Skip to content

Commit ead3923

Browse files
zhuzilinzhuzilin
committed
add ci
1 parent 4a1044c commit ead3923

File tree

2 files changed

+165
-35
lines changed

2 files changed

+165
-35
lines changed

‎slime/ray/rollout.py‎

Lines changed: 17 additions & 35 deletions
Original file line numberDiff line numberDiff line change
@@ -106,9 +106,9 @@ class SglangConfig:
106106
107107
Loaded from ``--sglang-config`` YAML file.
108108
109-
**Multi-model format**::
109+
**Config format**::
110110
111-
models:
111+
sglang:
112112
- name: actor
113113
model_path: /path/to/actor
114114
num_gpus_per_engine: 2
@@ -125,14 +125,6 @@ class SglangConfig:
125125
- worker_type: regular
126126
num_gpus: 4
127127
128-
**Legacy single-model format** (still supported)::
129-
130-
engine_groups:
131-
- worker_type: prefill
132-
num_gpus: 4
133-
- worker_type: decode
134-
num_gpus: 10
135-
136128
Each model gets its own router. ``placeholder`` groups reserve GPU
137129
slots without creating engines. ``overrides`` are ``ServerArgs``
138130
field names applied on top of the base ``--sglang-*`` CLI args.
@@ -145,26 +137,22 @@ def from_yaml(path: str) -> "SglangConfig":
145137
with open(path) as f:
146138
data = yaml.safe_load(f)
147139

148-
if "models" in data:
149-
models = []
150-
for m in data["models"]:
151-
groups = [EngineGroupConfig(**g) for g in m.get("engine_groups", [])]
152-
models.append(
153-
ModelConfig(
154-
name=m["name"],
155-
model_path=m.get("model_path"),
156-
num_gpus_per_engine=m.get("num_gpus_per_engine"),
157-
engine_groups=groups,
158-
)
140+
assert "sglang" in data, (
141+
f"sglang config must have a 'sglang' key, got {list(data.keys())}. "
142+
f"Wrap your engine_groups inside a model entry under 'sglang'."
143+
)
144+
models = []
145+
for m in data["sglang"]:
146+
groups = [EngineGroupConfig(**g) for g in m.get("engine_groups", [])]
147+
models.append(
148+
ModelConfig(
149+
name=m["name"],
150+
model_path=m.get("model_path"),
151+
num_gpus_per_engine=m.get("num_gpus_per_engine"),
152+
engine_groups=groups,
159153
)
160-
return SglangConfig(models=models)
161-
162-
# Backward compat: top-level engine_groups → single "default" model.
163-
assert (
164-
"engine_groups" in data
165-
), f"sglang config must have 'models' or 'engine_groups' key, got {list(data.keys())}"
166-
groups = [EngineGroupConfig(**g) for g in data["engine_groups"]]
167-
return SglangConfig(models=[ModelConfig(name="default", engine_groups=groups)])
154+
)
155+
return SglangConfig(models=models)
168156

169157
@staticmethod
170158
def from_prefill_num_servers(args) -> "SglangConfig":
@@ -193,12 +181,6 @@ def has_pd_disaggregation(self) -> bool:
193181
def total_num_gpus(self) -> int:
194182
return sum(m.total_num_gpus for m in self.models)
195183

196-
# Backward compat: single-model access.
197-
@property
198-
def engine_groups(self) -> list[EngineGroupConfig]:
199-
assert len(self.models) == 1, "engine_groups property requires single-model config"
200-
return self.models[0].engine_groups
201-
202184

203185
@dataclasses.dataclass
204186
class EngineGroup:

