skainet-plan: model-footprint analysis for GGUF, safetensors and ONNX with a fits-in-RAM verdict

[michalharakal]

There is no quick way to check whether a model will fit on an embedded device with limited memory and compute before a whole day is invested in converting it.

Goal

A standalone Kotlin/JVM CLI (extending skainet-apps/skainet-plan, the existing footprint tool) that:

• analyzes GGUF, safetensors and ONNX model files from the header/metadata only — no payload load, seconds not hours;
• reports the actual resident size of the model weights (as stored, and — where the format carries encodings — as they would be held after form resolution);
• renders a verdict against a usable-RAM limit of ~2.1 GB (the embedded-device budget): fits / does not fit, with the margin and the planner's suggestions where they apply.

Notes

• GGUF already works end to end (StreamingGGUFReader.planInput → resolveWeightForms → profile.plan().render()); the new work is the safetensors and ONNX front ends plus the budget flag/verdict ergonomics.
• safetensors: the header is a JSON table of names/dtypes/shapes at the file start — header-only sizing is cheap.
• ONNX: weights live in graph initializers; enumeration must avoid materializing tensor data.
• No architecture metadata exists in safetensors/ONNX, so KV-cache/forward estimates are GGUF-only; for the other formats the verdict is weights + headroom, stated as such.

Acceptance

• skainet-plan <file> auto-detects the format and prints the weights table + total + verdict for all three formats.
• A budget flag (default ~2.1 GB for the embedded-device scenario, overridable) drives the fits/doesn't-fit line.
• Header-only: analyzing a multi-GB file completes in seconds and never reads tensor payloads.
• Tests per format with synthetic fixtures.

[michalharakal]

Split per michalharakal's call: the library capability (planInput for safetensors/ONNX, the external_data fix, PlannerProfile.EDGE, and the tests that pin the counting behaviour through future memory-layout refactors) lands in core via #1170. The multi-format CLI moves to SKaiNET-research on feature/model-footprint-cli, where it waits for a stable core release carrying these APIs — tracked in SKaiNET-developers/SKaiNET-research#5. This issue closes when #1170 merges; the CLI's fate is tracked on the research side.

[michalharakal]

Done, per the agreed split:

• Capability in core (merged, Footprint capability: header-only planInput for safetensors/ONNX, external_data fix, EDGE profile #1170): header-only planInput for safetensors and ONNX, the ONNX external_data pricing fix with Long-safe sizes, PlannerProfile.EDGE, and the tests that pin the counting behaviour — per-format byte sums, the 3 GiB Long-correctness case, unknown-dtype pricing via offsets — so future memory-layout refactors cannot silently lose the capability.
• CLI in SKaiNET-research: the standalone multi-format model-footprint tool is parked on feature/model-footprint-cli, build-verified against mavenLocal and smoke-tested (--profile edge --budget 2.1G → correct verdict + exit code); it waits for the first stable core release carrying these APIs. Tracked in Model-footprint analysis: 'will it fit on the device?' in seconds, for GGUF / safetensors / ONNX SKaiNET-research#5.

Deferred follow-ups stay recorded in the issue body: HF config.json geometry for safetensors, sharded .index.json support, a /proc/meminfo DeviceMemory provider, ONNX TensorId name maps.