Skip to content

Starcoder2 15B VRAM requirements

How much VRAM Starcoder2 15B needs
4-bit weights: 9.0 GBWith 8K context: 10.4 GBWith 16K context: 11.0 GB8-bit, 8K context: 17.1 GB
Parameters
16B
Longest context
16K
Released
Feb 2024
License
bigcode-openrail-m
Strongest at
Not rated

Its licence (bigcode-openrail-m) is not a standard open-source one: check the terms before commercial use.

How much VRAM Starcoder2 15B needs

Memory Starcoder2 15B needs at each precision and context length
PrecisionWeightsWith 8KWith 16K
16-bit29.7 GB31.1 GB31.7 GB
8-bit15.8 GB17.1 GB17.8 GB
6-bit12.2 GB13.5 GB14.2 GB
5-bit10.6 GB11.9 GB12.6 GB
4-bit9.0 GB10.4 GB11.0 GB
3-bit7.3 GB8.6 GB9.2 GB
2-bit5.6 GB6.9 GB7.6 GB
1.6-bit3.1 GB4.4 GB5.0 GB

One conversation, 16-bit context cache, runtime overhead included.

GPUs and Macs that run Starcoder2 15B

The best way Starcoder2 15B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Starcoder2 15B on RTX 4060 report
RTX 5060Too bigPDF: Starcoder2 15B on RTX 5060 report
RTX 3070Too bigPDF: Starcoder2 15B on RTX 3070 report
RTX 3080 10 GB3-bit16K64 tok/s1.1 sPDF: Starcoder2 15B on RTX 3080 10 GB report
RTX 3060 12 GB4-bit16K26 tok/s2.7 sPDF: Starcoder2 15B on RTX 3060 12 GB report
Arc B5804-bit16K23 tok/s1.9 sPDF: Starcoder2 15B on Arc B580 report
RTX 40704-bit16K36 tok/s1.2 sPDF: Starcoder2 15B on RTX 4070 report
RTX 4070 Super4-bit16K36 tok/s1.0 sPDF: Starcoder2 15B on RTX 4070 Super report
RTX 50704-bit16K47 tok/s1.1 sPDF: Starcoder2 15B on RTX 5070 report
RTX 4060 Ti 16 GB6-bit16K15 tok/s1.6 sPDF: Starcoder2 15B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit16K24 tok/s1.4 sPDF: Starcoder2 15B on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit16K21 tok/s1.7 sPDF: Starcoder2 15B on Arc A770 16 GB report
Radeon RX 9070 XT6-bit16K28 tok/s1.0 sPDF: Starcoder2 15B on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit16K35 tok/s0.8 sPDF: Starcoder2 15B on RTX 4070 Ti Super report
RTX 4080 Super6-bit16K38 tok/s0.7 sPDF: Starcoder2 15B on RTX 4080 Super report
RTX 5070 Ti6-bit16K46 tok/s0.8 sPDF: Starcoder2 15B on RTX 5070 Ti report
RTX 50806-bit16K49 tok/s0.6 sPDF: Starcoder2 15B on RTX 5080 report
Radeon RX 7900 XTX8-bit16K32 tok/s1.5 sPDF: Starcoder2 15B on Radeon RX 7900 XTX report
RTX 30908-bit16K38 tok/s1.0 sPDF: Starcoder2 15B on RTX 3090 report
RTX 40908-bit16K40 tok/s0.4 sPDF: Starcoder2 15B on RTX 4090 report
Radeon AI PRO R97008-bit16K22 tok/s1.0 sPDF: Starcoder2 15B on Radeon AI PRO R9700 report
RTX 50908-bit16K69 tok/s0.3 sPDF: Starcoder2 15B on RTX 5090 report
2× RTX 30908-bit16K38 tok/s0.6 sPDF: Starcoder2 15B on 2× RTX 3090 report
2× RTX 40908-bit16K40 tok/s0.2 sPDF: Starcoder2 15B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit16K53 tok/s0.5 sPDF: Starcoder2 15B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit16K39 tok/s0.4 sPDF: Starcoder2 15B on RTX 6000 Ada report
L40S8-bit16K35 tok/s0.6 sPDF: Starcoder2 15B on L40S report
2× RTX 50908-bit16K69 tok/s0.2 sPDF: Starcoder2 15B on 2× RTX 5090 report
4× RTX 30908-bit16K38 tok/s0.3 sPDF: Starcoder2 15B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit16K69 tok/s0.3 sPDF: Starcoder2 15B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB3-bit16K12 tok/s22 sPDF: Starcoder2 15B on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit16K7.5 tok/s22 sPDF: Starcoder2 15B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit16K5.8 tok/s22 sPDF: Starcoder2 15B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit16K17 tok/s11 sPDF: Starcoder2 15B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit16K13 tok/s11 sPDF: Starcoder2 15B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit16K13 tok/s11 sPDF: Starcoder2 15B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit16K23 tok/s5.4 sPDF: Starcoder2 15B on M4 Max 128 GB report
M5 Max 128 GB8-bit16K25 tok/s1.6 sPDF: Starcoder2 15B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit16K29 tok/s3.0 sPDF: Starcoder2 15B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit16K29 tok/s3.0 sPDF: Starcoder2 15B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit16K29 tok/s3.0 sPDF: Starcoder2 15B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit16K41 tok/s1.0 sPDF: Starcoder2 15B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit16K9.1 tok/s3.1 sPDF: Starcoder2 15B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit16K9.1 tok/s3.1 sPDF: Starcoder2 15B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit16K12 tok/s1.1 sPDF: Starcoder2 15B on RTX Spark report
DGX Spark8-bit16K11 tok/s1.1 sPDF: Starcoder2 15B on DGX Spark report
Dual DGX Spark8-bit16K11 tok/s0.6 sPDF: Starcoder2 15B on Dual DGX Spark report
DGX Station GB3008-bit16K220 tok/s0.1 sPDF: Starcoder2 15B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit16K69 tok/s0.1 sPDF: Starcoder2 15B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit16K69 tok/s0.1 sPDF: Starcoder2 15B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit16K69 tok/s0.1 sPDF: Starcoder2 15B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit16K78 tok/s0.7 sPDF: Starcoder2 15B on A100 80 GB SXM report
H100 SXM8-bit16K121 tok/s0.2 sPDF: Starcoder2 15B on H100 SXM report
4× H200 NVL8-bit16K163 tok/s0.1 sPDF: Starcoder2 15B on 4× H200 NVL report
8× H200 NVL8-bit16K163 tok/s0.1 sPDF: Starcoder2 15B on 8× H200 NVL report

Best fit is the highest quality, up to 8-bit, that leaves 16K of context. First reply is the wait for a 2K prompt. Speeds are rough, plus or minus 25%.

How to run it

llama.cpp
llama-server -hf mradermacher/starcoder2-15b-i1-GGUF
vLLM
vllm serve bigcode/starcoder2-15b
MLX (Mac)
mlx_lm.chat --model mlx-community/bigcode-starcoder2-15b-4bit

Questions

How much VRAM does Starcoder2 15B need?

Starcoder2 15B needs 9.0 GB for its weights at 4-bit, 10.4 GB with an 8K context and 11.0 GB with 16K. At 8-bit, 17.1 GB.

What is the minimum VRAM for Starcoder2 15B?

About 8.6 GB at 3-bit with an 8K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 3080 10 GB.

How big is Starcoder2 15B?

16B parameters: 29.7 GB of weights at 16-bit, 9.0 GB at 4-bit.

Can Starcoder2 15B run on an RTX 4090?

Yes. At 8-bit it holds 16K of context and writes about 40 tokens a second.

More models