fix qwen2vl vision eager-attention (#33213)

* fix-qwen2vl-vision-eager-attention

* code-quality

* Update src/transformers/models/qwen2_vl/modeling_qwen2_vl.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* code-quality

---------

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

Author: simonJJJ

src/transformers/models/qwen2_vl/modeling_qwen2_vl.py

@@ -275,6 +275,7 @@ class VisionAttention(nn.Module):
     def __init__(self, dim: int, num_heads: int = 16) -> None:
         super().__init__()
         self.num_heads = num_heads
+        self.head_dim = dim // num_heads
         self.qkv = nn.Linear(dim, dim * 3, bias=True)
         self.proj = nn.Linear(dim, dim)
 
@@ -286,9 +287,11 @@ def forward(
         q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
         k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)
 
-        attention_mask = torch.zeros([1, seq_length, seq_length], device=q.device, dtype=torch.bool)
+        attention_mask = torch.full(
+            [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
+        )
         for i in range(1, len(cu_seqlens)):
-            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = 0
 
         q = q.transpose(0, 1)
         k = k.transpose(0, 1)