attention AMD support

wan/modules/attention.py

@@ -1,14 +1,17 @@
 # Copyright 2024-2025 The Alibaba Wan Team Authors. All rights reserved.
 import torch
+import math
 
 try:
     import flash_attn_interface
+
     FLASH_ATTN_3_AVAILABLE = True
 except ModuleNotFoundError:
     FLASH_ATTN_3_AVAILABLE = False
 
 try:
     import flash_attn
+
     FLASH_ATTN_2_AVAILABLE = True
 except ModuleNotFoundError:
     FLASH_ATTN_2_AVAILABLE = False
@@ -20,6 +23,93 @@ __all__ = [
     'attention',
 ]
 
+DEBUG_ATTENTION = True
+
+
+def log_debug(message):
+    if DEBUG_ATTENTION:
+        print(f"[DEBUG] {message}")
+
+
+def manual_attention(
+        q,
+        k,
+        v,
+        q_lens=None,
+        k_lens=None,
+        dropout_p=0.,
+        softmax_scale=None,
+        q_scale=None,
+        causal=False,
+        window_size=(-1, -1),
+        deterministic=False,
+        dtype=torch.bfloat16,
+):
+    """Attention manuelle optimisée pour tous les devices"""
+    # Déplacement immédiat sur le bon device
+    device = q.device
+    k = k.to(device)
+    v = v.to(device)
+    if q_lens is not None: q_lens = q_lens.to(device)
+    if k_lens is not None: k_lens = k_lens.to(device)
+
+    B, Lq, N, C = q.shape
+    _, Lk, _, _ = k.shape
+    original_dtype = q.dtype
+
+    # Conversion au dtype de calcul
+    q = q.to(dtype).transpose(1, 2)
+    k = k.to(dtype).transpose(1, 2)
+    v = v.to(dtype).transpose(1, 2)
+
+    # Scaling
+    scale_factor = softmax_scale or (1.0 / math.sqrt(C))
+    if q_scale is not None:
+        q = q * q_scale.view(1, -1, 1, 1)
+
+    # Calcul des scores d'attention
+    attn_scores = torch.matmul(q, k.transpose(-2, -1)) * scale_factor
+
+    # Création des masques
+    attn_mask = torch.zeros(B, 1, Lq, Lk, device=device, dtype=torch.float32)
+
+    # Masque de padding des clés
+    if k_lens is not None:
+        key_mask = torch.arange(Lk, device=device)[None, :] < k_lens[:, None]
+        attn_mask = attn_mask.masked_fill(~key_mask.view(B, 1, 1, Lk), float('-inf'))
+
+    # Masque causal
+    if causal:
+        causal_mask = torch.ones(Lq, Lk, device=device, dtype=torch.bool).tril()
+        attn_mask = attn_mask.masked_fill(~causal_mask, float('-inf'))
+
+    # Masque de fenêtre
+    if window_size != (-1, -1):
+        left, right = window_size
+        row = torch.arange(Lq, device=device)[:, None]
+        col = torch.arange(Lk, device=device)[None, :]
+        window_mask = (row - col >= -left) & (row - col <= right)
+        attn_mask = attn_mask.masked_fill(~window_mask, float('-inf'))
+
+    # Application du masque
+    attn_scores += attn_mask
+
+    # Softmax et dropout
+    attn_weights = torch.softmax(attn_scores, dim=-1)
+    if not deterministic and dropout_p > 0:
+        attn_weights = torch.dropout(attn_weights, dropout_p, True)
+
+    # Calcul de la sortie
+    out = torch.matmul(attn_weights, v)
+
+    # Masque de padding des requêtes
+    if q_lens is not None:
+        query_mask = torch.arange(Lq, device=device)[None, :] < q_lens[:, None]
+        out = out * query_mask.view(B, 1, Lq, 1).to(out.dtype)
+
+    # Retour au format original
+    return out.transpose(1, 2).contiguous().to(original_dtype)
+
 
