run_suite.py 11.4 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": [
Lianmin Zheng's avatar
Lianmin Zheng committed
16
        TestFile("models/lora/test_lora.py", 76),
17
        TestFile("models/lora/test_lora_backend.py", 99),
18
        TestFile("models/lora/test_multi_lora_backend.py", 60),
19
20
        TestFile("models/lora/test_lora_cuda_graph.py", 250),
        TestFile("models/test_embedding_models.py", 73),
21
        # TestFile("models/test_clip_models.py", 52),
Lianmin Zheng's avatar
Lianmin Zheng committed
22
        TestFile("models/test_compressed_tensors_models.py", 42),
Lianmin Zheng's avatar
Lianmin Zheng committed
23
        TestFile("models/test_generation_models.py", 103),
24
        # TestFile("models/test_gme_qwen_models.py", 45),
Lianmin Zheng's avatar
Lianmin Zheng committed
25
        # TestFile("models/test_grok_models.py", 60),  # Disabled due to illegal memory access
Lianmin Zheng's avatar
Lianmin Zheng committed
26
        TestFile("models/test_qwen_models.py", 82),
Lianmin Zheng's avatar
Lianmin Zheng committed
27
        TestFile("models/test_reward_models.py", 132),
28
        TestFile("models/test_vlm_models.py", 437),
29
        TestFile("models/test_transformers_models.py", 320),
Lianmin Zheng's avatar
Lianmin Zheng committed
30
        TestFile("test_abort.py", 51),
31
        TestFile("test_block_int8.py", 22),
Lianmin Zheng's avatar
Lianmin Zheng committed
32
        TestFile("test_create_kvindices.py", 2),
33
34
        TestFile("test_chunked_prefill.py", 313),
        TestFile("test_eagle_infer.py", 619),
Lianmin Zheng's avatar
Lianmin Zheng committed
35
        TestFile("test_ebnf_constrained.py", 108),
36
        TestFile("test_enable_thinking.py", 70),
Lianmin Zheng's avatar
Lianmin Zheng committed
37
38
        TestFile("test_embedding_openai_server.py", 141),
        TestFile("test_eval_fp8_accuracy.py", 303),
39
        TestFile("test_fa3.py", 376),
40
        TestFile("test_flashmla.py", 352),
41
        TestFile("test_fp8_kernel.py", 8),
42
        TestFile("test_function_call_parser.py", 10),
Lianmin Zheng's avatar
Lianmin Zheng committed
43
44
        TestFile("test_fused_moe.py", 30),
        TestFile("test_hicache.py", 116),
45
        TestFile("test_hicache_mla.py", 127),
46
        TestFile("test_hidden_states.py", 55),
47
        TestFile("test_int8_kernel.py", 8),
Lianmin Zheng's avatar
Lianmin Zheng committed
48
        TestFile("test_input_embeddings.py", 38),
49
50
51
        TestFile("test_json_constrained.py", 98),
        TestFile("test_large_max_new_tokens.py", 41),
        TestFile("test_metrics.py", 32),
52
53
54
55
56
        TestFile("test_mla.py", 167),
        TestFile("test_mla_deepseek_v3.py", 342),
        TestFile("test_mla_int8_deepseek_v3.py", 429),
        TestFile("test_mla_flashinfer.py", 302),
        TestFile("test_mla_fp8.py", 93),
Lianmin Zheng's avatar
Lianmin Zheng committed
57
        TestFile("test_no_chunked_prefill.py", 108),
58
        TestFile("test_no_overlap_scheduler.py", 234),
59
        TestFile("test_openai_adapter.py", 1),
60
        TestFile("test_openai_function_calling.py", 60),
61
        TestFile("test_openai_server.py", 149),
Lianmin Zheng's avatar
Lianmin Zheng committed
62
        TestFile("test_penalty.py", 41),
Lianmin Zheng's avatar
Lianmin Zheng committed
63
        TestFile("test_page_size.py", 60),
Lianmin Zheng's avatar
Lianmin Zheng committed
64
        TestFile("test_pytorch_sampling_backend.py", 66),
65
        TestFile("test_radix_attention.py", 105),
66
        TestFile("test_reasoning_content.py", 89),
Lianmin Zheng's avatar
Lianmin Zheng committed
67
68
69
70
71
        TestFile("test_regex_constrained.py", 64),
        TestFile("test_release_memory_occupation.py", 44),
        TestFile("test_request_length_validation.py", 31),
        TestFile("test_retract_decode.py", 54),
        TestFile("test_server_args.py", 1),
72
        TestFile("test_skip_tokenizer_init.py", 117),
73
        TestFile("test_srt_engine.py", 261),
Lianmin Zheng's avatar
Lianmin Zheng committed
74
        TestFile("test_srt_endpoint.py", 130),
75
        TestFile("test_tool_choice.py", 226),
Lianmin Zheng's avatar
Lianmin Zheng committed
76
        TestFile("test_torch_compile.py", 76),
Lianmin Zheng's avatar
Lianmin Zheng committed
77
        TestFile("test_torch_compile_moe.py", 172),
78
        TestFile("test_torch_native_attention_backend.py", 123),
Lianmin Zheng's avatar
Lianmin Zheng committed
79
80
        TestFile("test_torchao.py", 70),
        TestFile("test_triton_attention_kernels.py", 4),
81
        TestFile("test_triton_attention_backend.py", 150),
Lianmin Zheng's avatar
Lianmin Zheng committed
82
        TestFile("test_triton_moe_channel_fp8_kernel.py", 25),
83
        TestFile("test_triton_sliding_window.py", 250),
Lianmin Zheng's avatar
Lianmin Zheng committed
84
85
86
        TestFile("test_update_weights_from_disk.py", 114),
        TestFile("test_update_weights_from_tensor.py", 48),
        TestFile("test_vertex_endpoint.py", 31),
Lianmin Zheng's avatar
Lianmin Zheng committed
87
        TestFile("test_vision_chunked_prefill.py", 175),
88
        TestFile("test_vlm_input_format.py", 300),
89
90
        TestFile("test_vision_openai_server_a.py", 584),
        TestFile("test_vision_openai_server_b.py", 556),
Lianmin Zheng's avatar
Lianmin Zheng committed
91
        TestFile("test_w8a8_quantization.py", 46),
92
    ],
93
    "per-commit-amd": [
94
95
96
        TestFile("models/lora/test_lora_backend.py", 99),
        TestFile("models/lora/test_multi_lora_backend.py", 60),
        TestFile("models/lora/test_lora_cuda_graph.py", 250),
97
98
99
100
        TestFile("test_mla.py", 242),
        TestFile("test_mla_deepseek_v3.py", 221),
        TestFile("test_torch_compile.py", 76),
        TestFile("test_torch_compile_moe.py", 172),
101
102
103
104
105
106
        TestFile("models/test_qwen_models.py", 82),
        TestFile("models/test_reward_models.py", 132),
        TestFile("test_abort.py", 51),
        TestFile("test_block_int8.py", 22),
        TestFile("test_create_kvindices.py", 2),
        TestFile("test_chunked_prefill.py", 313),
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
        TestFile("test_embedding_openai_server.py", 141),
        TestFile("test_eval_fp8_accuracy.py", 303),
        TestFile("test_function_call_parser.py", 10),
        TestFile("test_input_embeddings.py", 38),
        TestFile("test_large_max_new_tokens.py", 41),
        TestFile("test_metrics.py", 32),
        TestFile("test_no_chunked_prefill.py", 108),
        TestFile("test_no_overlap_scheduler.py", 234),
        TestFile("test_penalty.py", 41),
        TestFile("test_page_size.py", 60),
        TestFile("test_pytorch_sampling_backend.py", 66),
        TestFile("test_radix_attention.py", 105),
        TestFile("test_reasoning_content.py", 89),
        TestFile("test_enable_thinking.py", 70),
        TestFile("test_request_length_validation.py", 31),
        TestFile("test_retract_decode.py", 54),
        TestFile("test_server_args.py", 1),
        TestFile("test_skip_tokenizer_init.py", 117),
        TestFile("test_torch_native_attention_backend.py", 123),
        TestFile("test_triton_attention_backend.py", 150),
        TestFile("test_update_weights_from_disk.py", 114),
        TestFile("test_vertex_endpoint.py", 31),
        TestFile("test_vision_chunked_prefill.py", 175),
130
    ],
131
    "per-commit-2-gpu": [
132
133
134
135
136
137
138
        TestFile("models/lora/test_lora_tp.py", 116),
        TestFile("test_data_parallelism.py", 73),
        TestFile("test_dp_attention.py", 137),
        TestFile("test_mla_tp.py", 170),
        TestFile("test_moe_ep.py", 181),
        TestFile("test_patch_torch.py", 19),
        TestFile("test_update_weights_from_distributed.py", 103),
139
        TestFile("test_verl_engine_2_gpu.py", 64),
140
    ],
141
    "per-commit-2-gpu-amd": [
142
143
        TestFile("models/lora/test_lora_tp.py", 116),
        TestFile("test_data_parallelism.py", 73),
144
        TestFile("test_mla_tp.py", 170),
145
146
        TestFile("test_patch_torch.py", 19),
        TestFile("test_update_weights_from_distributed.py", 103),
147
    ],
148
149
150
    "per-commit-4-gpu": [
        TestFile("test_local_attn.py", 250),
        TestFile("test_pp_single_node.py", 150),
151
        TestFile("test_verl_engine_4_gpu.py", 64),
152
    ],
153
154
155
    "per-commit-4-gpu-amd": [
        TestFile("test_pp_single_node.py", 150),
    ],
Stefan He's avatar
Stefan He committed
156
    "per-commit-8-gpu": [
157
158
159
160
161
        # Disabled deepep tests temporarily because it takes too much time.
        # TODO: re-enable them after reducing the test time with compilation cache and smaller models.
        # TestFile("test_deepep_intranode.py", 50),
        # TestFile("test_deepep_low_latency.py", 50),
        # TestFile("test_moe_deepep_eval_accuracy_large.py", 250),
162
163
164
        TestFile("test_disaggregation.py", 270),
        TestFile("test_disaggregation_different_tp.py", 155),
        TestFile("test_full_deepseek_v3.py", 463),
Stefan He's avatar
Stefan He committed
165
    ],
166
167
168
    "per-commit-8-gpu-amd": [
        TestFile("test_full_deepseek_v3.py", 250),
    ],
169
170
171
172
173
174
175
176
177
178
    "per-commit-cpu": [
        TestFile("cpu/test_activation.py"),
        TestFile("cpu/test_decode.py"),
        TestFile("cpu/test_extend.py"),
        TestFile("cpu/test_gemm.py"),
        TestFile("cpu/test_moe.py"),
        TestFile("cpu/test_norm.py"),
        TestFile("cpu/test_qkv_proj_with_rope.py"),
        TestFile("cpu/test_shared_expert.py"),
    ],
179
    "nightly": [
Lianmin Zheng's avatar
Lianmin Zheng committed
180
        TestFile("test_nightly_gsm8k_eval.py"),
181
    ],
182
183
184
    "nightly-amd": [
        TestFile("test_nightly_gsm8k_eval_amd.py"),
    ],
185
186
    "vllm_dependency_test": [
        TestFile("test_awq.py"),
187
        TestFile("test_bnb.py"),
188
189
        TestFile("test_gguf.py", 78),
        TestFile("test_gptqmodel_dynamic.py", 72),
190
        TestFile("test_vllm_dependency.py"),
191
    ],
192
193
}

Lianmin Zheng's avatar
Lianmin Zheng committed
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237

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]

238
239
240
241
242
243

if __name__ == "__main__":
    arg_parser = argparse.ArgumentParser()
    arg_parser.add_argument(
        "--timeout-per-file",
        type=int,
244
        default=1800,
245
246
247
248
249
250
251
252
253
        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",
    )
254
255
256
257
258
259
260
261
262
263
264
265
    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
266
267
268
269
270
271
272
273
274
275
    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.",
    )
276
    args = arg_parser.parse_args()
Lianmin Zheng's avatar
Lianmin Zheng committed
277
    print(f"{args=}")
278
279
280
281
282
283

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

Lianmin Zheng's avatar
Lianmin Zheng committed
284
285
286
287
    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]
288

Lianmin Zheng's avatar
Lianmin Zheng committed
289
    print("The running tests are ", [f.name for f in files])
290

291
292
    exit_code = run_unittest_files(files, args.timeout_per_file)
    exit(exit_code)