GPU & CUDA Configuration

Note: GPU inference is feature-gated behind cuda and real features. Real GPU benchmarks are deferred (no real-GPU run as of 2.0.9, ADR-067). Structural CPU benchmarks pass.

Building with CUDA Support

# Install CUDA toolkit (requires CUDA 12.x)
# Then build with CUDA features
cargo build --release --features "cuda,real"

# Or with the bench harness
cargo build --release -p justapi-bench --features cuda

Running on GPU

justapi serve --model my-model.gguf --gpu

GPU Memory Management

The inference engine uses PagedAttention for efficient GPU memory utilization:

justapi serve --model my-model.gguf --gpu \
  --pool-blocks 8192 \        # Total KV-cache blocks
  --max-seqs 64               # Maximum concurrent sequences

Block Pool Configuration

Flag	Default	Description
`--pool-blocks`	`4096`	Total blocks in the KV-cache pool
`--block-size`	`16`	Tokens per block
`--max-seqs`	`32`	Maximum concurrent sequences

Memory Calculation

Total KV-cache memory = pool_blocks * block_size * num_layers * 2 * dtype_size

For a 7B model with 32 layers, 16 block size, 4096 blocks:

4096 * 16 * 32 * 2 * 2 bytes (FP16) = 8 GB

Multi-GPU

Multi-GPU support is in development. Currently, models are loaded on a single GPU.

Quantization

JustAPI supports several quantization formats through GGUF:

Format	Bits	Description
`FP16`	16	Full precision
`Q4_0`	4	4-bit quants, no grouping
`Q4_K_M`	4	4-bit with medium grouping
`Q5_K_M`	5	5-bit with medium grouping
`Q8_0`	8	8-bit quants

justapi serve --model model-q4_k_m.gguf --gpu

LoRA Adapters

justapi serve --model base-model.gguf --gpu \
  --lora adapter.safetensors \
  --lora-scale 0.8

Multiple LoRAs can be loaded and swapped at runtime via the control plane API.