Fix PP in Gemma4 (#40786)

Signed-off-by: Rohit kumar Singh <rksingh@habana.ai>
Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk>
This commit is contained in:
Rohit Kumar Singh
2026-04-29 03:17:51 -07:00
committed by GitHub
co-authored by Cyrus Leung
parent 6d7d4da99e
commit 5371d6fb40
+9 -16
View File
@@ -1144,11 +1144,6 @@ class Gemma4Model(nn.Module, EagleModelMixin):
dtype=dtype,
device=device,
),
"residual": torch.zeros(
(batch_size, hidden_size),
dtype=dtype,
device=device,
),
}
if ple_dim and ple_dim > 0:
tensors["per_layer_inputs"] = torch.zeros(
@@ -1312,13 +1307,12 @@ class Gemma4Model(nn.Module, EagleModelMixin):
per_layer_inputs = self.project_per_layer_inputs(
hidden_states, per_layer_embeds
)
residual = None
else:
assert intermediate_tensors is not None
hidden_states = intermediate_tensors["hidden_states"]
residual = intermediate_tensors["residual"]
per_layer_inputs = intermediate_tensors.get("per_layer_inputs")
if per_layer_inputs is not None:
per_layer_inputs = intermediate_tensors["per_layer_inputs"]
residual = None
aux_hidden_states = self._maybe_add_hidden_state([], 0, hidden_states, residual)
for layer_idx, layer in enumerate(
islice(self.layers, self.start_layer, self.end_layer)
@@ -1342,13 +1336,12 @@ class Gemma4Model(nn.Module, EagleModelMixin):
aux_hidden_states, layer_idx + 1, hidden_states, residual
)
if not get_pp_group().is_last_rank:
return IntermediateTensors(
{
"hidden_states": hidden_states,
"residual": residual,
"per_layer_inputs": per_layer_inputs,
}
)
tensors: dict[str, torch.Tensor] = {
"hidden_states": hidden_states,
}
if per_layer_inputs is not None:
tensors["per_layer_inputs"] = per_layer_inputs
return IntermediateTensors(tensors)
# Gemma4 incorporates residual into hidden_states directly
# Apply norm without residual fusion when possible.
if residual is None: