Skip to content

Llama 7B VRAM requirements

How much VRAM Llama 7B needs
4-bit weights: 3.8 GBWith 2K context: 5.4 GB8-bit, 2K context: 8.3 GB
Parameters
6.7B
Longest context
2K
Released
Apr 2023
License
Custom
Strongest at
Not rated

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

How much VRAM Llama 7B needs

Memory Llama 7B needs at each precision and context length
PrecisionWeightsWith 2K
16-bit12.6 GB14.2 GB
8-bit6.7 GB8.3 GB
6-bit5.1 GB6.8 GB
5-bit4.5 GB6.1 GB
4-bit3.8 GB5.4 GB
3-bit3.1 GB4.7 GB
2-bit2.4 GB4.0 GB
1.6-bit1.3 GB2.9 GB

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

GPUs and Macs that run Llama 7B

The best way Llama 7B runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40606-bit2K34 tok/s1.0 sPDF: Llama 7B on RTX 4060 report
RTX 50606-bit2K55 tok/s0.8 sPDF: Llama 7B on RTX 5060 report
RTX 30706-bit2K55 tok/s0.7 sPDF: Llama 7B on RTX 3070 report
RTX 3080 10 GB8-bit2K71 tok/s0.5 sPDF: Llama 7B on RTX 3080 10 GB report
RTX 3060 12 GB8-bit2K35 tok/s1.1 sPDF: Llama 7B on RTX 3060 12 GB report
Arc B5808-bit2K31 tok/s0.8 sPDF: Llama 7B on Arc B580 report
RTX 40708-bit2K48 tok/s0.5 sPDF: Llama 7B on RTX 4070 report
RTX 4070 Super8-bit2K48 tok/s0.4 sPDF: Llama 7B on RTX 4070 Super report
RTX 50708-bit2K63 tok/s0.5 sPDF: Llama 7B on RTX 5070 report
RTX 4060 Ti 16 GB8-bit2K28 tok/s0.7 sPDF: Llama 7B on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit2K43 tok/s0.6 sPDF: Llama 7B on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit2K37 tok/s0.7 sPDF: Llama 7B on Arc A770 16 GB report
Radeon RX 9070 XT8-bit2K50 tok/s0.4 sPDF: Llama 7B on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit2K63 tok/s0.3 sPDF: Llama 7B on RTX 4070 Ti Super report
RTX 4080 Super8-bit2K69 tok/s0.3 sPDF: Llama 7B on RTX 4080 Super report
RTX 5070 Ti8-bit2K82 tok/s0.3 sPDF: Llama 7B on RTX 5070 Ti report
RTX 50808-bit2K88 tok/s0.3 sPDF: Llama 7B on RTX 5080 report
Radeon RX 7900 XTX8-bit2K73 tok/s0.6 sPDF: Llama 7B on Radeon RX 7900 XTX report
RTX 30908-bit2K86 tok/s0.4 sPDF: Llama 7B on RTX 3090 report
RTX 40908-bit2K92 tok/s0.2 sPDF: Llama 7B on RTX 4090 report
Radeon AI PRO R97008-bit2K50 tok/s0.4 sPDF: Llama 7B on Radeon AI PRO R9700 report
RTX 50908-bit2K153 tok/s0.1 sPDF: Llama 7B on RTX 5090 report
2× RTX 30908-bit2K86 tok/s0.2 sPDF: Llama 7B on 2× RTX 3090 report
2× RTX 40908-bit2K92 tok/s0.1 sPDF: Llama 7B on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit2K119 tok/s0.2 sPDF: Llama 7B on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit2K88 tok/s0.2 sPDF: Llama 7B on RTX 6000 Ada report
L40S8-bit2K80 tok/s0.3 sPDF: Llama 7B on L40S report
2× RTX 50908-bit2K153 tok/s0.1 sPDF: Llama 7B on 2× RTX 5090 report
4× RTX 30908-bit2K86 tok/s0.1 sPDF: Llama 7B on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit2K153 tok/s0.1 sPDF: Llama 7B on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit2K14 tok/s9.2 sPDF: Llama 7B on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit2K14 tok/s9.2 sPDF: Llama 7B on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit2K14 tok/s9.2 sPDF: Llama 7B on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit2K29 tok/s4.6 sPDF: Llama 7B on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit2K29 tok/s4.6 sPDF: Llama 7B on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit2K29 tok/s4.6 sPDF: Llama 7B on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit2K51 tok/s2.3 sPDF: Llama 7B on M4 Max 128 GB report
M5 Max 128 GB8-bit2K56 tok/s0.7 sPDF: Llama 7B on M5 Max 128 GB report
M3 Ultra 96 GB8-bit2K64 tok/s1.3 sPDF: Llama 7B on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit2K64 tok/s1.3 sPDF: Llama 7B on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit2K64 tok/s1.3 sPDF: Llama 7B on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit2K87 tok/s0.4 sPDF: Llama 7B on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit2K21 tok/s1.3 sPDF: Llama 7B on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit2K21 tok/s1.3 sPDF: Llama 7B on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit2K29 tok/s0.5 sPDF: Llama 7B on RTX Spark report
DGX Spark8-bit2K27 tok/s0.5 sPDF: Llama 7B on DGX Spark report
Dual DGX Spark8-bit2K27 tok/s0.3 sPDF: Llama 7B on Dual DGX Spark report
DGX Station GB3008-bit2K421 tok/s0.1 sPDF: Llama 7B on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit2K153 tok/s0.1 sPDF: Llama 7B on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit2K153 tok/s0.1 sPDF: Llama 7B on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit2K153 tok/s0.1 sPDF: Llama 7B on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit2K170 tok/s0.3 sPDF: Llama 7B on A100 80 GB SXM report
H100 SXM8-bit2K253 tok/s0.1 sPDF: Llama 7B on H100 SXM report
4× H200 NVL8-bit2K328 tok/s0.1 sPDF: Llama 7B on 4× H200 NVL report
8× H200 NVL8-bit2K328 tok/s0.1 sPDF: Llama 7B 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 tensorblock/llama-7b-GGUF
vLLM
vllm serve huggyllama/llama-7b

Questions

How much VRAM does Llama 7B need?

Llama 7B needs 3.8 GB for its weights at 4-bit, 5.4 GB with an 2K context and 5.4 GB with 2K. At 8-bit, 8.3 GB.

What is the minimum VRAM for Llama 7B?

About 4.7 GB at 3-bit with an 2K context, the lowest precision we recommend for a model this size. The smallest machine here that runs it without spilling is the RTX 4060.

How big is Llama 7B?

6.7B parameters: 12.6 GB of weights at 16-bit, 3.8 GB at 4-bit.

Can Llama 7B run on an RTX 4090?

Yes. At 8-bit it holds 2K of context and writes about 92 tokens a second.

More models