NASP/argus-cluster
84
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

V3.7 推理引擎从sglang 全量切换为vllm

Browse Source
This commit is contained in:
yuyr
2026-01-06 12:13:41 +08:00
parent a6c66995e3
commit 63963eba29
12 changed files with 484 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/mvp/configs/dev.yaml
View File

@@ -15,6 +15,8 @@ ray:
env_vars:
HF_ENDPOINT: "https://hf-mirror.com"
PYTHONUNBUFFERED: "1"
# v3.7: forbid HuggingFace Hub network access from Ray jobs (use cached snapshots).
HF_HUB_OFFLINE: "1"
# 用户自定义代码目录（可被 PYTHONPATH 注入）
user_code_path: "/private/user/code"

2
src/mvp/configs/dev_v30.yaml
View File

@@ -15,6 +15,8 @@ ray:
env_vars:
HF_ENDPOINT: "https://hf-mirror.com"
PYTHONUNBUFFERED: "1"
# v3.7: forbid HuggingFace Hub network access from Ray jobs (use cached snapshots).
HF_HUB_OFFLINE: "1"
# v3.0 先不支持 user code 执行
user_code_path: "/private/user/code"

11
src/mvp/docker-compose.yaml
View File

@@ -1,11 +1,6 @@
services:
ray_head:
build:
context: .
dockerfile: images/argus-ray-node/Dockerfile
args:
BASE_IMAGE: verlai/verl:sgl055.latest
image: argus/argus-ray-node:v2.5
image: argus/argus-ray-node:vllm011.latest
container_name: argus-ray-head
ports:
- "8265:8265"
@@ -96,7 +91,7 @@ services:
- argus-wandb
ray_worker_0:
image: argus/argus-ray-node:v2.5
image: argus/argus-ray-node:vllm011.latest
container_name: argus-ray-worker-0
volumes:
- ../../verl:/workspace/verl
@@ -128,7 +123,7 @@ services:
PYTHONUNBUFFERED: "1"
ray_worker_1:
image: argus/argus-ray-node:v2.5
image: argus/argus-ray-node:vllm011.latest
container_name: argus-ray-worker-1
volumes:
- ../../verl:/workspace/verl

2
src/mvp/images/argus-ray-node/Dockerfile
View File

@@ -1,4 +1,4 @@
ARG BASE_IMAGE=verlai/verl:sgl055.latest
ARG BASE_IMAGE=verlai/verl:vllm011.latest
FROM ${BASE_IMAGE}
SHELL ["/bin/bash", "-lc"]

2
src/mvp/py/argus/ray/builders.py
View File

@@ -38,7 +38,7 @@ def build_training_argv(spec: JobSpec, submission_id: str, job_dir: str) -> Buil
"actor_rollout_ref.actor.optim.lr=1e-6",
"actor_rollout_ref.actor.ppo_mini_batch_size=64",
"actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4",
"actor_rollout_ref.rollout.name=sglang",
"actor_rollout_ref.rollout.name=vllm",
"actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8",
"actor_rollout_ref.rollout.tensor_model_parallel_size=1",
"actor_rollout_ref.rollout.gpu_memory_utilization=0.4",

2
src/mvp/py/argus/service/ui.py
View File

@@ -416,7 +416,7 @@ command: |
actor_rollout_ref.actor.optim.lr=1e-6 \
actor_rollout_ref.actor.ppo_mini_batch_size=64 \
actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu=4 \
actor_rollout_ref.rollout.name=sglang \
actor_rollout_ref.rollout.name=vllm \
actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=8 \
actor_rollout_ref.rollout.tensor_model_parallel_size=1 \
actor_rollout_ref.rollout.gpu_memory_utilization=0.4 \

3
src/mvp/py/tests/test_builders.py
View File

@@ -29,6 +29,9 @@ def test_build_training_argv_ppo_smoke():
built = build_training_argv(spec, submission_id="sid", job_dir="/job")
assert built.argv[:3] == ["python3", "-m", "verl.trainer.main_ppo"]
assert "data.val_files=val.jsonl" in built.argv
# v3.7: default rollout backend switches from sglang -> vllm.
assert "actor_rollout_ref.rollout.name=vllm" in built.argv
assert "actor_rollout_ref.rollout.name=sglang" not in built.argv
assert "trainer.test_freq=-1" in built.argv

2
src/mvp/py/tests/test_ui.py
View File

@@ -93,6 +93,8 @@ def test_ui_new_task_contains_advanced_example_snippet(tmp_path, monkeypatch):
# workload is not needed for advanced in v3.5.
assert "# workload:" not in r.text
assert "actor_rollout_ref.actor.ppo_micro_batch_size_per_gpu" in r.text
# v3.7: default rollout backend switches from sglang -> vllm.
assert "actor_rollout_ref.rollout.name=vllm" in r.text
# v3.6: Advanced example uses platform-injected env vars so users don't need to edit W&B project/run.
assert "trainer.logger=${MVP_TRAINER_LOGGER}" in r.text
assert "trainer.project_name=${MVP_WANDB_PROJECT}" in r.text

