nodes/hao-ai-lab-FastVideo
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

[feat]: QAD gate — QAT student + dense scorers + guidance 2.0 (legacy loader-gated recipe)

Browse Source
This commit is contained in:
SolitaryThinker
2026-07-23 06:04:56 -07:00
parent 8808976943
commit a8205ea73b
2 changed files with 13 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastvideo2/train/gates/capture_train_main.py
+3 -3
View File
@@ -33,9 +33,9 @@ import sys
MODE = sys.argv[1] if len(sys.argv) > 1 and not sys.argv[1].startswith("-") else "finetune"
VSA = MODE in ("vsa", "vsa_dmd2")
DMD2 = MODE in ("dmd2", "vsa_dmd2")
DMD2 = MODE in ("dmd2", "vsa_dmd2", "qad")
os.environ["FASTVIDEO_ATTENTION_BACKEND"] = ("VIDEO_SPARSE_ATTN" if VSA else
"ATTN_QAT_TRAIN" if MODE == "qat" else "FLASH_ATTN")
"ATTN_QAT_TRAIN" if MODE in ("qat", "qad") else "FLASH_ATTN")
os.environ.setdefault("WANDB_MODE", "offline")
os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")
@@ -116,7 +116,7 @@ def main() -> None:
"--dmd_denoising_steps", "1000,757,522",
"--simulate_generator_forward",
"--generator_update_interval", "2",
"--real_score_guidance_scale", "3.5",
"--real_score_guidance_scale", "2.0" if MODE == "qad" else "3.5",
"--min_timestep_ratio", "0.02", "--max_timestep_ratio", "0.98",
"--flow_shift", "8.0"]
argv[argv.index("5e-5")] = "2e-6" # learning_rate per the shipped recipe
fastvideo2/train/gates/train_anchor.py
+10 -4
View File
@@ -162,8 +162,9 @@ def run_dmd2(mode: str = "dmd2") -> int:
steps = json.load(f)
device = "cuda"
is_vsa = mode == "vsa_dmd2"
# student: VSA blocks + FastWan weights; teacher/critic: DENSE blocks +
# base weights (main's scorers are dense-built and score via plain flash)
is_qad = mode == "qad"
# student: VSA/QAT blocks per recipe; teacher/critic: DENSE blocks + base
# weights (main's scorers are dense-built / loader-gated back to flash)
card = FASTWAN_T2V_1_3B if is_vsa else WAN21_T2V_1_3B
base_root = resolve_weights(WAN21_T2V_1_3B)
@@ -175,9 +176,14 @@ def run_dmd2(mode: str = "dmd2") -> int:
student = WanModelFVVSA.from_pretrained(resolve_weights(card),
torch_dtype=torch.float32,
subfolder="transformer").to(device)
elif is_qad:
from fastvideo2.wan21.model_fv import WanModelFVQAT
student = WanModelFVQAT.from_pretrained(base_root, torch_dtype=torch.float32,
subfolder="transformer").to(device)
else:
student = load_dense()
step_ctx = DMD2Step(student, load_dense(), load_dense())
step_ctx = DMD2Step(student, load_dense(), load_dense(),
guidance=2.0 if is_qad else 3.5)
pz = np.load(os.path.join(gold, "params.npz"))
uncond = torch.from_numpy(pz["neg_embeds"]).to(device, torch.bfloat16)
@@ -282,4 +288,4 @@ def run_dmd2(mode: str = "dmd2") -> int:
if __name__ == "__main__":
args = [a for a in sys.argv[1:] if not a.startswith("-")]
mode = args[0] if args else "finetune"
sys.exit(run_dmd2(mode) if mode in ("dmd2", "vsa_dmd2") else run(mode=mode))
sys.exit(run_dmd2(mode) if mode in ("dmd2", "vsa_dmd2", "qad") else run(mode=mode))