[Perf Regression] 34 config(s) regressed @ 83a22aa7

[github-actions]

Performance Regression Detected

Commit: 83a22aa7
Run: https://github.com/ROCm/ATOM/actions/runs/34143052526
Date: 2026-09-08T08:44:58.221467+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	419.6	422.3	-0.6%	9.27	9.24	0.3%
DeepSeek-R1-0528-MXFP4	1024/1024	8	723.3	727.7	-0.6%	10.60	10.65	-0.5%
DeepSeek-R1-0528-MXFP4	8192/1024	4	364.2	373.4	-2.5%	10.36	10.22	1.3%
DeepSeek-R1-0528-MXFP4	8192/1024	32	1221.0	1229.8	-0.7%	24.61	24.52	0.3%
DeepSeek-R1-0528-MXFP4 MTP3	1024/1024	8	955.9	979.0	-2.4%	7.79	7.80	-0.0%
DeepSeek-R1-0528-MXFP4 MTP3	1024/1024	16	1549.1	1573.7	-1.6%	9.71	9.59	1.3%
DeepSeek-V4-Pro DPA	8192/1024	1024	5166.0	5183.8	-0.3%	179.85	133.10	35.1%
DeepSeek-V4-Pro EPLB r0 MegaMoE MegaMoE	8192/1024	4096	4752.5	5631.5	-15.6%	653.75	509.98	28.2%
DeepSeek-V4-Pro MTP3	1024/1024	8	733.1	806.0	-9.1%	10.59	9.39	12.9%
DeepSeek-V4-Pro MTP3	1024/1024	128	3616.9	3611.0	0.2%	33.92	34.17	-0.8%
DeepSeek-V4-Pro TBO	8192/1024	256	1946.7	2541.2	-23.4%	122.18	95.12	28.4%
GLM-5.2-FP8	1024/1024	16	743.9	743.4	0.1%	20.94	20.99	-0.2%
GLM-5.2-FP8 MTP3	1024/1024	16	1167.4	1179.7	-1.0%	13.04	13.02	0.1%
GLM-5.2-MXFP4	1024/1024	8	597.3	584.4	2.2%	12.92	13.26	-2.6%
GLM-5.2-MXFP4	8192/1024	8	495.5	499.9	-0.9%	15.03	15.04	-0.1%
GLM-5.2-MXFP4 MTP3	1024/1024	32	2250.6	2372.3	-5.1%	13.56	12.90	5.1%
Kimi-K2.7-Code-MXFP4	1024/1024	64	2583.8	2586.4	-0.1%	24.02	24.03	-0.1%
Kimi-K3	1024/1024	16	613.0	607.6	0.9%	25.06	25.40	-1.4%
Kimi-K3	8192/1024	16	503.3	503.9	-0.1%	29.93	30.08	-0.5%
Kimi-K3	8192/1024	128	1185.7	1207.9	-1.8%	101.88	100.42	1.4%
Kimi-K3 DSpark	1024/1024	8	564.7	550.4	2.6%	13.22	13.10	0.9%
MiniMax-M3-MXFP4	1024/1024	64	3537.3	3623.8	-2.4%	17.16	16.92	1.4%
MiniMax-M3-MXFP4	8192/1024	8	816.9	859.8	-5.0%	9.14	8.81	3.8%
MiniMax-M3-MXFP4 EAGLE3	1024/1024	8	1572.2	1563.1	0.6%	4.68	4.82	-2.9%
MiniMax-M3-MXFP4 EAGLE3	1024/1024	64	5310.6	5321.6	-0.2%	11.25	11.37	-1.0%
MiniMax-M3-MXFP8	1024/1024	8	732.2	798.8	-8.3%	10.34	9.49	9.1%
MiniMax-M3-MXFP8	1024/1024	128	4332.9	4321.7	0.3%	28.33	28.49	-0.6%
MiniMax-M3-MXFP8	1024/1024	256	6377.5	6479.3	-1.6%	38.34	38.13	0.6%
MiniMax-M3-MXFP8 EAGLE3	1024/1024	8	1314.0	1291.7	1.7%	5.69	5.87	-3.0%
Qwen3.5-397B-A17B-MXFP4	8192/1024	4	370.4	371.2	-0.2%	10.35	10.37	-0.2%
gpt-oss-120b	1024/1024	8	1556.4	1525.2	2.0%	4.96	5.07	-2.1%
gpt-oss-120b	1024/1024	32	4098.8	4163.6	-1.6%	7.47	7.37	1.4%
gpt-oss-120b	8192/1024	4	852.3	832.6	2.4%	4.51	4.63	-2.6%
gpt-oss-120b	8192/1024	8	1430.0	1443.5	-0.9%	5.32	5.29	0.6%

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-34143052526 artifact
2. Open trace files in Perfetto UI
3. Compare kernel durations against previous traces
4. Identify bottleneck changes