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 7, 15:15 UTC
[Bugfix] Fix Qwen3NextForCausalLM packed_modules_mapping (#35413)
Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>
Jee Jee Li committed
Feb 27, 2026 at 11:46 UTC
1e5ad9b74f70f4690dff629598d414ee27116b85
1 file changed
+4
vllm/model_executor/models/qwen3_next.py
+4
@@ -412,6 +412,8 @@ class Qwen3NextGatedDeltaNet(nn.Module, MambaBase):
412
prefix=f"{prefix}.in_proj_qkvz",
413
)
414
# ba_proj doesn't support blockwise fp8 quantization.
415
+ # # in_proj_ba is defined as MergedColumnParallelLinear for
416
+ # compatibility with Qwen3_5.
417
self.in_proj_ba = MergedColumnParallelLinear(
418
input_size=self.hidden_size,
419
output_sizes=[self.num_v_heads] * 2,
@@ -1326,6 +1328,8 @@ class Qwen3NextForCausalLM(
1328
"v_proj",
1329
],
1330
"gate_up_proj": ["gate_proj", "up_proj"],
1331
+ "in_proj_qkvz": ["in_proj_qkvz"],
1332
+ "in_proj_ba": ["in_proj_ba"],
1333
}
1334
1335
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):