openai_whisper-large-v3-turbo-NVFP4
NVFP4 (NVFP4, W4A4) post-training quantization of openai/whisper-large-v3-turbo — architecture: whisper.
- Format:
nvfp4-pack-quantized(compressed-tensors). 4-bit FP4 weights, per-block FP8 (E4M3) scales, per-tensor FP32 global scales; activations dynamically quantized to FP4. - 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 on RTX 5080 Laptop (sm_120).
- Quantized layers: all Linear modules in the encoder/decoder (CTC
lm_head/proj_outleft full precision).
Eval — Reza2kn/persian-asr-eval-v0 (FLEURS-fa)
| Variant | WER ↓ | CER ↓ | clips | per-clip latency | peak VRAM |
|---|---|---|---|---|---|
| NVFP4 (this repo) | 24.18% | 8.30% | 200 | 652 ms | 2156 MiB |
Persian text normalization for WER/CER: NFKC, ZWNJ → space, ي→ی / ك→ک, digit folding, punctuation stripping, whitespace collapse.
Usage
import torch
import soundfile as sf
from transformers import AutoProcessor, AutoModel
repo = "Reza2kn/openai_whisper-large-v3-turbo-NVFP4"
processor = AutoProcessor.from_pretrained(repo)
# Load in bfloat16 — NVFP4 weights decompress to bf16 inside CompressedLinear.
model = AutoModel.from_pretrained(repo, dtype=torch.bfloat16).to("cuda").eval()
(See the original openai/whisper-large-v3-turbo model card for arch-specific decoding boilerplate.)
How it was made
llmcompressor QuantizationModifier(targets=["Linear"], scheme="NVFP4", ignore=...) →
compressed-tensors nvfp4-pack-quantized checkpoint.
License
Inherits the base model's license.