LLMScalelive Hugging Face data
ENTR

Llama 3.2 3B Instruct: how much VRAM does it need?

Llama 3.2 3B Instruct has 3.21B parameters in 28 layers. It uses grouped-query attention: 24 query heads share 8 key-value heads, so its KV cache is 3× smaller than full multi-head attention would need. It takes up to 131,072 tokens of context; one user at that length needs 14.00 GiB of KV cache at FP16 on top of the weights. Its repository is gated on Hugging Face, so LLMScale ships its architecture in a bundled database.

BF16 · 8k context · one user
8.28 GiB

5.98 GiB of weights, 896 MiB of KV cache and 1.42 GiB of vLLM's overhead. At INT4 the same workload needs 4.42 GiB.

  • 28 layers
  • hidden 3072
  • 24/8 heads
  • head_dim 128
  • 128k max context

Llama 3.2 3B Instruct memory by precision and context

Precision8k context32k context128k context
BF168.28 GiBKV cache 896 MiB10.90 GiBKV cache 3.50 GiB21.40 GiBKV cache 14.00 GiB
FP85.65 GiBKV cache 896 MiB8.28 GiBKV cache 3.50 GiB18.78 GiBKV cache 14.00 GiB
INT44.42 GiBKV cache 896 MiB7.05 GiBKV cache 3.50 GiB17.55 GiBKV cache 14.00 GiB

One user, FP16 KV cache and vLLM's overhead on one GPU at its defaults for a card under 70 GB. Quantizing shrinks the weights, not the KV cache, so the saving narrows as the context grows.

Which single GPUs hold it

At 8k context for one user, the smallest cards first. A discrete card is budgeted at vLLM's default 92% of its memory, a unified-memory device at the share its GPU can address, and each carries vLLM's overhead at its own defaults.

How many users fit

GPUBF16FP8INT4
RTX 4090161920
RTX 5090242728
RTX 6000 Ada414445
H100 SXM 80GB747778
RTX PRO 6000 Blackwell909395
AMD Ryzen AI Max+ 395 (128GB)929596
H200 141GB138141142
B200 180GB177180182

Users at 8k context each, on one card. ✗: the weights leave no room for even one.

Fine-tuning Llama 3.2 3B Instruct

MethodMemorySmallest GPU
Full fine-tune49.21 GiBNVIDIA Jetson AGX Orin (64GB)
LoRA (rank 16, attention)7.43 GiBRTX 3060
QLoRA (rank 16, attention)2.99 GiBRTX 3060

Sequence length 2,048, batch size 1, gradient checkpointing on, Adam. Full fine-tuning keeps 16 bytes per parameter before activations; QLoRA stores the frozen base in 4-bit NF4.

Where the numbers come from

Weights are parameters × bytes per parameter. The KV cache is 2 × layers × key-value heads × head dimension × bytes × context × users. For Llama 3.2 3B Instruct that is 112.0 KiB per token at FP16. These are capacity-planning figures: the serving engine — vLLM, TGI, llama.cpp — and its paged-attention efficiency still move the real number.

More on KV cache size and on GQA, MQA and MLA.

Work with Llama 3.2 3B Instruct

Other Llama models