LLMScalelive Hugging Face data
ENTR

Gemma 3 4B IT: how much VRAM does it need?

Gemma 3 4B IT has 4.30B parameters in 34 layers. It uses grouped-query attention: 8 query heads share 4 key-value heads, so its KV cache is 2× smaller than full multi-head attention would need. It takes up to 131,072 tokens of context; one user at that length needs 3.09 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
10.57 GiB

8.01 GiB of weights, 762 MiB of KV cache and 1.82 GiB of vLLM's overhead. At INT4 the same workload needs 5.61 GiB.

  • 34 layers
  • hidden 2560
  • 8/4 heads
  • head_dim 256
  • 128k max context

Gemma 3 4B IT memory by precision and context

Precision8k context32k context128k context
BF1610.57 GiBKV cache 762 MiB11.04 GiBKV cache 1.21 GiB12.91 GiBKV cache 3.09 GiB
FP87.19 GiBKV cache 762 MiB7.66 GiBKV cache 1.21 GiB9.53 GiBKV cache 3.09 GiB
INT45.61 GiBKV cache 762 MiB6.07 GiBKV cache 1.21 GiB7.95 GiBKV cache 3.09 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 4090162022
RTX 5090253032
RTX 6000 Ada455052
H100 SXM 80GB545758
RTX PRO 6000 Blackwell677071
AMD Ryzen AI Max+ 395 (128GB)687172
H200 141GB104108109
B200 180GB133136137

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

Fine-tuning Gemma 3 4B IT

MethodMemorySmallest GPU
Full fine-tune65.41 GiBRTX PRO 5000 Blackwell (72GB)
LoRA (rank 16, attention)9.47 GiBRTX 3060
QLoRA (rank 16, attention)3.52 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 Gemma 3 4B IT that is 136.0 KiB per token at FP16. 29 of its 34 layers attend over a sliding window of the last 1,024 tokens and hold no more than that, so past it each further token adds 20.0 KiB, the other 5 layers' share. 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 Gemma 3 4B IT

Other Gemma models