423A35C7/LLaMA-Factory
1
0
Fork 0
mirror of https://github.com/hiyouga/LLaMA-Factory.git synced 2025-10-14 23:58:11 +08:00
Code Issues Packages Projects Releases Wiki Activity

Use official Nvidia base image

Browse Source 
Note that the flash-attn library is installed in this image and the qwen model will use it automatically.
However, if the the host machine's GPU is not compatible with the library, an exception will be raised during the training process as follows:
FlashAttention only supports Ampere GPUs or newer.
So if the --flash_attn flag is not set, an additional patch for the qwen model's config is necessary to set the default value of use_flash_attn from "auto" to False.


Former-commit-id: cd2f5717d676e1a5afd2f4e7a38402d2e55e7479
This commit is contained in:
S3Studio 2024-03-14 18:03:33 +08:00 committed by liuzhao2
parent c6873211e9
commit 096869c7b6
2 changed files with 4 additions and 1 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
Dockerfile
View File

@ -1,4 +1,4 @@
FROM cnstark/pytorch:2.0.1-py3.9.17-cuda11.8.0-ubuntu20.04 FROM nvcr.io/nvidia/pytorch:24.01-py3
WORKDIR /app WORKDIR /app

3
src/llmtuner/model/patcher.py
View File

@ -283,6 +283,9 @@ def patch_config(
setattr(config, dtype_name, model_args.compute_dtype == dtype) setattr(config, dtype_name, model_args.compute_dtype == dtype)
_configure_attn_implementation(model_args, init_kwargs) _configure_attn_implementation(model_args, init_kwargs)
if getattr(config, "model_type", None) == "qwen" and init_kwargs["attn_implementation"] != "flash_attention_2":
config.use_flash_attn = False
_configure_rope(config, model_args, is_trainable) _configure_rope(config, model_args, is_trainable)
_configure_longlora(config, model_args, is_trainable) _configure_longlora(config, model_args, is_trainable)
_configure_quantization(config, tokenizer, model_args, init_kwargs) _configure_quantization(config, tokenizer, model_args, init_kwargs)