nemo_automodel._transformers.v4_patches.rotary#
Runtime RoPE patches for legacy v4-style remote-code models.
Module Contents#
Functions#
Unwrap DTensor to its local shard for numeric checks. |
|
Drop-in replacement for legacy rotary embedding forward methods. |
|
Return True when the legacy rotary workaround should run. |
|
Patch rotary embeddings to bypass fragile legacy HF runtime behavior. |
Data#
API#
- nemo_automodel._transformers.v4_patches.rotary.logger#
‘getLogger(…)’
- nemo_automodel._transformers.v4_patches.rotary._to_local(t)#
Unwrap DTensor to its local shard for numeric checks.
- nemo_automodel._transformers.v4_patches.rotary._safe_rope_forward(self, x, position_ids, **kwargs)#
Drop-in replacement for legacy rotary embedding forward methods.
- nemo_automodel._transformers.v4_patches.rotary._is_nemotron_flash_config(cfg)#
- nemo_automodel._transformers.v4_patches.rotary.should_fix_rotary_embeddings(model_parts)#
Return True when the legacy rotary workaround should run.
- nemo_automodel._transformers.v4_patches.rotary.fix_rotary_embeddings(model_parts)#
Patch rotary embeddings to bypass fragile legacy HF runtime behavior.
- nemo_automodel._transformers.v4_patches.rotary.__all__#
[‘fix_rotary_embeddings’, ‘should_fix_rotary_embeddings’]