chore(openapi): sync shared API contract from cloud@773d43b

comfy/ldm/cosmos/predict2.py

@@ -15,6 +15,15 @@ import comfy.patcher_extension
 from comfy.ldm.modules.attention import optimized_attention
 import comfy.ldm.common_dit
 
+def apply_rotary_pos_emb(
+    t: torch.Tensor,
+    freqs: torch.Tensor,
+) -> torch.Tensor:
+    t_ = t.reshape(*t.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).float()
+    t_out = freqs[..., 0] * t_[..., 0] + freqs[..., 1] * t_[..., 1]
+    t_out = t_out.movedim(-1, -2).reshape(*t.shape).type_as(t)
+    return t_out
+
 
 # ---------------------- Feed Forward Network -----------------------
 class GPT2FeedForward(nn.Module):
@@ -164,7 +173,8 @@ class Attention(nn.Module):
             k = self.k_norm(k)
             v = self.v_norm(v)
             if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
-                q, k = comfy.quant_ops.ck.apply_rope_split_half(q, k, rope_emb)
+                q = apply_rotary_pos_emb(q, rope_emb)
+                k = apply_rotary_pos_emb(k, rope_emb)
             return q, k, v
 
         q, k, v = apply_norm_and_rotary_pos_emb(q, k, v, rope_emb)

comfy/ldm/qwen_image/model.py

@@ -51,6 +51,15 @@ class FeedForward(nn.Module):
         return hidden_states
 
 
+def apply_rotary_emb(x, freqs_cis):
+    if x.shape[1] == 0:
+        return x
+
+    t_ = x.reshape(*x.shape[:-1], -1, 1, 2)
+    t_out = freqs_cis[..., 0] * t_[..., 0] + freqs_cis[..., 1] * t_[..., 1]
+    return t_out.reshape(*x.shape)
+
+
 class QwenTimestepProjEmbeddings(nn.Module):
     def __init__(self, embedding_dim, pooled_projection_dim, use_additional_t_cond=False, dtype=None, device=None, operations=None):
         super().__init__()

requirements.txt

@@ -22,7 +22,7 @@ alembic
 SQLAlchemy>=2.0.0
 filelock
 av>=16.0.0
-comfy-kitchen==0.2.10
+comfy-kitchen==0.2.9
 comfy-aimdo==0.4.5
 requests
 simpleeval>=1.0.0