README.md

metadata

base_model: Qwen/Qwen2-72B-Instruct
inference: false
language:
  - en
library_name: gguf
license: other
license_link: https://huggingface.co/Qwen/Qwen2-72B-Instruct/blob/main/LICENSE
license_name: tongyi-qianwen
pipeline_tag: text-generation
quantized_by: legraphista
tags:
  - chat
  - quantized
  - GGUF
  - imatrix
  - quantization
  - imat
  - imatrix
  - static
  - 16bit
  - 8bit
  - 6bit
  - 5bit
  - 4bit
  - 3bit
  - 2bit
  - 1bit

Qwen2-72B-Instruct-IMat-GGUF

Llama.cpp imatrix quantization of Qwen/Qwen2-72B-Instruct

Original Model: Qwen/Qwen2-72B-Instruct
Original dtype: BF16 (bfloat16)
Quantized by: llama.cpp b3091
IMatrix dataset: here

• Files
  • IMatrix
  • Common Quants
  • All Quants
• Downloading using huggingface-cli
• Inference
  • Simple chat template
  • Chat template with system prompt
  • Llama.cpp
• FAQ
  • Why is the IMatrix not applied everywhere?
  • How do I merge a split GGUF?

---

Files

IMatrix

Status: ✅ Available
Link: here

Common Quants

Filename	Quant type	File Size	Status	Uses IMatrix	Is Split
Qwen2-72B-Instruct.Q8_0/*	Q8_0	77.26GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q6_K/*	Q6_K	64.35GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q4_K/*	Q4_K	47.42GB	✅ Available	🟢 IMatrix	✂ Yes
Qwen2-72B-Instruct.Q3_K.gguf	Q3_K	37.70GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.Q2_K.gguf	Q2_K	29.81GB	✅ Available	🟢 IMatrix	📦 No

All Quants

Filename	Quant type	File Size	Status	Uses IMatrix	Is Split
Qwen2-72B-Instruct.BF16/*	BF16	145.42GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.FP16/*	F16	145.42GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q8_0/*	Q8_0	77.26GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q6_K/*	Q6_K	64.35GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q5_K/*	Q5_K	54.45GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q5_K_S/*	Q5_K_S	51.38GB	✅ Available	⚪ Static	✂ Yes
Qwen2-72B-Instruct.Q4_K/*	Q4_K	47.42GB	✅ Available	🟢 IMatrix	✂ Yes
Qwen2-72B-Instruct.Q4_K_S.gguf	Q4_K_S	43.89GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ4_NL.gguf	IQ4_NL	41.32GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ4_XS.gguf	IQ4_XS	39.71GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.Q3_K.gguf	Q3_K	37.70GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.Q3_K_L.gguf	Q3_K_L	39.51GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.Q3_K_S.gguf	Q3_K_S	34.49GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ3_M.gguf	IQ3_M	35.50GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ3_S.gguf	IQ3_S	34.49GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ3_XS.gguf	IQ3_XS	32.84GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ3_XXS.gguf	IQ3_XXS	31.85GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.Q2_K.gguf	Q2_K	29.81GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.Q2_K_S.gguf	Q2_K_S	29.57GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ2_M.gguf	IQ2_M	29.34GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ2_S.gguf	IQ2_S	27.94GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ2_XS.gguf	IQ2_XS	27.06GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ2_XXS.gguf	IQ2_XXS	25.49GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ1_M.gguf	IQ1_M	23.74GB	✅ Available	🟢 IMatrix	📦 No
Qwen2-72B-Instruct.IQ1_S.gguf	IQ1_S	22.69GB	✅ Available	🟢 IMatrix	📦 No

Downloading using huggingface-cli

If you do not have hugginface-cli installed:

pip install -U "huggingface_hub[cli]"

Download the specific file you want:

huggingface-cli download legraphista/Qwen2-72B-Instruct-IMat-GGUF --include "Qwen2-72B-Instruct.Q8_0.gguf" --local-dir ./

If the model file is big, it has been split into multiple files. In order to download them all to a local folder, run:

huggingface-cli download legraphista/Qwen2-72B-Instruct-IMat-GGUF --include "Qwen2-72B-Instruct.Q8_0/*" --local-dir ./
# see FAQ for merging GGUF's

---

Inference

Simple chat template

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
{user_prompt}<|im_end|>
<|im_start|>assistant
{assistant_response}<|im_end|>
<|im_start|>user
{next_user_prompt}<|im_end|>

Chat template with system prompt

<|im_start|>system
{system_prompt}<|im_end|>
<|im_start|>user
{user_prompt}<|im_end|>
<|im_start|>assistant
{assistant_response}<|im_end|>
<|im_start|>user
{next_user_prompt}<|im_end|>

Llama.cpp

llama.cpp/main -m Qwen2-72B-Instruct.Q8_0.gguf --color -i -p "prompt here (according to the chat template)"

---

FAQ

Why is the IMatrix not applied everywhere?

According to this investigation, it appears that lower quantizations are the only ones that benefit from the imatrix input (as per hellaswag results).

How do I merge a split GGUF?

1. Make sure you have gguf-split available
   • To get hold of gguf-split, navigate to https://github.com/ggerganov/llama.cpp/releases
   • Download the appropriate zip for your system from the latest release
   • Unzip the archive and you should be able to find gguf-split
2. Locate your GGUF chunks folder (ex: Qwen2-72B-Instruct.Q8_0)
3. Run gguf-split --merge Qwen2-72B-Instruct.Q8_0/Qwen2-72B-Instruct.Q8_0-00001-of-XXXXX.gguf Qwen2-72B-Instruct.Q8_0.gguf
   • Make sure to point gguf-split to the first chunk of the split.

---

Got a suggestion? Ping me @legraphista!