From fa1ff88b3145d1897558408a9001c030c39383b9 Mon Sep 17 00:00:00 2001 From: tc-mb <157115220+tc-mb@users.noreply.github.com> Date: Fri, 22 May 2026 13:44:06 +0800 Subject: [PATCH] [Model] Fix MiniCPM-V 4.6 vit_merger qkv weight loading (#43213) Signed-off-by: tc-mb --- vllm/model_executor/models/minicpmv4_6.py | 28 +++++++++++++++++++++++ 1 file changed, 28 insertions(+) diff --git a/vllm/model_executor/models/minicpmv4_6.py b/vllm/model_executor/models/minicpmv4_6.py index 48fe6e747b..d2d465b7e5 100644 --- a/vllm/model_executor/models/minicpmv4_6.py +++ b/vllm/model_executor/models/minicpmv4_6.py @@ -23,6 +23,7 @@ from vllm.model_executor.layers.mamba.mamba_utils import ( MambaStateShapeCalculator, ) from vllm.model_executor.layers.quantization import QuantizationConfig +from vllm.model_executor.model_loader.weight_utils import default_weight_loader from vllm.multimodal import MULTIMODAL_REGISTRY from vllm.multimodal.inputs import ( MultiModalFeatureSpec, @@ -590,6 +591,33 @@ class MiniCPMV4_6ViTWindowAttentionSelfAttn(nn.Module): out, _ = self.out_proj(attn_out) return out + def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]: + stacked_params_mapping = [ + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ] + params_dict = dict(self.named_parameters()) + loaded_params: set[str] = set() + for name, loaded_weight in weights: + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + mapped_name = name.replace(weight_name, param_name, 1) + if mapped_name not in params_dict: + continue + param = params_dict[mapped_name] + param.weight_loader(param, loaded_weight, shard_id) + break + else: + if name not in params_dict: + continue + param = params_dict[name] + weight_loader = getattr(param, "weight_loader", default_weight_loader) + weight_loader(param, loaded_weight) + loaded_params.add(name) + return loaded_params + class MiniCPMV4_6ViTWindowAttentionMerger(nn.Module): def __init__(