Make videos on the computer you already own. FreeVideo runs MiniMax H3 in as little as 8 GB of VRAM and 16 GB of RAM, and adapts its acceleration path to your hardware.
"""Isolated, hash-checked backport for FA4 b26's SM120 invalid-tile read."""
import hashlib
import importlib
import importlib.metadata
import json
import os
from pathlib import Path
import shutil
import sys
import tempfile
from .paths import data_root
ORIGINAL = 'e9d890e10611ce48dc57c40570a7b65d1ab802c508f785e5536fb10044db65f3'
PATCHED = '0b6c45fbacf31dfbc6bca600e4efa21892f95a4ad3e21a163c04a74ffa076a12'
def location():
return data_root() / 'vendor' / 'fa4-b26-valid-tile'
def prepare():
distribution = importlib.metadata.distribution('flash-attn-4')
source = Path(distribution.locate_file('flash_attn/cute/flash_fwd.py'))
content = source.read_bytes()
if distribution.version != '4.0.0b26' or hashlib.sha256(content).hexdigest() != ORIGINAL:
raise RuntimeError('This guard requires the verified unmodified FA4 4.0.0b26; no package was changed.')
target = location()
if target.exists():
validate(target)
return target
text = content.decode()
marker = ' m_block, num_head, batch_size, _ = work_tile.tile_idx\n'
if text.count(marker) != 1:
raise RuntimeError('Unexpected FA4 source layout')
start = text.index(marker)
end = text.index(' @cute.jit\n def compute_one_n_block', start)
text = text[:start] + ' if work_tile.is_valid_tile:\n' + ''.join(
' ' + line if line.strip() else line for line in text[start:end].splitlines(keepends=True)) + text[end:]
if hashlib.sha256(text.encode()).hexdigest() != PATCHED:
raise RuntimeError('FA4 guard output differs from the validated patch')
target.parent.mkdir(parents=True, exist_ok=True)
with tempfile.TemporaryDirectory(dir=target.parent, prefix='fa4-guard-') as temporary:
temp = Path(temporary) / 'overlay'
shutil.copytree(source.parent, temp / 'flash_attn/cute', ignore=shutil.ignore_patterns('__pycache__', '*.pyc'))
(temp / 'flash_attn/cute/flash_fwd.py').write_text(text, encoding='utf-8')
(temp / 'manifest.json').write_text(json.dumps({'distribution': distribution.version,
'source_sha256': ORIGINAL, 'patched_sha256': PATCHED, 'installed_package_modified': False,
'upstream_reference': 'https://github.com/Dao-AILab/flash-attention/blob/main/flash_attn/cute/flash_fwd.py'}, indent=2) + '\n', encoding='utf-8')
try:
os.rename(temp, target)
except OSError:
if not target.exists():
raise
validate(target) # Another setup process may have completed first.
return target
def validate(root):
target = Path(root) / 'flash_attn/cute/flash_fwd.py'
if hashlib.sha256(target.read_bytes()).hexdigest() != PATCHED:
raise RuntimeError('FA4 overlay failed its source check')
def activate():
root = location()
if not root.exists():
return None
validate(root)
if 'flash_attn.cute' in sys.modules:
current = Path(sys.modules['flash_attn.cute'].__file__).resolve()
if root.resolve() not in current.parents:
raise RuntimeError('FA4 was imported before the guard; start a fresh Engine process')
# Extend the package path instead of replacing FA2's __init__.py. This works
# with both the FA4 namespace package and a separately installed FA2 package.
package = importlib.import_module('flash_attn')
paths = [str(root / 'flash_attn'), *list(package.__path__)]
package.__path__ = list(dict.fromkeys(paths))
return root
def check_selected():
"""Refuse the known faulty SM120 kernel before any attention execution."""
import torch
if torch.cuda.get_device_capability()[0] != 12:
return
module = importlib.import_module('flash_attn.cute.flash_fwd')
digest = hashlib.sha256(Path(module.__file__).read_bytes()).hexdigest()
if digest == ORIGINAL:
raise RuntimeError('The attention speed-up component on this computer needs an update. Run freevideo setup --fa4-guard in a terminal, then restart FreeVideo.')