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 6, 21:45 UTC
fix(minicpmv): fix audio inference by handling meta device in init_re… (#36751)
Signed-off-by: caitianchi <caitianchi@modelbest.cn>
tc-mb committed
Mar 11, 2026 at 18:06 UTC
4286cc5ec24cf7a6d7c1a47e89dba914881be89a
1 file changed
+10
-8
vllm/model_executor/models/minicpmv.py
+10
-8
@@ -1453,10 +1453,11 @@ class MiniCPMV2_6(MiniCPMVBaseModel, SupportsLoRA):
1453
quant_config=quant_config,
1454
prefix=prefix,
1455
)
1456
-
1457
- return resampler.to(
1458
- device=current_platform.device_type, dtype=torch.get_default_dtype()
1459
- )
1456
+ target_device = current_platform.device_type
1457
+ target_dtype = torch.get_default_dtype()
1458
+ if any(p.is_meta for p in resampler.parameters()):
1459
+ return resampler.to_empty(device=target_device).to(dtype=target_dtype)
1460
+ return resampler.to(device=target_device, dtype=target_dtype)
1461
1462
def get_vision_hidden_states(self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
1463
pixel_values = data["pixel_values"]
@@ -1649,10 +1650,11 @@ class MiniCPMV4_5(MiniCPMVBaseModel, SupportsLoRA):
1650
quant_config=quant_config,
1651
prefix=prefix,
1652
)
1652
-
1653
- return resampler.to(
1654
- device=current_platform.device_type, dtype=torch.get_default_dtype()
1655
- )
1653
+ target_device = current_platform.device_type
1654
+ target_dtype = torch.get_default_dtype()
1655
+ if any(p.is_meta for p in resampler.parameters()):
1656
+ return resampler.to_empty(device=target_device).to(dtype=target_dtype)
1657
+ return resampler.to(device=target_device, dtype=target_dtype)
1658
1659
def get_vision_hidden_states(self, data: MiniCPMVImagePixelInputs) -> torch.Tensor:
1660
pixel_values = data["pixel_values"]