Training gets stuck after completing just one trajectory

I’m running a customized version of art-e with additional tools locally. Training proceeds normally until the first validation stage, where it gets stuck for hours.

I am running Qwen2.5-1.5B-Instruct on single A100. Below is the code I use to load the model.

    from art.local.backend import LocalBackend

    backend = LocalBackend(path="./.art")

    model = art.TrainableModel(
        name="Qwen2.5-1.5B",
        project="my-model",
        base_model="Qwen/Qwen2.5-1.5B-Instruct",
    )

    model._internal_config = art.dev.InternalModelConfig(
        init_args=art.dev.InitArgs(
            max_seq_length=4096,
        ),
        peft_args=art.dev.PeftArgs(
            r=8,
            lora_alpha=8,
            target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
        ),
        trainer_args=art.dev.TrainerArgs(
            per_device_train_batch_size=1,
            gradient_accumulation_steps=2,
        ),
        engine_args=art.dev.EngineArgs(
            gpu_memory_utilization=0.8,
            enforce_eager=True,
        ),
    )
    await model.register(backend)

The model shows this and stuck

    train:   0%|          | 0/3 [00:00<?, ?it/s]==((====))==  Unsloth - 2x faster free finetuning | Num GPUs used = 1
       \\   /|    Num examples = 10,000,000 | Num Epochs = 3 | Total steps = 15,000,000
    O^O/ \_/ \    Batch size per device = 2 | Gradient accumulation steps = 2
    \        /    Data Parallel GPUs = 1 | Total batch size (2 x 2 x 1) = 4
     "-____-"     Trainable parameters = 2,179,072 of 1,545,893,376 (0.14% trained)

I’ve been monitoring VLLM.log, and it’s showing the same entries.
From VLLM.log:

INFO:     127.0.0.1:46160 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46166 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46172 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46178 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46184 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46190 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46196 - "GET /metrics HTTP/1.1" 200
INFO:     127.0.0.1:46204 - "GET /metrics HTTP/1.1" 200

From debug_internal.log:

{"time":"2025-11","level":"ERROR","msg":"request failed","error":"Get \"http://localhost:8000/metrics\": dial tcp [::1]:8000: connect: connection refused","method":"GET","url":"http://localhost:8000/metrics"}
{"time":"2025-11","level":"ERROR","msg":"request failed","error":"Get \"http://localhost:8000/metrics\": dial tcp [::1]:8000: connect: connection refused","method":"GET","url":"http://localhost:8000/metrics"}
{"time":"2025-11","level":"ERROR","msg":"request failed","error":"Get \"http://localhost:8000/metrics\": dial tcp [::1]:8000: connect: connection refused","method":"GET","url":"http://localhost:8000/metrics"}
{"time":"2025-11","level":"ERROR","msg":"request failed","error":"Get \"http://localhost:8000/metrics\": dial tcp [::1]:8000: connect: connection refused","method":"GET","url":"http://localhost:8000/metrics"}
{"time":"2025-11","level":"ERROR","msg":"monitor: error sampling metrics: GET http://localhost:8000/metrics giving up after 4 attempt(s): Get \"http://localhost:8000/metrics\": dial tcp [::1]:8000: connect: connection refused"}

Similar issue to 329, I tried installing the latest version from GitHub, but the issue still persists. pointers or examples would be greatly appreciated!

Start with the LocalBackend registration and TrainableModel configuration shown in the report, then compare the behavior with issue 329. Inspect VLLM.log and debug_internal.log while reproducing the first validation stage, focusing on the repeated /metrics requests and connection-refused errors. Done means training progresses past validation without remaining stuck and the metrics-monitoring failure is explained or handled.