Repository navigation
[Draft] Add MatMulNBits estimation test matrices and profiling validation - #34211
Draft
Tommaso Adani (tadani3) wants to merge 36 commits into
Draft
Tommaso Adani (tadani3) wants to merge 36 commits into
Tommaso Adani (tadani3) wants to merge 36 commits into
Conversation
…components and counting categories
Forward the actual workload collector through eight GPT beam/greedy and T5/Whisper calls. Preserve the T5 encoder/decoder MatMulNBits regression sequences while covering collector propagation. Validation: CPU Debug builds successful; 25 collector tests passed with 1 LUT skipped; 26 search regression tests passed. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Enable logical core and bias reporting for successful CUDA submissions, explicitly exclude graph capture/replay, and add CPU/CUDA parity regressions. Fix the CI warning test's inherited logging severity. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…nt core work, bias additions, and nominal dequantization/scaling by selected CPU route. Preserve existing logical counts in a separate reporting section.
…A routes through append-only provider transport and estimate core work, bias additions, and nominal dequantization.
…and add CPU estimator validation, an 84-case matrix, and Linux perf collection with Markdown/CSV reporting, documenting measurement scope and accuracy limitations.
added 6 commits
November 13, 2026 19:39
…d K/M/N boundary probes. Document the experiment workflow and qualification strategy.
… numerical comparisons and evidence
…s and explanations
This branch has not been deployed
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.
Description
Add reproducible tooling and experimental evidence for validating
route-aware MatMulNBits estimates:
and reporting-enabled/disabled output equivalence.
Record CPU audits covering aligned shapes, bias/tails, partial K blocks,
explicit zero points, zero activation blocks, and matched FP16/FP32 inputs.
Compare hardware-measured recurring process FP work with independently
derived, bounded arithmetic models.
Production formula refinements remain deferred. CUDA execution/Nsight
qualification and broader CPU coverage are still pending.
Motivation and Context
Establish repeatable evidence for evaluating MatMulNBits estimates before
changing their formulas. Separate logical-count correctness, runtime capture
validation, and implementation-specific FP-work accuracy.
Document observed discrepancies, kernel-level explanations, held-out checks,
and explicit limitations rather than treating latency or passing estimator
tests as proof of executed-FLOP accuracy.