sglang 0.4.1.post2__tar.gz → 0.4.1.post3__tar.gz

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 (312) hide show
  1. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/PKG-INFO +1 -1
  2. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/pyproject.toml +2 -2
  3. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  4. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  5. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  6. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  7. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  8. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +218 -0
  9. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3072,device_name=NVIDIA_H100_80GB_HBM3.json +218 -0
  10. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  11. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  12. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  13. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  14. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  15. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1344,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  16. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1344,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  17. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  18. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=14336,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  19. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +218 -0
  20. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  21. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=2688,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  22. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=2688,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  23. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3072,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  24. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3072,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  25. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3200,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  26. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  27. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +218 -0
  28. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=6400,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  29. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json +146 -0
  30. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  31. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=int8_w8a16.json +146 -0
  32. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=16,N=800,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +130 -0
  33. sglang-0.4.1.post3/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
  34. sglang-0.4.1.post3/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
  35. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  36. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  37. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  38. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=1280,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  39. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=2560,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  40. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  41. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=320,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  42. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  43. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_A800-SXM4-80GB.json +146 -0
  44. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json +146 -0
  45. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  46. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=64,N=640,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  47. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=AMD_Instinct_MI300X.json +200 -0
  48. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=14336,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +138 -0
  49. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=AMD_Instinct_MI300X.json +200 -0
  50. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  51. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  52. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=1792,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  53. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  54. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  55. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=2048,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  56. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=AMD_Instinct_MI300X.json +200 -0
  57. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-40GB.json +146 -0
  58. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  59. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_GeForce_RTX_4090,dtype=fp8_w8a8.json +146 -0
  60. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  61. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  62. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=3584,device_name=NVIDIA_L40S.json +173 -0
  63. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +178 -0
  64. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  65. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  66. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  67. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=AMD_Instinct_MI300X.json +200 -0
  68. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  69. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  70. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=7168,device_name=NVIDIA_H100_80GB_HBM3.json +146 -0
  71. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=AMD_Instinct_MI300X,dtype=fp8_w8a8.json +175 -0
  72. sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=8,N=8192,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json +146 -0
  73. sglang-0.4.1.post3/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
  74. sglang-0.4.1.post3/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
  75. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=1536,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  76. sglang-0.4.1.post3/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
  77. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=2048,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  78. sglang-0.4.1.post3/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
  79. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=2304,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  80. sglang-0.4.1.post3/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
  81. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=24576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  82. sglang-0.4.1.post3/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
  83. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=3072,K=1536,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  84. sglang-0.4.1.post3/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
  85. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=3072,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  86. sglang-0.4.1.post3/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
  87. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=32768,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  88. sglang-0.4.1.post3/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
  89. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=36864,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  90. sglang-0.4.1.post3/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
  91. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=4096,K=512,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  92. sglang-0.4.1.post3/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
  93. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=4608,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  94. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=512,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  95. sglang-0.4.1.post3/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
  96. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=576,K=7168,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  97. sglang-0.4.1.post3/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
  98. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=1024,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  99. sglang-0.4.1.post3/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
  100. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=1152,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  101. sglang-0.4.1.post3/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
  102. sglang-0.4.1.post3/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
  103. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=16384,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  104. sglang-0.4.1.post3/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
  105. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=18432,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  106. sglang-0.4.1.post3/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
  107. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=2048,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  108. sglang-0.4.1.post3/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
  109. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=2304,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  110. sglang-0.4.1.post3/sglang/srt/layers/quantization/configs/N=7168,K=256,device_name=NVIDIA_H200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  111. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/detokenizer_manager.py +2 -0
  112. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/io_struct.py +11 -1
  113. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/scheduler.py +9 -1
  114. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/tokenizer_manager.py +7 -0
  115. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/server_args.py +7 -0
  116. sglang-0.4.1.post3/sglang/version.py +1 -0
  117. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang.egg-info/PKG-INFO +1 -1
  118. sglang-0.4.1.post3/sglang.egg-info/SOURCES.txt +308 -0
  119. sglang-0.4.1.post2/sglang/version.py +0 -1
  120. sglang-0.4.1.post2/sglang.egg-info/SOURCES.txt +0 -200
  121. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/LICENSE +0 -0
  122. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/README.md +0 -0
  123. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/setup.cfg +0 -0
  124. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/__init__.py +0 -0
  125. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/api.py +0 -0
  126. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/bench_latency.py +0 -0
  127. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/bench_offline_throughput.py +0 -0
  128. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/bench_one_batch.py +0 -0
  129. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/bench_one_batch_server.py +0 -0
  130. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/bench_serving.py +0 -0
  131. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/check_env.py +0 -0
  132. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/global_config.py +0 -0
  133. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/__init__.py +0 -0
  134. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/__init__.py +0 -0
  135. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/anthropic.py +0 -0
  136. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/base_backend.py +0 -0
  137. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/litellm.py +0 -0
  138. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/openai.py +0 -0
  139. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/runtime_endpoint.py +0 -0
  140. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/backend/vertexai.py +0 -0
  141. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/chat_template.py +0 -0
  142. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/choices.py +0 -0
  143. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/compiler.py +0 -0
  144. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/interpreter.py +0 -0
  145. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/ir.py +0 -0
  146. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/lang/tracer.py +0 -0
  147. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/launch_server.py +0 -0
  148. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/launch_server_llavavid.py +0 -0
  149. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/llama3_eval.py +0 -0
  150. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/_custom_ops.py +0 -0
  151. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/aio_rwlock.py +0 -0
  152. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/configs/__init__.py +0 -0
  153. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/configs/device_config.py +0 -0
  154. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/configs/exaone.py +0 -0
  155. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/configs/load_config.py +0 -0
  156. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/configs/model_config.py +0 -0
  157. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/configs/qwen2vl.py +0 -0
  158. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/constrained/__init__.py +0 -0
  159. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/constrained/base_grammar_backend.py +0 -0
  160. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/constrained/outlines_backend.py +0 -0
  161. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/constrained/outlines_jump_forward.py +0 -0
  162. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/constrained/xgrammar_backend.py +0 -0
  163. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/conversation.py +0 -0
  164. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/__init__.py +0 -0
  165. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/communication_op.py +0 -0
  166. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/__init__.py +0 -0
  167. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/cuda_wrapper.py +0 -0
  168. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/custom_all_reduce.py +0 -0
  169. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/custom_all_reduce_utils.py +0 -0
  170. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/hpu_communicator.py +0 -0
  171. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/pynccl.py +0 -0
  172. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/pynccl_wrapper.py +0 -0
  173. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/shm_broadcast.py +0 -0
  174. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/device_communicators/xpu_communicator.py +0 -0
  175. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/parallel_state.py +0 -0
  176. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/distributed/utils.py +0 -0
  177. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/hf_transformers_utils.py +0 -0
  178. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/activation.py +0 -0
  179. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/__init__.py +0 -0
  180. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/double_sparsity_backend.py +0 -0
  181. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/flashinfer_backend.py +0 -0
  182. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/torch_native_backend.py +0 -0
  183. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/triton_backend.py +0 -0
  184. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/triton_ops/decode_attention.py +0 -0
  185. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/triton_ops/double_sparsity_attention.py +0 -0
  186. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/triton_ops/extend_attention.py +0 -0
  187. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/attention/triton_ops/prefill_attention.py +0 -0
  188. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/custom_op_util.py +0 -0
  189. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/layernorm.py +0 -0
  190. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/linear.py +0 -0
  191. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/logits_processor.py +0 -0
  192. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/ep_moe/__init__.py +0 -0
  193. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/ep_moe/kernels.py +0 -0
  194. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/ep_moe/layer.py +0 -0
  195. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/fused_moe_native.py +0 -0
  196. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/fused_moe_triton/__init__.py +0 -0
  197. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/fused_moe_triton/fused_moe.py +0 -0
  198. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/fused_moe_triton/layer.py +0 -0
  199. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/moe/topk.py +0 -0
  200. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/pooler.py +0 -0
  201. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/quantization/__init__.py +0 -0
  202. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/quantization/base_config.py +0 -0
  203. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/quantization/fp8.py +0 -0
  204. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/quantization/fp8_kernel.py +0 -0
  205. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/quantization/fp8_utils.py +0 -0
  206. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/radix_attention.py +0 -0
  207. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/rotary_embedding.py +0 -0
  208. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/sampler.py +0 -0
  209. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/torchao_utils.py +0 -0
  210. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/layers/vocab_parallel_embedding.py +0 -0
  211. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/lora/lora.py +0 -0
  212. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/lora/lora_config.py +0 -0
  213. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/lora/lora_manager.py +0 -0
  214. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/data_parallel_controller.py +0 -0
  215. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/image_processor.py +0 -0
  216. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/schedule_batch.py +0 -0
  217. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/schedule_policy.py +0 -0
  218. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/session_controller.py +0 -0
  219. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/tp_worker.py +0 -0
  220. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/managers/tp_worker_overlap_thread.py +0 -0
  221. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/mem_cache/base_prefix_cache.py +0 -0
  222. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/mem_cache/chunk_cache.py +0 -0
  223. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/mem_cache/flush_cache.py +0 -0
  224. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/mem_cache/memory_pool.py +0 -0
  225. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/mem_cache/radix_cache.py +0 -0
  226. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/metrics/collector.py +0 -0
  227. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/metrics/func_timer.py +0 -0
  228. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/mm_utils.py +0 -0
  229. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_executor/cuda_graph_runner.py +0 -0
  230. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_executor/forward_batch_info.py +0 -0
  231. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_executor/model_runner.py +0 -0
  232. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_loader/__init__.py +0 -0
  233. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_loader/loader.py +0 -0
  234. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_loader/utils.py +0 -0
  235. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_loader/weight_utils.py +0 -0
  236. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/model_parallel.py +0 -0
  237. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/baichuan.py +0 -0
  238. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/chatglm.py +0 -0
  239. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/commandr.py +0 -0
  240. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/dbrx.py +0 -0
  241. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/deepseek.py +0 -0
  242. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/deepseek_v2.py +0 -0
  243. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/exaone.py +0 -0
  244. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/gemma.py +0 -0
  245. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/gemma2.py +0 -0
  246. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/gemma2_reward.py +0 -0
  247. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/gpt2.py +0 -0
  248. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/gpt_bigcode.py +0 -0
  249. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/granite.py +0 -0
  250. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/grok.py +0 -0
  251. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/internlm2.py +0 -0
  252. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/internlm2_reward.py +0 -0
  253. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llama.py +0 -0
  254. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llama_classification.py +0 -0
  255. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llama_eagle.py +0 -0
  256. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llama_embedding.py +0 -0
  257. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llama_reward.py +0 -0
  258. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llava.py +0 -0
  259. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/llavavid.py +0 -0
  260. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/minicpm.py +0 -0
  261. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/minicpm3.py +0 -0
  262. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/mistral.py +0 -0
  263. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/mixtral.py +0 -0
  264. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/mixtral_quant.py +0 -0
  265. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/mllama.py +0 -0
  266. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/olmo.py +0 -0
  267. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/olmo2.py +0 -0
  268. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/olmoe.py +0 -0
  269. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/phi3_small.py +0 -0
  270. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/qwen.py +0 -0
  271. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/qwen2.py +0 -0
  272. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/qwen2_moe.py +0 -0
  273. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/qwen2_vl.py +0 -0
  274. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/registry.py +0 -0
  275. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/stablelm.py +0 -0
  276. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/torch_native_llama.py +0 -0
  277. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/xverse.py +0 -0
  278. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/xverse_moe.py +0 -0
  279. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/models/yivl.py +0 -0
  280. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/openai_api/adapter.py +0 -0
  281. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/openai_api/protocol.py +0 -0
  282. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/penaltylib/__init__.py +0 -0
  283. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/penaltylib/orchestrator.py +0 -0
  284. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/penaltylib/penalizers/frequency_penalty.py +0 -0
  285. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/penaltylib/penalizers/min_new_tokens.py +0 -0
  286. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/penaltylib/penalizers/presence_penalty.py +0 -0
  287. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/penaltylib/penalizers/repetition_penalty.py +0 -0
  288. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/sampling_batch_info.py +0 -0
  289. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/sampling/sampling_params.py +0 -0
  290. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/server.py +0 -0
  291. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/speculative/spec_info.py +0 -0
  292. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/srt/utils.py +0 -0
  293. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/few_shot_gsm8k.py +0 -0
  294. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/few_shot_gsm8k_engine.py +0 -0
  295. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/run_eval.py +0 -0
  296. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/runners.py +0 -0
  297. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/simple_eval_common.py +0 -0
  298. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/simple_eval_gpqa.py +0 -0
  299. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/simple_eval_humaneval.py +0 -0
  300. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/simple_eval_math.py +0 -0
  301. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/simple_eval_mgsm.py +0 -0
  302. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/simple_eval_mmlu.py +0 -0
  303. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/srt/sampling/penaltylib/utils.py +0 -0
  304. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/test_activation.py +0 -0
  305. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/test_block_fp8.py +0 -0
  306. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/test_layernorm.py +0 -0
  307. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/test_programs.py +0 -0
  308. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/test/test_utils.py +0 -0
  309. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang/utils.py +0 -0
  310. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang.egg-info/dependency_links.txt +0 -0
  311. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang.egg-info/requires.txt +0 -0
  312. {sglang-0.4.1.post2 → sglang-0.4.1.post3}/sglang.egg-info/top_level.txt +0 -0
{sglang-0.4.1.post2 → sglang-0.4.1.post3}/PKG-INFO 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
{sglang-0.4.1.post2 → sglang-0.4.1.post3}/pyproject.toml RENAMED
@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
4
4
 
5
5
  [project]
6
6
  name = "sglang"
7
- version = "0.4.1.post2"
7
+ version = "0.4.1.post3"
8
8
  description = "SGLang is yet another fast serving framework for large language models and vision language models."
9
9
  readme = "README.md"
10
10
  requires-python = ">=3.8"
@@ -61,7 +61,7 @@ dev_hpu = ["sglang[all_hpu]", "sglang[test]"]
61
61
  "Bug Tracker" = "https://github.com/sgl-project/sglang/issues"
62
62
 
63
63
  [tool.setuptools.package-data]
64
- "sglang" = ["srt/layers/fused_moe_triton/configs/*.json"]
64
+ "sglang" = ["srt/layers/moe/fused_moe_triton/configs/*.json", "srt/layers/quantization/configs/*.json"]
65
65
 
66
66
  [tool.setuptools.packages.find]
67
67
  exclude = [
sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=14336,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.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": 1,
7
+ "num_warps": 4,
8
+ "num_stages": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 32,
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": 32,
21
+ "BLOCK_SIZE_K": 128,
22
+ "GROUP_SIZE_M": 1,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 3
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 32,
37
+ "BLOCK_SIZE_K": 128,
38
+ "GROUP_SIZE_M": 1,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 64,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 2
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 64,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
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": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
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": 64,
84
+ "BLOCK_SIZE_N": 64,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 16,
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": 64,
94
+ "GROUP_SIZE_M": 32,
95
+ "num_warps": 4,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 256,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 32,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 256,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 4
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 64,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 32,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ }
146
+ }
sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=14336,device_name=NVIDIA_A100-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": 256,
6
+ "GROUP_SIZE_M": 16,
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": 256,
14
+ "GROUP_SIZE_M": 32,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 64,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 16,
23
+ "num_warps": 4,
24
+ "num_stages": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 256,
30
+ "GROUP_SIZE_M": 32,
31
+ "num_warps": 4,
32
+ "num_stages": 4
33
+ },
34
+ "16": {
35
+ "BLOCK_SIZE_M": 16,
36
+ "BLOCK_SIZE_N": 64,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 64,
39
+ "num_warps": 4,
40
+ "num_stages": 4
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 32,
44
+ "BLOCK_SIZE_N": 64,
45
+ "BLOCK_SIZE_K": 128,
46
+ "GROUP_SIZE_M": 16,
47
+ "num_warps": 4,
48
+ "num_stages": 5
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 32,
52
+ "BLOCK_SIZE_N": 64,
53
+ "BLOCK_SIZE_K": 128,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 5
57
+ },
58
+ "48": {
59
+ "BLOCK_SIZE_M": 32,
60
+ "BLOCK_SIZE_N": 64,
61
+ "BLOCK_SIZE_K": 64,
62
+ "GROUP_SIZE_M": 16,
63
+ "num_warps": 4,
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": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 4
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
76
+ "BLOCK_SIZE_N": 256,
77
+ "BLOCK_SIZE_K": 64,
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": 128,
85
+ "BLOCK_SIZE_K": 64,
86
+ "GROUP_SIZE_M": 64,
87
+ "num_warps": 4,
88
+ "num_stages": 4
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 128,
92
+ "BLOCK_SIZE_N": 128,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 16,
95
+ "num_warps": 8,
96
+ "num_stages": 4
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 128,
100
+ "BLOCK_SIZE_N": 128,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 64,
103
+ "num_warps": 4,
104
+ "num_stages": 3
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 128,
108
+ "BLOCK_SIZE_N": 128,
109
+ "BLOCK_SIZE_K": 64,
110
+ "GROUP_SIZE_M": 64,
111
+ "num_warps": 4,
112
+ "num_stages": 3
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 128,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 16,
119
+ "num_warps": 4,
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": 32,
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": 16,
135
+ "num_warps": 4,
136
+ "num_stages": 3
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 128,
140
+ "BLOCK_SIZE_N": 128,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 16,
143
+ "num_warps": 4,
144
+ "num_stages": 3
145
+ }
146
+ }
sglang-0.4.1.post3/sglang/srt/layers/moe/fused_moe_triton/configs/E=1,N=1792,device_name=NVIDIA_A100-SXM4-80GB,dtype=int8_w8a16.json ADDED
@@ -0,0 +1,218 @@
1
+ {
2
+ "1": {
3
+ "BLOCK_SIZE_M": 16,
4
+ "BLOCK_SIZE_N": 32,
5
+ "BLOCK_SIZE_K": 256,
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": 32,
13
+ "BLOCK_SIZE_K": 256,
14
+ "GROUP_SIZE_M": 16,
15
+ "num_warps": 4,
16
+ "num_stages": 3
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 32,
21
+ "BLOCK_SIZE_K": 256,
22
+ "GROUP_SIZE_M": 32,
23
+ "num_warps": 4,
24
+ "num_stages": 3
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 32,
29
+ "BLOCK_SIZE_K": 256,
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": 32,
37
+ "BLOCK_SIZE_K": 256,
38
+ "GROUP_SIZE_M": 32,
39
+ "num_warps": 4,
40
+ "num_stages": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 32,
45
+ "BLOCK_SIZE_K": 256,
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": 32,
53
+ "BLOCK_SIZE_K": 256,
54
+ "GROUP_SIZE_M": 1,
55
+ "num_warps": 4,
56
+ "num_stages": 3
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": 8,
64
+ "num_stages": 3
65
+ },
66
+ "64": {
67
+ "BLOCK_SIZE_M": 64,
68
+ "BLOCK_SIZE_N": 64,
69
+ "BLOCK_SIZE_K": 64,
70
+ "GROUP_SIZE_M": 1,
71
+ "num_warps": 4,
72
+ "num_stages": 3
73
+ },
74
+ "96": {
75
+ "BLOCK_SIZE_M": 32,
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": 64,
84
+ "BLOCK_SIZE_N": 128,
85
+ "BLOCK_SIZE_K": 128,
86
+ "GROUP_SIZE_M": 1,
87
+ "num_warps": 8,
88
+ "num_stages": 3
89
+ },
90
+ "256": {
91
+ "BLOCK_SIZE_M": 64,
92
+ "BLOCK_SIZE_N": 64,
93
+ "BLOCK_SIZE_K": 64,
94
+ "GROUP_SIZE_M": 64,
95
+ "num_warps": 4,
96
+ "num_stages": 3
97
+ },
98
+ "512": {
99
+ "BLOCK_SIZE_M": 64,
100
+ "BLOCK_SIZE_N": 64,
101
+ "BLOCK_SIZE_K": 64,
102
+ "GROUP_SIZE_M": 1,
103
+ "num_warps": 4,
104
+ "num_stages": 4
105
+ },
106
+ "1024": {
107
+ "BLOCK_SIZE_M": 64,
108
+ "BLOCK_SIZE_N": 64,
109
+ "BLOCK_SIZE_K": 64,
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": 256,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 1,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 256,
125
+ "BLOCK_SIZE_K": 64,
126
+ "GROUP_SIZE_M": 32,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 256,
133
+ "BLOCK_SIZE_K": 64,
134
+ "GROUP_SIZE_M": 64,
135
+ "num_warps": 4,
136
+ "num_stages": 4
137
+ },
138
+ "4096": {
139
+ "BLOCK_SIZE_M": 64,
140
+ "BLOCK_SIZE_N": 256,
141
+ "BLOCK_SIZE_K": 64,
142
+ "GROUP_SIZE_M": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 4
145
+ },
146
+ "5120": {
147
+ "BLOCK_SIZE_M": 64,
148
+ "BLOCK_SIZE_N": 256,
149
+ "BLOCK_SIZE_K": 64,
150
+ "GROUP_SIZE_M": 16,
151
+ "num_warps": 4,
152
+ "num_stages": 4
153
+ },
154
+ "9216": {
155
+ "BLOCK_SIZE_M": 64,
156
+ "BLOCK_SIZE_N": 256,
157
+ "BLOCK_SIZE_K": 64,
158
+ "GROUP_SIZE_M": 32,
159
+ "num_warps": 4,
160
+ "num_stages": 4
161
+ },
162
+ "13312": {
163
+ "BLOCK_SIZE_M": 64,
164
+ "BLOCK_SIZE_N": 256,
165
+ "BLOCK_SIZE_K": 64,
166
+ "GROUP_SIZE_M": 16,
167
+ "num_warps": 4,
168
+ "num_stages": 4
169
+ },
170
+ "17408": {
171
+ "BLOCK_SIZE_M": 64,
172
+ "BLOCK_SIZE_N": 256,
173
+ "BLOCK_SIZE_K": 64,
174
+ "GROUP_SIZE_M": 32,
175
+ "num_warps": 4,
176
+ "num_stages": 4
177
+ },
178
+ "25600": {
179
+ "BLOCK_SIZE_M": 64,
180
+ "BLOCK_SIZE_N": 256,
181
+ "BLOCK_SIZE_K": 64,
182
+ "GROUP_SIZE_M": 16,
183
+ "num_warps": 4,
184
+ "num_stages": 4
185
+ },
186
+ "33792": {
187
+ "BLOCK_SIZE_M": 64,
188
+ "BLOCK_SIZE_N": 256,
189
+ "BLOCK_SIZE_K": 64,
190
+ "GROUP_SIZE_M": 16,
191
+ "num_warps": 4,
192
+ "num_stages": 4
193
+ },
194
+ "41984": {
195
+ "BLOCK_SIZE_M": 64,
196
+ "BLOCK_SIZE_N": 256,
197
+ "BLOCK_SIZE_K": 64,
198
+ "GROUP_SIZE_M": 16,
199
+ "num_warps": 4,
200
+ "num_stages": 4
201
+ },
202
+ "50176": {
203
+ "BLOCK_SIZE_M": 64,
204
+ "BLOCK_SIZE_N": 256,
205
+ "BLOCK_SIZE_K": 64,
206
+ "GROUP_SIZE_M": 32,
207
+ "num_warps": 4,
208
+ "num_stages": 4
209
+ },
210
+ "58368": {
211
+ "BLOCK_SIZE_M": 64,
212
+ "BLOCK_SIZE_N": 256,
213
+ "BLOCK_SIZE_K": 64,
214
+ "GROUP_SIZE_M": 16,
215
+ "num_warps": 4,
216
+ "num_stages": 4
217
+ }
218
+ }