English | 简体中文
A field-tested handbook for running MiniMax H3 local video generation on a 2080Ti 22G mod card (Turing / sm_75).
Every conclusion here comes from a real production pipeline, not paper math. Applies to all Turing (sm_75) GPUs — 2080Ti 22G / 11G both work; the closer your VRAM is to 22G, the closer you are to this configuration.
One main path: official tutorial reference (20–30 min/clip) → cu130 dequantization + W4A8 → three acceleration branches (Turbo / T8 / PDD) → PDD+T8 combo (210 s, 6/8 cache hits) → the six-workflow kit.
The same 5-second clip: 20–30 min/clip in a typical environment following the official tutorial, compressed to 4.7 min (final-shot standard) / 2.7 min (fast draft) / 3.5 min (PDD+T8 ultra draft) — 5–11× faster. Per-tier timings and data methodology: 08 / 09.
- ✅ Six ready-to-import workflows (t2v/i2v, each with a final-shot tier and a T8 fast-draft tier, plus a PDD ultra tier) — all verified on real hardware, with preview screenshots: workflows/README
- ✅ The Python library
h3-turing: all of the above codified - six tier presets, a retry-disciplined ComfyUI client, the A/B benchmark harness, pre-production checks. Zero dependencies (stdlib only) — see python/ - ✅ 13 field-tested FAQ entries: v3-node 400s on
/prompt, same-seed re-runs diverging, idle auto-exit killing batches, antivirus slowing model loading by 27 minutes, silent prompt_id dedup drops… every one paid for in real time → 06 - ✅ Verdict table for every speedup route: the full SageAttention crash autopsy, T8 BlockCache measured at −43%, TE-Speed permanently ruled out — which roads work and which are dead, so you don't have to try them again
- ✅ A/B measurement methodology: same-seed A/B control scripts ready to run for your own experiments → scripts/
- Hardware math: sm_75 has no BF16/FP8 tensor cores and 616 GB/s bandwidth — know the ceiling first → 01
- Quantization verdict: DiT must be INT8 (W4A8); W4A4 has 18× the reconstruction error and produces color tearing → 02
- The compat (degraded) workflows —
workflows/imports directly
- Launch flags
9235
against TDR black-screens / OOM:
--reserve-vram 2.5 --vram-headroom 0.5 --disable-pinned-memory→ scripts/ - Antivirus real-time scanning slowing model loading by 27 minutes → disable while rendering, re-enable after, plus a resume-from-checkpoint fallback → 06
- Ten more pitfalls cleared: prompt_id dedup, queue residue, big-file downloads → 06
- Resolution top-up: generate at 640×352 → RTX VSR frame upscale to 1080p (47 ms/frame)
- The cu130 dequantization bonus: the comfy-kitchen CUDA backend fully enabled — this is the root cause of 5.7 min vs the official 20–30 min. If W4A8 is absurdly slow on your box, check your runtime before blaming the GPU → 01
- SageAttention: Triton INT8 kernels fail to compile on sm_75 across all triton versions; the CUDA kernel runs standalone but crashes natively in the real pipeline → ruled out, fully rolled back → 03
- TE-Speed: semantic collapse at short step counts → permanently ruled out, no new version will save it → 06
- T8 BlockCache on real hardware: aggressive tier −43% = 2.7 min/clip; corrected the "requires ComfyUI ≥0.34" misdiagnosis (v0.33.1 works); default params are a negative optimization on the 4-step route (0 hits); the cost = same-seed runs are no longer reproducible → drafts yes, final shots no → 08
- PDD LoRA master backport (no waiting for v0.34.1+): single-file wasn't enough → full upgrade with three pitfalls solved (the comfy_api blind spot / PyAV / the T8 signature) → PDD 8-step at 600 s reproducible, and PDD+T8 combo at 210 s (−34%) with 6/8 hits → 09
- T8 node master-compat patch (self-adapting to the 7-arg FinalLayer signature), upstream feedback pending
- Ref2VA (i2v) PDD variant tested: 933 s reproducible / +T8 192 s (−51%, 6/8 hits) — fastest in the project
- T8 compat closed upstream: node pack v1.0.4 ships the explicit-signature fix; issue verified and closed
⚠️ Sep 20: PDD/T8 tiers temporarily benched — a pip env contamination (rembg/rapidocr installs) broke quant+LoRA loading; root-caused with full stack and reported upstream (#16420); final-shot blind test re-queued after the fix
- H3 Max (fal.ai post-trained edition): API-only, no open weights — unusable locally; pay per clip for critical shots, wait for an open follow-up → 07
| Doc | Contents |
|---|---|
| 01 Hardware limits | What sm_75 lacks, the bandwidth gap, the cu130 bonus, why W4A8 is the only route |
| 02 W4A8 vs W4A4 | Error data, production speed, another card's benchmark |
| 03 SageAttention crash autopsy | The 2.2.0 wheel pipeline crash → root-cause hunt → rollback verification |
| 04 Community tips verified | Three directly-copyable tips + their applicability conditions |
| 05 Workflows | compat t2v/i2v import, placeholder replacement, model download list |
| 06 FAQ | 13 entries: v3-node 400s, same-seed divergence, idle auto-exit, antivirus, prompt_id dedup, TE-Speed, audio clipping |
| 07 Upgrade-window watch | The full route verdict table, v0.34 assessment, PDD LoRA #15908, official T8 open-sourcing |
| 08 T8 on the 4-step route | The −43% measurement: zero hits at defaults, 2.7 min/clip aggressive, the same-seed reproducibility cost |
| 09 PDD without waiting | The master backport field report: PDD8+T8 210 s (−34%), 6/8 hits, three pitfalls solved |
| 10 Benchmark dataset | Every measurement on one page: all tiers, the hit matrix, mechanism constants, audio, version deltas, ruled-out routes |
| workflows/ | The six-workflow kit (with previews): t2v/i2v × final/draft + PDD ultra tiers, with import guide |
| scripts/ | Anti-black-screen launch template, T8 A/B harness, chart regen script |
- GPU: 2080Ti 22G mod (Turing, sm_75)
- ComfyUI: v0.33.1 baseline (the handbook also documents a full upgrade to master for PDD — see 09)
- Route: h3lite W4A8 compat + fl2v Turbo 4-step LoRA (optional T8 draft tier, see 08)
Both are auto-synced (Gitee is the authoritative source).
MIT