Revert "[skyrl-train] Updated RoPE Configuration for HF Models transformers) w. backward-compatible support for vLLM (#690)" #695
Conversation
…formers) w. backward-compatible support for vLLM (NovaSky-AI#690)" This reverts commit ad2b975.
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request reverts a previous change that introduced a breaking TypeError in the main branch. The revert appears to be correct and complete, touching documentation, configuration files, and multiple source code files to remove the rope_parameters configuration and restore the previous rope_scaling and rope_theta implementation. I've identified a minor type inconsistency in the reverted code and provided a suggestion to fix it. Otherwise, this PR looks good to merge to resolve the issue on main.
| if "rope_scaling" not in trainer_cfg: | ||
| return {} | ||
| if trainer_cfg.rope_scaling is None: | ||
| return None |
There was a problem hiding this comment.
The function get_rope_scaling_config is type-hinted to return dict[str, Any], but it can return None here, which is a type inconsistency. To fix this and align with the type hint, please return an empty dictionary instead of None. An empty dictionary is handled correctly by the callers to mean 'no rope scaling'.
| return None | |
| return {} |
…ormers) w. backward-compatible support for vLLM (NovaSky-AI#690)" (NovaSky-AI#695) NovaSky-AI#690 breaks main: ```python ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/tmp/ray/session_2025-11-10_23-41-08_808638_5823/runtime_resources/working_dir_files/_ray_pkg_70495ea0ae6ec02c/skyrl_train/workers/fsdp/fsdp_worker.py", line 66, in init_model wrapped_model = HFModelWrapper( ^^^^^^^^^^^^^^^ File "/tmp/ray/session_2025-11-10_23-41-08_808638_5823/runtime_resources/working_dir_files/_ray_pkg_70495ea0ae6ec02c/skyrl_train/model_wrapper.py", line 117, in __init__ self.model = model_class.from_pretrained( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ray/.cache/uv/builds-v0/.tmpxGZQ4v/lib/python3.12/site-packages/transformers/models/auto/auto_factory.py", line 604, in from_pretrained return model_class.from_pretrained( ^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/home/ray/.cache/uv/builds-v0/.tmpxGZQ4v/lib/python3.12/site-packages/transformers/modeling_utils.py", line 277, in _wrapper return func(*args, **kwargs) ^^^^^^^^^^^^^^^^^^^^^ File "/home/ray/.cache/uv/builds-v0/.tmpxGZQ4v/lib/python3.12/site-packages/transformers/modeling_utils.py", line 4971, in from_pretrained model = cls(config, *model_args, **model_kwargs) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ TypeError: Qwen2ForCausalLM.__init__() got an unexpected keyword argument 'rope_parameters' ```
1 participant
#690 breaks main: