[Perf Regression] 28 config(s) regressed @ f9127a88

[github-actions]

Performance Regression Detected

Commit: f9127a88
Run: https://github.com/ROCm/ATOM/actions/runs/33110140604
Date: 2026-08-28T07:55:08.175214+00:00

Regressed Configurations

Model	ISL/OSL	Conc	Tput (cur)	Tput (base)	Δ%	TPOT (cur)	TPOT (base)	Δ%
DeepSeek-R1-0528-MXFP4	1024/1024	4	417.1	418.4	-0.3%	9.31	9.33	-0.2%
DeepSeek-R1-0528-MXFP4	1024/1024	8	701.3	655.0	7.1%	11.00	11.81	-6.8%
DeepSeek-V4-Pro	8192/1024	128	1900.1	2080.4	-8.7%	60.74	58.07	4.6%
DeepSeek-V4-Pro DPA	8192/1024	1024	5222.1	5196.4	0.5%	140.76	168.38	-16.4%
DeepSeek-V4-Pro TBO	1024/1024	64	2223.8	2222.9	0.0%	27.18	27.31	-0.5%
DeepSeek-V4-Pro TBO	1024/1024	128	3281.6	3372.1	-2.7%	37.01	36.17	2.3%
GLM-5.2-FP8	1024/1024	4	283.5	285.1	-0.5%	13.78	13.75	0.3%
GLM-5.2-MXFP4 MTP3	1024/1024	16	1647.3	1465.3	12.4%	9.31	10.39	-10.4%
GLM-5.2-MXFP4 MTP3	8192/1024	64	1395.4	1612.6	-13.5%	43.31	37.27	16.2%
GLM-5.2-MXFP4 MTP3	8192/1024	256	2095.3	2207.1	-5.1%	116.04	109.52	6.0%
Kimi-K2.7-Code-MXFP4	1024/1024	64	2546.1	2568.5	-0.9%	24.39	24.20	0.8%
Kimi-K3	1024/1024	4	175.3	175.4	-0.0%	22.02	22.20	-0.8%
Kimi-K3 DSpark	8192/1024	4	293.4	312.2	-6.0%	11.91	11.62	2.5%
MiniMax-M3-MXFP4	8192/1024	16	1353.9	1435.1	-5.7%	11.05	10.58	4.5%
MiniMax-M3-MXFP4 EAGLE3	1024/1024	128	6944.9	7540.2	-7.9%	17.55	16.11	9.0%
MiniMax-M3-MXFP4 EAGLE3	8192/1024	16	1965.5	1875.1	4.8%	7.07	7.59	-6.9%
MiniMax-M3-MXFP4 EAGLE3	8192/1024	32	2523.8	2606.8	-3.2%	11.22	11.12	0.9%
Qwen3.5-397B-A17B-FP8	8192/1024	4	404.6	408.0	-0.8%	9.45	9.38	0.8%
Qwen3.5-397B-A17B-FP8	8192/1024	16	1078.9	1070.0	0.8%	14.12	14.29	-1.2%
Qwen3.5-397B-A17B-FP8 MTP3	8192/1024	4	801.0	807.9	-0.8%	4.41	4.42	-0.2%
Qwen3.5-397B-A17B-MXFP4	1024/1024	4	460.5	486.7	-5.4%	8.45	7.99	5.7%
Qwen3.5-397B-A17B-MXFP4	1024/1024	8	805.6	886.3	-9.1%	9.58	8.71	10.0%
Qwen3.5-397B-A17B-MXFP4	1024/1024	16	1321.1	1385.0	-4.6%	11.72	11.20	4.6%
Qwen3.5-397B-A17B-MXFP4	8192/1024	8	741.1	788.5	-6.0%	10.21	9.61	6.2%
gpt-oss-120b	1024/1024	4	886.4	875.0	1.3%	4.27	4.43	-3.7%
gpt-oss-120b	1024/1024	8	1497.2	1619.1	-7.5%	4.98	4.76	4.6%
gpt-oss-120b	1024/1024	16	2583.3	2716.8	-4.9%	5.97	5.71	4.6%
gpt-oss-120b	1024/1024	64	6259.9	6608.1	-5.3%	9.84	9.29	5.9%

Performance Summary

Summary not available

Profiler Traces

Download from workflow artifacts.
Open in Perfetto UI or Chrome chrome://tracing for analysis.

Next Steps

1. Download profiler-analysis-33110140604 artifact
2. Open trace files in Perfetto UI
3. Compare kernel durations against previous traces
4. Identify bottleneck changes