requirements-gpu.txt

# GPU / vLLM inference — OPTIONAL extra (#10251)
#
# NOT installed by the default backend image. vLLM provides high-performance
# in-process LLM inference (3-4x throughput, prefix caching) but requires a
# CUDA-compatible GPU and pulls the full CUDA toolkit (cuda-toolkit, nvidia-*,
# flashinfer, CUDA-built torch) — several GB. Installing it into the default
# CPU image was the cause of image bloat and intermittent CI disk-full failures.
#
# Install ONLY on GPU deployments that enable vLLM (config: llm.vllm.enabled = true):
#     pip install -r requirements-gpu.txt
#
# The vLLM provider (llm_shared/providers/vllm.py, vllm_base.py) lazy-/guard-imports
# vllm, so the backend runs normally without this file; the provider only activates
# when explicitly enabled and installed.
#
# Requires: CUDA-compatible GPU, Python 3.12.

vllm>=0.26.0

# vLLM pins these exact torch builds (CUDA). They intentionally override the
# CPU torch installed in the default image when this GPU extra is used.
torch==2.13.0
torchvision==0.28.0