test: harden cross-target verification to extreme/saturation operands - #1883
Merged
Conversation
Motivated by the bpseq silicon debug (hypothesis: training-grown weights push operands into a saturation range where the model and RTL might disagree). Tested: model smul/sadd vs the C emission over 1500 extreme operands (full offset span 0..127, both signs, saturation-adjacent) -- 0 mismatches. So the model is a faithful RTL reference across the whole range; the bpseq divergence is NOT an operand-range bug (confirmed timing). Turned the negative result into coverage: gen_pairs now draws from both the moderate range AND extreme raw GF-T u32 operands, so cross-target bit-exactness is proven on the full representable range. Refs #1764 Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Contributor
Contributor
|
📓 NotebookLM Notebook linked to this PR
This notebook contains session context, decisions, and artifacts for this work. |
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
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Motivated by the bpseq silicon debug (hypothesis: training-grown weights push operands into a saturation range where the Python model and the RTL might disagree — the cross-target proof only used moderate [-4,4] operands). Tested it: model smul/sadd vs the C emission over 1500 EXTREME operands (full offset span 0..127, both signs, saturation-adjacent) — 0 mismatches, ALL MATCH. So the model is a faithful RTL reference across the whole range; the bpseq silicon divergence is NOT an operand-range bug (confirmed timing).
Turned the negative result into a real coverage improvement: gen_pairs now draws from BOTH the moderate range AND extreme raw GF-T u32 operands (full offset span, both signs, saturation-adjacent) — overflow/underflow/carry edges the [-4,4] sweep never reached. Cross-target bit-exactness now proven on the full representable range. Refs #1764