.env.example

# imp Docker configuration
# Copy to .env and adjust as needed.

# Model to load on startup (path inside container)
IMP_MODEL=/models/Qwen3-8B-Q8_0.gguf

# Server port (default: 8080)
# IMP_PORT=8080

# OpenWebUI port (default: 3000)
# WEBUI_PORT=3000

# CUDA architectures for build (default: 90a;100;120a)
# For RTX 5090 only:
# CUDA_ARCHITECTURES=120a

# Models directory on host (default: ./models)
# IMP_MODELS_DIR=./models

# --- Optional engine tuning (translated to imp-server flags by
# --- docker-entrypoint.sh; leave unset for defaults)

# Default max tokens per generation
# IMP_MAX_TOKENS=

# Chat template: auto|none|chatml|llama2|llama3|nemotron|gemma|deepseek_r1|phi
# IMP_CHAT_TEMPLATE=auto

# KV cache quantization (1/true to enable)
# IMP_KV_FP8=
# IMP_KV_INT8=

# NVFP4 decode cache: 1 = on, 2 = decode-only (less VRAM), 0 = disable
# IMP_DECODE_NVFP4=

# Disable CUDA Graphs (1/true; diagnostic)
# IMP_NO_CUDA_GRAPHS=

# FP16 SSM state for recurrent models (1/true)
# IMP_SSM_FP16=

# Vision projector .gguf for multimodal models
# IMP_MMPROJ=

# Max tokens per prefill chunk (0 = unchunked)
# IMP_PREFILL_CHUNK_SIZE=

# Fraction of max_tokens reserved for <think> reasoning (0 = disabled)
# IMP_THINK_BUDGET=

# Monitoring stack (Prometheus + Grafana)
# PROMETHEUS_PORT=9090
# GRAFANA_PORT=3002
# GRAFANA_USER=admin
# GRAFANA_PASSWORD=admin