L2-Day12安装xtuner提示 xtuner: command not found

按照课件下载并安装xtuner。

然后更新课件提供的runtime.txt中的依赖，报错误提示：

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.

xtuner 0.2.0 requires mmengine==0.11.0rc0, but you have mmengine 0.10.6 which is incompatible.

xtuner 0.2.0 requires torch>=2.6.0, but you have torch 2.5.1 which is incompatible.

xtuner 0.2.0 requires transformers==4.57.0, but you have transformers 4.48.0 which is incompatible.

Successfully installed mmengine-0.10.6 nvidia-cublas-cu12-12.4.5.8 nvidia-cuda-cupti-cu12-12.4.127 nvidia-cuda-nvrtc-cu12-12.4.127 nvidia-cuda-runtime-cu12-12.4.127 nvidia-cudnn-cu12-9.1.0.70 nvidia-cufft-cu12-11.2.1.3 nvidia-curand-cu12-10.3.5.147 nvidia-cusolver-cu12-11.6.1.9 nvidia-cusparse-cu12-12.3.1.170 nvidia-nccl-cu12-2.21.5 nvidia-nvjitlink-cu12-12.4.127 nvidia-nvtx-cu12-12.4.127 sympy-1.13.1 tokenizers-0.21.4 torch-2.5.1 torchvision-0.20.1 transformers-4.48.0 triton-3.1.0

执行xtuner train XXX.py

提示： xtuner: command not found

执行：pip show xtuner，结果如下：

Name: xtuner

Version: 0.2.0

Summary: A Next-Generation Training Engine Built for Ultra-Large MoE Models

Home-page: https://github.com/InternLM/xtuner

Author:

Author-email: The XTuner Authors <openmmlab@gmail.com>

License: Apache License 2.0

Location: /usr/local/miniconda3/envs/xtuner/lib/python3.10/site-packages

Editable project location: /root/xtuner

Requires: astor, bitsandbytes, codetiming, cyclopts, datasets, einops, gitpython, imageio, loguru, mathruler, mmengine, opencv-python-headless, openpyxl, peft, py-libnuma, pydantic, pylatexenc, scikit-image, scipy, sentencepiece, tensorboard, tiktoken, timm, torch, torchvision, transformers, transformers-stream-generator, xxhash

Required-by:

看到其他同学发帖也碰到过类似的情况，不知道如何解决

2025年12月06日 | 264人阅读

回答 | 共 3 个

按点赞量排序

聚客AI-Li

按最新的官方文档上面的方法安装，有更新

2025年12月07日

请先登录 · 注册

zgt

请问老师，是按https://xtuner.readthedocs.io/zh-cn/latest/index.html这个文档说明安装吗？

安装后，之前的这个xtuner train 命令还能用吗？使用这个命令，还是提示：

xtuner: command not found。

该文档启用微调训练使用的torchrun命令，按给的示例方法执行：

torchrun --nproc-per-node 1 xtuner/v1/train/cli/sft.py --load-from /root/autodl-tmp/model/Qwen/Qwen3-0.6B --chat_template qwen3 --dataset /root/autodl-tmp/xtuner/tests/resource/openai_sft.jsonl --total-step 100 --work-dir /root/autodl-tmp/xtuner/work_dir

还是不成功，微调的参数如何设置也不清楚，还请老师解惑：

执行的输出如下：

Could not find the bitsandbytes CUDA binary at PosixPath('/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda128.so')

The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable.

Import FlashAttention 2 failed No module named 'flash_attn', Please install it manually.

/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/mmengine/utils/package_utils.py:17: UserWarning: pkg_resources is deprecated as an API. See https://setuptools.pypa.io/en/latest/pkg_resources.html. The pkg_resources package is slated for removal as early as 2025-11-30. Refrain from using this package or pin to Setuptools<81.

import pkg_resources

_bootstrap._exec(spec, module)

[XTuner][RANK 0][2025-12-08 09:46:12][WARNING] `resume_cfg` is deprecated, please use `auto_resume` and `load_checkpoint_cfg` instead

[XTuner][RANK 0][2025-12-08 09:46:12][INFO] set HF_MODULES_CACHE to /root/.cache/huggingface/modules_pid_8308 for current process 8308

[XTuner][RANK 0][2025-12-08 09:46:12][INFO]

