run_suite.py 16.7 KB
Newer Older
1
2
import argparse
import glob
Lianmin Zheng's avatar
Lianmin Zheng committed
3
from dataclasses import dataclass
4
5
6

from sglang.test.test_utils import run_unittest_files

Lianmin Zheng's avatar
Lianmin Zheng committed
7
8
9
10
11
12
13

@dataclass
class TestFile:
    name: str
    estimated_time: float = 60


14
suites = {
15
    "per-commit": [
16
17
18
19
20
21
22
23
        TestFile("hicache/test_hicache.py", 116),
        TestFile("hicache/test_hicache_mla.py", 127),
        TestFile("hicache/test_hicache_storage.py", 127),
        TestFile("lora/test_lora.py", 200),
        TestFile("lora/test_lora_eviction.py", 200),
        TestFile("lora/test_lora_backend.py", 99),
        TestFile("lora/test_multi_lora_backend.py", 60),
        TestFile("lora/test_lora_cuda_graph.py", 250),
24
        TestFile("lora/test_lora_update.py", 400),
25
        TestFile("lora/test_lora_qwen3.py", 97),
26
        TestFile("lora/test_lora_radix_cache.py", 100),
27
        TestFile("models/test_embedding_models.py", 73),
28
        # TestFile("models/test_clip_models.py", 52),
woodx's avatar
woodx committed
29
30
        TestFile("models/test_encoder_embedding_models.py", 100),
        TestFile("models/test_cross_encoder_models.py", 100),
Lianmin Zheng's avatar
Lianmin Zheng committed
31
        TestFile("models/test_compressed_tensors_models.py", 42),
Lianmin Zheng's avatar
Lianmin Zheng committed
32
        TestFile("models/test_generation_models.py", 103),
33
        # TestFile("models/test_gme_qwen_models.py", 45),
Lianmin Zheng's avatar
Lianmin Zheng committed
34
        # TestFile("models/test_grok_models.py", 60),  # Disabled due to illegal memory access
Lianmin Zheng's avatar
Lianmin Zheng committed
35
        TestFile("models/test_qwen_models.py", 82),
Lianmin Zheng's avatar
Lianmin Zheng committed
36
        TestFile("models/test_reward_models.py", 132),
37
        TestFile("models/test_vlm_models.py", 437),
38
        TestFile("models/test_transformers_models.py", 320),
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
        TestFile("openai_server/basic/test_protocol.py", 10),
        TestFile("openai_server/basic/test_serving_chat.py", 10),
        TestFile("openai_server/basic/test_serving_completions.py", 10),
        TestFile("openai_server/basic/test_serving_embedding.py", 10),
        TestFile("openai_server/basic/test_openai_embedding.py", 141),
        TestFile("openai_server/basic/test_openai_server.py", 149),
        TestFile("openai_server/features/test_enable_thinking.py", 70),
        TestFile("openai_server/features/test_json_constrained.py", 98),
        TestFile("openai_server/features/test_json_mode.py", 90),
        TestFile("openai_server/features/test_openai_server_ebnf.py", 95),
        TestFile("openai_server/features/test_openai_server_hidden_states.py", 240),
        TestFile("openai_server/features/test_reasoning_content.py", 89),
        TestFile("openai_server/function_call/test_openai_function_calling.py", 60),
        TestFile("openai_server/function_call/test_tool_choice.py", 226),
        TestFile("openai_server/validation/test_large_max_new_tokens.py", 41),
        TestFile("openai_server/validation/test_matched_stop.py", 60),
        TestFile("openai_server/validation/test_openai_server_ignore_eos.py", 85),
        TestFile("openai_server/validation/test_request_length_validation.py", 31),
57
58
59
        TestFile("quant/test_block_int8.py", 22),
        TestFile("quant/test_fp8_kernel.py", 8),
        TestFile("quant/test_int8_kernel.py", 8),
60
        TestFile("quant/test_triton_scaled_mm.py", 8),
61
62
63
        TestFile("quant/test_w8a8_quantization.py", 46),
        TestFile("rl/test_update_weights_from_disk.py", 114),
        TestFile("rl/test_update_weights_from_tensor.py", 48),
Lianmin Zheng's avatar
Lianmin Zheng committed
64
        TestFile("test_abort.py", 51),
Lianmin Zheng's avatar
Lianmin Zheng committed
65
        TestFile("test_create_kvindices.py", 2),
66
        TestFile("test_chunked_prefill.py", 313),
67
        TestFile("test_eagle_infer_a.py", 370),
68
        TestFile("test_eagle_infer_b.py", 700),
Lianmin Zheng's avatar
Lianmin Zheng committed
69
70
        TestFile("test_ebnf_constrained.py", 108),
        TestFile("test_eval_fp8_accuracy.py", 303),
71
        TestFile("test_fa3.py", 376),
72
        # TestFile("test_flashmla.py", 352),
73
        TestFile("test_function_call_parser.py", 10),
Lianmin Zheng's avatar
Lianmin Zheng committed
74
        TestFile("test_fused_moe.py", 30),
75
        TestFile("test_gpt_oss_1gpu.py", 600),
76
        TestFile("test_harmony_parser.py", 20),
77
        TestFile("test_hidden_states.py", 55),
78
        TestFile("test_hybrid_attn_backend.py", 100),
Lianmin Zheng's avatar
Lianmin Zheng committed
79
        TestFile("test_input_embeddings.py", 38),
80
        TestFile("test_io_struct.py", 8),
81
        TestFile("test_jinja_template_utils.py", 1),
82
        TestFile("test_metrics.py", 32),
83
        TestFile("test_metrics_utils.py", 1),
84
        TestFile("test_mla.py", 167),
85
        TestFile("test_mla_deepseek_v3.py", 700),
86
87
88
        TestFile("test_mla_int8_deepseek_v3.py", 429),
        TestFile("test_mla_flashinfer.py", 302),
        TestFile("test_mla_fp8.py", 93),
89
        TestFile("test_multi_tokenizer.py", 230),
Lianmin Zheng's avatar
Lianmin Zheng committed
90
        TestFile("test_no_chunked_prefill.py", 108),
91
        TestFile("test_no_overlap_scheduler.py", 234),
92
        TestFile("test_original_logprobs.py", 200),
Lianmin Zheng's avatar
Lianmin Zheng committed
93
        TestFile("test_penalty.py", 41),
Lianmin Zheng's avatar
Lianmin Zheng committed
94
        TestFile("test_page_size.py", 60),
Lianmin Zheng's avatar
Lianmin Zheng committed
95
        TestFile("test_pytorch_sampling_backend.py", 66),
96
        TestFile("test_radix_attention.py", 105),
Lianmin Zheng's avatar
Lianmin Zheng committed
97
        TestFile("test_regex_constrained.py", 64),
98
        TestFile("test_reasoning_parser.py", 5),
Lianmin Zheng's avatar
Lianmin Zheng committed
99
        TestFile("test_retract_decode.py", 54),
100
        TestFile("test_request_queue_validation.py", 30),
Lianmin Zheng's avatar
Lianmin Zheng committed
101
        TestFile("test_server_args.py", 1),
102
        TestFile("test_skip_tokenizer_init.py", 117),
103
        TestFile("test_srt_engine.py", 261),
Lianmin Zheng's avatar
Lianmin Zheng committed
104
        TestFile("test_srt_endpoint.py", 130),
105
        TestFile("test_start_profile.py", 60),
Lianmin Zheng's avatar
Lianmin Zheng committed
106
        TestFile("test_torch_compile.py", 76),
Lianmin Zheng's avatar
Lianmin Zheng committed
107
        TestFile("test_torch_compile_moe.py", 172),
108
        TestFile("test_torch_native_attention_backend.py", 123),
Lianmin Zheng's avatar
Lianmin Zheng committed
109
110
        TestFile("test_torchao.py", 70),
        TestFile("test_triton_attention_kernels.py", 4),
111
        TestFile("test_triton_attention_backend.py", 150),
Lianmin Zheng's avatar
Lianmin Zheng committed
112
        TestFile("test_triton_moe_channel_fp8_kernel.py", 25),
113
        TestFile("test_triton_sliding_window.py", 250),
114
        TestFile("test_utils_update_weights.py", 48),
Lianmin Zheng's avatar
Lianmin Zheng committed
115
        TestFile("test_vision_chunked_prefill.py", 175),
116
        TestFile("test_vlm_input_format.py", 300),
117
118
        TestFile("test_vision_openai_server_a.py", 403),
        TestFile("test_vision_openai_server_b.py", 446),
119
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
120
    "per-commit-2-gpu": [
121
122
        TestFile("lora/test_lora_tp.py", 116),
        TestFile("rl/test_update_weights_from_distributed.py", 103),
Lianmin Zheng's avatar
Lianmin Zheng committed
123
124
125
126
        TestFile("test_data_parallelism.py", 73),
        TestFile("test_dp_attention.py", 277),
        TestFile("test_patch_torch.py", 19),
        TestFile("test_release_memory_occupation.py", 127),
127
128
        TestFile("hicache/test_hicache_storage_e2e.py", 400),
        TestFile("hicache/test_hicache_storage_benchmark.py", 400),
Lianmin Zheng's avatar
Lianmin Zheng committed
129
130
131
132
133
134
135
136
137
    ],
    "per-commit-4-gpu": [
        TestFile("test_gpt_oss_4gpu.py", 600),
        TestFile("test_local_attn.py", 250),
        TestFile("test_pp_single_node.py", 372),
        TestFile("test_multi_instance_release_memory_occupation.py", 64),
    ],
    "per-commit-8-gpu": [
        # Disabled because it hangs on the CI.
138
        # TestFile("ep/test_moe_ep.py", 181),
139
        TestFile("lora/test_lora_llama4.py", 600),
Lianmin Zheng's avatar
Lianmin Zheng committed
140
141
142
143
144
145
        TestFile("test_disaggregation.py", 499),
        TestFile("test_disaggregation_different_tp.py", 155),
        TestFile("test_full_deepseek_v3.py", 333),
    ],
    "per-commit-8-gpu-b200": [
        # add more here
146
        TestFile("test_gpt_oss_4gpu.py", 600),
Lianmin Zheng's avatar
Lianmin Zheng committed
147
148
    ],
    "per-commit-4-gpu-deepep": [
149
        TestFile("ep/test_deepep_small.py", 531),
Lianmin Zheng's avatar
Lianmin Zheng committed
150
151
    ],
    "per-commit-8-gpu-deepep": [
152
        TestFile("ep/test_deepep_large.py", 338),
Lianmin Zheng's avatar
Lianmin Zheng committed
153
    ],
154
155
156
    "per-commit-8-gpu-h20": [
        TestFile("quant/test_w4a8_deepseek_v3.py", 371),
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
157
158
159
160
    "nightly": [
        TestFile("test_nightly_gsm8k_eval.py"),
    ],
    "vllm_dependency_test": [
161
        TestFile("quant/test_awq.py", 163),
Lianmin Zheng's avatar
Lianmin Zheng committed
162
163
164
        TestFile("test_bnb.py", 5),
        TestFile("test_gptqmodel_dynamic.py", 102),
        TestFile("test_vllm_dependency.py", 185),
165
        # TestFile("test_gguf.py", 96),
Lianmin Zheng's avatar
Lianmin Zheng committed
166
167
168
169
170
    ],
}

# Add AMD tests
suite_amd = {
171
    "per-commit-amd": [
172
173
174
        TestFile("hicache/test_hicache.py", 116),
        TestFile("hicache/test_hicache_mla.py", 127),
        TestFile("hicache/test_hicache_storage.py", 127),
175
176
        TestFile("lora/test_lora.py", 200),
        TestFile("lora/test_lora_eviction.py", 200),
177
178
179
        TestFile("lora/test_lora_backend.py", 99),
        TestFile("lora/test_multi_lora_backend.py", 60),
        TestFile("lora/test_lora_cuda_graph.py", 250),
180
181
182
        TestFile("lora/test_lora_qwen3.py", 97),
        TestFile("models/test_embedding_models.py", 73),
        TestFile("models/test_compressed_tensors_models.py", 42),
183
184
        TestFile("models/test_qwen_models.py", 82),
        TestFile("models/test_reward_models.py", 132),
185
186
187
188
189
        TestFile("models/test_transformers_models.py", 320),
        TestFile("openai_server/basic/test_protocol.py", 10),
        TestFile("openai_server/basic/test_serving_chat.py", 10),
        TestFile("openai_server/basic/test_serving_completions.py", 10),
        TestFile("openai_server/basic/test_serving_embedding.py", 10),
190
        TestFile("openai_server/basic/test_openai_embedding.py", 141),
191
        TestFile("openai_server/basic/test_openai_server.py", 149),
192
        TestFile("openai_server/features/test_enable_thinking.py", 70),
193
194
195
196
        TestFile("openai_server/features/test_json_constrained.py", 98),
        TestFile("openai_server/features/test_json_mode.py", 90),
        TestFile("openai_server/features/test_openai_server_ebnf.py", 95),
        # TestFile("openai_server/features/test_openai_server_hidden_states.py", 240),
197
        TestFile("openai_server/features/test_reasoning_content.py", 89),
198
199
        TestFile("openai_server/function_call/test_openai_function_calling.py", 60),
        TestFile("openai_server/function_call/test_tool_choice.py", 226),
200
        TestFile("openai_server/validation/test_large_max_new_tokens.py", 41),
201
202
        TestFile("openai_server/validation/test_matched_stop.py", 60),
        TestFile("openai_server/validation/test_openai_server_ignore_eos.py", 85),
203
        TestFile("openai_server/validation/test_request_length_validation.py", 31),
204
205
206
        TestFile("quant/test_block_int8.py", 22),
        TestFile("quant/test_awq_dequant.py", 2),
        TestFile("rl/test_update_weights_from_disk.py", 114),
207
        # TestFile("rl/test_update_weights_from_tensor.py", 48),
208
209
210
        TestFile("test_abort.py", 51),
        TestFile("test_create_kvindices.py", 2),
        TestFile("test_chunked_prefill.py", 313),
211
        TestFile("test_ebnf_constrained.py", 108),
212
213
        TestFile("test_eval_fp8_accuracy.py", 303),
        TestFile("test_function_call_parser.py", 10),
214
        TestFile("test_fused_moe.py", 30),
215
        TestFile("test_input_embeddings.py", 38),
216
217
218
        TestFile("test_io_struct.py", 8),
        TestFile("test_jinja_template_utils.py", 1),
        TestFile("test_metrics.py", 32),
219
        TestFile("test_metrics_utils.py", 1),
220
221
        TestFile("test_mla.py", 242),
        TestFile("test_mla_deepseek_v3.py", 221),
222
        TestFile("test_no_chunked_prefill.py", 108),
223
        # TestFile("test_no_overlap_scheduler.py", 234), # Disabled temporarily and track in #7703
224
225
226
227
        TestFile("test_penalty.py", 41),
        TestFile("test_page_size.py", 60),
        TestFile("test_pytorch_sampling_backend.py", 66),
        TestFile("test_radix_attention.py", 105),
228
        TestFile("test_regex_constrained.py", 64),
229
        TestFile("test_retract_decode.py", 54),
230
231
        TestFile("test_reasoning_parser.py", 5),
        TestFile("test_rope_rocm.py", 3),
232
233
        TestFile("test_server_args.py", 1),
        TestFile("test_skip_tokenizer_init.py", 117),
234
235
        TestFile("test_srt_engine.py", 261),
        TestFile("test_srt_endpoint.py", 130),
236
237
        TestFile("test_torch_compile.py", 76),
        TestFile("test_torch_compile_moe.py", 172),
238
239
        TestFile("test_torch_native_attention_backend.py", 123),
        TestFile("test_triton_attention_backend.py", 150),
240
        # TestFile("test_vision_chunked_prefill.py", 175), # Disabled temporarily and track in #7701
241
242
        TestFile("test_wave_attention_kernels.py", 2),
        TestFile("test_wave_attention_backend.py", 150),
243
244
    ],
    "per-commit-2-gpu-amd": [
245
246
        TestFile("lora/test_lora_tp.py", 116),
        TestFile("rl/test_update_weights_from_distributed.py", 103),
247
248
        TestFile("test_data_parallelism.py", 73),
        TestFile("test_patch_torch.py", 19),
249
    ],
250
251
252
    "per-commit-4-gpu-amd": [
        TestFile("test_pp_single_node.py", 150),
    ],
253
254
255
    "per-commit-8-gpu-amd": [
        TestFile("test_full_deepseek_v3.py", 250),
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
256
257
    "nightly-amd": [
        TestFile("test_nightly_gsm8k_eval_amd.py"),
fzyzcjy's avatar
fzyzcjy committed
258
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
259
260
261
262
}

# Add Intel Xeon tests
suite_xeon = {
263
264
    "per-commit-cpu": [
        TestFile("cpu/test_activation.py"),
265
        TestFile("cpu/test_binding.py"),
266
267
268
        TestFile("cpu/test_decode.py"),
        TestFile("cpu/test_extend.py"),
        TestFile("cpu/test_gemm.py"),
269
        TestFile("cpu/test_mla.py"),
270
271
272
        TestFile("cpu/test_moe.py"),
        TestFile("cpu/test_norm.py"),
        TestFile("cpu/test_qkv_proj_with_rope.py"),
273
        TestFile("cpu/test_rope.py"),
274
        TestFile("cpu/test_shared_expert.py"),
275
        TestFile("cpu/test_topk.py"),
YanbingJiang's avatar
YanbingJiang committed
276
        TestFile("test_intel_amx_attention_backend.py"),
277
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
278
279
280
281
282
}

# Add Ascend NPU tests
suite_ascend = {
    "per-commit-1-ascend-npu": [
283
        TestFile("ascend/test_ascend_tp1_bf16.py", 400),
284
        TestFile("ascend/test_ascend_graph_tp1_bf16.py", 400),
285
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
286
    "per-commit-2-ascend-npu": [
287
        TestFile("ascend/test_ascend_tp2_bf16.py", 400),
288
        TestFile("ascend/test_ascend_graph_tp2_bf16.py", 400),
289
290
        TestFile("ascend/test_ascend_tp2_fia_bf16.py", 400),
        TestFile("ascend/test_ascend_mla_fia_w8a8int8.py", 400),
291
    ],
Lianmin Zheng's avatar
Lianmin Zheng committed
292
    "per-commit-4-ascend-npu": [
293
        TestFile("ascend/test_ascend_mla_w8a8int8.py", 400),
294
    ],
295
296
}

Lianmin Zheng's avatar
Lianmin Zheng committed
297
298
299
300
suites.update(suite_amd)
suites.update(suite_xeon)
suites.update(suite_ascend)

Lianmin Zheng's avatar
Lianmin Zheng committed
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344

def auto_partition(files, rank, size):
    """
    Partition files into size sublists with approximately equal sums of estimated times
    using stable sorting, and return the partition for the specified rank.

    Args:
        files (list): List of file objects with estimated_time attribute
        rank (int): Index of the partition to return (0 to size-1)
        size (int): Number of partitions

    Returns:
        list: List of file objects in the specified rank's partition
    """
    weights = [f.estimated_time for f in files]

    if not weights or size <= 0 or size > len(weights):
        return []

    # Create list of (weight, original_index) tuples
    # Using negative index as secondary key to maintain original order for equal weights
    indexed_weights = [(w, -i) for i, w in enumerate(weights)]
    # Stable sort in descending order by weight
    # If weights are equal, larger (negative) index comes first (i.e., earlier original position)
    indexed_weights = sorted(indexed_weights, reverse=True)

    # Extract original indices (negate back to positive)
    indexed_weights = [(w, -i) for w, i in indexed_weights]

    # Initialize partitions and their sums
    partitions = [[] for _ in range(size)]
    sums = [0.0] * size

    # Greedy approach: assign each weight to partition with smallest current sum
    for weight, idx in indexed_weights:
        # Find partition with minimum sum
        min_sum_idx = sums.index(min(sums))
        partitions[min_sum_idx].append(idx)
        sums[min_sum_idx] += weight

    # Return the files corresponding to the indices in the specified rank's partition
    indices = partitions[rank]
    return [files[i] for i in indices]

345
346
347
348
349
350

if __name__ == "__main__":
    arg_parser = argparse.ArgumentParser()
    arg_parser.add_argument(
        "--timeout-per-file",
        type=int,
351
        default=1800,
352
353
354
355
356
357
358
359
360
        help="The time limit for running one file in seconds.",
    )
    arg_parser.add_argument(
        "--suite",
        type=str,
        default=list(suites.keys())[0],
        choices=list(suites.keys()) + ["all"],
        help="The suite to run",
    )
361
362
363
364
365
366
367
368
369
370
371
372
    arg_parser.add_argument(
        "--range-begin",
        type=int,
        default=0,
        help="The begin index of the range of the files to run.",
    )
    arg_parser.add_argument(
        "--range-end",
        type=int,
        default=None,
        help="The end index of the range of the files to run.",
    )
Lianmin Zheng's avatar
Lianmin Zheng committed
373
374
375
376
377
378
379
380
381
382
    arg_parser.add_argument(
        "--auto-partition-id",
        type=int,
        help="Use auto load balancing. The part id.",
    )
    arg_parser.add_argument(
        "--auto-partition-size",
        type=int,
        help="Use auto load balancing. The number of parts.",
    )
383
    args = arg_parser.parse_args()
Lianmin Zheng's avatar
Lianmin Zheng committed
384
    print(f"{args=}")
385
386
387
388
389
390

    if args.suite == "all":
        files = glob.glob("**/test_*.py", recursive=True)
    else:
        files = suites[args.suite]

Lianmin Zheng's avatar
Lianmin Zheng committed
391
392
393
394
    if args.auto_partition_size:
        files = auto_partition(files, args.auto_partition_id, args.auto_partition_size)
    else:
        files = files[args.range_begin : args.range_end]
395

Lianmin Zheng's avatar
Lianmin Zheng committed
396
    print("The running tests are ", [f.name for f in files])
397

398
399
    exit_code = run_unittest_files(files, args.timeout_per_file)
    exit(exit_code)