[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 <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -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
|
||||
|
||||
@@ -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,
|
||||
|
||||
Reference in New Issue
Block a user