Repository navigation
[WIP] support value head - #2099
Closed
PengchengShi00 wants to merge 3 commits into
Closed
PengchengShi00 wants to merge 3 commits into
PengchengShi00 wants to merge 3 commits into
Conversation
Closed
YanhuiDua
reviewed
Oct 15, 2026
| param_dtype: Annotated[torch.dtype, Parameter(help="Data type for model parameters")] = torch.bfloat16 | ||
| reduce_dtype: Annotated[torch.dtype, Parameter(help="Data type for reduction operations")] = torch.bfloat16 | ||
| fp32_lm_head: Annotated[bool, Parameter(help="Use float32 for language model head")] = False | ||
| fp32_head: Annotated[bool, Parameter(help="Use float32 for LM / value head")] = False |
Collaborator
There was a problem hiding this comment.
lmhead / value head是否等于fp32一定是一起配置的么？是不是增加fp32_value_head更加合理，这样没有BC changes
Collaborator
Author
There was a problem hiding this comment.
PPO 的配置文件，对 actor 和 critic 分别写一个 FSDPConfig ，分别进行配置，所以在配置参数里只保留了一个精度控制参数
This was referenced Oct 16, 2026
Collaborator
Author
|
see: #2188 |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
ValueHead(nn.Linear(hidden, 1)) selected byTransformerConfig.head_type="value_head". Runtime module name stayslm_head; HF save/load usesvalue_head.weight. Compose/VL only needstext_config.head_type.CriticLossConfig: SFT MSE whenold_valuesis omitted, OpenAI-style clipped value loss when it is provided.value_headrejects MTP and does not tie embeddings.FSDPConfig.fp32_lm_head→fp32_head(covers LM and value heads).RolloutConfig.fp32_lm_headis unchanged. Actor checkpoints load into a critic withstrict=False; missing value-head weights are randomly initialized on all ranks.Out of scope
lm_loss_cfgis stillCELossConfig; critic is not wired intoTrainEngine/Trainer._prepare_model_input.TrainingWorker.fitremains the actor (advantages / logprobs) path.WorkerConfig.loss_cfgacceptsCriticLossConfigfor typing only.Test plan
tests/engine/test_critic_sft.pywithQWEN3_5_MOE_PATH(Qwen3.5-35B-A3B): MTP reject, HF key remap, SFT MSE, clipped value loss, actor→critic load /save_hf/ reloadFSDPConfig(fp32_head=...)