[Qwen3.5] Qwen3.5-27B inference repeat bug fix (#19411)
This commit is contained in:
@@ -352,6 +352,7 @@ class Qwen3_5LinearDecoderLayer(nn.Module):
|
|||||||
input_layernorm=self.input_layernorm,
|
input_layernorm=self.input_layernorm,
|
||||||
post_attention_layernorm=self.post_attention_layernorm,
|
post_attention_layernorm=self.post_attention_layernorm,
|
||||||
allow_reduce_scatter=True,
|
allow_reduce_scatter=True,
|
||||||
|
is_last_layer=(layer_id == config.num_hidden_layers - 1),
|
||||||
)
|
)
|
||||||
|
|
||||||
def forward(
|
def forward(
|
||||||
@@ -542,6 +543,7 @@ class Qwen3_5AttentionDecoderLayer(nn.Module):
|
|||||||
input_layernorm=self.input_layernorm,
|
input_layernorm=self.input_layernorm,
|
||||||
post_attention_layernorm=self.post_attention_layernorm,
|
post_attention_layernorm=self.post_attention_layernorm,
|
||||||
allow_reduce_scatter=True,
|
allow_reduce_scatter=True,
|
||||||
|
is_last_layer=(layer_id == config.num_hidden_layers - 1),
|
||||||
)
|
)
|
||||||
|
|
||||||
self.alt_stream = alt_stream
|
self.alt_stream = alt_stream
|
||||||
|
|||||||
Reference in New Issue
Block a user