Fix RoPE related failures in Transformers nightly tests (#29333)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-11-25 16:23:45 +00:00
committed by GitHub
parent a1f2676879
commit 0353d2e162
5 changed files with 37 additions and 33 deletions
+1 -1
View File
@@ -233,7 +233,7 @@ class BaiChuanDecoderLayer(nn.Module):
hidden_size=self.hidden_size,
num_heads=config.num_attention_heads,
position_embedding=position_embedding,
rope_parameters=config.rope_parameters,
rope_parameters=getattr(config, "rope_parameters", None),
max_position_embeddings=max_position_embeddings,
cache_config=cache_config,
quant_config=quant_config,
+1 -1
View File
@@ -100,7 +100,7 @@ class GPTJAttention(nn.Module):
self.head_size,
rotary_dim=config.rotary_dim,
max_position=max_position_embeddings,
rope_parameters=config.rope_parameters,
rope_parameters=getattr(config, "rope_parameters", None),
is_neox_style=False,
)
self.attn = Attention(
+1 -1
View File
@@ -239,7 +239,7 @@ class Grok1DecoderLayer(nn.Module):
num_heads=config.num_attention_heads,
max_position=config.max_position_embeddings,
num_kv_heads=config.num_key_value_heads,
rope_parameters=config.rope_parameters,
rope_parameters=getattr(config, "rope_parameters", None),
cache_config=cache_config,
quant_config=quant_config,
prefix=f"{prefix}.attn",
+1 -1
View File
@@ -262,7 +262,7 @@ class LlamaAttention(nn.Module):
self.head_dim,
rotary_dim=self.head_dim,
max_position=self.max_position_embeddings,
rope_parameters=config.rope_parameters,
rope_parameters=getattr(config, "rope_parameters", None),
is_neox_style=is_neox_style,
partial_rotary_factor=self.partial_rotary_factor,
)