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
+1 -6
View File
@@ -161,7 +161,6 @@ class MixtralAttention(nn.Module):
num_heads: int,
num_kv_heads: int,
max_position: int = 4096 * 32,
rope_theta: float = 10000,
cache_config: CacheConfig | None = None,
quant_config: QuantizationConfig | None = None,
prefix: str = "",
@@ -189,7 +188,6 @@ class MixtralAttention(nn.Module):
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
self.rope_theta = rope_theta
self.qkv_proj = QKVParallelLinear(
hidden_size,
@@ -211,7 +209,7 @@ class MixtralAttention(nn.Module):
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position,
base=int(self.rope_theta),
rope_parameters=config.rope_parameters,
is_neox_style=True,
)
self.attn = Attention(
@@ -248,15 +246,12 @@ class MixtralDecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
# Requires transformers > 4.32.0
rope_theta = getattr(config, "rope_theta", 10000)
self.self_attn = MixtralAttention(
config=config,
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
max_position=config.max_position_embeddings,
num_kv_heads=config.num_key_value_heads,
rope_theta=rope_theta,
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.self_attn",