Pipeline block ③ (Compile: dataset → weights). Part of the compile-to-weights epic. Decision: full-parameter offline fine-tune (paper-faithful); stargraph does NOT own the GPU training loop — it owns the handoff.
Context. There is zero training infrastructure in stargraph today (ml/ is inference-only: registry + sklearn/xgboost/onnx loaders). The paper full-parameter fine-tunes a small model (e.g. Qwen3-8B) on the block-② dataset. Training is non-deterministic and heavyweight (torch, A100-class GPU) — it must live outside the deterministic runtime as an offline CI/CD "recompile" step.
Proposal. Define the compile job contract + registry handoff, not an in-process trainer.
- A documented external trainer (full-parameter fine-tune) consumes the block-② JSONL and produces a model directory / HF snapshot.
- A thin
stargraph side records the result into ModelRegistry.register (ml/registry.py:181) with provenance metadata: source graph_hash, dataset hash, base model id, run/sample count, training config hash. Stage via alias (registry.py:230) — staging first, never auto-production.
- If ever represented as a node, the training node is
ReplayPolicy.must_stub (non-deterministic; never replayed).
Where.
src/stargraph/ml/registry.py register:181 / alias:230 / ModelEntry.metadata:86 (currently unused — becomes the compile-provenance carrier).- Pairs with block ④ (LLM-weights runtime) for the registered artifact to be loadable.
Acceptance criteria.
Blockers (blocked). Depends on block ② (dataset) and block ④ (registry runtime). torch/vLLM are optional extras, never core deps.
Source: arXiv:2605.22502v1 — "Compiling Agentic Workflows into LLM Weights" (subterranean agents). Part of the end-to-end compile-to-weights pipeline epic; file refs verified against current main. Determinism: per-graph knob (fully-compiled | hybrid | rules-only) defaults to honor #117 (no learned core router).
Pipeline block ③ (Compile: dataset → weights). Part of the compile-to-weights epic. Decision: full-parameter offline fine-tune (paper-faithful); stargraph does NOT own the GPU training loop — it owns the handoff.
Context. There is zero training infrastructure in stargraph today (
ml/is inference-only: registry + sklearn/xgboost/onnx loaders). The paper full-parameter fine-tunes a small model (e.g. Qwen3-8B) on the block-② dataset. Training is non-deterministic and heavyweight (torch, A100-class GPU) — it must live outside the deterministic runtime as an offline CI/CD "recompile" step.Proposal. Define the compile job contract + registry handoff, not an in-process trainer.
stargraphside records the result intoModelRegistry.register(ml/registry.py:181) with provenancemetadata: sourcegraph_hash, dataset hash, base model id, run/sample count, training config hash. Stage viaalias(registry.py:230) —stagingfirst, never auto-production.ReplayPolicy.must_stub(non-deterministic; never replayed).Where.
src/stargraph/ml/registry.pyregister:181 /alias:230 /ModelEntry.metadata:86 (currently unused — becomes the compile-provenance carrier).Acceptance criteria.
graph_hash+ dataset hash + training-config hash inmetadata.stagingalias only; promotion toproductiongated by the eval harness (linked).Blockers (
blocked). Depends on block ② (dataset) and block ④ (registry runtime). torch/vLLM are optional extras, never core deps.Source: arXiv:2605.22502v1 — "Compiling Agentic Workflows into LLM Weights" (subterranean agents). Part of the end-to-end compile-to-weights pipeline epic; file refs verified against current
main. Determinism: per-graph knob (fully-compiled | hybrid | rules-only) defaults to honor #117 (no learned core router).