Update rope_scaling to rope_parameters in preparation for Transformers v5 (#28542)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-11-19 09:06:36 -08:00
committed by GitHub
parent d44e9df7d4
commit a8b70304d6
104 changed files with 542 additions and 910 deletions
+7 -6
View File
@@ -99,7 +99,6 @@ class Olmo2Attention(nn.Module):
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.max_position_embeddings = self.config.max_position_embeddings
self.rope_theta = self.config.rope_theta
# Attention input projection. Projects x -> (q, k, v)
self.qkv_proj = QKVParallelLinear(
@@ -139,15 +138,17 @@ class Olmo2Attention(nn.Module):
prefix=f"{prefix}.attn",
)
# Rotary embeddings. Rope scaling is only applied on full attention
# layers.
self.rope_scaling = self.config.rope_scaling if sliding_window is None else None
# Rotary embeddings. Rope scaling is only applied on full attention layers.
if sliding_window is None:
rope_parameters = self.config.rope_parameters
else:
rope_theta = self.config.rope_parameters["rope_theta"]
rope_parameters = {"rope_type": "default", "rope_theta": rope_theta}
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=self.max_position_embeddings,
base=self.rope_theta, # type: ignore
rope_scaling=self.rope_scaling,
rope_parameters=rope_parameters,
)
# Attention output projection.