LL

Llama3.1-70B-Chinese-Chat

LlamaGGUF by shenzhi-wangUpdated 2024-07-29
Download on Hugging Face
Parameters70.6B
Context128K
Layers80
Architecturellama
Smallest quant34.3 GB
Quick answer

Llama3.1-70B-Chinese-Chat needs about 45.7 GB of memory at Q4_K_M (a 42.5 GB file) with an 8K context window, so it fits on a 48 GB GPU. This repository has 4 quants, from 34.3 GB (Q3_K_M) to 141.1 GB. Longer context needs more: at 32K tokens the KV cache alone adds about 10.7 GB.

Llama3.1-70B-Chinese-Chat is a 70.6B-parameter Llama model in GGUF format, quantized by shenzhi-wang. Its trained context length is 128K tokens. The repository has 4 quant files, from 34.3 GB (Q3_K_M) to 141.1 GB.

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q3_K_M34.3 GB 37.5 GB
Too big Get file →
Q4_K_M42.5 GB 45.7 GB
Too big Get file →
Q8_075.0 GB (6 parts) 78.2 GB
Too big Get file →
—141.1 GB (4 parts) 144.3 GB
Too big Get file →

What GPU do you need for Llama3.1-70B-Chinese-Chat?

These figures add the file size, an fp16 KV cache at an 8K context window (worked out from this model’s own header) and about 0.5 GB for the runtime. They are estimates, so leave a little headroom, and expect to use less context or a smaller quant if you are right at the limit.

  • 4 GB VRAM: no quant fits at an 8K context.
  • 6 GB VRAM: no quant fits at an 8K context.
  • 8 GB VRAM: no quant fits at an 8K context.
  • 12 GB VRAM: no quant fits at an 8K context.
  • 16 GB VRAM: no quant fits at an 8K context.
  • 24 GB VRAM: no quant fits at an 8K context.
  • 32 GB VRAM: no quant fits at an 8K context.
  • 48 GB VRAM: the best quant that fits is Q4_K_M (42.5 GB file, about 45.7 GB in use).

Which Llama3.1-70B-Chinese-Chat quant should you choose?

Q4_K_M is the usual starting point for Llama3.1-70B-Chinese-Chat: it balances file size and quality well. Here it is a 42.5 GB file.

The Q8_0 file is 75.0 GB, about 1.8 times the size of Q4_K_M. It keeps the weights closer to the original, which only matters if you have the memory to spare.

If memory is tight, the smallest file is Q3_K_M at 34.3 GB, with the quality trade-offs that come with fewer bits. A reasonable choice when a 4-bit file will not fit, mostly on larger models. Quality loss is noticeable but often acceptable.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturellama
Parameters70.6B
Layers80
Trained context131,072 tokens
KV cache (fp16)320 KB per token

Source

Quantized byshenzhi-wang
FamilyLlama
Licensellama3.1
Downloads (HF)564
Last updated2024-07-29

Frequently asked questions about Llama3.1-70B-Chinese-Chat

How much VRAM does Llama3.1-70B-Chinese-Chat need?

About 45.7 GB at Q4_K_M with an 8K context window. That is the 42.5 GB file, plus the KV cache and roughly 0.5 GB of runtime overhead. Smaller quants need less and longer context needs more.

Can I run Llama3.1-70B-Chinese-Chat on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 37.5 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Llama3.1-70B-Chinese-Chat GGUF?

It depends on the quant: Q3_K_M is 34.3 GB, Q4_K_M is 42.5 GB, Q8_0 is 75.0 GB, file is 141.1 GB.

What context length does Llama3.1-70B-Chinese-Chat support?

The GGUF header lists a trained context length of 131,072 tokens (128K). The KV cache grows with the context you use: about 2.7 GB at 8K and 10.7 GB at 32K.

Where can I download Llama3.1-70B-Chinese-Chat GGUF?

The files are on Hugging Face in the shenzhi-wang/Llama3.1-70B-Chinese-Chat repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Llama3.1-70B-Chinese-Chat use?

Hugging Face lists the license as llama3.1. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.

Other GGUFs of Llama-3.1-70B-Instruct

Similar models

Browse all →

More from shenzhi-wang

All by shenzhi-wang →

More Llama models

All Llama →

Other large models

All large models →