Question 1

How much VRAM does starcoder2-15b need?

Accepted Answer

starcoder2-15b (16B) requires approximately 35.668661499023436 GB of VRAM at full precision (FP16), 17.8 GB at 8-bit quantization (Q8), and 8.9 GB at 4-bit quantization (Q4). The actual VRAM usage may be slightly higher due to KV cache and runtime overhead.

Question 2

What GPUs can run starcoder2-15b?

Accepted Answer

We have 10 compatible GPUs indexed for starcoder2-15b, including A40, A100 40GB, RTX A6000, A100 80GB PCIe, A100 80GB and 5 more. Compatibility is determined by whether the GPU's VRAM can fit the model's memory footprint, potentially using quantization to reduce memory usage.

Question 3

What is the cheapest GPU to run starcoder2-15b?

Accepted Answer

The cheapest GPU to run starcoder2-15b in the cloud is the A40 at $0.49/hr. Other affordable options include A100 40GB ($0.93/hr) and RTX A6000 ($1.09/hr). Prices reflect the lowest on-demand rate across cloud providers.

Question 4

Can starcoder2-15b run on consumer hardware?

Accepted Answer

Yes, starcoder2-15b can run on consumer GPUs like the RTX A6000 (48 GB VRAM) at full precision (FP16).

Question 5

What quantization options are available for starcoder2-15b?

Accepted Answer

starcoder2-15b supports the following quantization levels: FP16 (full precision): ~35.668661499023436 GB, Q8 (8-bit): ~17.8 GB, Q6 (6-bit): ~13.4 GB, Q4 (4-bit): ~8.9 GB. Lower quantization reduces VRAM usage at the cost of some quality loss. Q8 offers a good balance between quality and memory savings, while Q4 provides maximum compression.

GPU	VRAM	Best Quant	From	Rent
A40	48 GB	FP16	$0.49/hr	Rent
A100 40GB	40 GB	Q8	$0.93/hr	Rent
RTX A6000	48 GB	FP16	$1.09/hr
A100 80GB PCIe	80 GB	FP16	$1.59/hr	Rent
A100 80GB	80 GB	FP16	$1.59/hr	Rent
L40	48 GB	FP16	$1.64/hr	Rent
H100 MEGA	80 GB	FP16	$1.80/hr	Rent
H100 SXM	80 GB	FP16	$1.80/hr	Rent
H100 NVL	94 GB	FP16	$1.80/hr	Rent
H100 PCIe	80 GB	FP16	$1.80/hr	Rent

starcoder2-15b

VRAM Requirements

Compatible GPUs

starcoder2-15b FAQ

VRAM requirements for starcoder2-15b at different quantization levels
Quantization	VRAM Required
FP16	35.7 GB
Q8	17.8 GB
Q6	13.4 GB
Q4	8.9 GB