -23,6 +23,7 @@ import triton.language as tl
23
23
  from torch import nn
24
24
 
25
25
  from sglang.srt.distributed import (
26
+ get_tensor_model_parallel_rank,
26
27
  get_tensor_model_parallel_world_size,
27
28
  tensor_model_parallel_all_gather,
28
29
  )
sglang/srt/layers/moe/ep_moe/kernels.py CHANGED
@@ -6,8 +6,9 @@ import triton
6
6
  import triton.language as tl
7
7
 
8
8
  from sglang.srt.layers.quantization.fp8_kernel import per_token_group_quant_fp8
9
+ from sglang.srt.utils import is_cuda
9
10
 
10
- _is_cuda = torch.cuda.is_available() and torch.version.cuda
11
+ _is_cuda = is_cuda()
11
12
  if _is_cuda:
12
13
  from sglang.srt.layers.quantization.fp8_kernel import (
13
14
  sglang_per_token_group_quant_fp8,
sglang/srt/layers/moe/ep_moe/layer.py CHANGED
@@ -3,7 +3,7 @@ from typing import Callable, List, Optional, Tuple
3
3
 
4
4
  import torch
5
5
  from torch.nn import Module
6
- from vllm import _custom_ops as ops
6
+ from vllm import _custom_ops as vllm_ops
7
7
 
8
8
  from sglang.srt.custom_op import CustomOp
9
9
  from sglang.srt.distributed import (
@@ -26,10 +26,18 @@ from sglang.srt.layers.quantization.base_config import (
26
26
  QuantizeMethodBase,
27
27
  )
28
28
  from sglang.srt.layers.quantization.fp8 import Fp8Config, Fp8MoEMethod
29
- from sglang.srt.utils import is_hip, set_weight_attrs
29
+ from sglang.srt.utils import is_cuda, is_hip, set_weight_attrs
30
+
31
+ _is_cuda = is_cuda()
32
+
33
+ if _is_cuda:
34
+ from sglang.srt.custom_op import scaled_fp8_quant as sgl_scaled_fp8_quant
35
+
30
36
 
31
37
  logger = logging.getLogger(__name__)
32
38
 
39
+ _is_hip = is_hip()
40
+
33
41
 
34
42
  class GroupedGemmRunner(torch.nn.Module):
35
43
  flashinfer_gemm_warpper = None
@@ -703,7 +711,7 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
703
711
  # If checkpoint is fp16, quantize in place.
704
712
  if not self.quant_config.is_checkpoint_fp8_serialized:
705
713
  # If rocm, use float8_e4m3fnuz as dtype
706
- fp8_dtype = torch.float8_e4m3fnuz if is_hip() else torch.float8_e4m3fn
714
+ fp8_dtype = torch.float8_e4m3fnuz if _is_hip else torch.float8_e4m3fn
707
715
  w13_weight = torch.empty_like(layer.w13_weight.data, dtype=fp8_dtype)
708
716
  w2_weight = torch.empty_like(layer.w2_weight.data, dtype=fp8_dtype)
709
717
 
@@ -717,12 +725,20 @@ class Fp8EPMoEMethod(Fp8MoEMethod):
717
725
  )
718
726
 
719
727
  for expert in range(layer.num_experts_per_partition):
720
- w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
721
- ops.scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
722
- )
723
- w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
724
- ops.scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
725
- )
728
+ if _is_cuda:
729
+ w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
730
+ sgl_scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
731
+ )
732
+ w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
733
+ sgl_scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
734
+ )
735
+ else:
736
+ w13_weight[expert, :, :], layer.w13_weight_scale[expert] = (
737
+ vllm_ops.scaled_fp8_quant(layer.w13_weight.data[expert, :, :])
738
+ )
739
+ w2_weight[expert, :, :], layer.w2_weight_scale[expert] = (
740
+ vllm_ops.scaled_fp8_quant(layer.w2_weight.data[expert, :, :])
741
+ )
726
742
  layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False)
727
743
  layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
728
744
  return
sglang/srt/layers/moe/fused_moe_triton/configs/E=160,N=192,device_name=NVIDIA_A800-SXM4-80GB.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": 64,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 64,
14
+ "GROUP_SIZE_M": 64,
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": 64,
23
+ "num_warps": 4,
24
+ "num_stages": 5
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 128,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 2
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": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 2
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 2
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": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
69
+ "BLOCK_SIZE_K": 128,
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": 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": 16,
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": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 16,
103
+ "num_warps": 4,
104
+ "num_stages": 2
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": 4
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": 8,
120
+ "num_stages": 3
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 128,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 128,
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": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a8.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": 64,
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": 128,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 3
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": 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": 4,
40
+ "num_stages": 3
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": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 32,
55
+ "num_warps": 4,
56
+ "num_stages": 2
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": 2
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 16,
68
+ "BLOCK_SIZE_N": 128,
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": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 2
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 128,
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": 128,
118
+ "GROUP_SIZE_M": 64,
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": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 32,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 32,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8,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": 128,
6
+ "GROUP_SIZE_M": 32,
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": 128,
14
+ "GROUP_SIZE_M": 64,
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": 64,
23
+ "num_warps": 4,
24
+ "num_stages": 3
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": 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": 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": 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": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
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": 128,
70
+ "GROUP_SIZE_M": 1,
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": 128,
78
+ "GROUP_SIZE_M": 1,
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": 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": 64,
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": 64,
103
+ "num_warps": 4,
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": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 32,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 128,
118
+ "GROUP_SIZE_M": 64,
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": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 3
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 32,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 32,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_A800-SXM4-80GB,dtype=int8_w8a8.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": 64,
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": 128,
14
+ "GROUP_SIZE_M": 64,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 3
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": 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": 4,
40
+ "num_stages": 3
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": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 16,
55
+ "num_warps": 4,
56
+ "num_stages": 2
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 16,
60
+ "BLOCK_SIZE_N": 128,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
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": 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": 1,
79
+ "num_warps": 4,
80
+ "num_stages": 2
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 16,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 2
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 16,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 2
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 32,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 128,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 32,
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": 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": 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": 32,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 32,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }