run_suite.py 6.62 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
17
18
19
20
        TestFile("models/lora/test_lora.py", 76),
        TestFile("models/lora/test_lora_backend.py", 420),
        TestFile("models/lora/test_multi_lora_backend.py", 1),
        TestFile("models/test_embedding_models.py", 119),
        TestFile("models/test_generation_models.py", 103),
21
        TestFile("models/test_grok_models.py", 60),
Lianmin Zheng's avatar
Lianmin Zheng committed
22
23
24
25
        TestFile("models/test_qwen_models.py", 82),
        TestFile("models/test_reward_models.py", 83),
        TestFile("models/test_gme_qwen_models.py", 45),
        TestFile("test_abort.py", 51),
26
        TestFile("test_awq.py"),
27
        TestFile("test_block_int8.py", 22),
Lianmin Zheng's avatar
Lianmin Zheng committed
28
29
        TestFile("test_chunked_prefill.py", 336),
        TestFile("test_eagle_infer.py", 447),
30
31
        TestFile("test_ebnf_constrained.py"),
        TestFile("test_fp8_kernel.py", 2),
Lianmin Zheng's avatar
Lianmin Zheng committed
32
        TestFile("test_embedding_openai_server.py", 36),
33
        TestFile("test_expert_distribution.py", 31),
Lianmin Zheng's avatar
Lianmin Zheng committed
34
        TestFile("test_gguf.py", 78),
35
36
37
        TestFile("test_gptqmodel_dynamic.py", 72),
        TestFile("test_hidden_states.py", 55),
        TestFile("test_int8_kernel.py", 1),
Lianmin Zheng's avatar
Lianmin Zheng committed
38
        TestFile("test_input_embeddings.py", 38),
39
40
41
        TestFile("test_json_constrained.py", 98),
        TestFile("test_large_max_new_tokens.py", 41),
        TestFile("test_metrics.py", 32),
Lianmin Zheng's avatar
Lianmin Zheng committed
42
43
44
45
46
47
48
49
        TestFile("test_mla.py", 92),
        TestFile("test_mla_deepseek_v3.py", 221),
        TestFile("test_mla_flashinfer.py", 395),
        TestFile("test_mla_fp8.py", 93),
        TestFile("test_no_chunked_prefill.py", 126),
        TestFile("test_no_overlap_scheduler.py", 262),
        TestFile("test_openai_server.py", 124),
        TestFile("test_penalty.py", 41),
Lianmin Zheng's avatar
Lianmin Zheng committed
50
        TestFile("test_page_size.py", 60),
Lianmin Zheng's avatar
Lianmin Zheng committed
51
52
        TestFile("test_pytorch_sampling_backend.py", 66),
        TestFile("test_radix_attention.py", 167),
53
        TestFile("test_reasoning_content.py", 89),
Lianmin Zheng's avatar
Lianmin Zheng committed
54
55
56
57
58
59
60
61
62
63
64
65
66
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),
        TestFile("test_skip_tokenizer_init.py", 72),
        TestFile("test_srt_engine.py", 237),
        TestFile("test_srt_endpoint.py", 94),
        TestFile("test_torch_compile.py", 76),
        TestFile("test_torch_compile_moe.py", 85),
        TestFile("test_torch_native_attention_backend.py", 149),
        TestFile("test_torchao.py", 70),
        TestFile("test_triton_attention_kernels.py", 4),
        TestFile("test_triton_attention_backend.py", 134),
        TestFile("test_update_weights_from_disk.py", 114),
        TestFile("test_update_weights_from_tensor.py", 48),
        TestFile("test_vertex_endpoint.py", 31),
        TestFile("test_vision_chunked_prefill.py", 223),
72
        TestFile("test_vlm_accuracy.py", 60),
Lianmin Zheng's avatar
Lianmin Zheng committed
73
        TestFile("test_vision_openai_server.py", 344),
74
        TestFile("test_fim_completion.py", 120),
Lianmin Zheng's avatar
Lianmin Zheng committed
75
        TestFile("test_w8a8_quantization.py", 46),
HandH1998's avatar
HandH1998 committed
76
        TestFile("test_eval_fp8_accuracy.py", 172),
77
        TestFile("test_create_kvindices.py", 2),
78
79
        TestFile("test_hicache.py", 60),
        TestFile("test_hicache_mla.py", 90),
80
    ],
81
    "nightly": [
Lianmin Zheng's avatar
Lianmin Zheng committed
82
        TestFile("test_nightly_gsm8k_eval.py"),
83
    ],
84
85
}

Lianmin Zheng's avatar
Lianmin Zheng committed
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129

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]

130
131
132
133
134
135

if __name__ == "__main__":
    arg_parser = argparse.ArgumentParser()
    arg_parser.add_argument(
        "--timeout-per-file",
        type=int,
136
        default=1800,
137
138
139
140
141
142
143
144
145
        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",
    )
146
147
148
149
150
151
152
153
154
155
156
157
    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
158
159
160
161
162
163
164
165
166
167
    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.",
    )
168
    args = arg_parser.parse_args()
Lianmin Zheng's avatar
Lianmin Zheng committed
169
    print(f"{args=}")
170
171
172
173
174
175

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

Lianmin Zheng's avatar
Lianmin Zheng committed
176
177
178
179
    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]
180

Lianmin Zheng's avatar
Lianmin Zheng committed
181
    print("The running tests are ", [f.name for f in files])
182

183
184
    exit_code = run_unittest_files(files, args.timeout_per_file)
    exit(exit_code)