shaoxiongduan
ac919b1c87
- data_pipeline: SAM+CoTracker preprocessing, sharded Slurm launchers, OpenVid download/filter, and a data-parallel Wan2.2-T2V-A14B synthetic video generator with the first-frame over-brightness fix (over-generate 129 -> drop first 8 -> keep 121). - examples/train: OpenVid bidir/stage2/stage3 launchers + finetune YAMLs (sparse, merged, merged_bias, stage2v2 hi-lr, stage3A/B/C). - research_log: first-frame saturation root-cause writeup (DiT position-0 latent, not VAE decode). - fastvideo: trackwan model / track-encoder / preprocessing updates, dcp_to_diffusers export, FA4.
40 lines
2.0 KiB
Bash
40 lines
2.0 KiB
Bash
#!/usr/bin/env bash
|
|
# Replace the stale XOR-op fork FA4 with the OFFICIAL Dao-AILab flash_attn/cute,
|
|
# which pins the matching cutlass-dsl==4.6.0.dev0 + quack>=0.5.3 (coherent set).
|
|
set +e
|
|
export CUDA_HOME=/usr/local/cuda
|
|
export PATH=$CUDA_HOME/bin:/mnt/lustre/vlm-s4duan/bin:$PATH
|
|
export LD_LIBRARY_PATH=$CUDA_HOME/lib64:$LD_LIBRARY_PATH
|
|
export UV_CACHE_DIR=/mnt/lustre/vlm-s4duan/.uv/cache
|
|
export UV_PYTHON_INSTALL_DIR=/mnt/lustre/vlm-s4duan/.uv/python
|
|
cd /mnt/lustre/vlm-s4duan/FastVideo || exit 1
|
|
source .venv/bin/activate
|
|
|
|
echo "==================== remove fork FA4 + its cutlass/quack ===================="
|
|
uv pip uninstall flash-attn-cute quack-kernels nvidia-cutlass-dsl 2>&1 | tail -6
|
|
|
|
echo "==================== install OFFICIAL Dao-AILab flash_attn/cute (real name: flash-attn-4) ===================="
|
|
GIT="git+https://github.com/Dao-AILab/flash-attention.git#subdirectory=flash_attn/cute"
|
|
uv pip install --prerelease allow "$GIT" 2>&1 | tail -35
|
|
if [ "${PIPESTATUS[0]}" != "0" ]; then
|
|
echo ">> retry with --no-build-isolation + setuptools_scm"
|
|
uv pip install setuptools_scm setuptools wheel 2>&1 | tail -3
|
|
uv pip install --prerelease allow --no-build-isolation "$GIT" 2>&1 | tail -35
|
|
fi
|
|
|
|
echo "==================== resolved versions ===================="
|
|
python - <<'PY'
|
|
import importlib.metadata as md
|
|
for p in ["flash-attn-cute","nvidia-cutlass-dsl","nvidia-cutlass-dsl-libs-base",
|
|
"nvidia-cutlass-dsl-libs-core","nvidia-cutlass-dsl-libs-cu12",
|
|
"quack-kernels","torch-c-dlpack-ext","torch"]:
|
|
try: print(" ", p, md.version(p))
|
|
except Exception: print(" ", p, "absent")
|
|
PY
|
|
|
|
echo "==================== import tests ===================="
|
|
python -c "import cutlass.cute.core as c; print('ThrMma?', hasattr(c,'ThrMma'))" 2>&1
|
|
python -c "from flash_attn.cute.interface import _flash_attn_fwd,_flash_attn_bwd; print('FA4_CUTE_INTERFACE_OK')" 2>&1
|
|
python -c "import importlib; importlib.import_module('fastvideo.attention.backends.flash_attn'); print('FASTVIDEO_FLASH_ATTN_BACKEND_OK')" 2>&1
|
|
echo "==================== DONE fa4_official ===================="
|