sglang 0.4.1.post2__py3-none-any.whl → 0.4.1.post3__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 (120) hide show
  1. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  2. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  3. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  4. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  5. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  6. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +218 -0
  7. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json +218 -0
  8. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  9. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  10. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  11. sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  12. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  13. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  14. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  15. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  16. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  17. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  18. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  19. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  20. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  21. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  22. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  23. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  24. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  25. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  26. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  27. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  28. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  29. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  30. sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  31. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  32. sglang/srt/layers/moe/fused_moe_triton/configs/E=256,N=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  33. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  34. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  35. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  36. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  37. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  38. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  39. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  40. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  41. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  42. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json +146 -0
  43. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  44. sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  45. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json +200 -0
  46. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +138 -0
  47. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json +200 -0
  48. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  49. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  50. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  51. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  52. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  53. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  54. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json +200 -0
  55. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  56. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  57. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json +146 -0
  58. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  59. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  60. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_L40S.json +173 -0
  61. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +178 -0
  62. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  63. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  64. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  65. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json +200 -0
  66. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  67. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  68. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  69. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +175 -0
  70. sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  71. sglang/srt/layers/quantization/configs/N=1536,K=1536,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  72. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  73. sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  74. sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  75. sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  76. sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  77. sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  78. sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  79. sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  80. sglang/srt/layers/quantization/configs/N=256,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  81. sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  82. sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  83. sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  84. sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  85. sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  86. sglang/srt/layers/quantization/configs/N=36864,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  87. sglang/srt/layers/quantization/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  88. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  89. sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  90. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  91. sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  92. sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  93. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  94. sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  95. sglang/srt/layers/quantization/configs/N=7168,K=1024,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  96. sglang/srt/layers/quantization/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  97. sglang/srt/layers/quantization/configs/N=7168,K=1152,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  98. sglang/srt/layers/quantization/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  99. sglang/srt/layers/quantization/configs/N=7168,K=128,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  100. sglang/srt/layers/quantization/configs/N=7168,K=16384,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  101. sglang/srt/layers/quantization/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  102. sglang/srt/layers/quantization/configs/N=7168,K=18432,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  103. sglang/srt/layers/quantization/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  104. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  105. sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  106. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  107. sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  108. sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  109. sglang/srt/managers/detokenizer_manager.py +2 -0
  110. sglang/srt/managers/io_struct.py +11 -1
  111. sglang/srt/managers/scheduler.py +9 -1
  112. sglang/srt/managers/tokenizer_manager.py +7 -0
  113. sglang/srt/server_args.py +7 -0
  114. sglang/version.py +1 -1
  115. {sglang-0.4.1.post2.dist-info → sglang-0.4.1.post3.dist-info}/METADATA +1 -1
  116. sglang-0.4.1.post3.dist-info/RECORD +305 -0
  117. sglang-0.4.1.post2.dist-info/RECORD +0 -197
  118. {sglang-0.4.1.post2.dist-info → sglang-0.4.1.post3.dist-info}/LICENSE +0 -0
  119. {sglang-0.4.1.post2.dist-info → sglang-0.4.1.post3.dist-info}/WHEEL +0 -0
  120. {sglang-0.4.1.post2.dist-info → sglang-0.4.1.post3.dist-info}/top_level.txt +0 -0
sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=NVIDIA_H100_80GB_HBM3,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": 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": 64,
12
+ "BLOCK_SIZE_N": 64,
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": 64,
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": 64,
29
+ "BLOCK_SIZE_K": 128,
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": 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": 32,
47
+ "num_warps": 4,
48
+ "num_stages": 4
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": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 32,
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": 32,
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": 64,
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": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 16,
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": 32,
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": 64,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 1,
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": 1,
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": 64,
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": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang/srt/layers/quantization/configs/N=7168,K=2304,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": 64,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 3
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 64,
12
+ "BLOCK_SIZE_N": 64,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 32,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 64,
20
+ "BLOCK_SIZE_N": 64,
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": 64,
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": 64,
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": 4
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 64,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 4
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 32,
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": 32,
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": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 32,
79
+ "num_warps": 4,
80
+ "num_stages": 4
81
+ },
82
+ "128": {
83
+ "BLOCK_SIZE_M": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 64,
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": 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": 16,
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": 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": 64,
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": 64,
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=7168,K=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": 64,
5
+ "BLOCK_SIZE_K": 128,
6
+ "GROUP_SIZE_M": 16,
7
+ "num_warps": 4,
8
+ "num_stages": 3
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": 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": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 64,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 128,
30
+ "GROUP_SIZE_M": 16,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 64,
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": 64,
44
+ "BLOCK_SIZE_N": 64,
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": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 128,
62
+ "GROUP_SIZE_M": 1,
63
+ "num_warps": 4,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
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": 64,
77
+ "BLOCK_SIZE_K": 128,
78
+ "GROUP_SIZE_M": 16,
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": 32,
87
+ "num_warps": 4,
88
+ "num_stages": 5
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 128,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 4
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": 64,
109
+ "BLOCK_SIZE_K": 128,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
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": 128,
124
+ "BLOCK_SIZE_N": 64,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 1,
127
+ "num_warps": 4,
128
+ "num_stages": 2
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 128,
132
+ "BLOCK_SIZE_N": 64,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 2
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 64,
141
+ "BLOCK_SIZE_K": 128,
142
+ "GROUP_SIZE_M": 1,
143
+ "num_warps": 4,
144
+ "num_stages": 2
145
+ }
146
+ }
sglang/srt/managers/detokenizer_manager.py CHANGED
@@ -181,6 +181,8 @@ class DetokenizerManager:
181
181
  finished_reasons=recv_obj.finished_reasons,
182
182
  output_strs=output_strs,
183
183
  prompt_tokens=recv_obj.prompt_tokens,
184
+ origin_input_ids=recv_obj.origin_input_ids,
185
+ output_ids=recv_obj.output_ids,
184
186
  completion_tokens=recv_obj.completion_tokens,
185
187
  cached_tokens=recv_obj.cached_tokens,
186
188
  input_token_logprobs_val=recv_obj.input_token_logprobs_val,
sglang/srt/managers/io_struct.py CHANGED
@@ -323,7 +323,9 @@ class BatchTokenIDOut:
323
323
  decoded_texts: List[str]
324
324
  decode_ids: List[int]
325
325
  read_offsets: List[int]
326
- # Only used when `--skip-tokenizer-init`
326
+ # Only used when --return-token-ids` is set
327
+ origin_input_ids: Optional[List[int]]
328
+ # Only used when `--skip-tokenizer-init` or `--return-token-ids` is set
327
329
  output_ids: Optional[List[int]]
328
330
  # Detokenization configs
329
331
  skip_special_tokens: List[bool]
@@ -354,10 +356,18 @@ class BatchStrOut:
354
356
  # The output decoded strings
355
357
  output_strs: List[str]
356
358
 
359
+ # The token ids
360
+ origin_input_ids: Optional[List[int]]
361
+ output_ids: Optional[List[int]]
362
+
357
363
  # Token counts
364
+ # real input and output tokens can be get from
365
+ # origin_input_ids and output_ids by enabling --return_token_ids
366
+ # TODO (Shuai): Rename this to clarify the meaning.
358
367
  prompt_tokens: List[int]
359
368
  completion_tokens: List[int]
360
369
  cached_tokens: List[int]
370
+
361
371
  # Logprobs
362
372
  input_token_logprobs_val: List[float]
363
373
  input_token_logprobs_idx: List[int]
sglang/srt/managers/scheduler.py CHANGED
@@ -1218,6 +1218,7 @@ class Scheduler:
1218
1218
  decode_ids_list = []
1219
1219
  read_offsets = []
1220
1220
  output_ids = []
1221
+ origin_input_ids = []
1221
1222
 
1222
1223
  skip_special_tokens = []
1223
1224
  spaces_between_special_tokens = []
@@ -1266,8 +1267,14 @@ class Scheduler:
1266
1267
  decode_ids, read_offset = req.init_incremental_detokenize()
1267
1268
  decode_ids_list.append(decode_ids)
1268
1269
  read_offsets.append(read_offset)
1269
- if self.skip_tokenizer_init:
1270
+ if self.skip_tokenizer_init or self.server_args.return_token_ids:
1270
1271
  output_ids.append(req.output_ids)
1272
+ else:
1273
+ output_ids = None
1274
+ if self.server_args.return_token_ids:
1275
+ origin_input_ids.append(req.origin_input_ids)
1276
+ else:
1277
+ origin_input_ids = None
1271
1278
  skip_special_tokens.append(req.sampling_params.skip_special_tokens)
1272
1279
  spaces_between_special_tokens.append(
1273
1280
  req.sampling_params.spaces_between_special_tokens
@@ -1299,6 +1306,7 @@ class Scheduler:
1299
1306
  decoded_texts,
1300
1307
  decode_ids_list,
1301
1308
  read_offsets,
1309
+ origin_input_ids,
1302
1310
  output_ids,
1303
1311
  skip_special_tokens,
1304
1312
  spaces_between_special_tokens,
sglang/srt/managers/tokenizer_manager.py CHANGED
@@ -663,6 +663,13 @@ class TokenizerManager:
663
663
  "text": recv_obj.output_strs[i],
664
664
  "meta_info": meta_info,
665
665
  }
666
+ if self.server_args.return_token_ids:
667
+ out_dict.update(
668
+ {
669
+ "input_ids": recv_obj.origin_input_ids[i],
670
+ "output_ids": recv_obj.output_ids[i],
671
+ }
672
+ )
666
673
  elif isinstance(recv_obj, BatchTokenIDOut):
667
674
  out_dict = {
668
675
  "token_ids": recv_obj.output_ids[i],
sglang/srt/server_args.py CHANGED
@@ -54,6 +54,7 @@ class ServerArgs:
54
54
  chat_template: Optional[str] = None
55
55
  is_embedding: bool = False
56
56
  revision: Optional[str] = None
57
+ return_token_ids: bool = False
57
58
 
58
59
  # Port for the HTTP server
59
60
  host: str = "127.0.0.1"
@@ -280,6 +281,12 @@ class ServerArgs:
280
281
  action="store_true",
281
282
  help="If set, skip init tokenizer and pass input_ids in generate request",
282
283
  )
284
+ parser.add_argument(
285
+ "--return-token-ids",
286
+ action="store_true",
287
+ default=ServerArgs.return_token_ids,
288
+ help="Whether to return token IDs in the output, this may introduce additional overhead.",
289
+ )
283
290
  parser.add_argument(
284
291
  "--load-format",
285
292
  type=str,
sglang/version.py CHANGED
@@ -1 +1 @@
1
- __version__ = "0.4.1.post2"
1
+ __version__ = "0.4.1.post3"
{sglang-0.4.1.post2.dist-info → sglang-0.4.1.post3.dist-info}/METADATA RENAMED
@@ -1,6 +1,6 @@
1
1
  Metadata-Version: 2.1
2
2
  Name: sglang
3
- Version: 0.4.1.post2
3
+ Version: 0.4.1.post3
4
4
  Summary: SGLang is yet another fast serving framework for large language models and vision language models.
5
5
  License: Apache License
6
6
  Version 2.0, January 2004