starcoder2-15b needs about 14.3 GB of memory at Q6_K (a 13.1 GB file) with an 8K context window, so it fits on a 16 GB GPU. This repository has 5 quants, from 9.9 GB (Q4_K) to 31.9 GB (F16).
starcoder2-15b is a 16B-parameter StarCoder model in GGUF format, quantized by dranger003. Its trained context length is 16K tokens. The repository has 5 quant files, from 9.9 GB (Q4_K) to 31.9 GB (F16).
Quants & memory
| Quant | File | Est. memory | vs VRAM | Result | |
|---|---|---|---|---|---|
| Q4_K | 9.9 GB | 11.0 GB | Too big | Get file → | |
| Q5_K | 11.4 GB | 12.6 GB | Too big | Get file → | |
| Q6_K | 13.1 GB | 14.3 GB | Too big | Get file → | |
| Q8_0 | 17.0 GB | 18.1 GB | Too big | Get file → | |
| F16 | 31.9 GB | 33.1 GB | Too big | Get file → |
What GPU do you need for starcoder2-15b?
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 Q4_K (9.9 GB file, about 11.0 GB in use).
- 16 GB VRAM: the best quant that fits is Q6_K (13.1 GB file, about 14.3 GB in use).
- 24 GB VRAM: the best quant that fits is Q8_0 (17.0 GB file, about 18.1 GB in use).
- 32 GB VRAM: the best quant that fits is Q8_0 (17.0 GB file, about 18.1 GB in use).
- 48 GB VRAM: the best quant that fits is F16 (31.9 GB file, about 33.1 GB in use).
Which starcoder2-15b quant should you choose?
Q6_K (13.1 GB) is the middle option in this repository. Very close to the original quality for most uses, at a clearly larger size than 4-bit.
If memory is tight, the smallest file is Q4_K at 9.9 GB, with the quality trade-offs that come with fewer bits. The usual sweet spot between file size and quality, and the quant most people start with.
New to quant names? Read GGUF quantization explained.
Specifications
Architecture from the GGUF header
Source
Frequently asked questions about starcoder2-15b
How much VRAM does starcoder2-15b need?
About 14.3 GB at Q6_K with an 8K context window. That is the 13.1 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 starcoder2-15b on an 8 GB GPU?
Not at an 8K context. The smallest file needs about 11.0 GB, so you would need a larger GPU or CPU offloading.
What is the file size of starcoder2-15b GGUF?
It depends on the quant: Q4_K is 9.9 GB, Q5_K is 11.4 GB, Q6_K is 13.1 GB, Q8_0 is 17.0 GB, F16 is 31.9 GB.
What context length does starcoder2-15b support?
The GGUF header lists a trained context length of 16,384 tokens (16K). The KV cache grows with the context you use: about 0.7 GB at 8K and 1.3 GB at 16K.
Where can I download starcoder2-15b GGUF?
The files are on Hugging Face in the dranger003/starcoder2-15b-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 starcoder2-15b use?
Hugging Face lists the license as bigcode-openrail-m. Licenses can differ between the quantized files and the original model, so check the repository and the base model before commercial use.