Cccccz's picture
Add files using upload-large-folder tool
153dd42 verified
|
Raw
History Blame Contribute Delete
2.05 kB

AdaptiveDetailCache on MAGI-1 (dev6-adaptive)

基于 FlowCache4MAGI-1-dev4-detail 的扩展分支,在 MotionDetailCache(motion + detail 双度量)之上,增加 chunk 级 AR 难度预测 → 动态阈值 τ

与 dev4 的差异

维度 dev4 MotionDetailCache dev6 AdaptiveDetailCache
Phase 2 阈值 全局固定 τ 每个 chunk 独立 τ_eff
难度信号 motion / detail / delta / 历史 active / 历史 reuse
调参方式 手动 sweep rel_l1_thresh τ_base + beta + [τ_min, τ_max]

动态 τ 公式

每个 chunk 进入 Phase2 时(chunk 内固定,避免 step 间抖动):

d = weighted(motion, detail, delta, hist_active, hist_reuse)  ∈ [0, 1]
τ_eff = clamp(τ_base * exp(-β * (d - 0.5)), τ_min, τ_max)
active[p] = A[p] > τ_eff

难度高 → τ 降低 → 更常计算;难度低 → τ 升高 → 更敢 reuse。

默认超参

继承 dev4 best,并新增:

参数 默认值 说明
rel_l1_thresh 0.012 τ 基准(dev4 best)
use_adaptive_tau true 开启动态 τ
adaptive_tau_beta 0.8 难度→τ 灵敏度
adaptive_tau_min 0.008 τ 下界
adaptive_tau_max 0.020 τ 上界

快速运行

cd FlowCache4MAGI-1-dev6-adaptive
conda activate magi
export CUDA_VISIBLE_DEVICES=0
export MASTER_ADDR=localhost
bash scripts/single_run/adaptive_t2v.sh

dev4 vs dev6 对比

FRAMES=120 bash tools/run_compare_dev4_dev6.sh
FRAMES=240 bash tools/run_compare_dev4_dev6.sh

代码结构

inference/pipeline/cache/
├── motiondetailcache.py      # dev4 基类(含 τ hook)
└── adaptivedetailcache.py    # dev6 AdaptiveDetailCache

yaml_config/single_run/adaptive_config_best.yaml
scripts/single_run/adaptive_t2v.sh
tools/run_compare_dev4_dev6.sh

其余 sparse forward、KV cache、integrate 逻辑与 dev4 完全一致。