Backend Overview

FORGE Studio supports five reconstruction backends. Each backend corresponds to a different runtime environment.

Backend	Key	Runtime	Config format
Python	python	External Python 3	JSON
Julia	julia	External Julia	JSON
MATLAB	matlab	External MATLAB	JSON
MATLAB Runtime	runmat	Bundled MCR (read-only)	JSON
FORGE (C++/CUDA)	forge	Bundled binaries	CLI flags

---

How Backends Execute

Script Backends (Python, Julia, MATLAB)

Script backends use a persistent session model:

1. On the first stage for a backend, FORGE Studio spawns the interpreter with the forge_dispatcher script.
2. The dispatcher signals ready via the FORGE_RESP: protocol on stdout.
3. For each step, FORGE Studio writes a JSON config file and sends a run command to the session.
4. The dispatcher loads and calls your module's run() function.
5. Progress, logs, convergence metrics, and image previews are streamed back in real time via FORGE_RESP: events.
6. Subsequent stages on the same backend reuse the same session — cold-start latency only once per job.

FORGE Studio (Node.js)
  │
  │─── stdin: {"type":"run","module_path":"...","config_path":"..."} ───►
  │                                                                      │
  │◄── stdout: FORGE_RESP:{"type":"log","message":"starting"} ──────────│
  │◄── stdout: FORGE_RESP:{"type":"progress_tick",...} ─────────────────│
  │◄── stdout: FORGE_RESP:{"type":"ok","output_paths":["..."]} ─────────│
interpreter process

FORGE Backend

FORGE binaries are invoked directly as child processes with CLI flags — no persistent session. Each stage spawns a new process. Progress and metrics are read from stderr as JSONL events.

---

Config File

For script backends, FORGE Studio writes a JSON config file to the stage output directory before calling run(). Your module receives the path as its only argument.

{
  "job_id": "abc-123",
  "step_id": "espirit_calib",
  "stage_index": 0,
  "input_paths": ["/data/scan.dat"],
  "output_dir": "/output/stage_0_sensitivity_calibration",
  "input_format": "twix",
  "channels": 32,
  "params": {
    "calibration_size": 24,
    "use_gpu": true
  }
}

See Config Format Reference → for the full field list.

---

Stage Chaining

Each stage receives the outputs of the previous stage as its input_paths. Stage 0 receives the original raw file. This enables multi-stage pipelines where, for example:

• Stage 0 (Python): reads TWIX, produces calibration maps → .npy
• Stage 1 (FORGE): reads calibration maps, produces reconstruction → .nii
• Stage 2 (Julia): reads NIfTI, applies post-processing → .nii

input_format is detected from the file extension of input_paths[0]:

Extension	Format value
.dat	twix
.h5, .hdf5	ismrmrd
.mat	mat
.jld2	jld2
.npy, .npz	npy / npz

---

Configuring Backend Paths

Go to Settings → Backends to set or override the executable path for each backend. Paths are auto-detected at startup but can be overridden:

• Browse — opens a file picker to select the executable.
• Validate — checks the path and reports the detected version.

Detection order (macOS):

• MATLAB: /Applications/MATLAB_*.app, then PATH
• Python: python3 on PATH, then common Homebrew/pyenv paths
• Julia: /Applications/Julia-*.app, then PATH

---

Compute Backend (FORGE)

FORGE binaries are available in three variants, selected automatically based on hardware:

Variant	Platform
metal	Apple Silicon (macOS)
cuda	NVIDIA GPU (Linux)
cpu	Any (fallback)

Override in Settings → Compute Backend.