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 14, 05:30 UTC
[Bugfix] Fix Qwen2.5-VL quantized model weights loading (#23512)
Signed-off-by: Zifei Tong <zifeitong@gmail.com>
zifeitong committed
Aug 24, 2025 at 19:40 UTC
a71e4765cc0c1534f2a8891aaf628e1751f6df07
1 file changed
+5
-1
vllm/model_executor/models/qwen2_5_vl.py
+5
-1
@@ -135,7 +135,7 @@ class Qwen2_5_VLVideoPixelInputs(TypedDict):
135
136
second_per_grid_ts: torch.Tensor
137
"""
138
- The video time interval (in seconds) for each grid along the temporal
138
+ The video time interval (in seconds) for each grid along the temporal
139
dimension in the 3D position IDs. Returned when `videos` is not `None`.
140
"""
141
@@ -852,6 +852,10 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module, SupportsMultiModal,
852
SupportsLoRA, SupportsPP,
853
SupportsQuant):
854
855
+ packed_modules_mapping = {
856
+ "gate_up_proj": ["gate_proj", "up_proj"],
857
+ }
858
+
859
# To ensure correct weight loading and mapping.
860
hf_to_vllm_mapper = WeightsMapper(
861
orig_to_new_prefix={