fix flash attn

This commit is contained in:
Shenghai Yuan
2026-03-05 11:01:51 +08:00
parent 8ac4293d61
commit c0da835b5a
+4 -1
View File
@@ -214,7 +214,10 @@ def main():
transformer = replace_rmsnorm_with_fp32(transformer)
transformer = replace_all_norms_with_flash_norms(transformer)
replace_rope_with_flash_rope()
transformer.set_attention_backend("_flash_3_hub")
try:
transformer.set_attention_backend("_flash_3_hub")
except Exception:
transformer.set_attention_backend("flash_hub")
vae = AutoencoderKLWan.from_pretrained(
args.base_model_path,