Skip to content

Llava 1.5 13B HF VRAM requirements

How much VRAM Llava 1.5 13B HF needs
4-bit weights: 7.5 GBWith 4K context: 11.3 GB8-bit, 4K context: 17.0 GB
Parameters
13.4B
Longest context
4K
Released
Dec 2023
License
Llama 2 Community
Strongest at
Text and images

Its licence (Llama 2 Community) is not a standard open-source one: check the terms before commercial use.

How much VRAM Llava 1.5 13B HF needs

Memory Llava 1.5 13B HF needs at each precision and context length
PrecisionWeightsWith 4K
16-bit24.9 GB28.7 GB
8-bit13.2 GB17.0 GB
6-bit10.2 GB14.0 GB
5-bit8.8 GB12.6 GB
4-bit7.5 GB11.3 GB
3-bit6.1 GB9.9 GB
2-bit4.7 GB8.5 GB
1.6-bit2.6 GB6.4 GB

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

GPUs and Macs that run Llava 1.5 13B HF

The best way Llava 1.5 13B HF runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 4060Too bigPDF: Llava 1.5 13B HF on RTX 4060 report
RTX 5060Too bigPDF: Llava 1.5 13B HF on RTX 5060 report
RTX 3070Too bigPDF: Llava 1.5 13B HF on RTX 3070 report
RTX 3080 10 GBToo bigPDF: Llava 1.5 13B HF on RTX 3080 10 GB report
RTX 3060 12 GB4-bit4K30 tok/s2.3 sPDF: Llava 1.5 13B HF on RTX 3060 12 GB report
Arc B5804-bit4K27 tok/s1.6 sPDF: Llava 1.5 13B HF on Arc B580 report
RTX 40704-bit4K42 tok/s1.0 sPDF: Llava 1.5 13B HF on RTX 4070 report
RTX 4070 Super4-bit4K42 tok/s0.8 sPDF: Llava 1.5 13B HF on RTX 4070 Super report
RTX 50704-bit4K55 tok/s0.9 sPDF: Llava 1.5 13B HF on RTX 5070 report
RTX 4060 Ti 16 GB6-bit4K18 tok/s1.3 sPDF: Llava 1.5 13B HF on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB6-bit4K28 tok/s1.2 sPDF: Llava 1.5 13B HF on RTX 5060 Ti 16 GB report
Arc A770 16 GB6-bit4K24 tok/s1.4 sPDF: Llava 1.5 13B HF on Arc A770 16 GB report
Radeon RX 9070 XT6-bit4K33 tok/s0.8 sPDF: Llava 1.5 13B HF on Radeon RX 9070 XT report
RTX 4070 Ti Super6-bit4K41 tok/s0.7 sPDF: Llava 1.5 13B HF on RTX 4070 Ti Super report
RTX 4080 Super6-bit4K45 tok/s0.6 sPDF: Llava 1.5 13B HF on RTX 4080 Super report
RTX 5070 Ti6-bit4K54 tok/s0.7 sPDF: Llava 1.5 13B HF on RTX 5070 Ti report
RTX 50806-bit4K58 tok/s0.5 sPDF: Llava 1.5 13B HF on RTX 5080 report
Radeon RX 7900 XTX8-bit4K38 tok/s1.3 sPDF: Llava 1.5 13B HF on Radeon RX 7900 XTX report
RTX 30908-bit4K44 tok/s0.8 sPDF: Llava 1.5 13B HF on RTX 3090 report
RTX 40908-bit4K47 tok/s0.4 sPDF: Llava 1.5 13B HF on RTX 4090 report
Radeon AI PRO R97008-bit4K26 tok/s0.8 sPDF: Llava 1.5 13B HF on Radeon AI PRO R9700 report
RTX 50908-bit4K81 tok/s0.3 sPDF: Llava 1.5 13B HF on RTX 5090 report
2× RTX 30908-bit4K44 tok/s0.5 sPDF: Llava 1.5 13B HF on 2× RTX 3090 report
2× RTX 40908-bit4K47 tok/s0.2 sPDF: Llava 1.5 13B HF on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K62 tok/s0.4 sPDF: Llava 1.5 13B HF on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K45 tok/s0.3 sPDF: Llava 1.5 13B HF on RTX 6000 Ada report
L40S8-bit4K41 tok/s0.5 sPDF: Llava 1.5 13B HF on L40S report
2× RTX 50908-bit4K81 tok/s0.2 sPDF: Llava 1.5 13B HF on 2× RTX 5090 report
