FireRedTTS3-int8 (community INT8 ConvRot mirror)

INT8 ConvRot conversion of FireRedTeam/FireRedTTS3 for FireRedTTS3-ComfyUI, produced with the official comfy-kitchen quantizer (TensorWiseINT8Layout.quantize, registry quantize_int8_convrot_weight).

Format per quantized Linear (current ComfyUI representation):

weight — torch.int8, original [out, in] shape, contains the offline Hadamard-rotated weight (W @ H^T per 256-column group)
weight_scale — torch.float32, [out, 1] per-output-row scale
bias — original float bias
comfy_quant — uint8 JSON: {"format": "int8_tensorwise", "convrot": true, "convrot_groupsize": 256}

At inference the companion custom node rotates activations online via comfy_kitchen.int8_linear(..., convrot=True, convrot_groupsize=256) — dynamic per-row INT8 activation quantization + INT8 GEMM, rescaled by scale_x * scale_w. No whole-weight dequantization on the hot path.

What is quantized (safe profile, group size 256)

Component	Quantized	Kept float
`fireredtts3_base`	321/332 Linears (1.73B params, 81.5% of core): all `backbone_llm.layers.`, `patch_encoder.blocks.`, `dit.blocks.*`	embeddings, norms, `spk_proj_*`, `patch_encoder.in_proj/out_proj`, `dit_head`, `dit.in_proj` (1600 % 256 != 0), `dit.t_embedder`, `dit.final_layer`, `stop_head`, Conv1d
`fireredtts3_instruct`	321/331 Linears (1.73B params, 71.2% of core): same block families (`backbone_llm.model.layers.*`)	same exclusions
`redae`	nothing	everything
`campp`	nothing	everything

Sizes

Core	Official fp32	This repo
`fireredtts3_base`	8.48 GB	3.30 GB
`fireredtts3_instruct`	8.48 GB	3.30 GB
`redae` / `campp` / tokenizer	copied through unchanged

Validation (base variant, full suite; instruct smoke-tested)

Per-layer weight roundtrip (official quantize -> official dequantize): worst rel-L2 0.00967, worst cosine 0.999953 over 321 layers
Real-activation comparison vs fp32 through the same comfy_kitchen.int8_linear runtime: worst rel-L2 0.01162, worst cosine 0.999932
On-disk structure: all 677 original keys preserved, scales fp32 [N,1] and positive, Conv1d/RedAE/CAM++ untouched
Runtime proof: 321 ConvRotInt8Linear modules, >42k counted INT8 ConvRot kernel calls during generation, weights stay int8 across unload/reload
BF16 vs INT8 generation (same seed/settings): identical patch counts (200/200), finite latents, EN/ZH ASR-verified, speaker-similarity parity (0.9007 vs 0.8991)
Peak VRAM 13.1 -> 8.3 GiB; generation ~1.3x slower (memory optimization, honestly reported)

Usage Disclaimer

The project incorporates zero-shot voice cloning functionality; Please note that this capability is intended solely for academic research purposes.
DO NOT use this model for ANY illegal activities❗️❗️
The developers assume no liability for any misuse of this model.
If you identify any instances of abuse, misuse, or fraudulent activities related to this project, please report them to our team immediately.

Citation

@article{fireredtts3,
  title   = {FireRedTTS3: Unified Speech Generation and Editing with Semantically Enriched Speech Representations},
  author  = {FireRed Team},
  journal = {arXiv preprint},
  year    = {2026},
}

Acknowledgements

Qwen3 and Qwen2-Audio for the language model and audio understanding foundations
DiTAR for the patch-level diffusion autoregressive formulation
X-Codec for the discriminator design used in RedAE training
CAM++ for speaker embedding extraction
fastText for automatic language identification
WeTextProcessing (wetext) for the Chinese / English text normalization front-end

All credit to the FireRed Team — see the upstream repo and model card. Apache-2.0.

drbaph/FireRedTTS3-int8