Standardise get_rope to use rope_parameters["partial_rotary_factor"], not rotary_dim (#30389)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -95,12 +95,13 @@ class GPTJAttention(nn.Module):
|
||||
scaling = self.head_size**-0.5
|
||||
assert getattr(config, "rotary", True)
|
||||
assert config.rotary_dim % 2 == 0
|
||||
rope_parameters = getattr(config, "rope_parameters", {})
|
||||
rope_parameters["partial_rotary_factor"] = config.rotary_dim / self.head_size
|
||||
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
|
||||
self.rotary_emb = get_rope(
|
||||
self.head_size,
|
||||
rotary_dim=config.rotary_dim,
|
||||
max_position=max_position_embeddings,
|
||||
rope_parameters=getattr(config, "rope_parameters", None),
|
||||
rope_parameters=rope_parameters,
|
||||
is_neox_style=False,
|
||||
)
|
||||
self.attn = Attention(
|
||||
|
||||
Reference in New Issue
Block a user