14
src/mvp/scripts/run_all_v30_api.sh
View File

@@ -224,10 +224,18 @@ dexec "${HEAD_CONTAINER}" bash -lc "set -euo pipefail; \
(cp -f /private/common/datasets/gsm8k_sft/train.parquet '/private/users/${USER_ID}/datasets/gsm8k_sft/train.parquet' 2>/dev/null || cp -f /private/datasets/gsm8k_sft/train.parquet '/private/users/${USER_ID}/datasets/gsm8k_sft/train.parquet' 2>/dev/null || true); \
(cp -f /private/common/datasets/gsm8k_sft/test.parquet '/private/users/${USER_ID}/datasets/gsm8k_sft/test.parquet' 2>/dev/null || cp -f /private/datasets/gsm8k_sft/test.parquet '/private/users/${USER_ID}/datasets/gsm8k_sft/test.parquet' 2>/dev/null || true)"
echo "[host] resolve local model snapshot path (avoid HF mirror 429 for vllm rollout)"
LOCAL_MODEL_PATH="$(dexec "${HEAD_CONTAINER}" bash -lc "python3 - <<'PY'\nimport os\nfrom huggingface_hub import snapshot_download\nmodel_id=os.environ.get('MODEL_ID','Qwen/Qwen2.5-0.5B-Instruct')\nos.environ.setdefault('HF_HOME','/private/hf')\ntry:\n p=snapshot_download(repo_id=model_id, local_files_only=True)\n print(p)\nexcept Exception as e:\n raise SystemExit(f'ERROR: model snapshot not in cache; run 30_prepare_data_and_model.sh first. {e!r}')\nPY\n" MODEL_ID='Qwen/Qwen2.5-0.5B-Instruct' | tail -n 1)"
if [[ -z "${LOCAL_MODEL_PATH}" || "${LOCAL_MODEL_PATH}" != /* ]]; then
echo "ERROR: failed to resolve LOCAL_MODEL_PATH: ${LOCAL_MODEL_PATH}" >&2
exit 1
fi
echo "[host] local_model_path: ${LOCAL_MODEL_PATH}"
echo "[host] submit PPO/GRPO/SFT via API using user dataset paths"
PPO_TASK_ID="$(submit_taskspec_inline "${USER_TOKEN}" $'workload: ppo\nnnodes: 2\nn_gpus_per_node: 4\ncode_path: /private/common/code/verl/verl_repo\ntrain_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/train.parquet\nval_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/test.parquet\nmodel_id: Qwen/Qwen2.5-0.5B-Instruct\ntotal_epochs: 1\ntotal_training_steps: 10\nsave_freq: 10\n')"
GRPO_TASK_ID="$(submit_taskspec_inline "${USER_TOKEN}" $'workload: grpo\nnnodes: 2\nn_gpus_per_node: 4\ncode_path: /private/common/code/verl/verl_repo\ntrain_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/train.parquet\nval_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/test.parquet\nmodel_id: Qwen/Qwen2.5-0.5B-Instruct\ntotal_epochs: 1\ntotal_training_steps: 10\nsave_freq: 10\n')"
SFT_TASK_ID="$(submit_taskspec_inline "${USER_TOKEN}" $'workload: sft\nnnodes: 1\nn_gpus_per_node: 1\ncode_path: /private/common/code/verl/verl_repo\ntrain_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k_sft/train.parquet\nval_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k_sft/test.parquet\nmodel_id: Qwen/Qwen2.5-0.5B-Instruct\ntotal_epochs: 1\ntotal_training_steps: 10\nsave_freq: 10\n')"
PPO_TASK_ID="$(submit_taskspec_inline "${USER_TOKEN}" $'workload: ppo\nnnodes: 2\nn_gpus_per_node: 4\ncode_path: /private/common/code/verl/verl_repo\ntrain_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/train.parquet\nval_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/test.parquet\nmodel_id: '"${LOCAL_MODEL_PATH}"$'\ntotal_epochs: 1\ntotal_training_steps: 10\nsave_freq: 10\n')"
GRPO_TASK_ID="$(submit_taskspec_inline "${USER_TOKEN}" $'workload: grpo\nnnodes: 2\nn_gpus_per_node: 4\ncode_path: /private/common/code/verl/verl_repo\ntrain_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/train.parquet\nval_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k/test.parquet\nmodel_id: '"${LOCAL_MODEL_PATH}"$'\ntotal_epochs: 1\ntotal_training_steps: 10\nsave_freq: 10\n')"
SFT_TASK_ID="$(submit_taskspec_inline "${USER_TOKEN}" $'workload: sft\nnnodes: 1\nn_gpus_per_node: 1\ncode_path: /private/common/code/verl/verl_repo\ntrain_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k_sft/train.parquet\nval_file: /private/users/'"${USER_ID}"$'/datasets/gsm8k_sft/test.parquet\nmodel_id: '"${LOCAL_MODEL_PATH}"$'\ntotal_epochs: 1\ntotal_training_steps: 10\nsave_freq: 10\n')"
echo "[host] submitted task ids:"
echo " ppo=${PPO_TASK_ID}"