small bug fixes

comfy/clip_model.py

@@ -15,7 +15,7 @@ class SiglipMultiheadAttentionPoolingHead(torch.nn.Module):
         batch_size = hidden_state.shape[0]
         probe = self.probe.repeat(batch_size, 1, 1)
 
-        hidden_state = self.attention(probe, hidden_state, hidden_state)[0]
+        hidden_state = self.attention(probe, hidden_state, hidden_state)
 
         residual = hidden_state
         hidden_state = self.layernorm(hidden_state)

comfy/ldm/modules/attention.py

@@ -1164,12 +1164,14 @@ class MultiheadAttentionComfyv(nn.Module):
             error_msgs,
         )
 
-    def forward(self, src, attn_mask = None, key_padding_mask = None):
+    def forward(self, src, k = None, v = None, attn_mask = None, key_padding_mask = None):
 
         self._q_proj, self._k_proj, self._v_proj = [t.to(src.device).to(src.dtype) for t in (self._q_proj, self._k_proj, self._v_proj)]
         q = self._q_proj(src)
-        k = self._k_proj(src)
-        v = self._v_proj(src)
+        if k is None:
+            k = self._k_proj(src)
+        if v is None:
+            v = self._v_proj(src)
 
         output = optimized_attention(q, k, v, self.num_heads, mask = attn_mask)
         return self.out_proj(output)

comfy_extras/nodes_video.py

@@ -97,7 +97,7 @@ class EncodeVideo(io.ComfyNode):
                     except:
                         out = model.encode(chunk)
                 else:
-                    out = vae.encode_image(chunk)
+                    out = vae.encode_image(chunk, crop=False)
                     out = out["image_embeds"]
 
                 out_cpu = out.cpu()