Scicom-AI-Enterprise-Organization/titanet-vectors-fp16

TitaNet-L vectors output in FP16.

Python

3 commits

updated Jan 24, 2026

See the code

README

titanet-vectors-fp16

TitaNet-L vectors output in FP16.

How to

Install,

pip3 install git+https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16

Follow load.ipynb.

Small things that we learnt

STFT always returned in complex64, https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16/blob/main/titanet_vectors/nemo_featurization.py#L464
when you want to get the real value, it will always in fp32, https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16/blob/main/titanet_vectors/nemo_featurization.py#L472
Maintain stats precision in fp32, but after that downcast to expected dtype before attention, https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16/blob/main/titanet_vectors/nemo/tdnn_attention.py#L319

Similar repos

tiiuae/onebitllms

Lightweight toolkit package to train and fine-tune 1.58bit Language models

149

intel/neural-compressor

SOTA low-bit LLM quantization (INT8/FP8/MXFP8/INT4/MXFP4/NVFP4) & sparsity; leading model…

2,712

metebalci/kuantala

Quantize diffusers-format generative models to NVFP4/FP8 using NVIDIA Model Optimizer. Supports…

NVIDIA/TransformerEngine

A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit and 4-bit…

3,580

lucidrains/vector-quantize-pytorch

Vector (and Scalar) Quantization, in Pytorch

4,013

mit-han-lab/fouroversix

Code for the papers: “Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling”…

204

comfyanonymous/ControlNet-v1-1_fp16_safetensors

Model

Safetensors/FP16 versions of the new [ControlNet-v1-1](https://huggingface.co/lllyasviel/ControlNet…

532

vllm-project/humming

Humming is a high-performance, lightweight, and highly flexible JIT (Just-In-Time) compiled GEMM…

256

Scicom-AI-Enterprise-Organization/titanet-vectors-fp16

TitaNet-L vectors output in FP16.

Python

3 commits

updated Jan 24, 2026

See the code

README

titanet-vectors-fp16

TitaNet-L vectors output in FP16.

How to

Install,

pip3 install git+https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16

Follow load.ipynb.

Small things that we learnt

STFT always returned in complex64, https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16/blob/main/titanet_vectors/nemo_featurization.py#L464
when you want to get the real value, it will always in fp32, https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16/blob/main/titanet_vectors/nemo_featurization.py#L472
Maintain stats precision in fp32, but after that downcast to expected dtype before attention, https://github.com/Scicom-AI-Enterprise-Organization/titanet-vectors-fp16/blob/main/titanet_vectors/nemo/tdnn_attention.py#L319

Similar repos

tiiuae/onebitllms

Lightweight toolkit package to train and fine-tune 1.58bit Language models

149

intel/neural-compressor

SOTA low-bit LLM quantization (INT8/FP8/MXFP8/INT4/MXFP4/NVFP4) & sparsity; leading model…

2,712

metebalci/kuantala

Quantize diffusers-format generative models to NVFP4/FP8 using NVIDIA Model Optimizer. Supports…

NVIDIA/TransformerEngine

A library for accelerating Transformer models on NVIDIA GPUs, including using 8-bit and 4-bit…

3,580

lucidrains/vector-quantize-pytorch

Vector (and Scalar) Quantization, in Pytorch

4,013

mit-han-lab/fouroversix

Code for the papers: “Four Over Six: More Accurate NVFP4 Quantization with Adaptive Block Scaling”…

204

comfyanonymous/ControlNet-v1-1_fp16_safetensors

Model

Safetensors/FP16 versions of the new [ControlNet-v1-1](https://huggingface.co/lllyasviel/ControlNet…

532

vllm-project/humming

Humming is a high-performance, lightweight, and highly flexible JIT (Just-In-Time) compiled GEMM…

256