[Model] Align nemotron config with final HF state and fix lm-eval-small (#7611)
This commit is contained in:
@@ -774,7 +774,7 @@ def get_rope(
|
||||
is_neox_style: bool = True,
|
||||
rope_scaling: Optional[Dict[str, Any]] = None,
|
||||
dtype: Optional[torch.dtype] = None,
|
||||
rotary_percent: float = 1.0,
|
||||
partial_rotary_factor: float = 1.0,
|
||||
) -> RotaryEmbedding:
|
||||
if dtype is None:
|
||||
dtype = torch.get_default_dtype()
|
||||
@@ -787,8 +787,8 @@ def get_rope(
|
||||
rope_scaling_args = tuple(rope_scaling_tuple.items())
|
||||
else:
|
||||
rope_scaling_args = None
|
||||
if rotary_percent < 1.0:
|
||||
rotary_dim = int(rotary_dim * rotary_percent)
|
||||
if partial_rotary_factor < 1.0:
|
||||
rotary_dim = int(rotary_dim * partial_rotary_factor)
|
||||
key = (head_size, rotary_dim, max_position, base, is_neox_style,
|
||||
rope_scaling_args, dtype)
|
||||
if key in _ROPE_DICT:
|
||||
|
||||
Reference in New Issue
Block a user