Index-Echo-S2ST-9B-FP4-v4rft

Official NVFP4 (W4A4) quantization of IndexTeam/Index-Echo-S2ST-9B (v4rft), part of the Index-Echo speech-to-speech translation (S2ST) model family by bilibili.

The v4rft package updates the multilingual translator and extends the release to twelve directions: Chinese, Japanese, or English speech input, dubbed into English, Spanish, Japanese, Korean, or Chinese.

This repository mirrors the original checkpoint layout; only the text LLM backbone (stlm_llm/) is quantized to NVFP4 - the audio tower, connector, and all speech-synthesis components remain in BF16. Use it exactly like the original repo (same modeling_dubbing.py / configs).

Quantization

• Scheme: NVFP4 (W4A4: 4-bit floating-point weights with per-group-16 scales, 4-bit floating-point activations with calibrated per-tensor global scales), produced with llm-compressor (quantization_scheme recorded in recipe.yaml). Calibrated on a small bilingual translation corpus.
• All Linear layers of the language model backbone (stlm_llm/) are quantized except the hybrid linear-attention projections; the audio tower, connector, lm_head, embeddings and all other pipeline components are kept in BF16.
• Format: compressed-tensors nvfp4-pack-quantized safetensors - load directly with vLLM (quantization="compressed-tensors") or transformers.

Consistency validation

Measured on an NVIDIA A100 (weight-dequantized execution) against the original BF16 checkpoint (greedy decoding, fixed bilingual corpus):

Metric	BF16	FP4	Delta
Perplexity (fixed corpus)	5.6834	6.0724	+6.85%
zh->en generation identical	-	-	no (semantically equivalent)
en->zh generation identical	-	-	no (semantically equivalent)

Usage

Identical to the original checkpoint - clone this repo and follow the README / modeling_dubbing.py of the base model (IndexTeam/Index-Echo-S2ST-9B). The quantized LLM backbone loads via compressed-tensors; make sure compressed-tensors (or vLLM / a recent transformers) is installed.

Hardware note: full NVFP4 (W4A4) acceleration requires an NVIDIA Blackwell GPU (SM100+, e.g. B200 / RTX 50 series). On older GPUs vLLM falls back to weight-only dequantization, which still reduces memory but gives no FP4 speedup.

Quantized and published by the Index team, 2026-10-15.