# Copy to .env and edit. Every value has the same default in compose.yaml.

# Where the checkpoint lives on the host (scripts/download-model.sh fills it).
MODELS_DIR=./models
# Directory under MODELS_DIR that holds the checkpoint.
MODEL_NAME=dgemma
# FlashInfer JIT and torch compile cache; keep it between rebuilds.
CACHE_DIR=./cache

# Served canvas in tokens. A read only pays for its own width, so wide is
# cheap; the width bounds the answer template and a thought's block size.
CANVAS=256
# Generation block width by load: [low, high, width] over running plus
# waiting requests. 256 is best single-stream, 64 at 32 concurrent. No width
# may exceed CANVAS; empty uses CANVAS for every block.
CANVAS_SCHEDULE=[[1, 2, 256], [3, 6, 128], [7, 32, 64]]
MAX_SEQS=32
MAX_MODEL_LEN=4096
# Fraction of the box's memory vLLM may plan for. 0.40 leaves room on a
# 121 GB Spark for the sampler's transients; do not raise it next to another
# model.
GPU_UTIL=0.40
ATTN=TRITON_ATTN
PORT=8010
STRUCTURED_PORT=8011
# Nonzero adds an HTTPS listener on this port with a self-signed certificate
# (kept in the cache dir). Browsers only open a webcam on https or localhost.
TLS_PORT=0
# Appended to vllm serve. --async-scheduling is always on.
EXTRA_ARGS=
# The perf branches' server side: reads over the labels only, and a fixed
# sample count as one request. MAX_SAMPLES caps samples in both.
CONSTRAINED=1
ENGINE_SAMPLES=1
MAX_SAMPLES=32
# gemma4 strips the empty thought block from plain chat and parses tool
# calls. Empty turns either off.
REASONING_PARSER=gemma4
TOOL_CALL_PARSER=gemma4
# KV pool in GiB. 2 covers a few 4k prompts; see the 128k profile below.
KV_CACHE_GB=2
# Empty keeps vLLM's default prefill chunk.
MAX_NUM_BATCHED_TOKENS=
# The entrypoint refuses to start with less than this free after weights and
# the start-up transient.
HEADROOM_GB=12
# fp32 [MAX_SEQS x CANVAS, vocab] copies the start-up budget allows for.
TRANSIENT_COPIES=2
# Empty = no cap (upstream behaviour). 0.85 caps each worker at that fraction
# of device memory, so an overshoot fails the request instead of the host.
TORCH_MEM_FRACTION=
# 1 serves a playground page at http://<box>:8011/ (JSON request, image
# file or webcam frames), and /walk and /cube. Off by default.
TEST_PAGE=
# When set, POST routes on the structured port need "Authorization: Bearer <key>".
# GET /health and the playground page stay open; the page has a field for the key.
API_KEY=

# 128k-context profile for long documents. Uncomment together. A 128k
# request takes about 1.7 GiB of KV on this model, so 24 GiB holds thirteen
# at once; see README for what a long state costs in time.
#MAX_MODEL_LEN=131072
#KV_CACHE_GB=24
#GPU_UTIL=0.45