【CI】update vllm case

autotest/config-npu.yaml

@@ -213,11 +213,10 @@ case:
             parameters:
                 config: autotest/config/npu_rl_qwen3_30B_grpo_vllm.py
                 infer_backend: vllm
-                acceleator: NPU
+                accelerator: NPU
                 output_path: /mnt/hwfile/llmrazor/qa-llm-cic/qa-llm-cicd/test_output
             resource:
                 image: ccr-hw/910c:rl_s1_vllm
-                pip_package: ls
                 envs:
                     - MODEL_PATH=/mnt/hwfile/llmrazor/qa-llm-cicd/qa_test_models/Qwen3-30B-A3B
                     - DATA_PATH=/mnt/hwfile/llmrazor/qa-llm-cicd/xtuner_resource/datasets/gsm8k/train-mini.jsonl

autotest/config/npu_rl_qwen3_30B_grpo_vllm.py

@@ -0,0 +1,183 @@
+import os
+from copy import deepcopy
+from pathlib import Path
+
+from xtuner.v1.config import AdamWConfig, FSDPConfig, LRConfig
+from xtuner.v1.data_proto.rl_data import SampleParams
+from xtuner.v1.datasets.config import DataloaderConfig, DatasetConfig
+from xtuner.v1.datasets.rl_tokenize_fn import RLTextTokenizeFnConfig
+from xtuner.v1.model import get_model_config_from_hf
+from xtuner.v1.rl.advantage import GRPOAdvantageConfig
+from xtuner.v1.rl.agent_loop import SingleTurnAgentLoopConfig
+from xtuner.v1.rl.agent_loop_manager import AgentLoopManagerConfig, SamplerConfig, SyncProduceStrategyConfig, TaskSpecConfig
+from xtuner.v1.rl.evaluator import EvaluatorConfig
+from xtuner.v1.rl.judger import GSM8KJudgerConfig
+from xtuner.v1.rl.loss import GRPOLossConfig
+from xtuner.v1.rl.replay_buffer import SyncReplayBufferConfig
+from xtuner.v1.rl.rollout.worker import RolloutConfig
+from xtuner.v1.rl.trainer import WorkerConfig
+from xtuner.v1.rl.utils import AcceleratorResourcesConfig
+from xtuner.v1.train.rl_trainer import RLColocateTrainerConfig
+
+
+work_dir = os.environ["WORK_DIR"]
+model_path = os.environ["MODEL_PATH"]
+data_path = os.environ["DATA_PATH"]
+eval_data_path = os.environ["EVAL_DATA_PATH"]
+enable_return_routed_experts = os.environ.get("ENABLE_RETURN_ROUTED_EXPERTS", "0")
+enable_evaluate = eval_data_path != ""
+
+# basic settings
+experimental_name = "grpo_gsm8k"
+total_epochs = 2
+global_batch_size = 64
+prompt_repeat_k = 5
+rollout_tp_size = 4
+rollout_dp_size = 4
+rollout_ep_size = 1
+max_prompt_length = 512
+max_response_length = 1024
+pack_max_length = 32768
+train_optimizer_steps = 1
+hf_interval = 100
+enable_initial_evaluate = True
+evaluate_step = 10
+
+# 1. resources
+resources = AcceleratorResourcesConfig(
+    accelerator="NPU",
+    num_workers=16,
+    num_cpus_per_worker=6,
+    cpu_memory_per_worker=16 * 1024**3,
+)
+
+# 2. rollout
+rollout_config = RolloutConfig(
+    env=experimental_name,
+    device=resources.accelerator,
+    model_path=model_path,
+    dtype="bfloat16",
+    tensor_parallel_size=rollout_tp_size,
+    data_parallel_size=rollout_dp_size,
+    expert_parallel_size=rollout_ep_size,
+    gpu_memory_utilization=0.85,
+    context_length=max_response_length + max_prompt_length,
+    enable_return_routed_experts=enable_return_routed_experts == "1",
+)
+
+# sampling params
+training_sample_params = SampleParams(max_tokens=max_response_length)
+evaluation_sample_params = deepcopy(training_sample_params)
+evaluation_sample_params.top_p = 1.0
+evaluation_sample_params.temperature = 0.0
+evaluation_sample_params.top_k = 1
+
+# 3. datasets
+tokenizer_config = RLTextTokenizeFnConfig(max_length=max_prompt_length)
+train_dataset_cfg = [
+    {
+        "dataset": DatasetConfig(name=experimental_name, anno_path=data_path),
+        "tokenize_fn": tokenizer_config,
+    }
+]
+eval_dataset_cfg = [
+    {
+        "dataset": DatasetConfig(name=experimental_name, anno_path=eval_data_path if enable_evaluate else data_path),
+        "tokenize_fn": tokenizer_config,
+    }
+]
+dataloader_cfg = DataloaderConfig(
+    dataset_config_list=train_dataset_cfg,
