The Eight-Gigabyte Threshold: IndexTTS 2.5 Fully Homesteaded, the Hardware Ledger of 104 Seconds for 6.5 Seconds, and the Two Self-Inflicted Wounds of One Bat File

Bilibili’s voice-cloning marvel IndexTTS 2.5 ships code but no product — so today’s work was turning code into product: environment, weights, interface, launcher, and one honest hardware ledger.

I. The Homestead: From Clone to Smoke Test

The whole road followed the official way: the GitHub repository cloned as-is (including the patch for the end-of-clip click fixed two days ago), uv installing the environment from the lockfile while skipping the Windows-hostile DeepSpeed, PyTorch 2.8 plus CUDA 12.8 correctly recognizing the RTX 3070; model weights downloaded in batches through the proxy — 5.2GB of main models (the 3.2GB GPT backbone, codec, s2mel, the Qwen emotion model) plus 3.9GB auxiliary (w2v-bert, MaskGCT, CAMPPlus, BigVGAN). One cocoon was split on the way: the first start was interrupted, leaving w2v-bert-2.0 as a half-directory of config files only, and the second start’s program saw the directory, believed the model present, and crashed on load — remains deceive better than emptiness. Deleted, re-downloaded, passed in one go. The smoke test used the bundled sample voice as reference and synthesized 5.1 seconds of mixed Chinese-English speech into a valid 22050Hz WAV — the whole reference-audio→cloning→synthesis chain proven in the field.

II. BF16: The Correct Posture on an 8GB Card

The VRAM decision chain was traced: at startup the program detects 8GB below the 10GB threshold and enters low-VRAM mode — half precision plus skipping the QwenEmotion text-emotion model; which half precision is chosen is asked of the card (Ampere supports BF16), so it runs BF16 rather than FP16 — both 16-bit with identical memory cost, but BF16’s wider numeric range is stabler and the preferred choice for such models. The GPT backbone, s2mel and codec all run BF16. This is not a make-do downgrade; it is the optimum for this card — and slowness has nothing to do with precision, being purely the account of model compute against a nearly full VRAM.

III. The Hardware Ledger: 104 Seconds for 6.5 Seconds

The user’s complaint had data behind it: 6.5 seconds of audio cost about 104 seconds of waiting, the worst ratio 1:37 (seven seconds of audio waited four-plus minutes), the bottleneck in s2mel alone at 85 seconds. The verdict offers no excuse: 8GB of VRAM is the passing line for version 2.5, not the comfort zone — with the model resident the card is topped out, desktop and browser take their share, and when Windows spills into shared memory the speed drops an order of magnitude. A bundle repack runs the same model on the same card; packaging changes no compute — if a repack seems faster, it simply ships a lighter model (IndexTTS 1.5, GPT-SoVITS and their kind). The user tested a repack, agreed with the ledger, and chose to keep the official build.

IV. The Two Self-Inflicted Wounds of One Bat File

The closing vignette is a self-reflection textbook: the hand-written launcher bat left the site unable to start on double-click — first, the —inbrowser argument does not exist in the official webui.py, so Python exited on unrecognized arguments; second, the bat’s Chinese hints were UTF-8 while the Chinese Windows cmd parses GBK, mangling hint lines into phantom commands. Rewritten — argument removed, hints in pure ASCII, browser opening after a 40-second delay — and verified. Every deliverable for the user must be verified the way the user will use it — double-click is the user’s way; a background launch during testing is not.