ALWOAL/lscodec_25hz_int8

LSCodec-25 Hz — packed int8 weights

Int8 post-training quantization of cantabile-kwok/lscodec_25hz (encoder + vocoder + VQ codebook, 49.3 MB instead of 196.0 MB fp32), produced with the same quantizer the quantization lab Space uses for its lscodec-int8 row.

File	Contents	Size
encoder.safetensors	LSCodec encoder (conv frontend, conformer, quantizer)	17.6 MB
vocoder.safetensors	CTXVEC2WAV vocoder	31.6 MB
codebook.safetensors	1024-entry, 64-dim VQ codebook	0.07 MB

Method: per-output-channel absmax scale (per-tensor for 1-D params), round-to-nearest into signed int8, stored as int8 levels plus an fp32 _scale/<name> tensor per parameter. Dequantizing reproduces the lab's lscodec-int8 model bit-for-bit; integer buffers (num_batches_tracked) are stored unchanged.

Unchanged: the token stream is still 25 tokens/s × 10 bits = 0.25 kbps, and the WavLM-Large prompt encoder (microsoft/wavlm-large, first 6 layers) is neither included here nor quantized — it only embeds the prompt.

Measured on a 3 s LibriSpeech clip: STOI 0.694 / LSD 11.10 dB at int8 versus 0.725 / 11.18 dB at fp32 (fetched by the lab Space).

Native int8 — no dequantization

int8nn.py (shipped in this repo) swaps every Linear/Conv1d/ConvTranspose1d/Embedding for an int8 twin: weights stay int8 in memory, each GEMM quantizes its activations per row and computes int8 x int8 -> int32 via torch._int_mm (W8A8 dynamic). Measured natively: STOI 0.704 versus 0.694 through the dequantized fake-quant path (0.725 at fp32).

from huggingface_hub import hf_hub_download
from safetensors.torch import load_file
import int8nn  # from this repo

for module, name in ((encoder, "encoder"), (vocoder, "vocoder"), (codebook, "codebook")):
    packed = load_file(hf_hub_download("ALWOAL/lscodec_25hz_int8", f"{name}.safetensors", repo_type="model"))
    int8nn.apply_int8(module, packed)

Load

Requires the LSCodec-Inference repo to build the modules:

from huggingface_hub import hf_hub_download
from lscodec.utils import load_model, load_vocoder  # X-LANCE/LSCodec-Inference
from safetensors.torch import load_file

def dequantize(path: str) -> dict:
    raw = load_file(path)
    return {
        k: (v.float() if (s := raw.get("_scale/" + k)) is None else v.float() * s)
        for k, v in raw.items() if not k.startswith("_scale/")
    }

encoder = load_model(encoder_config, encoder_ckpt)
vocoder = load_vocoder(vocoder_config, vocoder_ckpt)
for module, name in ((encoder, "encoder"), (vocoder, "vocoder"), (codebook, "codebook")):
    path = hf_hub_download("ALWOAL/lscodec_25hz_int8", f"{name}.safetensors", repo_type="model")
    module.load_state_dict(dequantize(path))

Credits

LSCodec by Yiwei Guo et al. — MIT, X-LANCE/LSCodec-Inference. Quantization method and numbers from the FocalCodec quantization lab.