Use a faster int8 model on GeForce and RTX 30-series cards, with a one-click upgrade (0.3.0)

[KBRASK]

GeForce RTX 40/50 and RTX 30-series cards now run the prepared ConvRot int8 model, which is faster on these cards and closer to the original weights than FP8. Existing installations keep their model until the user upgrades from the launcher in one click: the new model downloads beside the current one, the switch happens between videos, and only then are the old model's own files removed. Release 0.3.0, which also ships the prompt enhancement (#75) and the installer retries and download cache cleanup (#73).

int8 on CUDA

• Activations get the cache's group-256 ConvRot rotation and one scale per row; a Triton GEMM multiplies them with the int8 weight rows, and the int32 products are exact before the scale epilogue.
• Chosen for GeForce Ada and Blackwell, where FP8 with FP32 accumulation runs at half rate, and for every Ampere card, which has no FP8 tensor cores. Workstation and datacenter cards keep FP8, which measured 5% faster on an RTX PRO 6000.
• int8 rows carry their own scales, so the planner reserves no FP8 activation stash and spends the room on larger attention batches.
• Setup checks the int8 path on the GPU before it finishes an int8 installation. Low-rank LoRAs run beside the int8 projections; adapters that must be merged keep the FP8 model.

Measured (Windows, 1344×768, Light 8+3, warm requests)

GPU	Video	FP8	int8
RTX 3090 24 GB	5 s	433.7 s	187.5 s (2.31×; sampling 329 → 113 s)
RTX 4070 12 GB	10 s	558 s (one out-of-memory retry)	433 s
RTX 5060 Ti 16 GB	15 s	945 s (FP8 staged to host memory)	655 s

On matched noise against BF16, int8 was closer than per-tensor FP8 in 7 of 7 scenes, row-wise FP8 in 6 of 7 and BF16 weight-only in 5 of 7.

One-click upgrade for existing installations

• An engine update keeps the installed model and never starts a download. The launcher shows a card with the expected speedup, the download and the space freed; nothing changes until the user clicks.
• The download is planned and approved exactly as shown, and runs beside the model in use with only the setup lease, so generation can continue.
• Between videos the launcher stops the ComfyUI it started, because its resident worker keeps SageAttention's DLL loaded and Windows refuses to replace it. Setup checks the int8 kernels before it marks the installation unfinished, so a GPU that fails them keeps its configuration.
• The retired variant is then removed file by file: only catalogued files at their catalogued sizes, their download sidecars, interrupted partials, staging folders and engine-generated AdaLN tables bound to their identity. Nothing is removed through links or junctions, nothing shared with a LoRA variant and never the model in use; the setup, engine and host-setup leases are held, every fingerprint is rechecked and a receipt is written.
• A move interrupted before the removal, or an int8 download the GPU could not run, is offered again on the card and in Settings → Storage and removed only after confirmation.

Verified on real machines

• RTX 3090, Windows 10: fresh v0.2.4 install, engine update, card shown without generating anything, 21.4 GiB download, switch with the int8 probe passing on sm_86, old model removed: 557 files, 21.355 GiB, 0 skipped. The inventory of 129,246 files before and after matched the receipt; the only other removals were the replaced engine package metadata. The leftover path removed a restored copy again after confirmation.
• RTX 4070: a command-line migration removed 559 files with none skipped, and removing the int8 variant again (302 files) matched the inventory diff exactly.

Setup runs from an open launcher (Install / repair, the switch) no longer fail on Windows after a video, since the launcher stops its idle ComfyUI first. Release notes, README News and the planning docs are updated.