LL

Meta Llama 3.1 405B Instruct 16KV heads

LlamaGGUF by ThomasBaruzierUpdated 2024-08-25
Download on Hugging Face
Parameters410B
Context128K
Layers126
Architecturellama
Smallest quant86.8 GB
Quick answer

Meta Llama 3.1 405B Instruct 16KV heads needs about 130.1 GB of memory at IQ2_XS (a 121.1 GB file) with an 8K context window, which is more than a 48 GB GPU holds. This repository has 8 quants, from 86.8 GB (IQ1_S) to 820.2 GB (F16). Longer context needs more: at 32K tokens the KV cache alone adds about 33.8 GB.

Meta Llama 3.1 405B Instruct 16KV heads is a 410B-parameter Llama model in GGUF format, quantized by ThomasBaruzier. Its trained context length is 128K tokens. The repository has 8 quant files, from 86.8 GB (IQ1_S) to 820.2 GB (F16).

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
IQ1_S86.8 GB (2 parts) 95.8 GB
Too big Get file →
IQ1_M95.1 GB (2 parts) 104.1 GB
Too big Get file →
IQ2_XXS109.0 GB (3 parts) 118.0 GB
Too big Get file →
IQ2_XS121.1 GB (3 parts) 130.1 GB
Too big Get file →
IQ2_S127.4 GB (3 parts) 136.4 GB
Too big Get file →
IQ2_M138.5 GB (3 parts) 147.5 GB
Too big Get file →
IQ3_XXS157.7 GB (4 parts) 166.7 GB
Too big Get file →
F16820.2 GB (18 parts) 829.1 GB
Too big Get file →

What GPU do you need for Meta Llama 3.1 405B Instruct 16KV heads?

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: no quant fits at an 8K context.

Which Meta Llama 3.1 405B Instruct 16KV heads quant should you choose?

IQ2_XS (121.1 GB) is the middle option in this repository. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

If memory is tight, the smallest file is IQ1_S at 86.8 GB, with the quality trade-offs that come with fewer bits. Use it only when memory is very tight. Expect a clear drop in quality, which is usually worse on small models and easier to live with on large ones.

New to quant names? Read GGUF quantization explained.

Specifications

Architecture from the GGUF header

Architecturellama
Parameters410B
Layers126
Trained context131,072 tokens
KV cache (fp16)1008 KB per token

Source

Quantized byThomasBaruzier
FamilyLlama
Licensellama3.1
Downloads (HF)152
Last updated2024-08-25

Frequently asked questions about Meta Llama 3.1 405B Instruct 16KV heads

How much VRAM does Meta Llama 3.1 405B Instruct 16KV heads need?

About 130.1 GB at IQ2_XS with an 8K context window. That is the 121.1 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 Meta Llama 3.1 405B Instruct 16KV heads on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 95.8 GB, so you would need a larger GPU or CPU offloading.

What is the file size of Meta Llama 3.1 405B Instruct 16KV heads GGUF?

It depends on the quant: IQ1_S is 86.8 GB, IQ1_M is 95.1 GB, IQ2_XXS is 109.0 GB, IQ2_XS is 121.1 GB, IQ2_S is 127.4 GB, IQ2_M is 138.5 GB and 2 more.

What context length does Meta Llama 3.1 405B Instruct 16KV heads support?

The GGUF header lists a trained context length of 131,072 tokens (128K). The KV cache grows with the context you use: about 8.5 GB at 8K and 33.8 GB at 32K.

Where can I download Meta Llama 3.1 405B Instruct 16KV heads GGUF?

The files are on Hugging Face in the ThomasBaruzier/Meta-Llama-3.1-405B-Instruct-16KV-heads-GGUF repository. The download buttons on this page open the file on Hugging Face. ggufmodels.com does not host any model files.

What license does Meta Llama 3.1 405B Instruct 16KV heads 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.

Similar models

Browse all →

More from ThomasBaruzier

All by ThomasBaruzier →

More Llama models

All Llama →

Other large models

All large models →