CPU / CUDA / MPS: one device resolver, memory-derived batches; docker-compose.cuda.yml for NVIDIA GPUs (branch distribution @ 75d5d80)
63ed286 verified | # The same trader on an NVIDIA GPU. Layer it over the base file: | |
| # | |
| # docker compose -f docker-compose.yml -f docker-compose.cuda.yml up -d --build | |
| # | |
| # Host prerequisites: an NVIDIA driver new enough for CUDA 12.6 (>= 560) and the NVIDIA Container Toolkit | |
| # (`nvidia-ctk runtime configure --runtime=docker`, then restart Docker). The CUDA runtime itself ships inside the torch | |
| # wheel, so no CUDA base image is needed. DEVICE=auto picks the GPU when the container sees one; DEVICE_MEMORY_GB caps | |
| # the batch sizes derived from its free memory when the card is shared. | |
| services: | |
| fly: | |
| build: | |
| args: | |
| TORCH_INDEX: cu126 # or cu130 for the newest driver | |
| environment: | |
| DEVICE: auto | |
| deploy: | |
| resources: | |
| reservations: | |
| devices: | |
| - driver: nvidia | |
| count: all | |
| capabilities: [gpu] | |