[diffusion] fix: fix ZImage SP sharding for 5D latents and unpad frames (#16418)

Signed-off-by: Chi <chixie.mcisaac@gmail.com>
This commit is contained in:
Chi McIsaac
2026-01-07 13:24:04 +08:00
committed by GitHub
parent 52de807dd7
commit f066036c8b
3 changed files with 52 additions and 0 deletions
@@ -529,6 +529,31 @@
"expected_avg_denoise_ms": 94.15,
"expected_median_denoise_ms": 102.03
},
"zimage_image_t2i_2_gpus": {
"stages_ms": {
"InputValidationStage": 0.08,
"TextEncodingStage": 118.66,
"ConditioningStage": 0.01,
"TimestepPreparationStage": 1.5,
"LatentPreparationStage": 0.12,
"DenoisingStage": 1304.07,
"DecodingStage": 37.83
},
"denoise_step_ms": {
"0": 49.76,
"1": 155.22,
"2": 155.98,
"3": 156.16,
"4": 157.04,
"5": 156.54,
"6": 156.29,
"7": 157.36,
"8": 156.05
},
"expected_e2e_ms": 1464.87,
"expected_avg_denoise_ms": 144.49,
"expected_median_denoise_ms": 156.16
},
"zimage_image_t2i_warmup": {
"stages_ms": {
"InputValidationStage": 0.02,
@@ -578,6 +578,16 @@ TWO_GPU_CASES_B = [
),
T2I_sampling_params,
),
DiffusionTestCase(
"zimage_image_t2i_2_gpus",
DiffusionServerArgs(
model_path="Tongyi-MAI/Z-Image-Turbo",
modality="image",
num_gpus=2,
ulysses_degree=2,
),
T2I_sampling_params,
),
DiffusionTestCase(
"flux_image_t2i_2_gpus",
DiffusionServerArgs(