Environment and files

Environment variables#

MLX_DSPARK_MODEL_DIRS

Extra folders to search for MLX models before downloading, :-separated, ~ expanded. Each folder is tried as publisher/model, publisher_model and a bare model folder. They appear in mlx-dspark models and the app's list.

MLX_DSPARK_TRUST_REMOTE_CODE=1

Allow checkpoints that ship their own Python (model_file / auto_map in their config) to be imported, for the whole process. Same as --trust-remote-code. Off by default.

MLX_DSPARK_CPU_SPLIT_FP32=0

Run the CPU share of CPU co-prefill in bf16 through BNNS instead of fp32 through Accelerate BLAS. Faster, but the BNNS route was behind every co-prefill crash ever reported; only for machines that never reach memory pressure.

MLX_DSPARK_FORCE_SMALL_M=1

Enable the small-M verify kernel on M5 and newer, where it's gated off. For comparison runs only.

MLX_DSPARK_FORCE_MULTIROW=1

Enable the multi-row attention kernel on M5 and newer, where it's gated off until verified. For comparison runs only.

MLX_DSPARK_STREAM_KEEPALIVE_S

Seconds between keep-alive frames on quiet streams (default 15). Lower it for aggressive proxies.

MLX_DSPARK_SLOW_ROUND_LOG_S

Log any gap between rounds longer than this many seconds to stderr, with mode, cap and context (default 10; 0 disables). Useful when reporting a stall.

HF_HOME, HF_HUB_CACHE

Where the Hugging Face cache lives. Honoured everywhere, including the "is it installed?" checks.

Files#

~/.cache/mlx_dspark/calibration.json

This Mac's measured curves per model, quantization and MLX version: verify cost by width, drafter cost, depth slope, kernel races and the CPU co-prefill fraction. Safe to delete; it re-measures on the next run. See the calibration cache.

~/.cache/mlx_dspark/models/

Plain model folders the engine looks in before Hugging Face: converted drafters (like the Bonsai GGUF repack) and anything you put there by hand.

~/.lmstudio/models/, ~/.cache/lm-studio/models/

LM Studio's folders. Its MLX downloads load in place.

--prefix-cache-dir

Where the optional SSD tier of the prefix cache spills, when you enable it.

The Mac app keeps its runtime, logs and chats under its own Application Support folder, outside the app bundle and separate from any pip install.