 def flash_attention(
         q,
@@ -36,98 +126,78 @@ def flash_attention(
         dtype=torch.bfloat16,
         version=None,
 ):
-    """
+    """Wrapper pour FlashAttention avec fallback manuel"""
-    q:              [B, Lq, Nq, C1].
+    # Fallback si FlashAttention non disponible
-    k:              [B, Lk, Nk, C1].
+    if not (FLASH_ATTN_2_AVAILABLE or FLASH_ATTN_3_AVAILABLE):
-    v:              [B, Lk, Nk, C2]. Nq must be divisible by Nk.
+        return manual_attention(
-    q_lens:         [B].
+            q, k, v, q_lens, k_lens, dropout_p, softmax_scale,
-    k_lens:         [B].
+            q_scale, causal, window_size, deterministic, dtype
-    dropout_p:      float. Dropout probability.
-    softmax_scale:  float. The scaling of QK^T before applying softmax.
-    causal:         bool. Whether to apply causal attention mask.
-    window_size:    (left right). If not (-1, -1), apply sliding window local attention.
-    deterministic:  bool. If True, slightly slower and uses more memory.
-    dtype:          torch.dtype. Apply when dtype of q/k/v is not float16/bfloat16.
-    """
-    half_dtypes = (torch.float16, torch.bfloat16)
-    assert dtype in half_dtypes
-    assert q.device.type == 'cuda' and q.size(-1) <= 256
-
-    # params
-    b, lq, lk, out_dtype = q.size(0), q.size(1), k.size(1), q.dtype
-
-    def half(x):
-        return x if x.dtype in half_dtypes else x.to(dtype)
-
-    # preprocess query
-    if q_lens is None:
-        q = half(q.flatten(0, 1))
-        q_lens = torch.tensor(
-            [lq] * b, dtype=torch.int32).to(
-                device=q.device, non_blocking=True)
-    else:
-        q = half(torch.cat([u[:v] for u, v in zip(q, q_lens)]))
-
-    # preprocess key, value
-    if k_lens is None:
-        k = half(k.flatten(0, 1))
-        v = half(v.flatten(0, 1))
-        k_lens = torch.tensor(
-            [lk] * b, dtype=torch.int32).to(
-                device=k.device, non_blocking=True)
-    else:
-        k = half(torch.cat([u[:v] for u, v in zip(k, k_lens)]))
-        v = half(torch.cat([u[:v] for u, v in zip(v, k_lens)]))
-
-    q = q.to(v.dtype)
-    k = k.to(v.dtype)
-
-    if q_scale is not None:
-        q = q * q_scale
-
-    if version is not None and version == 3 and not FLASH_ATTN_3_AVAILABLE:
-        warnings.warn(
-            'Flash attention 3 is not available, use flash attention 2 instead.'
         )
 
-    # apply attention
+    # Paramètres GPU
-    if (version is None or version == 3) and FLASH_ATTN_3_AVAILABLE:
+    device = q.device
-        # Note: dropout_p, window_size are not supported in FA3 now.
+    b, lq, lk = q.size(0), q.size(1), k.size(1)
+    out_dtype = q.dtype
+
+    # Préparation des séquences
+    if q_lens is None:
+        q_lens = torch.full((b,), lq, dtype=torch.int32, device=device)
+        q_flat = q.flatten(0, 1)
+    else:
+        q_lens = q_lens.to(device)
+        q_flat = torch.cat([u[:l] for u, l in zip(q, q_lens)])
+
+    if k_lens is None:
+        k_lens = torch.full((b,), lk, dtype=torch.int32, device=device)
+        k_flat = k.flatten(0, 1)
+        v_flat = v.flatten(0, 1)
+    else:
+        k_lens = k_lens.to(device)
+        k_flat = torch.cat([u[:l] for u, l in zip(k, k_lens)])
+        v_flat = torch.cat([u[:l] for u, l in zip(v, k_lens)])
+
+    # Conversion de type
+    q_flat = q_flat.to(dtype)
+    k_flat = k_flat.to(dtype)
+    v_flat = v_flat.to(dtype)
+
+    # Application de q_scale
+    if q_scale is not None:
+        q_flat = q_flat * q_scale
+
+    # Préparation des séquences cumulatives
+    cu_seqlens_q = torch.cat([torch.tensor([0], device=device), q_lens.cumsum(0)])
+    cu_seqlens_k = torch.cat([torch.tensor([0], device=device), k_lens.cumsum(0)])
+
+    # Appel à FlashAttention
+    try:
+        if FLASH_ATTN_3_AVAILABLE and (version is None or version == 3):
             x = flash_attn_interface.flash_attn_varlen_func(
-            q=q,
+                q_flat, k_flat, v_flat,
-            k=k,
+                cu_seqlens_q, cu_seqlens_k,
-            v=v,
+                max_seqlen_q=lq, max_seqlen_k=lk,
-            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),
-            seqused_q=None,
-            seqused_k=None,
-            max_seqlen_q=lq,
-            max_seqlen_k=lk,
                 softmax_scale=softmax_scale,
                 causal=causal,
-            deterministic=deterministic)[0].unflatten(0, (b, lq))
+                deterministic=deterministic
+            )[0]
         else:
-        assert FLASH_ATTN_2_AVAILABLE
             x = flash_attn.flash_attn_varlen_func(
-            q=q,
+                q_flat, k_flat, v_flat,
-            k=k,
+                cu_seqlens_q, cu_seqlens_k,
-            v=v,
+                max_seqlen_q=lq, max_seqlen_k=lk,
-            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))
+                deterministic=deterministic
-
+            )
-    # output
+        return x.unflatten(0, (b, lq)).to(out_dtype)
-    return x.type(out_dtype)
+    except Exception as e:
+        warnings.warn(f"FlashAttention failed: {e}, using manual attention")
+        return manual_attention(
+            q, k, v, q_lens, k_lens, dropout_p, softmax_scale,
+            q_scale, causal, window_size, deterministic, dtype
+        )
 
 
 def attention(
@@ -145,35 +215,22 @@ def attention(
         dtype=torch.bfloat16,
         fa_version=None,
 ):
+    """Fonction d'attention unifiée"""
+    # Synchronisation des devices
+    device = q.device
+    k = k.to(device)
+    v = v.to(device)
+    if q_lens is not None: q_lens = q_lens.to(device)
+    if k_lens is not None: k_lens = k_lens.to(device)
+
+    # Sélection de l'implémentation
     if FLASH_ATTN_2_AVAILABLE or FLASH_ATTN_3_AVAILABLE:
         return flash_attention(
-            q=q,
+            q, k, v, q_lens, k_lens, dropout_p, softmax_scale,
-            k=k,
+            q_scale, causal, window_size, deterministic, dtype, fa_version
-            v=v,
-            q_lens=q_lens,
-            k_lens=k_lens,
-            dropout_p=dropout_p,
-            softmax_scale=softmax_scale,
-            q_scale=q_scale,
-            causal=causal,
-            window_size=window_size,
-            deterministic=deterministic,
-            dtype=dtype,
-            version=fa_version,
         )
     else:
-        if q_lens is not None or k_lens is not None:
+        return manual_attention(
-            warnings.warn(
+            q, k, v, q_lens, k_lens, dropout_p, softmax_scale,
-                'Padding mask is disabled when using scaled_dot_product_attention. It can have a significant impact on performance.'
+            q_scale, causal, window_size, deterministic, dtype
         )
-        attn_mask = None
-
-        q = q.transpose(1, 2).to(dtype)
-        k = k.transpose(1, 2).to(dtype)
-        v = v.transpose(1, 2).to(dtype)
-
-        out = torch.nn.functional.scaled_dot_product_attention(
-            q, k, v, attn_mask=attn_mask, is_causal=causal, dropout_p=dropout_p)
-
-        out = out.transpose(1, 2).contiguous()
-        return out