Qwen2.5 7B Instruct: how much VRAM does it need?

Qwen2.5 7B Instruct has 7.62B parameters in 28 layers. It uses grouped-query attention: 28 query heads share 4 key-value heads, so its KV cache is 7× smaller than full multi-head attention would need. It takes up to 32,768 tokens of context; one user at that length needs 1.75 GiB of KV cache at FP16 on top of the weights.

BF16 · 8k context · one user

16.37 GiB

14.19 GiB of weights, 448 MiB of KV cache and 1.75 GiB of vLLM's overhead. At INT4 the same workload needs 7.45 GiB.

28 layers
hidden 3584
28/4 heads
head_dim 128
32k max context

Qwen2.5 7B Instruct memory by precision and context

Precision	8k context	32k context
BF16	16.37 GiBKV cache 448 MiB	17.69 GiBKV cache 1.75 GiB
FP8	10.30 GiBKV cache 448 MiB	11.61 GiBKV cache 1.75 GiB
INT4	7.45 GiBKV cache 448 MiB	8.76 GiBKV cache 1.75 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

BF16 (16.37 GiB): NVIDIA L4, Apple M4, NVIDIA A30, RTX 3090 / Ti
FP8 (10.30 GiB): RTX 3060, NVIDIA T4, RTX 4060 Ti 16GB, RTX 4070 Ti Super
INT4 (7.45 GiB): RTX 3060, NVIDIA T4, RTX 4060 Ti 16GB, RTX 4070 Ti Super

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

GPU	BF16	FP8	INT4
RTX 4090	13	27	33
RTX 5090	29	43	49
RTX 6000 Ada	63	77	83
H100 SXM 80GB	128	142	149
RTX PRO 6000 Blackwell	161	175	182
AMD Ryzen AI Max+ 395 (128GB)	164	178	184
H200 141GB	256	270	276
B200 180GB	333	347	354

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

Fine-tuning Qwen2.5 7B Instruct

Method	Memory	Smallest GPU
Full fine-tune	114.9 GiB	NVIDIA DGX Spark (GB10)
LoRA (rank 16, attention)	15.70 GiB	NVIDIA L4
QLoRA (rank 16, attention)	5.17 GiB	RTX 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 Qwen2.5 7B Instruct that is 56.0 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.