[Bugfix] Fix Minicpm-O-int4 GPTQ model inference

vllm/model_executor/models/minicpmo.py

@@ -28,12 +28,16 @@
 
 import torch
 from torch import nn
-from transformers import BatchFeature
+from transformers import BatchFeature, PretrainedConfig
 from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.models.whisper.modeling_whisper import (
     ACT2FN, WHISPER_ATTENTION_CLASSES, WhisperConfig, WhisperEncoder)
 
 from vllm.config import VllmConfig
+from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.quantization.gptq import GPTQConfig
+from vllm.model_executor.layers.quantization.gptq_marlin import (
+    GPTQMarlinConfig)
 from vllm.multimodal import MULTIMODAL_REGISTRY, MultiModalKwargs
 from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
                                     NestedTensors)
@@ -512,6 +516,36 @@ def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
 
         self.audio_token_id = None
 
+    def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):
+        # GPTQ configs do not have a list of ignored modules, however AutoGPTQ
+        # seems to avoid vision encoder sections for some models.
+        # See: https://huggingface.co/openbmb/MiniCPM-o-2_6-int4
+        if isinstance(quant_config, (GPTQConfig, GPTQMarlinConfig)):
+            return None
+        return quant_config
+
+    def init_vision_module(
+        self,
+        config: PretrainedConfig,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> nn.Module:
+        # MiniCPMO GPTQ model leave vpm unquantized.
+        quant_config = self._maybe_ignore_quant_config(quant_config)
+        return super().init_vision_module(config, quant_config, prefix)
+
+    def init_resampler(
+        self,
+        embed_dim: int,
+        vision_dim: int,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+    ) -> nn.Module:
+        # MiniCPMO GPTQ model leave resampler unquantized.
+        quant_config = self._maybe_ignore_quant_config(quant_config)
+        return super().init_resampler(embed_dim, vision_dim, quant_config,
+                                      prefix)
+
     def init_audio_module(self, *, vllm_config: VllmConfig, prefix: str = ""):
         # Do not use parameters temporarily
         audio_config = self.config.audio_config

vllm/model_executor/models/minicpmv.py

@@ -1181,7 +1181,7 @@ def init_llm(
     def init_vision_module(
         self,
         config: PretrainedConfig,
-        quant_config: Optional[QuantizationConfig],
+        quant_config: Optional[QuantizationConfig] = None,
         prefix: str = "",
     ) -> nn.Module:
         model = Idefics2VisionTransformer(config.vision_config,