[ci][xpu] Record device time in the multimodal_gen perf lane (#39956)

This commit is contained in:
ashwini rathi
2026-09-21 15:39:29 +08:00
committed by GitHub
parent b63f8416b3
commit c2f860af1c
2 changed files with 161 additions and 154 deletions
+3 -1
View File
@@ -251,7 +251,9 @@ jobs:
- name: Run diffusion server tests (1-GPU)
timeout-minutes: 60
run: |
docker exec ci_sglang_xpu bash -c "source /opt/venv/bin/activate && cd /sglang-checkout/python && python3 sglang/multimodal_gen/test/run_suite.py --suite 1-gpu-xpu"
# xpu_b60.json is seeded with this flag on; without it the SYCL queue
# is deep enough that step timings record host enqueue, not device time.
docker exec -e SGLANG_DIFFUSION_SYNC_STAGE_PROFILING=1 ci_sglang_xpu bash -c "source /opt/venv/bin/activate && cd /sglang-checkout/python && python3 sglang/multimodal_gen/test/run_suite.py --suite 1-gpu-xpu"
- name: Cleanup container
if: always()
@@ -2,8 +2,8 @@
"metadata": {
"model": "Diffusion Server",
"hardware": "CI Intel Arc Pro B60 (24 GiB) pool: bmg-multigen-models",
"description": "Reference numbers seeded from XPU multimodal_gen CI run 32736878259 job 97461821112 (PR #36100). wan2_1_t2v_1.3b re-seeded from run 35074312985 job 104723059209 (PR #36825); its expected_avg_denoise_ms still carries the step-0 lazy-init cost.",
"last_updated": "2026-09-17"
"description": "All scenarios re-seeded from XPU multimodal_gen CI run 35563559050 job 106220948594 (PR #39956), with SGLANG_DIFFUSION_SYNC_STAGE_PROFILING=1 so step timings record device time rather than host enqueue. Each value is the max over the 7 attempts of that job. Denoise numbers are stable across attempts (<=2% spread); expected_load_ms is host I/O bound and ranged 21.4-48.0s across attempts, so it is guarded only by the wide e2e tolerance.",
"last_updated": "2026-09-21"
},
"tolerances": {
"long_term": {
@@ -42,181 +42,186 @@
"zimage_image_t2i": {
"estimated_full_test_time_s": 65.1,
"stages_ms": {
"InputValidationStage": 0.03,
"TextEncodingStage": 87.64,
"LatentPreparationStage": 0.21,
"TimestepPreparationStage": 136.9,
"DenoisingStage": 8361.1,
"DecodingStage": 8.39
"InputValidationStage": 0.1,
"TextEncodingStage": 89.87,
"LatentPreparationStage": 0.28,
"TimestepPreparationStage": 2.44,
"DenoisingStage": 8614.3,
"DecodingStage": 293.13
},
"denoise_step_ms": {
"0": 99.57,
"1": 927.61,
"2": 926.72,
"3": 928.32,
"4": 929.29,
"5": 928.8,
"6": 925.2,
"7": 927.22,
"8": 931.28
"0": 862.67,
"1": 859.76,
"2": 861.38,
"3": 860.12,
"4": 859.95,
"5": 859.53,
"6": 859.63,
"7": 860.1,
"8": 859.77
},
"expected_e2e_ms": 9312.75,
"expected_avg_denoise_ms": 836.0,
"expected_median_denoise_ms": 927.61
