# PrismML Bonsai 2 27B ternary GGUF. PQ2_0 is recommended on Apple Silicon:
# Prism's Apple Silicon measurements favor its prompt-processing speed; PTQ1_0 is a
# smaller alternative. Use a model-specific path so both files can coexist.
MODEL_REPO=prism-ml/Ternary-Bonsai-2-27B-gguf
MODEL_FILE=Ternary-Bonsai-2-27B-PQ2_0.gguf
MODEL_ALIAS=bonsai-2-27b-pq2-0
MODEL_PATH=models/Ternary-Bonsai-2-27B-PQ2_0.gguf
MODEL_URL=https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf/resolve/main/Ternary-Bonsai-2-27B-PQ2_0.gguf

# Binary built by scripts/bootstrap.sh from PrismML's llama.cpp fork.
LLAMA_CPP_REPO=https://github.com/PrismML-Eng/llama.cpp.git
LLAMA_CPP_REF=prism
LLAMA_SERVER_BIN=vendor/llama.cpp/build/bin/llama-server

# Endpoint. Use 127.0.0.1 for local-only agents. Use 0.0.0.0 for LAN.
HOST=127.0.0.1
PORT=8080

# llama-server flags for Bonsai 2 on Apple Silicon.
CTX_SIZE=262144
GPU_LAYERS=99
PARALLEL=1
JINJA=on

# Optional llama.cpp runtime passthroughs (blank uses llama-server defaults; off disables supported toggles):
# BATCH_SIZE=2048
# UBATCH_SIZE=512
# CACHE_RAM=8192
# CACHE_REUSE=256
# KV_UNIFIED=on
# CACHE_IDLE_SLOTS=on
# MLOCK=on
# MMAP=off
# THREADS=8
# THREADS_BATCH=8
# CPU_MASK=0xff
# CPU_RANGE=0-7
# CPU_STRICT=1
# PRIORITY=1
# POLL=50
# POLL_BATCH=1
FLASH_ATTN=on
CACHE_TYPE_K=f16
CACHE_TYPE_V=f16

# Bonsai 2 has no MTP drafter in this workflow; leave speculative flags blank.
SPEC_TYPE=
SPEC_DRAFT_N_MAX=

# Optional speculative-decoding knobs for alternate model profiles:
# SPEC_DRAFT_N_MIN=
# SPEC_DRAFT_P_MIN=0.75
# SPEC_DRAFT_P_SPLIT=0.10
# SPEC_DRAFT_TYPE_K=
# SPEC_DRAFT_TYPE_V=

# Keep model reasoning enabled by default. Set REASONING=off and pass
# CHAT_TEMPLATE_KWARGS='{"enable_thinking":false}' for concise agent output.
REASONING=auto
CHAT_TEMPLATE_KWARGS=

# Optional auth. Leave blank for local-only endpoint.
API_KEY=

# Runtime files
LOG_FILE=logs/llama-server.log
PID_FILE=.llama-server.pid