.env.example

# diffusers-workflow environment knobs.
#
# Copy to .env and uncomment what you need. dw loads .env from the directory
# a process starts in (load_dotenv() in dw/__init__.py), before diffusers or
# torch is imported, so every line here reaches dw.serve, dw.run and each
# spawned worker. A variable already set in the environment wins over this file.
# Everything below is commented out, so copying this file as is changes nothing.
# Workers read .env when they start, so restart dw.serve (or its idle workers)
# after editing it.
#
# Never commit .env: it is git-ignored because tokens go in it.

# --- dw: server and clients --------------------------------------------------

# Bearer token for dw.serve. Set, every API, UI and MCP request must carry it.
# dw.run and the stdio MCP server read the same variable, so one value
# configures both ends. Required before binding to anything but localhost.
# DW_API_TOKEN=some-long-random-string

# This server's public origin, for links handed to someone not at the box
# (docs/REMOTE.md). Overrides the 'public_url' setting.
# DW_PUBLIC_URL=http://lem:8765

# Where dw.run and the stdio MCP server find dw.serve.
# DW_MCP_URL=http://127.0.0.1:8765

# The workspace the stdio MCP session starts in - a name on the server.
# DW_MCP_WORKSPACE=default

# The cap on one wait_for_job call, in seconds (default 55). Raise it only if
# your MCP client holds a tool call open longer than that.
# DW_MCP_MAX_WAIT_SECONDS=55

# --- dw: engine ----------------------------------------------------------------

# The device for a CLI run: cuda, cuda:1, mps or cpu (default: detected).
# dw.serve picks cards with --devices instead.
# DW_DEVICE=cuda

# The workspace directory: workflows, assets, outputs, prompts
# (default ~/diffusers-workspace; dw.serve --workspace wins over it).
# DW_WORKSPACE=~/diffusers-workspace

# 'flat' restores the pre-run-directory output layout for scripts that glob
# the output folder (default: one directory per run).
# DW_OUTPUT_LAYOUT=flat

# Where settings.json, the job database and the log live
# (default ~/.diffusers_helper).
# DIFFUSERS_HELPER_ROOT=~/.diffusers_helper

# --- Defaults dw sets itself (shown so you know they exist) -----------------
# Each is set at import only when absent; set it here to override.

# Lets the CUDA allocator grow segments instead of fragmenting fixed ones.
# PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True

# Loads sharded checkpoints on several threads (default true; false on macOS,
# where it segfaulted an SDNQ load on MPS).
# HF_ENABLE_PARALLEL_LOADING=true

# MPS only: let PyTorch use all unified memory.
# PYTORCH_MPS_HIGH_WATERMARK_RATIO=0.0

# torch.compile's cache, and with it SDNQ's Triton autotune results - the
# first-step stall of a new size or frame count is paid once per box while
# this survives reboots (default $XDG_CACHE_HOME/dw/torchinductor, or
# ~/.cache/dw/torchinductor; PyTorch's own default under /tmp does not).
# TORCHINDUCTOR_CACHE_DIR=~/.cache/dw/torchinductor

# --- Hugging Face Hub (all of diffusers and transformers download through it) --

# Token for gated models such as FLUX.1-dev (or run `hf auth login` once).
# HF_TOKEN=hf_...

# Where models are cached (default ~/.cache/huggingface). Point HF_HUB_CACHE
# at a bigger disk without moving the token and settings in HF_HOME.
# HF_HOME=~/.cache/huggingface
# HF_HUB_CACHE=/mnt/models/hub

# Never touch the network: every model must already be cached. Also stops
# diffusers checking Hub kernel publishers (see DIFFUSERS_TRUST_REMOTE_KERNELS).
# HF_HUB_OFFLINE=1

# Faster large downloads through the Xet backend, at the cost of more CPU
# and memory while downloading.
# HF_XET_HIGH_PERFORMANCE=1

# Seconds before a slow download is abandoned (default 10).
# HF_HUB_DOWNLOAD_TIMEOUT=60

# HF_HUB_DISABLE_TELEMETRY=1

# --- diffusers -----------------------------------------------------------------

# Allow attention kernels from Hub publishers diffusers does not list as
# trusted. Needed for "attention_backend": "sage_hub" (SageAttention 2 from
# SageAttention/sage-attention) - about 30% faster per MiniMax-H3 step on an
# RTX 3090. It downloads and runs remote code from any such publisher, not
# only that one, so enable it deliberately. CUDA only; no use on MPS.
# DIFFUSERS_TRUST_REMOTE_KERNELS=true

# The attention backend for every model that sets none (default native). A
# workflow's per-component "attention_backend" overrides it. Same values as
# that setting: native, flash_hub, sage_hub, _native_cudnn, ... Leave it
# native: it reaches every model on the server, and sage cannot serve all of
# them (Music3's head_dim-256 decoder, LTX's attention mask). Pin sage_hub per
# workflow instead - the MiniMax-H3 templates take attention_backend=sage_hub.
# DIFFUSERS_ATTN_BACKEND=native

# Refuse any pipeline or component that would load custom code from the Hub.
# dw's own trust gate (--trust-workflows) already covers workflow files; this
# closes the same door at the diffusers layer.
# DIFFUSERS_DISABLE_REMOTE_CODE=true

# --- transformers --------------------------------------------------------------

# Log level for transformers (debug, info, warning, error, critical).
# TRANSFORMERS_VERBOSITY=error

# Silence the tokenizers fork-safety warning after a tokenizer has been used.
# TOKENIZERS_PARALLELISM=false

# --- SDNQ quantization (CUDA) -----------------------------------------------------

# SDNQ's int matmuls run on autotuned Triton kernels by default. 0 falls back
# to PyTorch's kernels: no first-step autotune stall at a new shape, but
# slower steps. Measure before turning it off.
# SDNQ_USE_TRITON_MM=0

# Narrow the autotune search to shorten that stall, at the risk of a slower
# kernel. Comma lists; the defaults are shown.
# SDNQ_TRITON_MM_BLOCK_SIZE_M_LIST=64,128,256
# SDNQ_TRITON_MM_BLOCK_SIZE_N_LIST=64,128,256
# SDNQ_TRITON_MM_BLOCK_SIZE_K_LIST=32,64