4× RTX 30908-bit4K44 tok/s0.3 sPDF: Llava 1.5 13B HF on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K81 tok/s0.2 sPDF: Llava 1.5 13B HF on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB3-bit4K15 tok/s18 sPDF: Llava 1.5 13B HF on Mac mini M4 16 GB report
Mac mini M4 24 GB6-bit4K8.8 tok/s18 sPDF: Llava 1.5 13B HF on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit4K6.9 tok/s18 sPDF: Llava 1.5 13B HF on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB6-bit4K19 tok/s9.2 sPDF: Llava 1.5 13B HF on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit4K15 tok/s9.2 sPDF: Llava 1.5 13B HF on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K15 tok/s9.2 sPDF: Llava 1.5 13B HF on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K27 tok/s4.6 sPDF: Llava 1.5 13B HF on M4 Max 128 GB report
M5 Max 128 GB8-bit4K30 tok/s1.3 sPDF: Llava 1.5 13B HF on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K34 tok/s2.6 sPDF: Llava 1.5 13B HF on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K34 tok/s2.6 sPDF: Llava 1.5 13B HF on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K34 tok/s2.6 sPDF: Llava 1.5 13B HF on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K47 tok/s0.8 sPDF: Llava 1.5 13B HF on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K11 tok/s2.6 sPDF: Llava 1.5 13B HF on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K11 tok/s2.6 sPDF: Llava 1.5 13B HF on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K15 tok/s0.9 sPDF: Llava 1.5 13B HF on RTX Spark report
DGX Spark8-bit4K13 tok/s0.9 sPDF: Llava 1.5 13B HF on DGX Spark report
Dual DGX Spark8-bit4K13 tok/s0.5 sPDF: Llava 1.5 13B HF on Dual DGX Spark report
DGX Station GB3008-bit4K249 tok/s0.1 sPDF: Llava 1.5 13B HF on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K81 tok/s0.1 sPDF: Llava 1.5 13B HF on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K81 tok/s0.1 sPDF: Llava 1.5 13B HF on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K81 tok/s0.1 sPDF: Llava 1.5 13B HF on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K91 tok/s0.6 sPDF: Llava 1.5 13B HF on A100 80 GB SXM report
H100 SXM8-bit4K139 tok/s0.2 sPDF: Llava 1.5 13B HF on H100 SXM report
4× H200 NVL8-bit4K186 tok/s0.1 sPDF: Llava 1.5 13B HF on 4× H200 NVL report
8× H200 NVL8-bit4K186 tok/s0.1 sPDF: Llava 1.5 13B HF 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/llava-1.5-13b-hf-i1-GGUF
vLLM
vllm serve llava-hf/llava-1.5-13b-hf

Questions

How much VRAM does Llava 1.5 13B HF need?

Llava 1.5 13B HF needs 7.5 GB for its weights at 4-bit, 11.3 GB with an 4K context and 11.3 GB with 4K. At 8-bit, 17.0 GB.

What is the minimum VRAM for Llava 1.5 13B HF?

About 9.9 GB at 3-bit with an 4K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 3060 12 GB.

How big is Llava 1.5 13B HF?

13.4B parameters: 24.9 GB of weights at 16-bit, 7.5 GB at 4-bit.

Can Llava 1.5 13B HF run on an RTX 4090?

Yes. At 8-bit it holds 4K of context and writes about 47 tokens a second.

More models