Record adaptive-MTP and DFlash7 decode diagnostic - #132
Closed
FujitsuPolycom wants to merge 2 commits into
Closed
Conversation
Add a research-only GLM-5.3 server-log record comparing the observed adaptive-MTP and retained DFlash7 decode intervals. Preserve exact runtime, model, image, timestamp, depth, acceptance, and generation-gauge evidence while marking the absent prompt/output receipt and user-reported coding peak as unaudited. State that SparkCache emitted no work in either decode interval, so the observation does not measure cache behavior. Validation: 9 focused fixture, link, and terminology tests passed; Ruff E/F/W checks passed.
This was referenced Aug 30, 2026
Owner
Author
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Result
Status: research-only.
Add a bounded GLM-5.3 decode-log diagnostic for the adaptive embedded-MTP runtime and the retained DFlash7 runtime. The record preserves exact target, runtime, image, draft, topology, and UTC interval identities.
The adaptive interval observed depths 2–4, draft acceptance from 48.9% to 81.2%, and server generation gauges from 24.5 to 38.7 tok/s. The retained DFlash7 interval observed generation gauges from 40.6 to 136.0 tok/s. A separate operator-reported matched coding peak of approximately 70 versus 33.5 tok/s is labeled unaudited.
The prompt, generated output, request shape, and client receipt were not preserved. The record therefore makes no benchmark, quality, or isolated speculator claim. The two runtime compositions differ in multiple components.
Neither interval contains a SparkCache store, restore, placement, publication, or maintenance event. The evidence does not measure or implicate SparkCache work.
Validation
python -m pytest performance/test_glm53_decode_log_diagnostic.py scripts/test_sparkcache_terminology.py -q: 9 passedruff check --select E,F,W --ignore E501 performance scripts: passedThis draft is stacked on #131 because the adaptive runtime identity uses the CUDA-terminology profile from that branch.