fix minicpmo4.5: fix attn_mask in vit attn && fix resampler pos_emb i… (#34127)

Signed-off-by: tc-mb <caitianchi@modelbest.cn>
Co-authored-by: hezhihui <hezhihui@modelbest.cn>

vllm/model_executor/models/idefics2_vision_model.py

@@ -22,6 +22,7 @@ from collections.abc import Iterable
 
 import torch
 from torch import nn
+from torch.nn import functional as F
 from transformers.models.idefics2.configuration_idefics2 import (
     Idefics2Config,
     Idefics2VisionConfig,
@@ -172,14 +173,41 @@ class Idefics2VisionAttention(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
     ) -> torch.Tensor:
         qkv, _ = self.qkv_proj(
             hidden_states
         )  # batch_size, q_len, 3 * num_heads_per_partition * head_dim
         query_states, key_states, value_states = qkv.chunk(3, dim=-1)
 
-        # Use unified MMEncoderAttention implementation
+        # If attention_mask is provided, prefer Torch SDPA so the mask is
-        out = self.attn(query_states, key_states, value_states)
+        # correctly applied (aligns with HuggingFace NaViT SigLIP behavior).
+        if attention_mask is None:
+            # Use unified MMEncoderAttention implementation
+            out = self.attn(query_states, key_states, value_states)
+        else:
+            bsz, q_len = query_states.size()[:2]
+            kv_len = key_states.size(1)
+
+            query = query_states.view(
+                bsz, q_len, self.num_heads_per_partition, self.head_dim
+            ).transpose(1, 2)
+            key = key_states.view(
+                bsz, kv_len, self.num_heads_per_partition, self.head_dim
+            ).transpose(1, 2)
+            value = value_states.view(
+                bsz, kv_len, self.num_heads_per_partition, self.head_dim
+            ).transpose(1, 2)
+
+            out = F.scaled_dot_product_attention(
+                query,
+                key,
+                value,
+                attn_mask=attention_mask,
+                dropout_p=0.0,
+                scale=self.scale,
+            )
+            out = out.transpose(1, 2).reshape(bsz, q_len, -1)
         attn_output, _ = self.out_proj(out)
         return attn_output
 
@@ -245,6 +273,7 @@ class Idefics2EncoderLayer(nn.Module):
     def forward(
         self,
         hidden_states: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
     ) -> torch.Tensor:
         """
         Args:
@@ -254,7 +283,7 @@ class Idefics2EncoderLayer(nn.Module):
         """
         residual = hidden_states
         hidden_states = self.layer_norm1(hidden_states)
-        hidden_states = self.self_attn(hidden_states)
+        hidden_states = self.self_attn(hidden_states, attention_mask=attention_mask)
         hidden_states += residual
         residual = hidden_states
         hidden_states = self.layer_norm2(hidden_states)
@@ -304,6 +333,7 @@ class Idefics2Encoder(nn.Module):
     def forward(
         self,
         inputs_embeds: torch.Tensor,
+        attention_mask: torch.Tensor | None = None,
     ) -> torch.Tensor:
         r"""
         Args:
@@ -316,7 +346,7 @@ class Idefics2Encoder(nn.Module):
         """
         hidden_states = inputs_embeds
         for encoder_layer in self.layers:
-            layer_outputs = encoder_layer(hidden_states)
+            layer_outputs = encoder_layer(hidden_states, attention_mask=attention_mask)
             hidden_states = layer_outputs
         return hidden_states
 
@@ -370,15 +400,47 @@ class Idefics2VisionTransformer(nn.Module):
         patch_attention_mask: torch.BoolTensor | None = None,
         tgt_sizes: torch.IntTensor | None = None,
     ) -> torch.Tensor:
+        batch_size = pixel_values.size(0)
+
+        if patch_attention_mask is None:
+            # No mask provided - create default all-ones mask for embeddings
+            # and skip attention masking (no padding to mask)
+            patch_attention_mask = torch.ones(
+                size=(
+                    batch_size,
+                    pixel_values.size(2) // self.config.patch_size,
+                    pixel_values.size(3) // self.config.patch_size,
+                ),
+                dtype=torch.bool,
+                device=pixel_values.device,
+            )
+            flat_patch_mask = None
+        else:
+            flat_patch_mask = patch_attention_mask.view(batch_size, -1)
+
         hidden_states = self.embeddings(
             pixel_values=pixel_values,
             patch_attention_mask=patch_attention_mask,
             tgt_sizes=tgt_sizes,
         )
+
+        # Align with HuggingFace NaViT SigLIP in MiniCPMV/O:
+        # - if patch_attention_mask was None, skip attention masking
+        # - if any padding exists, create an additive 4D mask and pass it
+        #   to attention; else skip mask for performance.
+        if flat_patch_mask is None or not torch.any(~flat_patch_mask):
+            attention_mask = None
+        else:
+            # Additive mask: masked positions receive a large negative value.
+            # Shape: (B, 1, 1, L) broadcastable to (B, H, Q, K).
+            min_val = torch.finfo(hidden_states.dtype).min
+            attention_mask = (~flat_patch_mask).to(dtype=hidden_states.dtype) * min_val
+            attention_mask = attention_mask[:, None, None, :]
+
         if self.use_data_parallel:
             encoder_outputs = run_dp_sharded_vision_model(hidden_states, self.encoder)
         else:
-            encoder_outputs = self.encoder(hidden_states)
+            encoder_outputs = self.encoder(hidden_states, attention_mask=attention_mask)
         last_hidden_state = self.post_layernorm(encoder_outputs)
         return last_hidden_state
 

vllm/model_executor/models/minicpmv.py

@@ -387,8 +387,8 @@ class Resampler4_5(Resampler2_5):
             pos_embed_2d, batch_first=True, padding_value=0.0
         ).permute(1, 0, 2)  # BLD => L * B * D
 
-        k = x
+        k = x + pos_embed_2d
-        v = x + pos_embed_2d
+        v = x
         if pos_embed_temporal:
             k += torch.stack(pos_embed_temporal, dim=0)
             bs = len(temporal_ids)