Skip to content

ALMA 7B R VRAM requirements

How much VRAM ALMA 7B R needs
4-bit weights: 3.8 GBWith 4K context: 6.4 GB8-bit, 4K context: 9.3 GB
Parameters
6.7B
Longest context
4K
Released
Jan 2024
License
MIT
Strongest at
Not rated

How much VRAM ALMA 7B R needs

Memory ALMA 7B R needs at each precision and context length
PrecisionWeightsWith 4K
16-bit12.6 GB15.2 GB
8-bit6.7 GB9.3 GB
6-bit5.1 GB7.8 GB
5-bit4.5 GB7.1 GB
4-bit3.8 GB6.4 GB
3-bit3.1 GB5.7 GB
2-bit2.4 GB5.0 GB
1.6-bit1.3 GB3.9 GB

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

GPUs and Macs that run ALMA 7B R

The best way ALMA 7B R runs on common GPUs, Macs and AI boxes
MachineBest fitContextSpeedFirst replyReport
RTX 40605-bit4K39 tok/s1.0 sPDF: ALMA 7B R on RTX 4060 report
RTX 50605-bit4K63 tok/s0.8 sPDF: ALMA 7B R on RTX 5060 report
RTX 30705-bit4K63 tok/s0.7 sPDF: ALMA 7B R on RTX 3070 report
RTX 3080 10 GB8-bit4K71 tok/s0.5 sPDF: ALMA 7B R on RTX 3080 10 GB report
RTX 3060 12 GB8-bit4K35 tok/s1.1 sPDF: ALMA 7B R on RTX 3060 12 GB report
Arc B5808-bit4K31 tok/s0.8 sPDF: ALMA 7B R on Arc B580 report
RTX 40708-bit4K48 tok/s0.5 sPDF: ALMA 7B R on RTX 4070 report
RTX 4070 Super8-bit4K48 tok/s0.4 sPDF: ALMA 7B R on RTX 4070 Super report
RTX 50708-bit4K63 tok/s0.5 sPDF: ALMA 7B R on RTX 5070 report
RTX 4060 Ti 16 GB8-bit4K28 tok/s0.7 sPDF: ALMA 7B R on RTX 4060 Ti 16 GB report
RTX 5060 Ti 16 GB8-bit4K43 tok/s0.6 sPDF: ALMA 7B R on RTX 5060 Ti 16 GB report
Arc A770 16 GB8-bit4K37 tok/s0.7 sPDF: ALMA 7B R on Arc A770 16 GB report
Radeon RX 9070 XT8-bit4K50 tok/s0.4 sPDF: ALMA 7B R on Radeon RX 9070 XT report
RTX 4070 Ti Super8-bit4K63 tok/s0.3 sPDF: ALMA 7B R on RTX 4070 Ti Super report
RTX 4080 Super8-bit4K69 tok/s0.3 sPDF: ALMA 7B R on RTX 4080 Super report
RTX 5070 Ti8-bit4K82 tok/s0.3 sPDF: ALMA 7B R on RTX 5070 Ti report
RTX 50808-bit4K88 tok/s0.3 sPDF: ALMA 7B R on RTX 5080 report
Radeon RX 7900 XTX8-bit4K73 tok/s0.6 sPDF: ALMA 7B R on Radeon RX 7900 XTX report
RTX 30908-bit4K86 tok/s0.4 sPDF: ALMA 7B R on RTX 3090 report
RTX 40908-bit4K92 tok/s0.2 sPDF: ALMA 7B R on RTX 4090 report
Radeon AI PRO R97008-bit4K50 tok/s0.4 sPDF: ALMA 7B R on Radeon AI PRO R9700 report
RTX 50908-bit4K153 tok/s0.1 sPDF: ALMA 7B R on RTX 5090 report
2× RTX 30908-bit4K86 tok/s0.2 sPDF: ALMA 7B R on 2× RTX 3090 report
2× RTX 40908-bit4K92 tok/s0.1 sPDF: ALMA 7B R on 2× RTX 4090 report
RTX PRO 5000 Blackwell 48 GB8-bit4K119 tok/s0.2 sPDF: ALMA 7B R on RTX PRO 5000 Blackwell 48 GB report
RTX 6000 Ada8-bit4K88 tok/s0.2 sPDF: ALMA 7B R on RTX 6000 Ada report
L40S8-bit4K80 tok/s0.3 sPDF: ALMA 7B R on L40S report
2× RTX 50908-bit4K153 tok/s0.1 sPDF: ALMA 7B R on 2× RTX 5090 report
4× RTX 30908-bit4K86 tok/s0.1 sPDF: ALMA 7B R on 4× RTX 3090 report
RTX PRO 6000 Blackwell8-bit4K153 tok/s0.1 sPDF: ALMA 7B R on RTX PRO 6000 Blackwell report
Mac mini M4 16 GB8-bit4K14 tok/s9.2 sPDF: ALMA 7B R on Mac mini M4 16 GB report
Mac mini M4 24 GB8-bit4K14 tok/s9.2 sPDF: ALMA 7B R on Mac mini M4 24 GB report
Mac mini M4 32 GB8-bit4K14 tok/s9.2 sPDF: ALMA 7B R on Mac mini M4 32 GB report
Mac mini M4 Pro 24 GB8-bit4K29 tok/s4.6 sPDF: ALMA 7B R on Mac mini M4 Pro 24 GB report
Mac mini M4 Pro 48 GB8-bit4K29 tok/s4.6 sPDF: ALMA 7B R on Mac mini M4 Pro 48 GB report
Mac mini M4 Pro 64 GB8-bit4K29 tok/s4.6 sPDF: ALMA 7B R on Mac mini M4 Pro 64 GB report
M4 Max 128 GB8-bit4K51 tok/s2.3 sPDF: ALMA 7B R on M4 Max 128 GB report
M5 Max 128 GB8-bit4K56 tok/s0.7 sPDF: ALMA 7B R on M5 Max 128 GB report
M3 Ultra 96 GB8-bit4K64 tok/s1.3 sPDF: ALMA 7B R on M3 Ultra 96 GB report
M3 Ultra 256 GB8-bit4K64 tok/s1.3 sPDF: ALMA 7B R on M3 Ultra 256 GB report
M3 Ultra 512 GB8-bit4K64 tok/s1.3 sPDF: ALMA 7B R on M3 Ultra 512 GB report
M5 Ultra 512 GB8-bit4K87 tok/s0.4 sPDF: ALMA 7B R on M5 Ultra 512 GB report
Ryzen AI Max+ 395 64 GB8-bit4K21 tok/s1.3 sPDF: ALMA 7B R on Ryzen AI Max+ 395 64 GB report
Ryzen AI Max+ 395 128 GB8-bit4K21 tok/s1.3 sPDF: ALMA 7B R on Ryzen AI Max+ 395 128 GB report
RTX Spark8-bit4K29 tok/s0.5 sPDF: ALMA 7B R on RTX Spark report
DGX Spark8-bit4K27 tok/s0.5 sPDF: ALMA 7B R on DGX Spark report
Dual DGX Spark8-bit4K27 tok/s0.3 sPDF: ALMA 7B R on Dual DGX Spark report
DGX Station GB3008-bit4K421 tok/s0.1 sPDF: ALMA 7B R on DGX Station GB300 report
2× RTX PRO 6000 Blackwell8-bit4K153 tok/s0.1 sPDF: ALMA 7B R on 2× RTX PRO 6000 Blackwell report
4× RTX PRO 6000 Blackwell8-bit4K153 tok/s0.1 sPDF: ALMA 7B R on 4× RTX PRO 6000 Blackwell report
8× RTX PRO 6000 Blackwell8-bit4K153 tok/s0.1 sPDF: ALMA 7B R on 8× RTX PRO 6000 Blackwell report
A100 80 GB SXM8-bit4K170 tok/s0.3 sPDF: ALMA 7B R on A100 80 GB SXM report
H100 SXM8-bit4K253 tok/s0.1 sPDF: ALMA 7B R on H100 SXM report
4× H200 NVL8-bit4K328 tok/s0.1 sPDF: ALMA 7B R on 4× H200 NVL report
8× H200 NVL8-bit4K328 tok/s0.1 sPDF: ALMA 7B R 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 cstr/ALMA-7B-R-GGUF
vLLM
vllm serve haoranxu/ALMA-7B-R

Questions

How much VRAM does ALMA 7B R need?

ALMA 7B R needs 3.8 GB for its weights at 4-bit, 6.4 GB with an 4K context and 6.4 GB with 4K. At 8-bit, 9.3 GB.

What is the minimum VRAM for ALMA 7B R?

About 5.7 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 4060.

How big is ALMA 7B R?

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

Can ALMA 7B R run on an RTX 4090?

Yes. At 8-bit it holds 4K of context and writes about 92 tokens a second.

More models