forked from Karylab-cklius/vllm
@@ -374,8 +374,9 @@ class DeepseekV4MultiHeadLatentAttentionWrapper(PluggableLayer):
|
||||
)
|
||||
elif self.compressor is not None:
|
||||
# Compressor on default, kv_insert on aux.
|
||||
compressor = self.compressor
|
||||
maybe_execute_in_parallel(
|
||||
lambda: self.compressor(hidden_states, positions, self.rotary_emb),
|
||||
lambda: compressor(hidden_states, positions, self.rotary_emb),
|
||||
lambda: self._fused_qnorm_rope_kv_insert(
|
||||
q, kv, positions, attn_metadata
|
||||
),
|
||||
|
||||
Reference in New Issue
Block a user