+    pack_max_length=pack_max_length,
+    collator="fake_collator",
+    pack_level="none",
+)
+eval_dataloader_cfg = DataloaderConfig(
+    dataset_config_list=eval_dataset_cfg,
+    pack_max_length=pack_max_length,
+    collator="fake_collator",
+    pack_level="none",
+)
+
+# 4. judger
+judger_config = GSM8KJudgerConfig(judger_name="openai/gsm8k")
+
+# 5. train worker
+model_cfg = get_model_config_from_hf(Path(model_path))
+optim_cfg = AdamWConfig(lr=1e-6, foreach=False)
+loss_cfg = GRPOLossConfig(
+    policy_loss_cfg=dict(
+        cliprange_high=0.2,
+        cliprange_low=0.2,
+        loss_type="vanilla",
+    ),
+    ignore_idx=-100,
+    use_kl_loss=False,
+    kl_loss_coef=0.001,
+    kl_loss_type="low_var_kl",
+    mode="chunk",
+    chunk_size=512,
+)
+lr_cfg = LRConfig(lr_type="constant", warmup_ratio=0, lr_min=1e-6)
+fsdp_cfg = FSDPConfig(torch_compile=False, cpu_offload=False, ep_size=1)
+train_worker_cfg = WorkerConfig(
+    model_cfg=model_cfg,
+    load_from=model_path,
+    optim_cfg=optim_cfg,
+    loss_cfg=loss_cfg,
+    lr_cfg=lr_cfg,
+    fsdp_cfg=fsdp_cfg,
+    sp_size=1,
+    optimizer_steps=train_optimizer_steps,
+    pack_max_length=pack_max_length,
+)
+
+# 6. agent loop managers
+agent_loop_config = SingleTurnAgentLoopConfig(
+    hf_checkpoint=model_path,
+    sample_params=training_sample_params,
+)
+agent_loop_manager_cfg = AgentLoopManagerConfig(
+    tasks=TaskSpecConfig(
+        task_name="train_task",
+        agent_loop_config=agent_loop_config,
+        judger_config=judger_config,
+        produce_strategy_config=SyncProduceStrategyConfig(),
+        sampler_config=SamplerConfig(dataloader_cfg=dataloader_cfg, prompt_repeat_k=prompt_repeat_k),
+    ),
+)
+
+eval_agent_loop_config = SingleTurnAgentLoopConfig(
+    hf_checkpoint=model_path,
+    sample_params=evaluation_sample_params,
+)
+eval_agent_loop_manager_cfg = AgentLoopManagerConfig(
+    tasks=TaskSpecConfig(
+        task_name="eval_task",
+        agent_loop_config=eval_agent_loop_config,
+        judger_config=judger_config,
+        sampler_config=SamplerConfig(dataloader_cfg=eval_dataloader_cfg, prompt_repeat_k=1),
+    ),
+)
+
+# 7. trainer
+trainer = RLColocateTrainerConfig(
+    resources=resources,
+    train_worker_cfg=train_worker_cfg,
+    rollout_config=rollout_config,
+    tokenizer_path=model_path,
+    replay_buffer_config=SyncReplayBufferConfig(),
+    agent_loop_manager_cfg=agent_loop_manager_cfg,
+    eval_agent_loop_manager_cfg=eval_agent_loop_manager_cfg,
+    evaluator_config=EvaluatorConfig(compute_metric_func=None),
+    load_from=model_path,
+    total_epochs=total_epochs,
+    train_batch_size=global_batch_size,
+    advantage_estimator_config=GRPOAdvantageConfig(eps=1e-8),
+    enable_evaluate=enable_evaluate,
+    enable_initial_evaluate=enable_evaluate and enable_initial_evaluate,
+    evaluate_step=evaluate_step,
+    work_dir=work_dir,
+    hf_interval=hf_interval,
+    exp_tracker="jsonl",
+)

autotest/module/train.py

@@ -34,7 +34,7 @@ def get_cmd(config):
             config["work_dir"] = work_dir
 
             # Print runtime image package list before training for version diagnosis.
-            image_version_check = 'echo "===== runtime pip list ====="; pip list; '
+            image_version_check = "echo '===== runtime pip list ====='; pip list; "
 
             # this patch is for torch 2.9.1 Conv3d memory issue fix
             cudnn_patch = (

autotest/utils/ci_run_rl.sh

@@ -156,7 +156,7 @@ elif [ "$ACCELERATOR" = "NPU" ]; then
         --dashboard-host=0.0.0.0 \
         --dashboard-port=$RAY_DASHBOARD_PORT \
         --include-dashboard=true \
-        --disable-usage-stats \
+        --disable-usage-stats 
     else
         RAY_RESOLVED_IP=$(nslookup $MASTER_ADDR | awk '/^Address: / { addr=$2 } END { print addr }')
         if [ -z "$RAY_RESOLVED_IP" ]; then