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, 15:15 UTC
[Bugfix] Adopt `ChunkGatedDeltaRule` for Qwen3.5 (#34198)
Signed-off-by: Roger Wang <hey@rogerw.io>
Roger Wang committed
Feb 9, 2026 at 19:47 UTC
047a457fa4af2010303ba775ae6f3ee9c1852c2c
1 file changed
+3
vllm/model_executor/models/qwen3_5.py
+3
@@ -99,6 +99,7 @@ from .interfaces import (
99
)
100
from .qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP
101
from .qwen3_next import (
102
+ ChunkGatedDeltaRule,
103
Qwen3NextAttention,
104
Qwen3NextDecoderLayer,
105
Qwen3NextGatedDeltaNet,
@@ -268,6 +269,8 @@ class Qwen3_5GatedDeltaNet(Qwen3NextGatedDeltaNet):
269
prefix=f"{prefix}.out_proj",
270
)
271
272
+ self.chunk_gated_delta_rule = ChunkGatedDeltaRule()
273
+
274
compilation_config = get_current_vllm_config().compilation_config
275
if prefix in compilation_config.static_forward_context:
276
raise ValueError(f"Duplicate layer name: {prefix}")