Stage types are LLM-specific — forge-alloy should be domain-agnostic

[joelteply]

Problem

The current stage types (prune, train, lora, quant, eval, etc.) are hardcoded for LLM forging. But forge-alloy is meant to be a universal pipeline contract for ANY model domain — vision, audio, diffusion, robotics, whatever.

Current (LLM-only)

Stage	Domain
prune	LLM
train	LLM
lora	LLM
compact	LLM
quant	LLM (GGUF, MLX)
eval	LLM (HumanEval, MMLU)
expert-prune	LLM (MoE)
context-extend	LLM (RoPE)
modality	LLM (add encoder)

The Beyond

Forge-alloy should support domain-specific stage types as extensions of a base stage system. The schema should define WHAT a stage is (input/transform/output with typed config) and let domains register their own.

Camera/Vision domain stages

• augment — data augmentation (crop, rotate, color jitter)
• backbone-swap — replace feature extractor (ResNet → EfficientNet)
• detection-head — add/modify object detection head (YOLO, DETR)
• calibrate — camera calibration, lens correction
• eval — mAP, IoU, FPS benchmarks (not HumanEval)

Diffusion domain stages

• unet-prune — prune U-Net attention
• scheduler-swap — DDPM → DPM++ → Euler
• vae-tune — fine-tune VAE decoder
• eval — FID, CLIP score, aesthetic score

Audio domain stages

• codec-swap — replace audio codec (Encodec → DAC)
• speaker-adapt — voice cloning LoRA
• eval — WER, MOS, speaker similarity

Robotics domain stages

• sim-to-real — domain adaptation from simulator
• policy-distill — compress large policy to deployable size
• eval — task success rate, latency, safety bounds

Design

The stage type system should be a registry, not a hardcoded enum:

{
  "domains": {
    "llm": { "stages": ["prune", "train", "lora", "quant", "eval", ...] },
    "vision": { "stages": ["augment", "backbone-swap", "detection-head", ...] },
    "diffusion": { "stages": ["unet-prune", "scheduler-swap", "vae-tune", ...] }
  }
}

Each domain registers its stage types with their schemas. The alloy executor discovers domain-specific executors at runtime.

For Now

The README should mention that the current stages are the LLM domain, and that forge-alloy is designed to be extended to other domains. Don't promise what doesn't exist yet, but plant the flag.

Priority

Low for implementation. High for README/docs positioning.