MI

causallm_8x7b-moe-test-not-mixtral

MistralGGUF by TheBlokeUpdated 2023-12-17
Download on Hugging Face
Parameters38B
Context32K
Layers32
Architecturellama
Smallest quant13.6 GB
Quick answer

causallm_8x7b-moe-test-not-mixtral needs about 26.9 GB of memory at Q4_K_M (a 22.1 GB file) with an 8K context window, so it fits on a 32 GB GPU. This repository has 8 quants, from 13.6 GB (Q2_K) to 40.4 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 17.2 GB.

causallm_8x7b-moe-test-not-mixtral is a 38B-parameter Mistral model in GGUF format, quantized by TheBloke. Its trained context length is 32K tokens. The repository has 8 quant files, from 13.6 GB (Q2_K) to 40.4 GB (Q8_0).

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q2_K13.6 GB 18.4 GB
Too big Get file →
Q3_K_M17.3 GB 22.1 GB
Too big Get file →
Q4_022.1 GB 26.9 GB
Too big Get file →
Q4_K_M22.1 GB 26.9 GB
Too big Get file →
Q5_026.6 GB 31.4 GB
Too big Get file →
Q5_K_M26.6 GB 31.4 GB
Too big Get file →
Q6_K31.5 GB 36.3 GB
Too big Get file →
Q8_040.4 GB 45.2 GB
Too big Get file →

What GPU do you need for causallm_8x7b-moe-test-not-mixtral?

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: the best quant that fits is Q3_K_M (17.3 GB file, about 22.1 GB in use).
  • 32 GB VRAM: the best quant that fits is Q5_0 (26.6 GB file, about 31.4 GB in use).
  • 48 GB VRAM: the best quant that fits is Q8_0 (40.4 GB file, about 45.2 GB in use).

Which causallm_8x7b-moe-test-not-mixtral quant should you choose?

Q4_K_M is the usual starting point for causallm_8x7b-moe-test-not-mixtral: it balances file size and quality well. Here it is a 22.1 GB file.

The Q8_0 file is 40.4 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 Q2_K at 13.6 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
Parameters38B (approximate, mixture of experts)
Layers32
Trained context32,768 tokens
KV cache (fp16)512 KB per token

Source

Quantized byTheBloke
FamilyMistral
Licensegpl-3.0
Downloads (HF)485
Last updated2023-12-17

Frequently asked questions about causallm_8x7b-moe-test-not-mixtral

How much VRAM does causallm_8x7b-moe-test-not-mixtral need?

About 26.9 GB at Q4_K_M with an 8K context window. That is the 22.1 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 causallm_8x7b-moe-test-not-mixtral on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 18.4 GB, so you would need a larger GPU or CPU offloading.

What is the file size of causallm_8x7b-moe-test-not-mixtral GGUF?

It depends on the quant: Q2_K is 13.6 GB, Q3_K_M is 17.3 GB, Q4_0 is 22.1 GB, Q4_K_M is 22.1 GB, Q5_0 is 26.6 GB, Q5_K_M is 26.6 GB and 2 more.

What context length does causallm_8x7b-moe-test-not-mixtral support?

The GGUF header lists a trained context length of 32,768 tokens (32K). The KV cache grows with the context you use: about 4.3 GB at 8K and 17.2 GB at 32K.

Where can I download causallm_8x7b-moe-test-not-mixtral GGUF?

The files are on Hugging Face in the TheBloke/8x7B-MoE-test-NOT-MIXTRAL-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 causallm_8x7b-moe-test-not-mixtral use?

Hugging Face lists the license as gpl-3.0. 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 TheBloke

All by TheBloke →

More Mistral models

All Mistral →

Other large models

All large models →