这是indexloc提供的服务,不要输入任何密码
Skip to content

Conversation

@nirbo
Copy link
Contributor

@nirbo nirbo commented Oct 2, 2025

Summary

  • Add rwkv and rwkv7 entries to the LoRA target-module mapping so Transformers-based RWKV checkpoints automatically receive adapters.
  • Document the new behaviour in the README with a short RWKV snippet.
  • Introduce tests/test_rwkv_lora.py, a tiny forward/backward regression guarded by PEFT_RUN_RWKV_TESTS=1 so the Triton compile remains opt-in.

Testing

  make quality

  python -m pytest tests/test_rwkv_lora.py   # (skipped unless PEFT_RUN_RWKV_TESTS=1)

  # To exercise the RWKV-specific test:
  PEFT_RUN_RWKV_TESTS=1 python -m pytest tests/test_rwkv_lora.py

Copy link
Member

@BenjaminBossan BenjaminBossan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for adding default target modules for RWKV. I have some comments, please take a look.

README.md Outdated
# prints something like: Preheat the oven to 350 degrees and place the cookie dough in a baking dish [...]
```

> [!NOTE]
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

IMO this README entry can be removed, as we generally don't highlight individual models here.

@@ -0,0 +1,49 @@
# Copyright 2025-present the HuggingFace Inc. team.
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

There isn't really a need to add this test here, we don't have tests just to ensure that the default target modules are being set. I did, however, confirm that the test passes locally.

I could see an argument to add it to the general test suite, since RWKV has a different architecture:

PEFT_DECODER_MODELS_TO_TEST = [
"hf-internal-testing/tiny-random-OPTForCausalLM",
"hf-internal-testing/tiny-random-GPT2LMHeadModel",
"hf-internal-testing/tiny-random-BloomForCausalLM",
"hf-internal-testing/tiny-random-gpt_neo",
"hf-internal-testing/tiny-random-GPTJForCausalLM",
"hf-internal-testing/tiny-random-GPTBigCodeForCausalLM",
"trl-internal-testing/tiny-random-LlamaForCausalLM",
"peft-internal-testing/tiny-dummy-qwen2",
"hf-internal-testing/tiny-random-Gemma3ForCausalLM",
]

However, at the moment, the PEFT CI is already stressing the rate limit of HF Hub, so adding yet another model would not be a good idea. I think that if this situation relaxes and if we find that there is a big demand for RWKV finetuning with PEFT, we can consider that option.

@BenjaminBossan
Copy link
Member

gentle ping @nirbo

@nirbo
Copy link
Contributor Author

nirbo commented Oct 15, 2025

Sorry for the delay, been busy and this slipped my mind.
I've pushed the changes you requested.

@BenjaminBossan

@HuggingFaceDocBuilderDev

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

Copy link
Member

@BenjaminBossan BenjaminBossan left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for adding default target modules for RWKV, LGTM. The failing test is unrelated.

@BenjaminBossan BenjaminBossan merged commit 182f4c9 into huggingface:main Oct 16, 2025
12 of 13 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants