Add fp32_lm_head flag for vLLM precision parity#526
Draft
jlamypoirier wants to merge 1 commit into
Draft
Conversation
When True, upcasts the LM head linear's input and weight to FP32 before the matmul, matching vLLM's bf16_last_layer_fp32 quantization. This lets the trainer compute log-probabilities at the same numerical precision as the actor's sampling, so the importance-sampling ratio starts near 1.0 instead of being inflated by trainer/actor precision mismatch. The detached FP32 weight has requires_grad=False, which makes output_parallel_linear_backward skip the weight-grad path. The FSDP gradient contract is restored by computing grad_weight explicitly and accumulating into the original BF16 param's grad_buffer via accumulate_gradient. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1 task
jlamypoirier
added a commit
that referenced
this pull request
May 28, 2026
Adds an `fp32_lm_head` field on `LanguageModelHeadConfig`. When `True`, the LM head linear's input and weight are upcast to FP32 before the matmul, matching vLLM's `bf16_last_layer_fp32` quantization. This lets the trainer compute log-probabilities at the same numerical precision as the actor's sampling, so the importance-sampling ratio starts near 1.0 instead of being artificially inflated by a trainer/actor precision mismatch. The detached FP32 weight has `requires_grad=False`, which makes `output_parallel_linear_backward` skip the weight-grad path. The FSDP gradient contract is restored by computing `grad_weight = grad.t() @ saved_input` explicitly and accumulating into the original BF16 param's `grad_buffer` via `accumulate_gradient`. Off by default — disabled path is byte-identical to before. Cherry-picked from #526 to unblock the precision-evaluation tool's GSPO smoke test, which compares fp32_lm_head=true vs false. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2 tasks
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
1 participant
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
Adds an
fp32_lm_headfield onLanguageModelHeadConfig. WhenTrue, the LM head linear's input and weight are upcast to FP32 before the matmul, matching vLLM'sbf16_last_layer_fp32quantization. This lets the trainer compute log-probabilities at the same numerical precision as the actor's sampling, so the importance-sampling ratio starts near 1.0 instead of being artificially inflated by a trainer/actor precision mismatch.The detached FP32 weight has
requires_grad=False, which makesoutput_parallel_linear_backwardskip the weight-grad path. The FSDP gradient contract is restored by computinggrad_weight = grad.t() @ saved_inputexplicitly and accumulating into the original BF16 param'sgrad_bufferviaaccumulate_gradient.Off by default — disabled path is byte-identical to before.
Test plan
pytest tests/layers/test_lm_head.py— passesOriginally part of #502.