[Bugfix][Model] Gemma4: register aliased embedding scalars as buffers (#54213)

Signed-off-by: Yannick Schnider <Yannick.Schnider1@ibm.com>
This commit is contained in:
Yannick Schnider
2026-09-28 10:56:31 +00:00
committed by GitHub
parent d6cce94fd4
commit 8cc9aa5ad3
+13 -4
View File
@@ -846,13 +846,22 @@ class Gemma4SelfDecoderLayers(nn.Module):
# Shared references to modules owned by Gemma4Model — must be
# inside this nn.Module so torch.compile captures them.
self.embed_tokens = embed_tokens
self.normalizer = normalizer
self.embed_tokens_per_layer = embed_tokens_per_layer
self.embed_scale_per_layer = embed_scale_per_layer
self.per_layer_model_projection = per_layer_model_projection
self.per_layer_projection_norm = per_layer_projection_norm
self.per_layer_input_scale = per_layer_input_scale
self.per_layer_projection_scale = per_layer_projection_scale
# Register the shared scalars as buffers so they follow the model on
# .to(device); plain attributes stay on CPU.
self.register_buffer("normalizer", normalizer, persistent=False)
self.register_buffer(
"embed_scale_per_layer", embed_scale_per_layer, persistent=False
)
self.register_buffer(
"per_layer_input_scale", per_layer_input_scale, persistent=False
)
self.register_buffer(
"per_layer_projection_scale", per_layer_projection_scale, persistent=False
)
def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
return self.embed_tokens(input_ids) * self.normalizer