##
## vLLM environment file
##
## Injected into every vLLM container launched by llmsnap (via --env-file in
## config.yaml). Installed with 0600 root:root permissions as it holds a secret.
##

# Hugging Face token, used to download gated / private model weights.
#HF_TOKEN=hf_replace_me

# Optional: speed up weight downloads (requires hf_transfer in the image).
#HF_HUB_ENABLE_HF_TRANSFER=1

# Keep vLLM's INTERNAL inter-worker traffic off the LAN. With Network=host, a
# multi-worker instance (tensor/pipeline parallel) otherwise exposes a fan of
# `VLLM::Worker` listeners on the primary routable IP. Two independent knobs are
# needed because vLLM binds these ports through two different backends, and ONLY
# the second one is what you actually see in `netstat`:
#
#   1. VLLM_HOST_IP -> vLLM's own ZMQ path (get_ip(), MessageQueue). In a
#      single-node run its remote socket is not even bound (local readers use
#      IPC), so this alone has NO visible effect — but it is the documented knob
#      and correct to pin. https://docs.vllm.ai/en/stable/configuration/env_vars/
#
#   2. GLOO_SOCKET_IFNAME / NCCL_SOCKET_IFNAME -> the torch distributed process
#      group (gloo on CPU, nccl on GPU). vLLM builds ProcessGroupGloo WITHOUT an
#      interface (distributed/utils.py: init_gloo_process_group), so gloo falls
#      back to resolving the hostname and binds the default-route interface (the
#      LAN IP). These are torch/NCCL env vars, NOT routed through VLLM_HOST_IP —
#      this is THE fix for the `VLLM::Worker` listeners. `lo` = loopback iface.
#
# All valid ONLY because this is a SINGLE-NODE deployment (every worker on this
# host, loopback reachable). Do NOT set these on a real multi-node cluster.
# Distinct from the API server bind, which is --host/--port in
# vllm-model@.container (COOKBOOK_VLLM_MODEL_PORT).
VLLM_HOST_IP=127.0.0.1
GLOO_SOCKET_IFNAME=lo
NCCL_SOCKET_IFNAME=lo

# VLLM container image
#
# CPU inference image: the OFFICIAL vLLM CI build (multi-arch amd64/arm64), used
# by the shipped config.yaml / model Quadlets to validate the whole chain WITHOUT
# a GPU. Its ENTRYPOINT is `vllm serve`, so the model Quadlets only append flags.
#
# For a real GPU deployment, swap this for the CUDA image, whose tag is
# architecture-specific (`docker.io/vllm/vllm-openai:v0.24.0` on amd64,
# `:v0.24.0-aarch64` on arm64) — Podman cannot template it at generation time.
# See SPECS.md for the production (GPU) reference.
COOKBOOK_VLLM_IMAGE=public.ecr.aws/q9t5s3a7/vllm-cpu-release-repo:v0.24.0