[bugfix] fix MOSS-Audio deepstack_input_embeds initialization in PP (#47607)
Signed-off-by: Yan Ma <yan.ma@intel.com>
Yan Ma committed
Jul 6, 2026 at 17:15 UTC
90ce3a09bef2fd7203369b3f7aeabee15ea6f0f8
1 file changed
+4
vllm/model_executor/models/moss_audio.py
+4
@@ -862,6 +862,10 @@ class MossQwen3ForCausalLM(Qwen3ForCausalLM):
862
batch_size, dtype, device
863
)
864
for layer_idx in self.deepstack_inject_layer_indices:
865
+ # Non-first PP ranks only receive DeepStack payloads for layers
866
+ # at or after their local start layer.
867
+ if layer_idx < self.model.start_layer:
868
+ continue
869
intermediate_tensors[f"deepstack_input_embeds_{layer_idx}"] = torch.zeros(
870
(batch_size, self.config.hidden_size),
871
dtype=dtype,