256c6e4884Replaces prefer_flex_attn_if_supported (which only returned flex_attention or None) with determine_attention_implementation, a centralized hierarchy: FA2 > Flex > SDPA > Eager. Changes: - New determine_attention_implementation function in _utils.py with clear priority chain - _set_attn_impl helper to stamp config consistently - _FLEX_EXCLUDED_MODELS / _FLEX_EXCLUDED_PREFIXES for model-specific exclusions - Gemma3N explicit eager override in vision.py (timm vision towers) - Preserved sdpa fallback for unmapped/remote-code vision configs - Config re-stamped to eager when supports_sdpa guard fires Co-authored-by: Datta Nimmaturi <Datta0@users.noreply.github.com> |
||
|---|---|---|
| .. | ||
| dataprep | ||
| kernels | ||
| models | ||
| optimizers | ||
| registry | ||
| utils | ||
| __init__.py | ||
| _auto_install.py | ||
| chat_templates.py | ||
| device_type.py | ||
| import_fixes.py | ||
| ollama_template_mappers.py | ||
| save.py | ||
| tokenizer_utils.py | ||
| trainer.py | ||