.env.example

# vllm.cpp developer environment.
#
# Copy this file to `.env` (untracked) and fill in what your setup actually
# has. The protocol docs (AGENTS.md, .agents/*.md) reference values as
# `${VLLM_SOURCE}`-style placeholders; this file is where those values come
# from. Scripts read them from the process environment (e.g.
# scripts/upstream-inventory.py), so load it with
#   set -a; . ./.env; set +a
# or your direnv equivalent.
#
# Leave a variable empty when your setup does not have the thing. Empty means
# UNAVAILABLE: the gates that need it stay PENDING for you, and nothing in the
# protocol may substitute another developer's paths. Do not put credentials or
# other secrets in this file.
#
# When GATE_HOST is set, the oracle/gate-build variables (VLLM_ORACLE,
# DEPENDENCY_SOURCE, CUTLASS_DIR, GPU_LOCK, DEVICE_*) are paths on that host,
# since that is where gate commands run. Without a GATE_HOST they are local
# paths. Per-box quirks (PATH prepends, services to stop, lock discipline)
# live in the matching profile in .agents/environment.md.

# --- Upstream reference checkouts (read-only source mirrors) ---
# The pinned vLLM checkout used for 1:1 porting and every parity citation.
VLLM_SOURCE=
# Secondary references.
SGLANG_SOURCE=
LLAMACPP_SOURCE=

# --- Model checkpoints ---
# Where model weights are meant to live. Point it at shared or network storage
# (a NAS mount, for example) so a large checkpoint is fetched ONCE and reused by
# every host, worktree and agent instead of filling each box's system disk — a
# 30B bf16 checkpoint is ~60 GB, and gate boxes run out of disk long before they
# run out of anything else.
#
# This declares an INTENT, not a behaviour. Nothing in the tree reads
# CHECKPOINT_ROOT today: it does not redirect a download and it does not resolve
# a bare directory name on its own. It exists so whoever fetches a checkpoint —
# a person, a script, an agent — puts it in the agreed place and passes the full
# path onward. Empty means UNAVAILABLE. Nothing here authorizes a download;
# large asset fetches still need explicit authority.
#
# Keep a checkpoint's provenance with it: fetch by an explicit revision rather
# than a bare branch name, since publishers do re-quantize in place under an
# unchanged repo name.
CHECKPOINT_ROOT=

# --- Oracle ---
# The pinned vLLM oracle used for correctness and performance gates: a venv
# root (or python executable) that can run teacher-forcing and `vllm bench`.
VLLM_ORACLE=
# Installed dependency source for kernel provenance: the site-packages root
# holding flashinfer/, triton/, ... (usually inside the oracle venv).
DEPENDENCY_SOURCE=

# --- Gate environment ---
# SSH target of the hardware-gate box your developer profile enables (see
# .agents/environment.md). Empty = no gate hardware: hardware gates stay
# PENDING and you run honest local diagnostics only.
GATE_HOST=
# CUTLASS checkout passed as -DVLLM_CPP_CUTLASS_DIR on CUDA gate builds.
CUTLASS_DIR=

# --- Device ---
# Whole-series GPU exclusion: a file mutex every GPU job must hold when the
# box is shared. Run GPU work as `flock $GPU_LOCK -c '<command>'`, or take the
# lock once around a whole benchmark series so arms are never interleaved.
GPU_LOCK=
# Build target architecture(s) for your accelerator, vendor naming:
#   CUDA: 121a (sm_121a), 110, 87 ...    ROCm: gfx942, gfx1100 ...
DEVICE_ARCH=
# Device toolchain root and compiler, when not already on PATH.
#   CUDA: /usr/local/cuda + nvcc         ROCm: /opt/rocm + hipcc
DEVICE_TOOLKIT_ROOT=
DEVICE_COMPILER=