A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 01:15 UTC
[Bugfix] Qwen3.5 kv-scale weight remapping (#34719)
Signed-off-by: Linda-Stadter <57756729+Linda-Stadter@users.noreply.github.com>
Linda committed
Feb 19, 2026 at 13:13 UTC
6fff24f30fe2554f43871978ef59feaa87f245c0
1 file changed
+7
vllm/model_executor/models/qwen3_5.py
+7
@@ -57,6 +57,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
57
)
58
from vllm.model_executor.model_loader.weight_utils import (
59
default_weight_loader,
60
+ maybe_remap_kv_scale_name,
61
)
62
from vllm.multimodal import MULTIMODAL_REGISTRY
63
from vllm.sequence import IntermediateTensors
@@ -397,6 +398,12 @@ class Qwen3_5Model(Qwen3NextModel):
398
if name.startswith("mtp."):
399
continue
400
401
+ # Remapping the name of FP8 kv-scale.
402
+ if name.endswith("scale"):
403
+ name = maybe_remap_kv_scale_name(name, params_dict)
404
+ if name is None:
405
+ continue
406
+
407
for param_name, weight_name, shard_id in stacked_params_mapping:
408
if "experts.gate_up_proj" in name or "experts.down_proj" in name:
409
is_fused_expert = True