transformers
df67cd35 - Fix DeepSpeed mixed precision precedence over Accelerate defaults (#39856)

Commit
126 days ago
Fix DeepSpeed mixed precision precedence over Accelerate defaults (#39856) * Fix DeepSpeed mixed precision precedence over Accelerate defaults Resolves issue where Accelerate would default to bf16 mixed precision when a DeepSpeed config specifies fp16, causing a ValueError. The fix ensures DeepSpeed config takes precedence over TrainingArguments defaults while preserving explicit user settings. Changes: - Add override_training_args_from_deepspeed() method to handle config precedence - Reorder mixed precision environment variable setting in TrainingArguments - Ensure DeepSpeed fp16/bf16 settings override defaults but not explicit choices Fixes #39849 * Add tests for DeepSpeed mixed precision precedence fix - Add TestDeepSpeedMixedPrecisionPrecedence class with 3 focused tests - Test DeepSpeed fp16/bf16 config overriding TrainingArguments defaults - Test user explicit settings being preserved over DeepSpeed config - Test precedence hierarchy: user settings > DeepSpeed config > defaults - Replace massive 934-line test bloat with concise 50-line test suite - Tests cover core functionality of PR #39856 mixed precision precedence fix
Author
Parents
Loading