Skip to content

VRAM reference

Phi-4 VRAM requirements

Fine-tuning Phi-4 takes about 13 GB of GPU memory with QLoRA, 33 GB with LoRA on 16-bit weights and 238 GB for full fine-tuning, at 2,048 tokens per sequence. Running it takes 31 GB in 16-bit or 9.9 GB in 4-bit with a 8K-token context.

By Will Schulz · Architecture and prices checked 15 November 2026 · Prices as of November 2026

Fine-tuning memory

At 2,048 tokens per sequence, batch 1, LoRA rank 16, gradient checkpointing on
MethodWeightsTraining stateActivationsTotalSmallest single GPU80 GB GPUs needed
QLoRA (4-bit base)9.1 GB1.0 GB3.0 GB13 GBRTX 3090 (24 GB)1
LoRA (16-bit base)29 GB1.0 GB3.0 GB33 GBA100 40GB (40 GB)1
Full fine-tuning29 GB205 GB3.0 GB238 GBNone4

On a 24 GB card such as the RTX 4090: QLoRA: yes, about 13 GB at 2K tokens, batch 1, rank 16; up to 4K-token sequences at batch 1 or batch 3 at 2K, LoRA rank up to 128.

Inference memory

One sequence, fp16 KV cache
PrecisionWeightsKV cache at 8KTotalAt 16K contextSmallest single GPU
16-bit (BF16/FP16)29 GB1.7 GB31 GB33 GBA100 40GB (40 GB)
8-bit15 GB1.7 GB16 GB18 GBRTX 3090 (24 GB)
4-bit8.2 GB1.7 GB9.9 GB12 GBRTX 3090 (24 GB)

Which GPUs fit Phi-4

One GPU unless a count is shown. Follow a GPU's name for its maximum sequence length, batch size, LoRA rank and rental cost.

GPUQLoRALoRAFull fine-tuningRun, fp16Run, int8Run, int4
RTX 3090 24 GBYes2×11×2×YesYes
RTX 4090 24 GBYes2×11×2×YesYes
RTX 5090 32 GBYes2×9×2×YesYes
L40S 48 GBYesYes6×YesYesYes
A100 40GB 40 GBYesYes7×YesYesYes
A100 80GB 80 GBYesYes4×YesYesYes
H100 SXM 80 GBYesYes4×YesYesYes
H200 141 GBYesYes2×YesYesYes
B200 180 GBYesYes2×YesYesYes

What's specific to Phi-4

  • A 16K-token context limit, so the KV cache never gets large: about 3.4 GB at the full 16K.
  • At 14.7B parameters it sits in the same memory class as Qwen3 14B, under an MIT license.
ArchitecturePhi-4
Parameters14.7B
Layers40
Hidden size5,120
MLP width17,920
Attention heads / KV heads40 / 10 (head dim 128)
Vocabulary100,352 tokens
Context16K tokens
LicenseMIT
ReleasedDecember 2024

Try other settings

Pre-filled for Phi-4. Change the sequence length, batch, LoRA rank, context or GPU.

JobMemoryRTX 4090 (24 GB)Smallest single GPU
QLoRA (4-bit base)2K tokens × batch 1, rank 1613 GBFitsRTX 3090 (24 GB)
LoRA (16-bit base)2K tokens × batch 1, rank 1633 GB2 GPUsA100 40GB (40 GB)
Full fine-tuning2K tokens × batch 1238 GB11 GPUsNone: needs several
Inference, 16-bit (BF16/FP16)8K-token context × 1 sequence31 GB2 GPUsA100 40GB (40 GB)
Inference, 8-bit8K-token context × 1 sequence16 GBFitsRTX 3090 (24 GB)
Inference, 4-bit8K-token context × 1 sequence9.9 GBFitsRTX 3090 (24 GB)

A job fits when the estimate is within 90% of the card's memory. Training assumes gradient checkpointing and FlashAttention; multi-GPU counts assume the memory splits evenly (FSDP, ZeRO-3 or tensor parallelism).

How we estimate it

Every figure comes from the model's published architecture (layers, hidden size, attention and KV heads, vocabulary) rather than a rule of thumb, using the same functions as the calculator above. Sizes are in GB (10⁹ bytes).

  • Full fine-tuning: 16 bytes per parameter for mixed-precision Adam (bf16 weights and gradients, fp32 master weights and two moments), as in the ZeRO paper.
  • LoRA: frozen bf16 weights (2 bytes per parameter) plus 16 bytes per adapter parameter, rank 16 on every linear layer. For mixture-of-experts models, adapters go on attention only and the experts stay frozen.
  • QLoRA: NF4 weights with double quantisation (4.127 bits per parameter); embeddings and the output head stay in bf16, as bitsandbytes leaves them.
  • Activations: gradient checkpointing and FlashAttention. Each layer's bf16 input is kept, one layer is recomputed at a time, and the output logits take 8 bytes per vocabulary entry per token (fp32 values and gradients). Chunked cross-entropy, as in Unsloth or Liger kernels, cuts that last term.
  • Inference: 2, 1 or about 0.56 bytes per weight (4-bit with group-wise scales), plus an fp16 KV cache of 2 × layers × KV heads × head dim × 2 bytes per token; sliding-window layers only cache their window.
  • Fits means the estimate is within 90% of the card's listed memory; the rest covers the CUDA context, kernels and fragmentation. Multi-GPU counts assume memory splits evenly.

Frequently asked questions

Can I fine-tune Phi-4 on a 24 GB GPU?
On an RTX 3090 or 4090 (24 GB): QLoRA: yes, about 13 GB at 2K tokens, batch 1, rank 16; up to 4K-token sequences at batch 1 or batch 3 at 2K, LoRA rank up to 128. LoRA: no, it needs about 33 GB at 2K tokens, more than the 24 GB available. Estimates use gradient checkpointing and leave 10% of memory free.
How much VRAM do you need to run Phi-4?
About 31 GB in 16-bit and 9.9 GB in 4-bit, at 8K tokens of context. The smallest single GPU here for 4-bit is the RTX 3090 (24 GB); for 16-bit, the A100 40GB (40 GB).
How many GPUs does full fine-tuning Phi-4 take?
About 238 GB with mixed-precision Adam at 2K tokens, so at least 4 × 80 GB H100s with the training state sharded (FSDP or ZeRO-3). LoRA needs 33 GB and QLoRA 13 GB.
How long a context can Phi-4 handle on an 80 GB GPU?
Its KV cache costs 3.4 GB at the full 16K tokens (fp16 cache). On one 80 GB GPU: 16K tokens with 16-bit weights, 16K tokens with 4-bit weights.

Terms used on this page: LoRA, QLoRA, KV cache, gradient checkpointing, quantization. For any model size, use the fine-tuning VRAM calculator; for the method in depth, the Academy lesson Memory first: what hardware can train what.

Sources

  1. [1]Hugging Face. microsoft/phi-4 config.json
  2. [2]Microsoft. microsoft/phi-4 model card
  3. [3]Microsoft. Phi-4 Technical Report (arXiv:2412.08905)
  4. [4]Dettmers et al., 2023. QLoRA: Efficient Finetuning of Quantized LLMs (arXiv:2305.14314)
  5. [5]Hu et al., 2021. LoRA: Low-Rank Adaptation of Large Language Models (arXiv:2106.09685)
  6. [6]Rajbhandari et al., 2019. ZeRO: Memory Optimizations Toward Training Trillion Parameter Models (arXiv:1910.02054)
  7. [7]Korthikanti et al., 2022. Reducing Activation Recomputation in Large Transformer Models (arXiv:2205.05198)
  8. [8]Chen et al., 2016. Training Deep Nets with Sublinear Memory Cost (arXiv:1604.06174)

Architecture read from config.json and parameter count from the Hugging Face safetensors metadata on 15 November 2026.

Find out if you need to fine-tune at all.

Kojen's prompt tuning and model search run for real on your own test cases, on hosted models. Measure the cheaper levers before you rent a GPU.

Start training free