Skip to content

PR #6837 staging CI#35

Open
Imagineer99 wants to merge 20 commits into
mainfrom
fix/pr-6837-ci
Open

PR #6837 staging CI#35
Imagineer99 wants to merge 20 commits into
mainfrom
fix/pr-6837-ci

Conversation

@Imagineer99

Copy link
Copy Markdown
Owner

Staging-only CI validation for unslothai#6837.

This PR exists only to run fork GitHub Actions checks and should be closed after validation.

Ayushman Paul and others added 20 commits July 2, 2026 12:39
When doing full finetuning (FFT) of a bfloat16 model, the fp16/bf16
mismatch validation fires before the corrective logic runs, causing a
misleading error even though the code would properly handle it downstream.
Skip the validation when full_finetuning is active.

Fixes unslothai#6731
…idation

Instead of entirely skipping validation (which could let mismatches
through when mixed_precision_dtype is float32), auto-correct explicit
fp16/bf16 settings that conflict with the model's dtype for FFT. This
way the existing validation still catches real mismatches for non-FFT
cases, and the corrective logic below handles the normalized settings.

Fixes the issue raised in Codex review of PR unslothai#6813.
Detect installed ROCm torch directly before applying the torchao override so Windows ROCm environments never install the crashing torchao package even if the earlier ROCm-installed flag is missing.
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Tolerate stray stdout noise when probing Windows ROCm torch installs by checking the last non-empty output line, matching the existing torch version probe behavior. Also keep args.fp16 and args.bf16 synchronized with the full-finetuning precision auto-corrections in the RL trainer patch so downstream eval settings see a consistent TrainingArguments state.
Patch imported MLXTrainer and MLXTrainingConfig objects to preserve the expected dataclass field ordering and to provide a _train_dataset_for_batches fallback when older trainers or test doubles only expose train_dataset. Also add focused worker tests covering both compatibility paths.
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.

2 participants