sglang 0.5.1.post1__py3-none-any.whl → 0.5.1.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 (69) hide show
  1. sglang/bench_one_batch_server.py +79 -53
  2. sglang/bench_serving.py +186 -14
  3. sglang/profiler.py +0 -1
  4. sglang/srt/conversation.py +38 -5
  5. sglang/srt/disaggregation/decode.py +4 -0
  6. sglang/srt/disaggregation/prefill.py +4 -0
  7. sglang/srt/entrypoints/engine.py +2 -2
  8. sglang/srt/entrypoints/openai/protocol.py +27 -24
  9. sglang/srt/entrypoints/openai/serving_chat.py +50 -9
  10. sglang/srt/entrypoints/openai/serving_completions.py +15 -0
  11. sglang/srt/entrypoints/tool.py +7 -7
  12. sglang/srt/function_call/deepseekv31_detector.py +222 -0
  13. sglang/srt/function_call/function_call_parser.py +2 -0
  14. sglang/srt/function_call/gpt_oss_detector.py +144 -256
  15. sglang/srt/harmony_parser.py +588 -0
  16. sglang/srt/hf_transformers_utils.py +16 -7
  17. sglang/srt/layers/attention/ascend_backend.py +218 -111
  18. sglang/srt/layers/attention/flashattention_backend.py +241 -7
  19. sglang/srt/layers/attention/flashinfer_backend.py +5 -2
  20. sglang/srt/layers/attention/flashinfer_mla_backend.py +76 -91
  21. sglang/srt/layers/attention/utils.py +15 -94
  22. sglang/srt/layers/communicator.py +1 -2
  23. sglang/srt/layers/moe/cutlass_moe.py +0 -15
  24. sglang/srt/layers/moe/ep_moe/layer.py +1 -7
  25. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=256,device_name=NVIDIA_B200,dtype=fp8_w8a8,block_shape=[128, 128].json +146 -0
  26. sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=257,N=64,device_name=NVIDIA_A100-SXM4-80GB.json +146 -0
  27. sglang/srt/layers/moe/topk.py +1 -1
  28. sglang/srt/layers/quantization/deep_gemm_wrapper/compile_utils.py +133 -235
  29. sglang/srt/layers/quantization/deep_gemm_wrapper/configurer.py +5 -7
  30. sglang/srt/layers/quantization/deep_gemm_wrapper/entrypoint.py +5 -23
  31. sglang/srt/layers/quantization/fp8.py +2 -1
  32. sglang/srt/layers/quantization/fp8_kernel.py +2 -2
  33. sglang/srt/layers/quantization/fp8_utils.py +2 -2
  34. sglang/srt/layers/quantization/modelopt_quant.py +2 -2
  35. sglang/srt/layers/quantization/mxfp4.py +16 -23
  36. sglang/srt/layers/quantization/mxfp4_tensor.py +3 -1
  37. sglang/srt/layers/utils.py +0 -14
  38. sglang/srt/lora/lora_manager.py +29 -12
  39. sglang/srt/managers/cache_controller.py +223 -156
  40. sglang/srt/managers/detokenizer_manager.py +5 -0
  41. sglang/srt/managers/io_struct.py +30 -0
  42. sglang/srt/managers/scheduler.py +58 -7
  43. sglang/srt/managers/scheduler_metrics_mixin.py +15 -0
  44. sglang/srt/managers/tokenizer_manager.py +36 -3
  45. sglang/srt/mem_cache/hicache_storage.py +31 -20
  46. sglang/srt/mem_cache/hiradix_cache.py +12 -3
  47. sglang/srt/mem_cache/memory_pool.py +73 -14
  48. sglang/srt/mem_cache/memory_pool_host.py +3 -2
  49. sglang/srt/mem_cache/radix_cache.py +1 -0
  50. sglang/srt/mem_cache/storage/hf3fs/storage_hf3fs.py +5 -13
  51. sglang/srt/mem_cache/storage/mooncake_store/mooncake_store.py +85 -81
  52. sglang/srt/metrics/collector.py +5 -5
  53. sglang/srt/model_executor/cuda_graph_runner.py +2 -2
  54. sglang/srt/model_executor/model_runner.py +1 -1
  55. sglang/srt/models/deepseek_v2.py +12 -3
  56. sglang/srt/models/gpt_oss.py +2 -1
  57. sglang/srt/models/qwen2_5_vl.py +1 -0
  58. sglang/srt/offloader.py +115 -0
  59. sglang/srt/reasoning_parser.py +56 -300
  60. sglang/srt/server_args.py +10 -5
  61. sglang/srt/tokenizer/tiktoken_tokenizer.py +6 -1
  62. sglang/srt/utils.py +59 -12
  63. sglang/test/test_cutlass_moe.py +33 -28
  64. sglang/version.py +1 -1
  65. {sglang-0.5.1.post1.dist-info → sglang-0.5.1.post3.dist-info}/METADATA +6 -5
  66. {sglang-0.5.1.post1.dist-info → sglang-0.5.1.post3.dist-info}/RECORD +69 -65
  67. {sglang-0.5.1.post1.dist-info → sglang-0.5.1.post3.dist-info}/WHEEL +0 -0
  68. {sglang-0.5.1.post1.dist-info → sglang-0.5.1.post3.dist-info}/licenses/LICENSE +0 -0
  69. {sglang-0.5.1.post1.dist-info → sglang-0.5.1.post3.dist-info}/top_level.txt +0 -0
sglang/srt/layers/moe/fused_moe_triton/configs/triton_3_4_0/E=256,N=256,device_name=NVIDIA_B200,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": 4
9
+ },
10
+ "2": {
11
+ "BLOCK_SIZE_M": 16,
12
+ "BLOCK_SIZE_N": 128,
13
+ "BLOCK_SIZE_K": 128,
14
+ "GROUP_SIZE_M": 1,
15
+ "num_warps": 4,
16
+ "num_stages": 4
17
+ },
18
+ "4": {
19
+ "BLOCK_SIZE_M": 16,
20
+ "BLOCK_SIZE_N": 128,
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": 128,
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": 16,
36
+ "BLOCK_SIZE_N": 128,
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": 3
49
+ },
50
+ "32": {
51
+ "BLOCK_SIZE_M": 16,
52
+ "BLOCK_SIZE_N": 128,
53
+ "BLOCK_SIZE_K": 64,
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": 256,
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": 16,
68
+ "BLOCK_SIZE_N": 128,
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": 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": 1,
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": 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": 1,
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": 32,
119
+ "num_warps": 4,
120
+ "num_stages": 4
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 64,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 128,
126
+ "GROUP_SIZE_M": 16,
127
+ "num_warps": 4,
128
+ "num_stages": 4
129
+ },
130
+ "3072": {
131
+ "BLOCK_SIZE_M": 64,
132
+ "BLOCK_SIZE_N": 128,
133
+ "BLOCK_SIZE_K": 128,
134
+ "GROUP_SIZE_M": 1,
135
+ "num_warps": 4,
136
+ "num_stages": 4
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/moe/fused_moe_triton/configs/triton_3_4_0/E=257,N=64,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": 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": 4
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": 4
25
+ },
26
+ "8": {
27
+ "BLOCK_SIZE_M": 16,
28
+ "BLOCK_SIZE_N": 64,
29
+ "BLOCK_SIZE_K": 64,
30
+ "GROUP_SIZE_M": 1,
31
+ "num_warps": 4,
32
+ "num_stages": 3
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": 3
41
+ },
42
+ "24": {
43
+ "BLOCK_SIZE_M": 16,
44
+ "BLOCK_SIZE_N": 128,
45
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 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": 16,
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": 16,
76
+ "BLOCK_SIZE_N": 128,
77
+ "BLOCK_SIZE_K": 64,
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": 64,
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": 64,
94
+ "GROUP_SIZE_M": 1,
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": 64,
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": 64,
110
+ "GROUP_SIZE_M": 1,
111
+ "num_warps": 4,
112
+ "num_stages": 2
113
+ },
114
+ "1536": {
115
+ "BLOCK_SIZE_M": 64,
116
+ "BLOCK_SIZE_N": 128,
117
+ "BLOCK_SIZE_K": 64,
118
+ "GROUP_SIZE_M": 32,
119
+ "num_warps": 4,
120
+ "num_stages": 2
121
+ },
122
+ "2048": {
123
+ "BLOCK_SIZE_M": 32,
124
+ "BLOCK_SIZE_N": 128,
125
+ "BLOCK_SIZE_K": 64,
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": 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": 64,
143
+ "num_warps": 4,
144
+ "num_stages": 2
145
+ }
146
+ }
sglang/srt/layers/moe/topk.py CHANGED
@@ -304,7 +304,7 @@ class TopK(CustomOp):
304
304
  global_num_experts = router_logits.shape[-1]
305
305
 
306
306
  # NOTE: now npu_moe_gating_top_k can only support `group_count=256` pattern
307
- if global_num_experts == 256 and self.topk_config.renormalize is False:
307
+ if global_num_experts == 256 and self.topk_config.renormalize is True:
308
308
 
309
309
  routed_scaling_factor = self.topk_config.routed_scaling_factor or 1
310
310
  router_logits = router_logits.to(torch.float32)