From 90ce3a09bef2fd7203369b3f7aeabee15ea6f0f8 Mon Sep 17 00:00:00 2001 From: Yan Ma Date: Mon, 6 Jul 2026 17:15:50 +0800 Subject: [PATCH] [bugfix] fix MOSS-Audio deepstack_input_embeds initialization in PP (#47607) Signed-off-by: Yan Ma --- vllm/model_executor/models/moss_audio.py | 4 ++++ 1 file changed, 4 insertions(+) diff --git a/vllm/model_executor/models/moss_audio.py b/vllm/model_executor/models/moss_audio.py index fdfe982fca..bef8057d2f 100644 --- a/vllm/model_executor/models/moss_audio.py +++ b/vllm/model_executor/models/moss_audio.py @@ -862,6 +862,10 @@ class MossQwen3ForCausalLM(Qwen3ForCausalLM): batch_size, dtype, device ) for layer_idx in self.deepstack_inject_layer_indices: + # Non-first PP ranks only receive DeepStack payloads for layers + # at or after their local start layer. + if layer_idx < self.model.start_layer: + continue intermediate_tensors[f"deepstack_input_embeds_{layer_idx}"] = torch.zeros( (batch_size, self.config.hidden_size), dtype=dtype,