Repository navigation
bench/results: RTX A3000 12GB laptop + i7-12850HX (sm_86): hugepages, PCIe probe, calibration - #489
Closed
yannickloth wants to merge 2 commits into
Closed
yannickloth wants to merge 2 commits into
yannickloth wants to merge 2 commits into
Conversation
… PCIe probe, calibration Host tuning and calibration for a 12 GB Ampere laptop part with a 40 GiB CPU-side expert arena: the hugetlb pool must cover the whole MAP_HUGETLB mapping (a 34 GiB pool silently fell back to 4 KiB for a 39.97 GiB arena), the PCIe probe read this x16 Gen4 link at 5.8-18.5 GB/s across starts, and tools/calibrate.py keeps --pcie-frac 0.35 / --spec-min-p 0.70 / --pool-workers 8. Raw calibration in calibrate.json.
…round at the real hit rate One decode round on the fixed bench prompt, split from --stats: at the fresh-decode cache hit rate (0.354) the CPU expert pool is 52-53% of the round, the GPU 34-35%, MTP 4-5%, host 2-4% and the verify copy-stream commit 0.03 ms (~0%). At the warm 75% hit rate the pool falls to 25% and the GPU becomes the largest term. Also records that --pool-workers 8 on the default all affinity lands on 7 P + 1 E-core, that PCIe in use is ~4 GB/s (not bandwidth-bound), and that the native IQ pack cannot run --gpu-stages/--gpu-only-full/--no-pool and reports 0 for the /metrics blob fields.
This was referenced Oct 2, 2026
Contributor
Author
|
Withdrawing this per the measurement-records policy (#443). The probe finding it rested on (#485) is fixed in 0.1.37, the Linux largepages part (#488) is in 0.1.38, and the decode profile was taken as a report in #494; the numbers stay in my forks notes. Glad to fold anything you want into docs/ instead of a And thank you for the effort on 0.1.38 and the patient triage here — appreciated. |
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.
A results entry for a 12 GB Ampere laptop part with a 40 GiB CPU-side expert arena (engine 0.1.34, Qwen3.8-Flash-Next Swift 1.5 IQ3_XXS native pack).
vm.nr_hugepagessilently gave 4 KiB pages for the 39.97 GiB arena; 44 GiB giveshugetlb 2 MB pages. UnlimitedRLIMIT_MEMLOCKis required too (the systemd user manager caps units at 8 MiB).--pcie-frac 0.35over the 0.55 default.tools/calibrate.pykeeps--pcie-frac 0.35,--spec-min-p 0.70,--pool-workers 8(8 P + 8 E). Raw output incalibrate.json; decode numbers and their caveats are in the README.Hardware/OS/engine details only; no usernames, hostnames or paths.