Repository navigation
Restore PI recipes and shared token diagnostics - #152
Merged
Merged
Conversation
This was referenced Sep 24, 2026
Collaborator
This was referenced Oct 2, 2026
ElmoPA
marked this pull request as ready for review
October 2, 2026 19:05
This was referenced Oct 7, 2026
ElmoPA
changed the base branch from
codex/graph-integration-20260924/02-hpt-parity
to
graphite-base/152
October 7, 2026 21:08
ElmoPA
force-pushed
the
codex/graph-integration-20260924/03-pi-parity
branch
from
October 7, 2026 21:10
daeaf92 to
cbd8233
Compare
ElmoPA
added a commit
that referenced
this pull request
Oct 7, 2026
Restore recorded-data workflows through configured DataModule and evaluator capabilities. The standalone language/trajectory tool now writes complete episode videos plus sampled PNG/NPZ artifacts without constructing a policy or fitting normalization. A configurable field audit replaces the EVA-only zero check and retains exact episode/frame identities. HDF5 conversion delegates to the maintained EVA converter, preserves source bytes, refuses existing outputs, and imports without cloud/model dependencies. Retained visualization config names use shared renderers. Migration decisions, including the audited debug converter’s broken import, are recorded in `docs/integration/DATA_TOOLS.md`. Validation: 1,392 CPU tests passed in an isolated checkout, including actual HDF5 conversion, encoded video frame counts/FPS/annotations, overwrite rejection, field-audit identities, and installed-wheel checks. See `docs/integration/evidence/data-tools-cpu-receipt.json`. Combined-tree and real-data validation remain open; no legacy cutover is included. Stack: #150 → #151 → #152 → this PR.
ElmoPA
added a commit
that referenced
this pull request
Oct 7, 2026
Checkpoint restoration could previously pair matching tensor shapes with a different preprocessing or normalization state. Bind new checkpoints and inference artifacts to the resolved graph, model-owned declaration and complete data context; reject mismatches before model construction. The Zarr adapter records frame/keymap/transform semantics while allowing relocated validation data and held-out episode selection. Robot rollout and the new no-hardware `InferenceSession` share strict bound loading, decoding and typed runtime controls. Restoration constructs the architecture without reloading external initialization weights. Fresh initialization remains explicit and rejects overlapping targets. Historical unbound checkpoints are preserved with their original runtime; no automatic converter is claimed. Validation: **1,404 tests passed** in an isolated checkout, including eight HPT/stub-PI source recipes with real optimizer steps, checkpoint export/reload, canonical sequence decoding, and corrupt-context rejection. See `docs/integration/evidence/bound-inference-cpu-receipt.json` for the exact tested tree and log hash. Stack: #150 → #151 → #152 → #153 → this PR. Real-weight/data OSMO GPU gates and assembled EgoVerse integration remain pending. No legacy runtime cutover or deletion.
ElmoPA
added a commit
that referenced
this pull request
Oct 7, 2026
Add required lock, pinned lint/format, recursive YAML and constructor, CPU, and installed-wheel checks to CI. All 253 shipped YAMLs are composed and their graph/data/evaluator components are constructed offline; ready declarations must produce their native output from the declared observations. Meta tensors and explicit tokenizer/architecture fixtures keep these checks separate from real-weight training. The constructor gate caught eager Scale API access and a stale EVA wrist keymap. Defer annotation queries until resolution, preserving scoped memoization, and use the canonical Cartesian keymap with the frame choice in transforms. Finish the shared prediction-metric entrypoint used by latent diagnostics and canonical Yam camera naming. Broad formatting enforces the existing rules while preserving byte-pinned Yam mapper/IK sources. Validation: lock and Ruff checks pass; **1,408 tests passed** in an isolated checkout, including 253 compose contexts and 253 constructor contexts. Exact tree and log hash: `docs/integration/evidence/ci-cpu-receipt.json`. Constructor substitutions and outstanding gates are documented in `docs/integration/VALIDATION.md`. Stack: #150 → #151 → #152 → #153 → #154 → this PR. Real-weight/data OSMO and assembled EgoVerse validation remain pending. No legacy-runtime removal.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Retained PI source recipes now run through the graph with their original action encodings and cosine schedules. Attention extraction uses a configured model provider; the shared evaluator writes bounded PCA/UMAP/t-SNE archives with episode/frame identities and supports rebuilding CSVs without a model. The cotrain port also fixes shape inference overwriting EVA’s 14D normalization schema with the human source’s 12D schema.
OpenPI installation verifies the fork’s required Transformers patches, preserves originals, and avoids modifying uv cache hardlinks. Weights-only initialization requires an explicit source hash and namespace.
Validation: 1,388 CPU tests passed in an isolated checkout, including two-source optimization, actual language-overlay video encoding, source selection, projection/rebuilds, strict initialization, and installed-wheel checks. See
docs/integration/evidence/pi-cpu-receipt.jsonfor the tested tree and log hash. Real-weight OSMO and assembled-EgoVerse gates remain open; this draft does not authorize legacy removal.Stack: #150 → #151 → this PR.