Lower cosmos diffusion model memory usage.

comfy/ldm/cosmos/blocks.py

@@ -168,14 +168,18 @@ class Attention(nn.Module):
         k = self.to_k[1](k)
         v = self.to_v[1](v)
         if self.is_selfattn and rope_emb is not None:  # only apply to self-attention!
-            q = apply_rotary_pos_emb(q, rope_emb)
-            k = apply_rotary_pos_emb(k, rope_emb)
-        return q, k, v
+            # apply_rotary_pos_emb inlined
+            q_shape = q.shape
+            q = q.reshape(*q.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2)
+            q = rope_emb[..., 0] * q[..., 0] + rope_emb[..., 1] * q[..., 1]
+            q = q.movedim(-1, -2).reshape(*q_shape).to(x.dtype)
 
-    def cal_attn(self, q, k, v, mask=None):
-        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True)
-        out = rearrange(out, " b n s c -> s b (n c)")
-        return self.to_out(out)
+            # apply_rotary_pos_emb inlined
+            k_shape = k.shape
+            k = k.reshape(*k.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2)
+            k = rope_emb[..., 0] * k[..., 0] + rope_emb[..., 1] * k[..., 1]
+            k = k.movedim(-1, -2).reshape(*k_shape).to(x.dtype)
+        return q, k, v
 
     def forward(
         self,
@@ -191,7 +195,10 @@ class Attention(nn.Module):
             context (Optional[Tensor]): The key tensor of shape [B, Mk, K] or use x as context [self attention] if None
         """
         q, k, v = self.cal_qkv(x, context, mask, rope_emb=rope_emb, **kwargs)
-        return self.cal_attn(q, k, v, mask)
+        out = optimized_attention(q, k, v, self.heads, skip_reshape=True, mask=mask, skip_output_reshape=True)
+        del q, k, v
+        out = rearrange(out, " b n s c -> s b (n c)")
+        return self.to_out(out)
 
 
 class FeedForward(nn.Module):
@@ -788,10 +795,7 @@ class GeneralDITTransformerBlock(nn.Module):
         crossattn_mask: Optional[torch.Tensor] = None,
         rope_emb_L_1_1_D: Optional[torch.Tensor] = None,
         adaln_lora_B_3D: Optional[torch.Tensor] = None,
-        extra_per_block_pos_emb: Optional[torch.Tensor] = None,
     ) -> torch.Tensor:
-        if extra_per_block_pos_emb is not None:
-            x = x + extra_per_block_pos_emb
         for block in self.blocks:
             x = block(
                 x,

comfy/ldm/cosmos/model.py

@@ -168,7 +168,7 @@ class GeneralDIT(nn.Module):
             operations=operations,
         )
 
-        self.build_pos_embed(device=device)
+        self.build_pos_embed(device=device, dtype=dtype)
         self.block_x_format = block_x_format
         self.use_adaln_lora = use_adaln_lora
         self.adaln_lora_dim = adaln_lora_dim
@@ -210,7 +210,7 @@ class GeneralDIT(nn.Module):
             operations=operations,
         )
 
-    def build_pos_embed(self, device=None):
+    def build_pos_embed(self, device=None, dtype=None):
         if self.pos_emb_cls == "rope3d":
             cls_type = VideoRopePosition3DEmb
         else:
@@ -242,6 +242,7 @@ class GeneralDIT(nn.Module):
             kwargs["w_extrapolation_ratio"] = self.extra_w_extrapolation_ratio
             kwargs["t_extrapolation_ratio"] = self.extra_t_extrapolation_ratio
             kwargs["device"] = device
+            kwargs["dtype"] = dtype
             self.extra_pos_embedder = LearnablePosEmbAxis(
                 **kwargs,
             )
@@ -476,6 +477,8 @@ class GeneralDIT(nn.Module):
             inputs["original_shape"],
         )
         extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D = inputs["extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D"].to(x.dtype)
+        del inputs
+
         if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
             assert (
                 x.shape == extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D.shape
@@ -486,6 +489,8 @@ class GeneralDIT(nn.Module):
                 self.blocks["block0"].x_format == block.x_format
             ), f"First block has x_format {self.blocks[0].x_format}, got {block.x_format}"
 
+            if extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D is not None:
+                x += extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D
             x = block(
                 x,
                 affline_emb_B_D,
@@ -493,7 +498,6 @@ class GeneralDIT(nn.Module):
                 crossattn_mask,
                 rope_emb_L_1_1_D=rope_emb_L_1_1_D,
                 adaln_lora_B_3D=adaln_lora_B_3D,
-                extra_per_block_pos_emb=extra_pos_emb_B_T_H_W_D_or_T_H_W_B_D,
             )
 
         x_B_T_H_W_D = rearrange(x, "T H W B D -> B T H W D")

comfy/ldm/cosmos/position_embedding.py

@@ -173,6 +173,7 @@ class LearnablePosEmbAxis(VideoPositionEmb):
         len_w: int,
         len_t: int,
         device=None,
+        dtype=None,
         **kwargs,
     ):
         """
@@ -184,9 +185,9 @@ class LearnablePosEmbAxis(VideoPositionEmb):
         self.interpolation = interpolation
         assert self.interpolation in ["crop"], f"Unknown interpolation method {self.interpolation}"
 
-        self.pos_emb_h = nn.Parameter(torch.empty(len_h, model_channels, device=device))
-        self.pos_emb_w = nn.Parameter(torch.empty(len_w, model_channels, device=device))
-        self.pos_emb_t = nn.Parameter(torch.empty(len_t, model_channels, device=device))
+        self.pos_emb_h = nn.Parameter(torch.empty(len_h, model_channels, device=device, dtype=dtype))
+        self.pos_emb_w = nn.Parameter(torch.empty(len_w, model_channels, device=device, dtype=dtype))
+        self.pos_emb_t = nn.Parameter(torch.empty(len_t, model_channels, device=device, dtype=dtype))
 
 
     def generate_embeddings(self, B_T_H_W_C: torch.Size, fps=Optional[torch.Tensor], device=None) -> torch.Tensor: