diff --git a/tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json b/tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json index a86a6a3d5..94599ca8c 100644 --- a/tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json +++ b/tests/dfx/perf/tests/test_hunyuan_image_tp2_cfgp2.json @@ -43,6 +43,12 @@ "latency_p99": 4.7469, "peak_memory_mb_max": 101100, "peak_memory_mb_mean": 101100 + }, + "A3": { + "throughput_qps": 0.1855, + "latency_p99": 5.6304, + "peak_memory_mb_max": 47194.0, + "peak_memory_mb_mean": 47194.0 } } } diff --git a/tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json b/tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json index e49cd0543..bfe8c9582 100644 --- a/tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json +++ b/tests/dfx/perf/tests/test_hunyuan_image_tp2_sp2.json @@ -43,6 +43,12 @@ "latency_p99": 5.1025, "peak_memory_mb_max": 97402, "peak_memory_mb_mean": 97402 + }, + "A3": { + "throughput_qps": 0.1733, + "latency_p99": 5.9271, + "peak_memory_mb_max": 47196.0, + "peak_memory_mb_mean": 47196.0 } } } diff --git a/tests/dfx/perf/tests/test_hunyuan_image_tp4.json b/tests/dfx/perf/tests/test_hunyuan_image_tp4.json index 54631471a..913c2804b 100644 --- a/tests/dfx/perf/tests/test_hunyuan_image_tp4.json +++ b/tests/dfx/perf/tests/test_hunyuan_image_tp4.json @@ -43,6 +43,13 @@ "latency_p99": 4.8601, "peak_memory_mb_max": 56912.0, "peak_memory_mb_mean": 56912.0 + }, + "A3": { + "throughput_qps": 0.1691, + "latency_mean": 5.9169, + "latency_p99": 6.1671, + "peak_memory_mb_max": 45454.0, + "peak_memory_mb_mean": 45454.0 } } } diff --git a/tests/dfx/perf/tests/test_minicpmo_4_5.json b/tests/dfx/perf/tests/test_minicpmo_4_5.json index 7d7ca08ad..510715b22 100644 --- a/tests/dfx/perf/tests/test_minicpmo_4_5.json +++ b/tests/dfx/perf/tests/test_minicpmo_4_5.json @@ -1,6 +1,7 @@ [ { "test_name": "test_minicpmo_4_5", + "description": "H100 baseline refreshed 2026-09-29: equal-weight nightly means for 2026-09-22 through 2026-09-28 (7 runs each for C1/P32, C4/P64, C8/P128), rounded to 4 decimals without an added margin. Uses request reference audio after #6628; duplex_seed_tts excluded. A3 unchanged.", "mark": [ { "hardware_marks": { @@ -56,56 +57,56 @@ "baseline": { "H100": { "request_throughput": [ - 0.9731, - 1.1986, - 1.686 + 0.9615, + 1.4882, + 1.6245 ], "mean_ttft_ms": [ - 152.5665, - 270.2559, - 301.455 + 156.5006, + 272.8529, + 283.2613 ], "mean_e2el_ms": [ - 1027.2953, - 3306.3735, - 4686.5366 + 1042.7617, + 2669.8806, + 4882.2845 ], "mean_audio_ttfp_ms": [ - 458.1553, - 2030.2335, - 2068.6233 + 585.7387, + 1396.1575, + 2508.4384 ], "mean_audio_rtf": [ - 0.2357, - 0.791, - 1.0324 + 0.2504, + 0.6281, + 1.1787 ] }, "A3": { "request_throughput": [ - 0.5383, - 0.6042, - 0.7547 + 0.4813, + 0.6381, + 0.8728 ], "mean_ttft_ms": [ - 333.2633, - 533.8656, - 547.3388 + 328.182, + 470.0053, + 506.3193 ], "mean_e2el_ms": [ - 1857.2154, - 6600.5101, - 10450.6811 + 2086.7028, + 6480.6866, + 9103.7395 ], "mean_audio_ttfp_ms": [ - 986.4666, - 3411.0754, - 3352.7515 + 1023.3301, + 2734.1486, + 3151.4020 ], "mean_audio_rtf": [ - 0.4423, - 1.5734, - 2.3024 + 0.4659, + 1.4737, + 2.1758 ] } } diff --git a/tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json b/tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json index 1292a79e5..1864bf027 100644 --- a/tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json +++ b/tests/dfx/perf/tests/test_qwen3_omni_async_chunk.json @@ -49,6 +49,11 @@ "mean_audio_rtf": [ 0.1098 ] + }, + "A3": { + "mean_audio_rtf": [ + 0.7014 + ] } } }, @@ -82,6 +87,11 @@ "mean_audio_rtf": [ 0.1335 ] + }, + "A3": { + "mean_audio_rtf": [ + 1.2352 + ] } } }, @@ -117,6 +127,11 @@ "mean_audio_rtf": [ 0.1798 ] + }, + "A3": { + "mean_audio_rtf": [ + 2.5023 + ] } } }, @@ -153,6 +168,11 @@ "mean_e2el_ms": [ 756.8415 ] + }, + "A3": { + "mean_e2el_ms": [ + 2436.1048 + ] } } }, @@ -191,6 +211,11 @@ "mean_e2el_ms": [ 694.9777 ] + }, + "A3": { + "mean_e2el_ms": [ + 2534.4575 + ] } } }, @@ -231,6 +256,11 @@ "mean_e2el_ms": [ 1006.1663 ] + }, + "A3": { + "mean_e2el_ms": [ + 3460.0385 + ] } } } diff --git a/tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json b/tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json index 9fb31e273..012532fef 100644 --- a/tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json +++ b/tests/dfx/perf/tests/test_qwen3_omni_no_async_chunk.json @@ -95,6 +95,11 @@ "mean_audio_rtf": [ 0.138 ] + }, + "A3": { + "mean_audio_rtf": [ + 0.6604 + ] } } }, @@ -128,6 +133,11 @@ "mean_audio_rtf": [ 0.1609 ] + }, + "A3": { + "mean_audio_rtf": [ + 1.0521 + ] } } }, @@ -163,6 +173,11 @@ "mean_audio_rtf": [ 0.2048 ] + }, + "A3": { + "mean_audio_rtf": [ + 1.9804 + ] } } } diff --git a/tests/dfx/perf/tests/test_tts.json b/tests/dfx/perf/tests/test_tts.json index c627c01e8..5fac5d9ef 100644 --- a/tests/dfx/perf/tests/test_tts.json +++ b/tests/dfx/perf/tests/test_tts.json @@ -46,6 +46,14 @@ "median_audio_rtf": [ 0.1591 ] + }, + "A3": { + "median_audio_ttfp_ms": [ + 220.3239 + ], + "median_audio_rtf": [ + 0.3995 + ] } } }, @@ -85,6 +93,23 @@ 40.3552, 43.2572 ] + }, + "A3": { + "median_audio_ttfp_ms": [ + 345.5784, + 451.6758, + 1844.1087 + ], + "median_audio_rtf": [ + 0.6414, + 0.7981, + 1.7064 + ], + "audio_throughput": [ + 12.2961, + 18.9478, + 31.6908 + ] } } }, diff --git a/tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json b/tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json index 0aa060b4c..4ef4ecbed 100644 --- a/tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json +++ b/tests/dfx/perf/tests/test_wan22_i2v_vllm_omni.json @@ -125,6 +125,11 @@ "request_throughput": 0.05263, "mean_e2el_ms": 18941.5, "mean_peak_memory_mb": 50053.7 + }, + "A3": { + "request_throughput": 0.0155, + "mean_e2el_ms": 64626.4676, + "mean_peak_memory_mb": 59932.5714 } } }, @@ -161,6 +166,11 @@ "request_throughput": 0.009595, "mean_e2el_ms": 104105.8, "mean_peak_memory_mb": 59718.0 + }, + "A3": { + "request_throughput": 0.0021, + "mean_e2el_ms": 469604.3521, + "mean_peak_memory_mb": 59889.7714 } } }