Commit Graph

Select branches

Hide Pull Requests

feature/musa-benchmark-support

feature/musa-rlc-product-harness

main

Mono Color

• 2786eb9886 bench: orchestrate a multi-card campaign without touching the measurement path main Aharrypotter 2026-09-10 19:43:37 +08:00
• 03fdcbab28 bench: partition operators across cards by measured cost, and rescore in place Aharrypotter 2026-09-10 19:25:34 +08:00
• c6a7f834cb bench: ignore the switch's own annotations when fingerprinting the IR Aharrypotter 2026-09-10 19:13:51 +08:00
• 4ac2a9acd8 bench: report when a foreign vendor's FlagGems implementation gets loaded Aharrypotter 2026-09-10 18:51:13 +08:00
• 2e456bdc74 bench: let the verdict govern the report, and put the caliber in the table Aharrypotter 2026-09-10 17:10:59 +08:00
• a11863399a bench: bound each screening pair and record what it cost Aharrypotter 2026-09-10 17:08:08 +08:00
• 1fd3abd56d bench: rename NO_KERNEL to NO_TRITON and say what it rules out Aharrypotter 2026-09-10 17:00:45 +08:00
• 4d69320139 bench: let ab_rlc_test.sh decide which operators are worth six rounds Aharrypotter 2026-09-10 16:54:26 +08:00
• 2fbfe1dd16 bench: make screening a first-class step with the same caliber as the A/B Aharrypotter 2026-09-10 16:49:32 +08:00
• 5236d0745f bench: fingerprint only the IR, and refuse to compare when there is none Aharrypotter 2026-09-10 16:47:39 +08:00
• 551ffbbb77 bench: exit the sampler quietly when its reader goes away Aharrypotter 2026-09-10 16:10:15 +08:00
• 4329b0c835 docs: stop claiming non-CUDA backends never attempt graph capture Aharrypotter 2026-09-10 16:00:44 +08:00
• ee5f70fa89 bench: record device clocks, temperature and power for the whole campaign Aharrypotter 2026-09-10 16:00:01 +08:00
• d45806eacd bench: log every A/B cell as it finishes, and cap it with a timeout Aharrypotter 2026-09-10 15:55:36 +08:00
• 92f3d150c7 bench: preflight the interpreter, not just the paths Aharrypotter 2026-09-10 15:18:00 +08:00
• c95474a432 bench: refuse to report a comparison whose two sides have different calibers AharrypotterandClaude Opus 5 2026-09-10 14:32:49 +08:00
• aa147c938e feat: repeat an A/B over rounds and report whether the verdict holds AharrypotterandClaude Opus 5 2026-09-10 13:11:37 +08:00
• 9665d36a7c feat: allow the direct timer to repeat and take the median AharrypotterandClaude Opus 5 2026-09-10 12:56:08 +08:00
• d7cb7efc94 fix: have the campaign worker ask zl_bench for its plugin list AharrypotterandClaude Opus 5 2026-09-10 10:34:05 +08:00
• 75405cc555 Merge the S5000 MUSA RLC product campaign AharrypotterandClaude Opus 5 2026-09-10 10:32:33 +08:00
• 29af59be9b feat: absorb the MThreads compat plugin and the A/B range report AharrypotterandClaude Opus 5 2026-09-10 10:31:59 +08:00
• c9f9b7885f feat: check the run's preconditions, and record which tuner cache it used AharrypotterandClaude Opus 5 2026-09-10 10:31:36 +08:00
• ac818291a6 feat: turn on graph capture for MUSA AharrypotterandClaude Opus 5 2026-09-09 17:56:56 +08:00
• 6fa5f5d7a2 refactor: let any backend own graph capture and its failure modes AharrypotterandClaude Opus 5 2026-09-09 17:56:32 +08:00
• 783c0655bc fix: never let the parallel sweep cover only part of the shape set AharrypotterandClaude Opus 5 2026-09-09 16:40:12 +08:00
• 92dae82c7f feat: expose the upstream benchmark level as BENCH_LEVEL AharrypotterandClaude Opus 5 2026-09-09 16:39:45 +08:00
• c8ea586d0f feat: make the harness backend-agnostic AharrypotterandClaude Opus 5 2026-09-09 16:39:14 +08:00
• 35a928a516 chore: pin shipped text formats to LF AharrypotterandClaude Opus 5 2026-09-09 16:38:33 +08:00
• abf9475ff3 feat: report a per-operator speedup range across two runs feature/musa-benchmark-support Aharrypotter 2026-09-02 12:19:26 +08:00
• 7522143747 feat: drop stale CUDA-only skip guards on MThreads Aharrypotter 2026-09-02 12:19:26 +08:00
• 3c9529dab2 fix: take graph-capture timing from Triton's device interface Aharrypotter 2026-09-02 12:19:07 +08:00
• 64baa11a23 chore: pin shipped text formats to LF Aharrypotter 2026-09-02 12:19:06 +08:00
• e970ed94c2 docs: document the MUSA RLC campaign feature/musa-rlc-product-harness Aharrypotter 2026-09-02 11:36:33 +08:00
• 8a685638bb feat: re-account a sealed MUSA RLC run without re-running it Aharrypotter 2026-09-02 11:35:43 +08:00
• 8bbea472bb feat: add the S5000 MUSA RLC product campaign Aharrypotter 2026-09-02 11:35:30 +08:00
• 22139e4b87 chore: pin shipped text formats to LF Aharrypotter 2026-09-02 11:34:48 +08:00
• 3b54f458e6 feat: add a MThreads benchmark compatibility plugin Aharrypotter 2026-09-02 11:34:04 +08:00
• da22885645 Add multi-op batch screening layer; fix cudagraph fallback regressions zhoulin 2026-08-12 19:04:09 +00:00
• 95895cea0e Adapt to upstream FlagGems repo split and record pre_hook configs for libtuner zhoulin 2026-08-12 10:55:32 +00:00
• 1e1d612031 Add opt-in multi-GPU sweep parallelism; fix GPU pinning and run dir races zhoulin 2026-07-29 12:14:12 +00:00
• 7e0e8648f1 Fix autotune record/replay silently skipping libtuner kernels zhoulin 2026-07-29 00:38:14 +00:00
• f98ec10fd0 Renew scripts zhoulin 2026-07-24 10:17:33 +00:00
• c8a42b66ee Move trace shape sets to top-level shape/ and frame ab_fold_test.sh as A/B reference example zhoulin 2026-07-20 16:53:14 +00:00
• 00902e6c67 Default USE_FLAGTUNE to 0 and add fold_scale A/B script with trace shapes zhoulin 2026-07-20 16:50:59 +00:00
• 26b071c6e1 Stabilize first-run cudagraph timing and colorize terminal output zhoulin 2026-07-19 19:21:13 +00:00
• 8ace0d790c Add FlagGems single-op perf benchmark harness zhoulinandClaude Fable 5 2026-07-16 03:52:07 +00:00
• 242d74af0e Initial commit zhoulin 2026-07-15 17:20:27 +08:00