NVFP4 (W4A4) post-training quantization of ALIGHASEMI931/w2v-bert-2.0-persian-colab-CV16.0, itself a Common Voice 16.0 fine-tune of facebook/w2v-bert-2.0 for Persian ASR.

Format: nvfp4-pack-quantized (compressed-tensors). 4-bit FP4 (E2M1) weights, per-block FP8 (E4M3) scales, per-tensor FP32 global scales; activations dynamically quantized to FP4 (W4A4).
Quantized layers: all 199 Linear layers in the encoder + adapter. CTC lm_head left full precision.
Calibration: 32 Persian clips from Reza2kn/persian-asr-eval-v0 (held out from the WER eval set).
Hardware target: NVIDIA Blackwell tensor cores (sm_100+). Quantized + verified on RTX 5080 Laptop (sm_120).

Eval — `Reza2kn/persian-asr-eval-v0` (FLEURS-fa, 839 clips)

Variant	WER ↓	CER ↓	Disk	Peak VRAM	per-clip latency
NVFP4 (this repo)	35.19%	9.61%	620 MB	2590 MiB	199 ms

Persian text normalization for WER/CER: NFKC, ZWNJ→space, ي→ی / ك→ک, digit folding, punctuation stripping, whitespace collapse.

Usage

python

import torch, soundfile as sf
from transformers import AutoFeatureExtractor, AutoTokenizer, Wav2Vec2BertForCTC

repo = "Reza2kn/ALIGHASEMI931_w2v-bert-2.0-persian-colab-CV16.0-NVFP4"
fe = AutoFeatureExtractor.from_pretrained(repo)
tok = AutoTokenizer.from_pretrained(repo)
model = Wav2Vec2BertForCTC.from_pretrained(repo, dtype=torch.bfloat16).to("cuda").eval()

wav, sr = sf.read("clip.wav", dtype="float32")
feat = fe(wav, sampling_rate=sr, return_tensors="pt")
feat = {k: (v.to("cuda").to(torch.bfloat16) if v.is_floating_point() else v.to("cuda")) for k, v in feat.items()}

with torch.no_grad():
    logits = model(**feat).logits

ids = logits.argmax(-1)[0].tolist()
out, prev = [], -1
for i in ids:
    if i != prev and i != tok.pad_token_id:
        out.append(i)
    prev = i
print(tok.decode(out, skip_special_tokens=True))

License

Inherits the base model's license.

Reza2kn/ALIGHASEMI931_w2v-bert-2.0-persian-colab-CV16.0-NVFP4

w2v-bert-2.0 Persian — NVFP4

Eval — Reza2kn/persian-asr-eval-v0 (FLEURS-fa, 839 clips)

Usage

License

Reza2kn/
ALIGHASEMI931_w2v-bert-2.0-persian-colab-CV16.0-NVFP4

Eval — `Reza2kn/persian-asr-eval-v0` (FLEURS-fa, 839 clips)