Llama 3.1 8B VRAM and GPU Requirements

Which GPUs fit

VRAM by quantization

Totals at 8,192 tokens. Q4_K_M is the usual balance of size and quality. FP16 weights are the parameter count times 2 bytes. The FP16 column in the table is the full estimate: weights, cache, and 1 GB.

Context length

Q4_K_M total as the context changes. Cache is the part that grows.

• 2,048 5.46 GB
• 8,192 6.21 GB
• 32,768 9.21 GB
• 131,072 21.21 GB

Llama 3.1 8B hardware notes

• Unified memory counts 75 percent of the listed pool as usable. Discrete cards use the full listed memory. The memory column still shows the listed figure.
• Last updated 2024-09-25.
• Weights, cache, and the 1 GB overhead are estimates. How this is calculated.
• Architecture numbers come from unsloth/Llama-3.1-8B-Instruct because the official download is gated. The parameter count is the official Hugging Face record.

Llama 3.1 8B VRAM FAQ

How much VRAM does Llama 3.1 8B need at Q4_K_M?

6.21 GB at 8,192 tokens. Weights are 4.21 GB, the cache is 1.00 GB, and overhead is 1 GB.

Does a 24 GB GPU fit Llama 3.1 8B at Q4_K_M?

GeForce RTX 3090 24GB is Fits, 17.79 GB left. The quant table uses 8,192 tokens, the smaller of 8,192 and the configured maximum of 131,072.

What context length is configured for Llama 3.1 8B?

The configured context length is 131,072 tokens. The quant table uses 8,192 tokens, the smaller of 8,192 and the configured maximum of 131,072.

Which source supplies the parameter count for Llama 3.1 8B?

8.03B parameters, from the Hugging Face model record for meta-llama/Llama-3.1-8B-Instruct. Architecture numbers come from the public copy at unsloth/Llama-3.1-8B-Instruct, because the official download is gated. The parameter count is the official record.

Related models