============XTuner Training Environment============

XTUNER_DETERMINISTIC: None

XTUNER_GC_ENABLE: None

XTUNER_FILE_OPEN_CONCURRENCY: None

XTUNER_TOKENIZE_CHUNK_SIZE: None

XTUNER_TOKENIZE_WORKERS: None

XTUNER_ACTIVATION_OFFLOAD: None

XTUNER_USE_FA3: None

XTUNER_DISPATCHER_DEBUG: None

XTUNER_ROUTER_DEBUG: None

XTUNER_DECORD_VIDEO_THREADS: None

XTUNER_USE_CUTLASS_GROUP_GEMM: None

GROUPED_GEMM_USE_CUTLASS: None

XTUNER_USE_NATIVE_RMSNORM: None

==================================================

[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0

[XTuner][RANK 0][2025-12-08 09:46:12][WARNING] `dataset_cfg` is deprecated, please use `dataloader_cfg.dataset_config_list` instead

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] [Dataset] Start loading [default]/root/autodl-tmp/xtuner/tests/resource/openai_sft.jsonl with sample_ratio=1.0.

[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] [Dataset] (Original) default/openai_sft.jsonl: 128 samples.

[XTuner][RANK 0][2025-12-08 09:46:13][SUCCESS] [Build Datasets] Elapsed time 0.28 seconds, peak gpu memory 0.0G

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] [Dataset] Start packing data of ExpandSoftPackDataset.

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] Using 8 pack workers for packing datasets.

1it [00:00, 325.64it/s]

[XTuner][RANK 0][2025-12-08 09:46:13][SUCCESS] [Pack Datasets] Elapsed time 0.12 seconds, peak gpu memory 0.0G

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] [Dataset] (Original) 128 samples.

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] [Dataset] (Packed) 3 samples.

[Gloo] Rank 0 is connected to 0 peer ranks. Expected number of connected peer ranks is : 0

[FSDP Sharding]: 0%| | 0/28 [00:00<?, ?it/s]/root/autodl-tmp/xtuner/xtuner/v1/model/utils/checkpointing.py:92: FutureWarning: Please specify CheckpointImpl.NO_REENTRANT as CheckpointImpl.REENTRANT will soon be removed as the default and eventually deprecated.

return ptd_checkpoint_wrapper(module, *args, **kwargs)

[FSDP Sharding]: 100%|███████████████████████████████████████████████████████████████████████████████████████████████| 28/28 [00:00<00:00, 264.99it/s]

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] FSDPQwen3Dense(

(norm): FSDPRMSNorm((1024,), eps=1e-06)

(lm_head): FSDPLMHead(in_features=1024, out_features=151936, bias=False)

(layers): ModuleDict(

(0): FSDPCheckpointWrapper(

(_checkpoint_wrapped_module): DenseDecoderLayer(

(self_attn): MultiHeadAttention(

(q_proj): _Linear(in_features=1024, out_features=2048, bias=False)

(k_proj): _Linear(in_features=1024, out_features=1024, bias=False)

(v_proj): _Linear(in_features=1024, out_features=1024, bias=False)

(o_proj): _Linear(in_features=2048, out_features=1024, bias=False)

(q_norm): RMSNorm((128,), eps=1e-06)

(k_norm): RMSNorm((128,), eps=1e-06)

)

(mlp): DenseMLP(

(gate_proj): _Linear(in_features=1024, out_features=3072, bias=False)

(up_proj): _Linear(in_features=1024, out_features=3072, bias=False)

(down_proj): _Linear(in_features=3072, out_features=1024, bias=False)

)

(input_layernorm): RMSNorm((1024,), eps=1e-06)

(post_attention_layernorm): RMSNorm((1024,), eps=1e-06)

)

(1): FSDPCheckpointWrapper(

(_checkpoint_wrapped_module): DenseDecoderLayer(

(self_attn): MultiHeadAttention(

(q_proj): _Linear(in_features=1024, out_features=2048, bias=False)

(k_proj): _Linear(in_features=1024, out_features=1024, bias=False)

(v_proj): _Linear(in_features=1024, out_features=1024, bias=False)

(o_proj): _Linear(in_features=2048, out_features=1024, bias=False)

(q_norm): RMSNorm((128,), eps=1e-06)

(k_norm): RMSNorm((128,), eps=1e-06)

)

(mlp): DenseMLP(

(gate_proj): _Linear(in_features=1024, out_features=3072, bias=False)

(up_proj): _Linear(in_features=1024, out_features=3072, bias=False)

(down_proj): _Linear(in_features=3072, out_features=1024, bias=False)

)

(input_layernorm): RMSNorm((1024,), eps=1e-06)

(post_attention_layernorm): RMSNorm((1024,), eps=1e-06)

)

................

(27): FSDPCheckpointWrapper(

(_checkpoint_wrapped_module): DenseDecoderLayer(

(self_attn): MultiHeadAttention(

(q_proj): _Linear(in_features=1024, out_features=2048, bias=False)

(k_proj): _Linear(in_features=1024, out_features=1024, bias=False)

(v_proj): _Linear(in_features=1024, out_features=1024, bias=False)

(o_proj): _Linear(in_features=2048, out_features=1024, bias=False)

(q_norm): RMSNorm((128,), eps=1e-06)

(k_norm): RMSNorm((128,), eps=1e-06)

)

(mlp): DenseMLP(

(gate_proj): _Linear(in_features=1024, out_features=3072, bias=False)

(up_proj): _Linear(in_features=1024, out_features=3072, bias=False)

(down_proj): _Linear(in_features=3072, out_features=1024, bias=False)

)

(input_layernorm): RMSNorm((1024,), eps=1e-06)

(post_attention_layernorm): RMSNorm((1024,), eps=1e-06)

)

(rotary_emb): RotaryEmbedding()

(embed_tokens): FSDPEmbedding(151936, 1024)

)

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] Total trainable parameters: 596.0M, total parameters: 596.0M

[XTuner][RANK 0][2025-12-08 09:46:13][INFO] Untrainable parameters names: []

[rank0]: Traceback (most recent call last):

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/train/cli/sft.py", line 43, in <module>

[rank0]: app(exit_on_error=False)

[rank0]: File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/cyclopts/core.py", line 1865, in __call__

[rank0]: result = _run_maybe_async_command(command, bound, resolved_backend)

[rank0]: File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/cyclopts/_run.py", line 50, in _run_maybe_async_command

[rank0]: return command(*bound.args, **bound.kwargs)

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/train/cli/sft.py", line 35, in main

[rank0]: trainer = Trainer.from_config(trainer_cfg)

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/train/trainer.py", line 642, in from_config

[rank0]: self = cls(

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/train/trainer.py", line 594, in __init__

[rank0]: self._engine = self.build_engine(

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/train/trainer.py", line 1018, in build_engine

[rank0]: engine.from_hf(hf_path=model_path, strict=strict)

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/engine/train_engine.py", line 355, in from_hf

[rank0]: self.model.from_hf(hf_path=hf_path, strict=strict)

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/model/dense/dense.py", line 144, in from_hf

[rank0]: loaded_keys, unloaded_keys, missing_keys = super().from_hf(hf_path, strict)

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/model/base.py", line 217, in from_hf

[rank0]: loaded_keys, unloaded_keys, missing_keys = self._load_params(hf_loader, strict=strict)

[rank0]: File "/root/autodl-tmp/xtuner/xtuner/v1/model/base.py", line 906, in _load_params

[rank0]: raise RuntimeError(f"Unexpected keys in checkpoint: {unexpected_keys}. ")

[rank0]: RuntimeError: Unexpected keys in checkpoint: {'lm_head.weight'}.

[rank0]:[W1208 09:46:14.098953152 ProcessGroupNCCL.cpp:1524] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())

E1208 09:46:14.855000 8291 site-packages/torch/distributed/elastic/multiprocessing/api.py:882] failed (exitcode: 1) local_rank: 0 (pid: 8308) of binary: /root/miniconda3/envs/xtuner/bin/python3.10

Traceback (most recent call last):

File "/root/miniconda3/envs/xtuner/bin/torchrun", line 7, in <module>

sys.exit(main())

File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 357, in wrapper

return f(*args, **kwargs)

File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/torch/distributed/run.py", line 936, in main

run(args)

File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/torch/distributed/run.py", line 927, in run

elastic_launch(

File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 156, in __call__

return launch_agent(self._config, self._entrypoint, list(args))

File "/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/torch/distributed/launcher/api.py", line 293, in launch_agent

raise ChildFailedError(

torch.distributed.elastic.multiprocessing.errors.ChildFailedError:

============================================================

xtuner/v1/train/cli/sft.py FAILED

2025年12月08日

请先登录 · 注册

聚客AI-Li

新建环境，按官方文档安装，最新的xtuner版本可能存在部分函数使用会发生变化，你使用时按报错百度下，或者在文档上查找就行；
Could not find the bitsandbytes CUDA binary at PosixPath('/root/miniconda3/envs/xtuner/lib/python3.10/site-packages/bitsandbytes/libbitsandbytes_cuda128.so')
The installed version of bitsandbytes was compiled without GPU support. 8-bit optimizers, 8-bit multiplication, and GPU quantization are unavailable.
Import FlashAttention 2 failed No module named 'flash_attn', Please install it manually.
这个报错，表示你安装的bitsandbytes 、FlashAttention 不对，很可能就是你xtuner安装错误导致的

2025年12月09日

请先登录 · 注册

游客