Spaces:
Running on Zero
Running on Zero
Update app.py
Browse files
app.py
CHANGED
|
@@ -71,23 +71,6 @@ from ltx_pipelines.utils.media_io import decode_audio_from_file, encode_video
|
|
| 71 |
from ltx_core.loader.primitives import LoraPathStrengthAndSDOps
|
| 72 |
from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
|
| 73 |
|
| 74 |
-
from ltx_core.model.transformer import attention as _attn_mod
|
| 75 |
-
|
| 76 |
-
print(f"[ATTN] Before patch: memory_efficient_attention={_attn_mod.memory_efficient_attention}")
|
| 77 |
-
try:
|
| 78 |
-
from xformers.ops import memory_efficient_attention as _mea
|
| 79 |
-
from xformers.ops.fmha import cutlass
|
| 80 |
-
|
| 81 |
-
def _cutlass_memory_efficient_attention(*args, **kwargs):
|
| 82 |
-
# Force CUTLASS and avoid FlashAttention paths that are crashing.
|
| 83 |
-
kwargs["op"] = (cutlass.FwOp, cutlass.BwOp)
|
| 84 |
-
return _mea(*args, **kwargs)
|
| 85 |
-
|
| 86 |
-
_attn_mod.memory_efficient_attention = _cutlass_memory_efficient_attention
|
| 87 |
-
print(f"[ATTN] After patch: memory_efficient_attention={_attn_mod.memory_efficient_attention}")
|
| 88 |
-
except Exception as e:
|
| 89 |
-
print(f"[ATTN] xformers patch FAILED: {type(e).__name__}: {e}")
|
| 90 |
-
|
| 91 |
logging.getLogger().setLevel(logging.INFO)
|
| 92 |
|
| 93 |
MAX_SEED = np.iinfo(np.int32).max
|
|
|
|
| 71 |
from ltx_core.loader.primitives import LoraPathStrengthAndSDOps
|
| 72 |
from ltx_core.loader.sd_ops import LTXV_LORA_COMFY_RENAMING_MAP
|
| 73 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
logging.getLogger().setLevel(logging.INFO)
|
| 75 |
|
| 76 |
MAX_SEED = np.iinfo(np.int32).max
|