EA

EastAsia 4x7B Moe experiment

EastasiaGGUF by mradermacherUpdated 2024-12-02
Download on Hugging Face
Parameters18.5B
Context32K
Layers32
Architecturellama
Smallest quant6.8 GB
Quick answer

EastAsia 4x7B Moe experiment needs about 12.8 GB of memory at Q4_K_M (a 11.2 GB file) with an 8K context window, so it fits on a 16 GB GPU. This repository has 11 quants, from 6.8 GB (Q2_K) to 19.7 GB (Q8_0). Longer context needs more: at 32K tokens the KV cache alone adds about 4.3 GB.

EastAsia 4x7B Moe experiment is a 18.5B-parameter Eastasia model in GGUF format, quantized by mradermacher. Its trained context length is 32K tokens. The repository has 11 quant files, from 6.8 GB (Q2_K) to 19.7 GB (Q8_0).

Quants & memory

QuantFileEst. memoryvs VRAMResultDownload
Q2_K6.8 GB 8.4 GB
Too big Get file →
Q3_K_S8.0 GB 9.6 GB
Too big Get file →
Q3_K_M8.9 GB 10.5 GB
Too big Get file →
Q3_K_L9.6 GB 11.2 GB
Too big Get file →
IQ4_XS10.0 GB 11.6 GB
Too big Get file →
Q4_K_S10.5 GB 12.1 GB
Too big Get file →
Q4_K_M11.2 GB 12.8 GB
Too big Get file →
Q5_K_S12.8 GB 14.3 GB
Too big Get file →
Q5_K_M13.1 GB 14.7 GB
Too big Get file →
Q6_K15.2 GB 16.8 GB
Too big Get file →
Q8_019.7 GB 21.3 GB
Too big Get file →

What GPU do you need for EastAsia 4x7B Moe experiment?

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: the best quant that fits is IQ4_XS (10.0 GB file, about 11.6 GB in use).
  • 16 GB VRAM: the best quant that fits is Q5_K_M (13.1 GB file, about 14.7 GB in use).
  • 24 GB VRAM: the best quant that fits is Q8_0 (19.7 GB file, about 21.3 GB in use).

Which EastAsia 4x7B Moe experiment quant should you choose?

Q4_K_M is the usual starting point for EastAsia 4x7B Moe experiment: it balances file size and quality well. Here it is a 11.2 GB file.

The Q8_0 file is 19.7 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 6.8 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
Parameters18.5B (approximate, mixture of experts)
Layers32
Trained context32,768 tokens
KV cache (fp16)128 KB per token

Source

Quantized bymradermacher
FamilyEastasia
Licenseapache-2.0
Downloads (HF)136
Last updated2024-12-02

Frequently asked questions about EastAsia 4x7B Moe experiment

How much VRAM does EastAsia 4x7B Moe experiment need?

About 12.8 GB at Q4_K_M with an 8K context window. That is the 11.2 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 EastAsia 4x7B Moe experiment on an 8 GB GPU?

Not at an 8K context. The smallest file needs about 8.4 GB, so you would need a larger GPU or CPU offloading.

What is the file size of EastAsia 4x7B Moe experiment GGUF?

It depends on the quant: Q2_K is 6.8 GB, Q3_K_S is 8.0 GB, Q3_K_M is 8.9 GB, Q3_K_L is 9.6 GB, IQ4_XS is 10.0 GB, Q4_K_S is 10.5 GB and 5 more.

What context length does EastAsia 4x7B Moe experiment support?

The GGUF header lists a trained context length of 32,768 tokens (32K). The KV cache grows with the context you use: about 1.1 GB at 8K and 4.3 GB at 32K.

Where can I download EastAsia 4x7B Moe experiment GGUF?

The files are on Hugging Face in the mradermacher/EastAsia-4x7B-Moe-experiment-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 EastAsia 4x7B Moe experiment use?

Hugging Face lists the license as apache-2.0. 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 EastAsia-4x7B-Moe-experiment

Similar models

Browse all →

More from mradermacher

All by mradermacher →

More Eastasia models

All Eastasia →

Other medium models

All medium models →