‎tests/test_qwen2.5_0.5B_sglang_config.py‎

Lines changed: 148 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,148 @@
1+
import os
2+
import tempfile
3+
4+
import slime.utils.external_utils.command_utils as U
5+
6+
TIGHT_DEVICE_MEMORY = U.get_bool_env_var("SLIME_TEST_TIGHT_DEVICE_MEMORY", "1")
7+
8+
MODEL_NAME = "Qwen2.5-0.5B-Instruct"
9+
MODEL_TYPE = "qwen2.5-0.5B"
10+
NUM_GPUS = 4
11+
12+
# Inline sglang config: same model, 3 engine groups with different parallelism.
13+
# Group 1: 2 GPUs, 2 GPUs/engine (tp=2) → 1 engine
14+
# Group 2: 1 GPU, 1 GPU/engine (tp=1) → 1 engine
15+
# Group 3: 1 GPU, placeholder → reserves 1 GPU slot, no engine created
16+
SGLANG_CONFIG_YAML = """\
17+
sglang:
18+
- name: default
19+
engine_groups:
20+
- worker_type: regular
21+
num_gpus: 2
22+
num_gpus_per_engine: 2
23+
- worker_type: regular
24+
num_gpus: 1
25+
num_gpus_per_engine: 1
26+
- worker_type: placeholder
27+
num_gpus: 1
28+
"""
29+
30+
31+
def prepare():
32+
U.exec_command("mkdir -p /root/models /root/datasets")
33+
U.exec_command(f"huggingface-cli download Qwen/{MODEL_NAME} --local-dir /root/models/{MODEL_NAME}")
34+
U.hf_download_dataset("zhuzilin/gsm8k")
35+
36+
37+
def execute():
38+
# Write inline sglang config to a temp file
39+
config_file = tempfile.NamedTemporaryFile(mode="w", suffix=".yaml", prefix="sglang_config_", delete=False)
40+
config_file.write(SGLANG_CONFIG_YAML)
41+
config_file.flush()
42+
config_path = config_file.name
43+
44+
ckpt_args = f"--hf-checkpoint /root/models/{MODEL_NAME}/ " f"--ref-load /root/models/{MODEL_NAME}/ "
45+
46+
rollout_args = (
47+
"--prompt-data /root/datasets/gsm8k/train.parquet "
48+
"--input-key messages "
49+
"--label-key label "
50+
"--apply-chat-template "
51+
"--rollout-shuffle "
52+
"--rm-type math "
53+
"--num-rollout 3 "
54+
"--rollout-batch-size 8 "
55+
"--n-samples-per-prompt 4 "
56+
"--rollout-max-response-len 1024 "
57+
"--rollout-temperature 0.8 "
58+
"--over-sampling-batch-size 16 "
59+
"--dynamic-sampling-filter-path slime.rollout.filter_hub.dynamic_sampling_filters.check_reward_nonzero_std "
60+
"--global-batch-size 32 "
61+
)
62+
63+
eval_args = (
64+
"--eval-interval 20 "
65+
"--eval-prompt-data gsm8k /root/datasets/gsm8k/test.parquet "
66+
"--n-samples-per-eval-prompt 1 "
67+
"--eval-max-response-len 1024 "
68+
"--eval-top-k 1 "
69+
)
70+
71+
perf_args = (
72+
"--tensor-model-parallel-size 1 "
73+
"--sequence-parallel "
74+
"--pipeline-model-parallel-size 1 "
75+
"--context-parallel-size 1 "
76+
"--expert-model-parallel-size 1 "
77+
"--expert-tensor-parallel-size 1 "
78+
"--use-dynamic-batch-size "
79+
"--max-tokens-per-gpu 9216 "
80+
)
81+
82+
grpo_args = (
83+
"--advantage-estimator grpo "
84+
"--use-kl-loss "
85+
"--kl-loss-coef 0.00 "
86+
"--kl-loss-type low_var_kl "
87+
"--entropy-coef 0.00 "
88+
"--eps-clip 0.2 "
89+
"--eps-clip-high 0.28 "
90+
)
91+
92+
optimizer_args = (
93+
"--optimizer adam "
94+
"--lr 1e-6 "
95+
"--lr-decay-style constant "
96+
"--weight-decay 0.1 "
97+
"--adam-beta1 0.9 "
98+
"--adam-beta2 0.98 "
99+
)
100+
101+
sglang_args = (
102+
"--rollout-num-gpus-per-engine 1 "
103+
f"--sglang-mem-fraction-static {0.6 if TIGHT_DEVICE_MEMORY else 0.7} "
104+
"--sglang-enable-metrics "
105+
f"--sglang-config {config_path} "
106+
)
107+
108+
ci_args = "--ci-test "
109+
110+
misc_args = (
111+
"--attention-dropout 0.0 "
112+
"--hidden-dropout 0.0 "
113+
"--accumulate-allreduce-grads-in-fp32 "
114+
"--attention-softmax-in-fp32 "
115+
"--attention-backend flash "
116+
"--actor-num-nodes 1 "
117+
"--actor-num-gpus-per-node 4 "
118+
"--colocate "
119+
"--megatron-to-hf-mode bridge "
120+
)
121+
122+
train_args = (
123+
f"{ckpt_args} "
124+
f"{rollout_args} "
125+
f"{optimizer_args} "
126+
f"{grpo_args} "
127+
f"{U.get_default_wandb_args(__file__)} "
128+
f"{perf_args} "
129+
f"{eval_args} "
130+
f"{sglang_args} "
131+
f"{ci_args} "
132+
f"{misc_args} "
133+
)
134+
135+
U.execute_train(
136+
train_args=train_args,
137+
num_gpus_per_node=NUM_GPUS,
138+
megatron_model_type=MODEL_TYPE,
139+
)
140+
141+
142+
if __name__ == "__main__":
143+
prepare()
144+
os.environ.pop("http_proxy", None)
145+
os.environ.pop("https_proxy", None)
146+
os.environ.pop("HTTP_PROXY", None)
147+
os.environ.pop("HTTPS_PROXY", None)
148+
execute()

0 commit comments

Comments
 (0)