Upload 11 files

hyvideo/modules/attenion.py

@@ -178,7 +178,7 @@ def parallel_attention(
         joint_tensor_value=v[:,img_kv_len:cu_seqlens_kv[1]],
         joint_strategy="rear",
     )
-    if flash_attn.__version__ >= "2.7.0":
+    if flash_attn.__version__ >= '2.7.0':
         attn2, *_ = _flash_attn_forward(
             q[:,cu_seqlens_q[1]:],
             k[:,cu_seqlens_kv[1]:],

hyvideo/modules/fp8_optimization.py

@@ -83,7 +83,7 @@ def convert_fp8_linear(module, dit_weight_path, original_dtype, params_to_keep={
     setattr(module, "fp8_matmul_enabled", True)
 
     # loading fp8 mapping file
-    fp8_map_path = dit_weight_path.replace(".pt", "_map.pt")
+    fp8_map_path = dit_weight_path.replace('.pt', '_map.pt')
     if os.path.exists(fp8_map_path):
         fp8_map = torch.load(fp8_map_path, map_location=lambda storage, loc: storage)
     else:
@@ -91,7 +91,7 @@ def convert_fp8_linear(module, dit_weight_path, original_dtype, params_to_keep={
 
     fp8_layers = []
     for key, layer in module.named_modules():
-        if isinstance(layer, nn.Linear) and ("double_blocks" in key or "single_blocks" in key):
+        if isinstance(layer, nn.Linear) and ('double_blocks' in key or 'single_blocks' in key):
             fp8_layers.append(key)
             original_forward = layer.forward
             layer.weight = torch.nn.Parameter(layer.weight.to(torch.float8_e4m3fn))