You can deploy this model and start earning money today!

RichardErkhov/
arnavgrg_-_llama-2-7b-nf4-fp16-upscaled-gguf

Name: arnavgrg_-_llama-2-7b-nf4-fp16-upscaled-gguf
Availability: InStock

endpoints_compatible

gguf

template

Views

No views yet

Quantization made by Richard Erkhov.

Github

Discord

Request more models

llama-2-7b-nf4-fp16-upscaled - GGUF

Model creator: https://huggingface.co/arnavgrg/
Original model: https://huggingface.co/arnavgrg/llama-2-7b-nf4-fp16-upscaled/

Name	Quant method	Size
llama-2-7b-nf4-fp16-upscaled.Q2_K.gguf	Q2_K	2.36GB
llama-2-7b-nf4-fp16-upscaled.IQ3_XS.gguf	IQ3_XS	2.6GB
llama-2-7b-nf4-fp16-upscaled.IQ3_S.gguf	IQ3_S	2.75GB
llama-2-7b-nf4-fp16-upscaled.Q3_K_S.gguf	Q3_K_S	2.75GB
llama-2-7b-nf4-fp16-upscaled.IQ3_M.gguf	IQ3_M	2.9GB
llama-2-7b-nf4-fp16-upscaled.Q3_K.gguf	Q3_K	3.07GB
llama-2-7b-nf4-fp16-upscaled.Q3_K_M.gguf	Q3_K_M	3.07GB
llama-2-7b-nf4-fp16-upscaled.Q3_K_L.gguf	Q3_K_L	3.35GB
llama-2-7b-nf4-fp16-upscaled.IQ4_XS.gguf	IQ4_XS	3.4GB
llama-2-7b-nf4-fp16-upscaled.Q4_0.gguf	Q4_0	3.56GB
llama-2-7b-nf4-fp16-upscaled.IQ4_NL.gguf	IQ4_NL	3.58GB
llama-2-7b-nf4-fp16-upscaled.Q4_K_S.gguf	Q4_K_S	3.59GB
llama-2-7b-nf4-fp16-upscaled.Q4_K.gguf	Q4_K	3.8GB
llama-2-7b-nf4-fp16-upscaled.Q4_K_M.gguf	Q4_K_M	3.8GB
llama-2-7b-nf4-fp16-upscaled.Q4_1.gguf	Q4_1	3.95GB
llama-2-7b-nf4-fp16-upscaled.Q5_0.gguf	Q5_0	4.33GB
llama-2-7b-nf4-fp16-upscaled.Q5_K_S.gguf	Q5_K_S	4.33GB
llama-2-7b-nf4-fp16-upscaled.Q5_K.gguf	Q5_K	4.45GB
llama-2-7b-nf4-fp16-upscaled.Q5_K_M.gguf	Q5_K_M	4.45GB
llama-2-7b-nf4-fp16-upscaled.Q5_1.gguf	Q5_1	4.72GB
llama-2-7b-nf4-fp16-upscaled.Q6_K.gguf	Q6_K	5.15GB
llama-2-7b-nf4-fp16-upscaled.Q8_0.gguf	Q8_0	6.67GB

Original model description:

license: apache-2.0 tags:

text-generation-inference

This is an upscaled fp16 variant of the original Llama-2-7b base model by Meta after it has been loaded with nf4 4-bit quantization via bitsandbytes. The main idea here is to upscale the linear4bit layers to fp16 so that the quantization/dequantization cost doesn't have to paid for each forward pass at inference time.

Note: The quantization operation to nf4 is not lossless, so the model weights for the linear layers are lossy, which means that this model will not work as well as the official base model.

To use this model, you can just load it via transformers in fp16:

python

import torch
from transformers import AutoModelForCausalLM

model = AutoModelForCausalLM.from_pretrained(
  "arnavgrg/llama-2-7b-nf4-fp16-upscaled",
  device_map="auto",
  torch_dtype=torch.float16,
)

RichardErkhov/arnavgrg_-_llama-2-7b-nf4-fp16-upscaled-gguf

Original model description:

RichardErkhov/
arnavgrg_-_llama-2-7b-nf4-fp16-upscaled-gguf