issue-868: extraction.py docstring — last-layer final-RMSNorm caveat + PEFT fallback trap (doc-only)#666
Draft
superkaiba wants to merge 12 commits into
Draft
issue-868: extraction.py docstring — last-layer final-RMSNorm caveat + PEFT fallback trap (doc-only)#666superkaiba wants to merge 12 commits into
superkaiba wants to merge 12 commits into
Conversation
…ubagent=implement [task.py]
[task.py]
…ubagent=implement [task.py]
…eep-dispatch] wat [task.py]
# Code Review: Issue #816 — Reru [task.py]
…ommitted wildchat [task.py]
…+ PEFT fallback trap (doc-only) Correct the "byte-identical to the full-tuple read" overclaim: hook on model.model.layers[L_max] captures the RAW pre-final-RMSNorm block output while output_hidden_states[-1] is POST-norm (transformers 4.57.x check_model_inputs(tie_last_hidden_states=True) replaces the collected raw last output with self.norm(hidden_states)); the helper's fallback path returns the post-norm value at L_max, so the two internal paths disagree there. Also name the peft.PeftModel silent-fallback trap and scope the .detach() no-op claim to gradient semantics (value equality, never object identity). Zero behavior change (AST-identical modulo docstrings). Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
superkaiba
added a commit
that referenced
this pull request
Jul 3, 2026
…gate Determinism pin (plan v2 §5): the repo-root-faking fixtures in the active-consumer / parity / symlinks / vm_disk_guard test files gain monkeypatch.setattr(ced, "_off_main_checkout_root", lambda: None) so a hypothetical fresh-clone-on-a-branch test runner cannot flip the new probe. Additionally sandboxes the #773 active-consumer gate in the 8 tests (+2 accidentally-passing idempotency tests) that walked the LIVE tasks/ tree: real active tasks (#661/#666/#742/#911) now reference data/issue_658/, so the un-sandboxed gate correctly kept their synthetic issue-658 caches — a pre-existing live-state-dependent failure confirmed identical on pristine main, the exact trap the sibling files' fake_repo fixtures document. Assertions unchanged. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
4 tasks
superkaiba
added a commit
that referenced
this pull request
Jul 7, 2026
…val/debug (L4) intent Mechanizes plan-compute-sizing § Activation-capture HBM sizing (the #666/#744 OOM class; founding false negative #825 v17). FAIL (experiment) / WARN (analysis), conditional; reuses c26's intent machinery; intent sets derived from the _C26_INTENT_GPU mirror (BIG by exclusion); four permissive skips (RunPod pin, pod.py provision, window H100/H200, big-intent absolution); N/A escape 'N/A — no 7B activation capture'. Scale-qualified, self-DISARM-safe FAIL detail (no H100/H200/backend-runpod tokens). Docstring: catalog row, conditional enumeration + N/A phrase list. Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
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
1 participant
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.
Closes task #868 (doc-only; zero behavior change; AST-equality-verified).