Reduce Peak WAN inference VRAM usage

comfy/ldm/flux/math.py

@@ -35,11 +35,10 @@ def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
     out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
     return out.to(dtype=torch.float32, device=pos.device)
 
+def apply_rope1(x: Tensor, freqs_cis: Tensor):
+    x_ = x.to(dtype=freqs_cis.dtype).reshape(*x.shape[:-1], -1, 1, 2)
+    x_out = freqs_cis[..., 0] * x_[..., 0] + freqs_cis[..., 1] * x_[..., 1]
+    return x_out.reshape(*x.shape).type_as(x)
 
 def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor):
-    xq_ = xq.to(dtype=freqs_cis.dtype).reshape(*xq.shape[:-1], -1, 1, 2)
-    xk_ = xk.to(dtype=freqs_cis.dtype).reshape(*xk.shape[:-1], -1, 1, 2)
-    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
-    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
-    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)
-
+    return apply_rope1(xq, freqs_cis), apply_rope1(xk, freqs_cis)

comfy/ldm/wan/model.py

@@ -8,7 +8,7 @@
 
 from comfy.ldm.modules.attention import optimized_attention
 from comfy.ldm.flux.layers import EmbedND
-from comfy.ldm.flux.math import apply_rope
+from comfy.ldm.flux.math import apply_rope1
 import comfy.ldm.common_dit
 import comfy.model_management
 import comfy.patcher_extension
@@ -60,20 +60,24 @@ def forward(self, x, freqs, transformer_options={}):
         """
         b, s, n, d = *x.shape[:2], self.num_heads, self.head_dim
 
-        # query, key, value function
-        def qkv_fn(x):
+        def qkv_fn_q(x):
             q = self.norm_q(self.q(x)).view(b, s, n, d)
+            return apply_rope1(q, freqs)
+
+        def qkv_fn_k(x):
             k = self.norm_k(self.k(x)).view(b, s, n, d)
-            v = self.v(x).view(b, s, n * d)
-            return q, k, v
+            return apply_rope1(k, freqs)
 
-        q, k, v = qkv_fn(x)
-        q, k = apply_rope(q, k, freqs)
+        #These two are VRAM hogs, so we want to do all of q computation and
+        #have pytorch garbage collect the intermediates on the sub function
+        #return before we touch k
+        q = qkv_fn_q(x)
+        k = qkv_fn_k(x)
 
         x = optimized_attention(
             q.view(b, s, n * d),
             k.view(b, s, n * d),
-            v,
+            self.v(x).view(b, s, n * d),
             heads=self.num_heads,
             transformer_options=transformer_options,
         )