"expected_e2e_ms": 9799.49,
"expected_load_ms": 33283.41,
"expected_avg_denoise_ms": 860.1,
"expected_median_denoise_ms": 859.76
},
"flux_2_klein_image_t2i": {
"estimated_full_test_time_s": 106.8,
"stages_ms": {
"InputValidationStage": 0.04,
"TextEncodingStage": 37.83,
"ImageVAEEncodingStage": 0.0,
"LatentPreparationStage": 0.33,
"TimestepPreparationStage": 351.26,
"DenoisingStage": 2518.52,
"DecodingStage": 9.13
"InputValidationStage": 0.11,
"TextEncodingStage": 88.26,
"ImageVAEEncodingStage": 0.06,
"LatentPreparationStage": 0.47,
"TimestepPreparationStage": 2.37,
"DenoisingStage": 2803.24,
"DecodingStage": 292.44
},
"denoise_step_ms": {
"0": 83.44,
"1": 16.73,
"2": 18.04,
"3": 17.11
"0": 591.79,
"1": 536.64,
"2": 537.71,
"3": 537.7
},
"expected_e2e_ms": 3509.48,
"expected_avg_denoise_ms": 33.83,
"expected_median_denoise_ms": 17.58
"expected_e2e_ms": 3754.98,
"expected_load_ms": 26566.83,
"expected_avg_denoise_ms": 550.96,
"expected_median_denoise_ms": 537.7
},
"flux_2_klein_base_image_t2i": {
"stages_ms": {
"InputValidationStage": 0.03,
"TextEncodingStage": 36.3,
"ImageVAEEncodingStage": 0.0,
"LatentPreparationStage": 0.28,
"TimestepPreparationStage": 355.5,
"DenoisingStage": 55237.43,
"DecodingStage": 11.09
"InputValidationStage": 0.11,
"TextEncodingStage": 87.34,
"ImageVAEEncodingStage": 0.06,
"LatentPreparationStage": 0.44,
"TimestepPreparationStage": 2.49,
"DenoisingStage": 54080.81,
"DecodingStage": 287.66
},
"denoise_step_ms": {
"0": 106.73,
"1": 34.25,
"2": 34.59,
"3": 34.59,
"4": 35.08,
"5": 34.24,
"6": 34.03,
"7": 34.47,
"8": 34.17,
"9": 44.04,
"10": 34.83,
"11": 34.64,
"12": 35.38,
"13": 34.32,
"14": 35.51,
"15": 36.16,
"16": 35.32,
"17": 34.75,
"18": 35.28,
"19": 50.88,
"20": 35.41,
"21": 35.36,
"22": 34.53,
"23": 35.42,
"24": 34.75,
"25": 34.78,
"26": 34.95,
"27": 34.84,
"28": 35.38,
"29": 34.7,
"30": 34.42,
"31": 35.01,
"32": 34.98,
"33": 34.61,
"34": 34.78,
"35": 34.76,
"36": 34.61,
"37": 37.59,
"38": 35.0,
"39": 34.6,
"40": 35.52,
"41": 35.14,
"42": 34.84,
"43": 35.16,
"44": 33.86,
"45": 34.52,
"46": 34.94,
"47": 33.84,
"48": 35.14,
"49": 34.75
"0": 1123.84,
"1": 1068.88,
"2": 1068.05,
"3": 1067.09,
"4": 1068.74,
"5": 1068.64,
"6": 1067.83,
"7": 1068.84,
"8": 1068.23,
"9": 1068.88,
"10": 1067.99,
"11": 1068.93,
"12": 1069.2,
"13": 1067.82,
"14": 1068.2,
"15": 1068.84,
"16": 1068.15,
"17": 1067.99,
"18": 1068.51,
"19": 1068.77,
"20": 1068.34,
"21": 1068.12,
"22": 1069.19,
"23": 1068.35,
"24": 1069.71,
"25": 1068.26,
"26": 1068.36,
"27": 1069.13,
"28": 1068.15,
"29": 1068.11,
"30": 1069.23,
"31": 1069.6,
"32": 1069.44,
"33": 1068.34,
"34": 1068.38,
"35": 1069.22,
"36": 1068.4,
"37": 1067.92,
"38": 1068.22,
"39": 1068.85,
"40": 1068.95,
"41": 1069.67,
"42": 1068.07,
"43": 1068.93,
"44": 1069.75,
"45": 1067.77,
"46": 1068.09,
"47": 1068.11,
"48": 1068.67,
"49": 1068.8
},
"expected_e2e_ms": 56248.35,
"expected_avg_denoise_ms": 36.83,
"expected_median_denoise_ms": 34.84
"expected_e2e_ms": 55028.04,
"expected_load_ms": 31539.71,
"expected_avg_denoise_ms": 1069.67,
"expected_median_denoise_ms": 1068.45,
"estimated_full_test_time_s": 103.3
},
"wan2_1_t2v_1.3b": {
"estimated_full_test_time_s": 331.5,
"stages_ms": {
"InputValidationStage": 0.05,
"TextEncodingStage": 761.88,
"LatentPreparationStage": 1.12,
"TimestepPreparationStage": 1.66,
"DenoisingStage": 73363.89,
"DecodingStage": 295.76,
"InputValidationStage": 0.11,
"TextEncodingStage": 764.42,
"LatentPreparationStage": 1.26,
"TimestepPreparationStage": 0.97,
"DenoisingStage": 70658.77,
"DecodingStage": 3102.93,
"per_frame_generation": null
},
"denoise_step_ms": {
"0": 3421.66,
"1": 103.21,
"2": 101.69,
"3": 102.97,
"4": 102.86,
"5": 104.51,
"6": 103.23,
"7": 102.74,
"8": 101.86,
"9": 102.28,
"10": 103.18,
"11": 105.43,
"12": 103.01,
"13": 102.33,
"14": 101.37,
"15": 102.32,
"16": 102.29,
"17": 103.88,
"18": 103.13,
"19": 103.36,
"20": 101.79,
"21": 101.59,
"22": 103.18,
"23": 106.52,
"24": 104.18,
"25": 103.53,
"26": 104.71,
"27": 103.92,
"28": 104.26,
"29": 105.03,
"30": 104.62,
"31": 104.0,
"32": 104.37,
"33": 103.74,
"34": 104.16,
"35": 105.51,
"36": 102.68,
"37": 101.05,
"38": 103.52,
"39": 101.97,
"40": 102.62,
"41": 108.73,
"42": 102.88,
"43": 102.2,
"44": 102.84,
"45": 102.31,
"46": 103.4,
"47": 102.44,
"48": 106.32,
"49": 101.64
"0": 1677.96,
"1": 1401.03,
"2": 1400.7,
"3": 1401.14,
"4": 1400.29,
"5": 1401.15,
"6": 1400.76,
"7": 1400.83,
"8": 1399.92,
"9": 1401.37,
"10": 1400.79,
"11": 1401.01,
"12": 1400.59,
"13": 1400.05,
"14": 1400.17,
"15": 1400.44,
"16": 1400.18,
"17": 1400.41,
"18": 1399.86,
"19": 1400.23,
"20": 1400.58,
"21": 1400.54,
"22": 1400.3,
"23": 1400.26,
"24": 1400.32,
"25": 1400.66,
"26": 1400.31,
"27": 1400.01,
"28": 1400.15,
"29": 1400.61,
"30": 1400.58,
"31": 1400.79,
"32": 1400.66,
"33": 1400.92,
"34": 1401.43,
"35": 1400.94,
"36": 1400.91,
"37": 1400.93,
"38": 1401.66,
"39": 1400.7,
"40": 1400.75,
"41": 1401.24,
"42": 1400.85,
"43": 1400.86,
"44": 1400.98,
"45": 1400.81,
"46": 1400.27,
"47": 1400.9,
"48": 1400.3,
"49": 1400.39
},
"expected_e2e_ms": 75344.43,
"expected_avg_denoise_ms": 169.74,
"expected_median_denoise_ms": 103.18
"expected_e2e_ms": 75427.76,
"expected_load_ms": 48009.32,
"expected_avg_denoise_ms": 1406.13,
"expected_median_denoise_ms": 1400.58
}
}
}