From e7cf27cc6f77f136d1435896da15c34ab8d312b5 Mon Sep 17 00:00:00 2001 From: MOVZX Date: Wed, 14 Oct 2026 08:11:55 +0000 Subject: [PATCH] [Model] Fix runtime abliteration for Qwen3.5 models and Qwen3-Next MTP Qwen3_5Model.__init__ skips Qwen3NextModel.__init__, so the abliteration projector was never installed and embedding lookup raised AttributeError when VLLM_ABLITERATION was set. Install it in Qwen3_5Model. Project the MTP token embeddings of Qwen3NextMultiTokenPredictor too, matching the abliterated releases, which edit the MTP embed_tokens rows. Co-Authored-By: Hinata Signed-off-by: MOVZX --- vllm/model_executor/models/qwen3_5.py | 1 + vllm/model_executor/models/qwen3_next_mtp.py | 4 +++- 2 files changed, 4 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/qwen3_5.py b/vllm/model_executor/models/qwen3_5.py index fb9724db08..5ee6655054 100644 --- a/vllm/model_executor/models/qwen3_5.py +++ b/vllm/model_executor/models/qwen3_5.py @@ -280,6 +280,7 @@ class Qwen3_5Model(Qwen3NextModel): self.norm = PPMissingLayer() self.aux_hidden_state_layers: tuple[int, ...] = () + self.abliterate = projector(config.hidden_size) def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: # FSE must match construction (Qwen3NextSparseMoeBlock): reroute the diff --git a/vllm/model_executor/models/qwen3_next_mtp.py b/vllm/model_executor/models/qwen3_next_mtp.py index 4ae8f21840..72008ecd5a 100644 --- a/vllm/model_executor/models/qwen3_next_mtp.py +++ b/vllm/model_executor/models/qwen3_next_mtp.py @@ -31,6 +31,7 @@ from vllm.model_executor.models.qwen3_next import ( from vllm.model_executor.models.utils import sequence_parallel_chunk from vllm.sequence import IntermediateTensors +from .abliteration import projector from .utils import ( AutoWeightsLoader, make_empty_intermediate_tensors_factory, @@ -111,9 +112,10 @@ class Qwen3NextMultiTokenPredictor(nn.Module): self.pre_fc_norm_embedding = Qwen3NextRMSNorm( config.hidden_size, eps=config.rms_norm_eps ) + self.abliterate = projector(config.hidden_size) def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor: - return self.embed_tokens(input_ids) + return self.abliterate(self.embed_tokens(input_ids)) def forward( self,