IIcyoungv0.1.0 scaffold — MiniMax H3 TeaCache

76e66b95创建于 8月4日历史提交

文件	最后提交记录	最后更新时间
README.md	v0.1.0 scaffold — MiniMax H3 TeaCache - ModelPatcher wrapper approach (no core file modification) - Rel-L1 threshold-based reuse decision - Guard rails: skip first/last N steps - Benchmark harness with 10 stock prompts × N seeds × threshold sweep - Design notes explaining departure from TE-Speed - MIT license	2 个月前
prompts.jsonl	v0.1.0 scaffold — MiniMax H3 TeaCache - ModelPatcher wrapper approach (no core file modification) - Rel-L1 threshold-based reuse decision - Guard rails: skip first/last N steps - Benchmark harness with 10 stock prompts × N seeds × threshold sweep - Design notes explaining departure from TE-Speed - MIT license	2 个月前
run.py	v0.1.0 scaffold — MiniMax H3 TeaCache - ModelPatcher wrapper approach (no core file modification) - Rel-L1 threshold-based reuse decision - Guard rails: skip first/last N steps - Benchmark harness with 10 stock prompts × N seeds × threshold sweep - Design notes explaining departure from TE-Speed - MIT license	2 个月前

Benchmark harness

Two things we measure:

Speed — wall-clock time per sampling step, and total end-to-end.
Quality — pixel-space LPIPS and VBench subset scores against a TeaCache-off baseline.

Fixed test setup

Model: minimax_h3_fl2va_pruned_int8_convrot.safetensors
Encoder: qwen3vl_32b_minimax_h3_int8_convrot.safetensors
VAE (video): minimax_h3_video_vae_fp16.safetensors
VAE (audio): minimax_h3_audio_vae_fp32.safetensors
Sampler: res_multistep
Scheduler: beta, 20 steps, denoise 1.0
Prompts: prompts.jsonl (10 prompts, mixed styles)
Seeds: [0, 1, 2] (three seeds per prompt)
Resolution: 1024 × 576, length 121 (5 s @ 24 fps)

Sweeps

`rel_l1_thresh`	Expected speedup
0.00 (off)	baseline
0.05	~15%
0.10	~25%
0.15	~40%
0.20	~50%
0.25	~55%

Run

cd benchmarks
python run.py \
    --comfy-url http://hp-z4-server:8188 \
    --workflow ../workflows/fl2va-bench.json \
    --thresholds 0,0.05,0.10,0.15,0.20,0.25 \
    --output output/

Output is a CSV with columns: prompt_id, seed, thresh, real_steps, reuse_steps, wall_time_s, lpips_vs_off.

Hardware note

The 170HX in the reference machine is a mining card with mod-VRAM (64GB total). Its FP16 throughput is ~85% of an RTX 3090, so absolute wall-clock numbers are conservative. Speedup ratios transfer well to 4090/5090/H100.