AbhishekGhosh1998 commited on
Commit
dc746d6
·
verified ·
1 Parent(s): 2562d81

granularity results tierb-fixed-moonep-full @ 2026-08-30T00:10:12Z from sp-aiinfra-interactive-aghosh-7jtjl-8xb200-moe-cg-0 (part 2)

Browse files
Files changed (1) hide show
  1. tierb-fixed-moonep-full/study.log +59 -0
tierb-fixed-moonep-full/study.log CHANGED
@@ -437,3 +437,62 @@ Already logged in to Hugging Face as: AbhishekGhosh1998
437
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
438
 
439
  === ROLL-UP
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
437
  https://huggingface.co/MoE-CUDA-Graph-Analysis/tierb-b200-gbs8/tree/main
438
 
439
  === ROLL-UP
440
+ 28 runs ok=24 oom=0 -> /workspace/tierb-fixed-moonep-full/collected_all.tsv
441
+
442
+ phase config seq gbs policy rc cg asy backend iter_ms vs_best
443
+ -------------------------------------------------------------------------------------
444
+ tierb DSV2_LITE_L24 8192 8 matched none 1 1 moonep 595.97 1.00x
445
+
446
+ tierb QWEN2_57B 4096 8 matched none 1 1 moonep 184.20 1.00x
447
+
448
+ tierb QWEN3_235B 8192 8 matched none 1 1 moonep 322.10 1.00x
449
+
450
+ tierb QWEN3_30B 4096 8 matched none 1 1 moonep 360.76 1.00x
451
+
452
+ tierbmem DSV2_LITE_L24 8192 8 matched none 1 1 moonep 622.80 1.00x
453
+
454
+ tierbmem QWEN2_57B 4096 8 matched none 1 1 moonep 196.60 1.00x
455
+
456
+ tierbmem QWEN3_235B 8192 8 matched none 1 1 moonep 341.70 1.00x
457
+
458
+ tierbmem QWEN3_30B 4096 8 matched none 1 1 moonep 435.30 1.00x
459
+
460
+ tierbprobeDSV2_LITE_L24 8192 8 matched none 0 1 moonep 673.94 1.00x
461
+
462
+ tierbprobeQWEN2_57B 4096 8 matched none 0 1 moonep 238.91 1.00x
463
+
464
+ tierbprobeQWEN3_235B 8192 8 matched none 0 1 moonep 384.76 1.00x
465
+
466
+ tierbprobeQWEN3_30B 4096 8 matched none 0 1 moonep 824.23 1.00x
467
+
468
+ 8 probe runs -> /workspace/tierb-fixed-moonep-full/probe_summary_all.tsv
469
+ run arm cap rawExpI preRankI postRankI drop% reqCap
470
+ --------------------------------------------------------------------------------------------------------------------
471
+ moonep/tierbprobe/tierbprobe_gbs8/DSV2_LITE_L24__moo natural none 7.5443* 2.1316 1.1058
472
+ moonep/tierbprobe/tierbprobe_gbs8/DSV2_LITE_L24__moo forced none 1.0286 1.0061 1.0026
473
+ moonep/tierbprobe/tierbprobe_gbs8/QWEN2_57B__moonep_ natural none 7.4132* 2.3019 1.0979
474
+ moonep/tierbprobe/tierbprobe_gbs8/QWEN2_57B__moonep_ forced none 1.0348 1.0073 1.0033
475
+ moonep/tierbprobe/tierbprobe_gbs8/QWEN3_235B__moonep natural none 15.9227* 2.5851 1.0712
476
+ moonep/tierbprobe/tierbprobe_gbs8/QWEN3_235B__moonep forced none 1.0386 1.0053 1.0025
477
+ moonep/tierbprobe/tierbprobe_gbs8/QWEN3_30B__moonep_ natural none 15.7078* 2.4672 1.0635
478
+ moonep/tierbprobe/tierbprobe_gbs8/QWEN3_30B__moonep_ forced none 1.0565 1.0077 1.0036
479
+
480
+ * At least one globally aggregated snapshot reached the E/K expert-Imax ceiling.
481
+ Inspect raw_expert_imax_mean and hottest_expert_token_fraction_mean before generalising.
482
+
483
+ ECHO note: echo_logged_pre_rank_metric_valid=0 and echo_logged_load_reduction_valid=0 are intentional.
484
+ The branch's built-in fields mix source-rank and destination-rank axes; this tool reconstructs both globally.
485
+ 8 memory-dump runs -> /workspace/tierb-fixed-moonep-full/memory_summary_all.tsv
486
+ model backend cg allocation capacity status peak GiB dumps
487
+ ----------------------------------------------------------------------------------
488
+ DSV2_LITE_L24 moonep 1 none none OK 130.2 48/48
489
+ DSV2_LITE_L24 moonep 0 none none OK 122.2 40/40
490
+ QWEN2_57B moonep 1 none none OK 94.5 48/48
491
+ QWEN2_57B moonep 0 none none OK 93.6 40/40
492
+ QWEN3_235B moonep 1 none none OK 117.5 48/48
493
+ QWEN3_235B moonep 0 none none OK 117.3 40/40
494
+ QWEN3_30B moonep 1 none none OK 114.0 48/48
495
+ QWEN3_30B moonep 0 none none OK 113.7 40/40
496
+
497
+ ECHO/UltraEP note: graph-on uses static capacity + device-initiated GEMM while graph-off uses dynamic allocation + host-derived sizes. Their difference is not labelled as pure graph cost.
498
+ Already logged in to Hugging Face as: AbhishekGhosh1998