updates .gitignore and excludes flash-attn

.gitignore

@@ -4,4 +4,12 @@ __pycache__/
 tmp_examples*
 new_checkpoint*
 batch_test*
-nohup*
+nohup*
+wan.egg-info
+build
+Wan2.2-T2V-A14B
+2.4.0
+1.23.5
+Wan2.2-TI2V-5B
+input
+out

requirements.txt

@@ -12,5 +12,5 @@ easydict
 ftfy
 dashscope
 imageio-ffmpeg
-flash_attn
+#flash_attn
 numpy>=1.23.5,<2

wan/modules/attention.py

@@ -1,5 +1,6 @@
 # Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
 import torch
+import torch.nn.functional as F
 
 try:
     import flash_attn_interface
@@ -109,25 +110,43 @@ def half(x):
             causal=causal,
             deterministic=deterministic)[0].unflatten(0, (b, lq))
     else:
-        assert FLASH_ATTN_2_AVAILABLE
-        x = flash_attn.flash_attn_varlen_func(
-            q=q,
-            k=k,
-            v=v,
-            cu_seqlens_q=torch.cat([q_lens.new_zeros([1]), q_lens]).cumsum(
-                0, dtype=torch.int32).to(q.device, non_blocking=True),
-            cu_seqlens_k=torch.cat([k_lens.new_zeros([1]), k_lens]).cumsum(
-                0, dtype=torch.int32).to(q.device, non_blocking=True),
-            max_seqlen_q=lq,
-            max_seqlen_k=lk,
-            dropout_p=dropout_p,
-            softmax_scale=softmax_scale,
-            causal=causal,
-            window_size=window_size,
-            deterministic=deterministic).unflatten(0, (b, lq))
-
-    # output
-    return x.type(out_dtype)
+        # ---- SDPA Fallback (ohne Flash-Attn v2) ----
+        # q, k, v: (b*lq, h, d), (b*lk, h, d)
+        bld_q, h, d = q.shape
+        bld_k = k.shape[0]
+        lq = bld_q // b
+        lk = bld_k // b
+
+        # -> (b, h, L, d)
+        q_ = q.unflatten(0, (b, lq)).transpose(1, 2)  # (b, h, lq, d)
+        k_ = k.unflatten(0, (b, lk)).transpose(1, 2)  # (b, h, lk, d)
+        v_ = v.unflatten(0, (b, lk)).transpose(1, 2)  # (b, h, lk, d)
+
+        if softmax_scale is not None:
+            q_ = q_ * softmax_scale
+
+        # bool-Maske: True = ignorieren
+        attn_mask = None
+        if (q_lens is not None) or (k_lens is not None):
+            attn_mask = torch.zeros((b, 1, lq, lk), dtype=torch.bool, device=q_.device)
+            if q_lens is not None:
+                for i in range(b):
+                    if q_lens[i] < lq:
+                        attn_mask[i, 0, q_lens[i]:, :] = True
+            if k_lens is not None:
+                for i in range(b):
+                    if k_lens[i] < lk:
+                        attn_mask[i, 0, :, k_lens[i]:] = True
+
+        x_sdpa = F.scaled_dot_product_attention(
+            q_, k_, v_,
+            attn_mask=attn_mask,
+            dropout_p=0.0,
+            is_causal=bool(causal),
+        )  # (b, h, lq, d)
+
+        x = x_sdpa.permute(0, 2, 1, 3).contiguous()
+        return x
 
 
 def attention(