Skip to content

Commit feaa9e7

Browse files
[diffusion] fix: replace TextEncoderConfig with Qwen3TextConfig for Z-Image (sgl-project#18560)
Co-authored-by: Mick <mickjagger19@icloud.com>
1 parent c929766 commit feaa9e7

File tree

2 files changed

+12
-14
lines changed

2 files changed

+12
-14
lines changed

python/sglang/multimodal_gen/configs/pipeline_configs/zimage.py

Lines changed: 3 additions & 5 deletions
Original file line numberDiff line numberDiff line change
@@ -7,10 +7,8 @@
77

88
from sglang.multimodal_gen.configs.models import DiTConfig, EncoderConfig, VAEConfig
99
from sglang.multimodal_gen.configs.models.dits.zimage import ZImageDitConfig
10-
from sglang.multimodal_gen.configs.models.encoders import (
11-
BaseEncoderOutput,
12-
TextEncoderConfig,
13-
)
10+
from sglang.multimodal_gen.configs.models.encoders import BaseEncoderOutput
11+
from sglang.multimodal_gen.configs.models.encoders.qwen3 import Qwen3TextConfig
1412
from sglang.multimodal_gen.configs.models.vaes.flux import FluxVAEConfig
1513
from sglang.multimodal_gen.configs.pipeline_configs.base import (
1614
ImagePipelineConfig,
@@ -49,7 +47,7 @@ class ZImagePipelineConfig(ImagePipelineConfig):
4947
dit_config: DiTConfig = field(default_factory=ZImageDitConfig)
5048
vae_config: VAEConfig = field(default_factory=FluxVAEConfig)
5149
text_encoder_configs: tuple[EncoderConfig, ...] = field(
52-
default_factory=lambda: (TextEncoderConfig(),)
50+
default_factory=lambda: (Qwen3TextConfig(),)
5351
)
5452

5553
preprocess_text_funcs: tuple[Callable, ...] = field(

python/sglang/multimodal_gen/test/server/perf_baselines.json

Lines changed: 9 additions & 9 deletions
Original file line numberDiff line numberDiff line change
@@ -525,7 +525,7 @@
525525
"stages_ms": {
526526
"ConditioningStage": 0.01,
527527
"DecodingStage": 33.93,
528-
"TextEncodingStage": 104.21,
528+
"TextEncodingStage": 409.74,
529529
"InputValidationStage": 0.03,
530530
"DenoisingStage": 808.69,
531531
"LatentPreparationStage": 0.14,
@@ -542,14 +542,14 @@
542542
"7": 99.3,
543543
"8": 100.83
544544
},
545-
"expected_e2e_ms": 952.09,
545+
"expected_e2e_ms": 1461.08,
546546
"expected_avg_denoise_ms": 89.36,
547547
"expected_median_denoise_ms": 99.67
548548
},
549549
"zimage_image_t2i_multi_lora": {
550550
"stages_ms": {
551551
"InputValidationStage": 0.04,
552-
"TextEncodingStage": 103.81,
552+
"TextEncodingStage": 413.69,
553553
"ConditioningStage": 0.01,
554554
"TimestepPreparationStage": 1.3,
555555
"LatentPreparationStage": 0.11,
@@ -567,14 +567,14 @@
567567
"7": 100.86,
568568
"8": 103.87
569569
},
570-
"expected_e2e_ms": 955.14,
570+
"expected_e2e_ms": 1464.31,
571571
"expected_avg_denoise_ms": 89.96,
572572
"expected_median_denoise_ms": 99.72
573573
},
574574
"zimage_image_t2i_2_gpus": {
575575
"stages_ms": {
576576
"InputValidationStage": 0.08,
577-
"TextEncodingStage": 118.66,
577+
"TextEncodingStage": 420.74,
578578
"ConditioningStage": 0.01,
579579
"TimestepPreparationStage": 1.5,
580580
"LatentPreparationStage": 0.12,
@@ -599,7 +599,7 @@
599599
"zimage_image_t2i_warmup": {
600600
"stages_ms": {
601601
"InputValidationStage": 0.02,
602-
"TextEncodingStage": 100.65,
602+
"TextEncodingStage": 456.16,
603603
"ConditioningStage": 0.01,
604604
"TimestepPreparationStage": 0.98,
605605
"LatentPreparationStage": 0.06,
@@ -617,7 +617,7 @@
617617
"7": 110.54,
618618
"8": 115.24
619619
},
620-
"expected_e2e_ms": 1029.96,
620+
"expected_e2e_ms": 1184.45,
621621
"expected_avg_denoise_ms": 98.46,
622622
"expected_median_denoise_ms": 109.65
623623
},
@@ -2001,7 +2001,7 @@
20012001
"fsdp-inference": {
20022002
"stages_ms": {
20032003
"InputValidationStage": 0.04,
2004-
"TextEncodingStage": 128.3,
2004+
"TextEncodingStage": 411.12,
20052005
"ConditioningStage": 0.01,
20062006
"TimestepPreparationStage": 1.44,
20072007
"LatentPreparationStage": 0.1,
@@ -2019,7 +2019,7 @@
20192019
"7": 178.53,
20202020
"8": 178.08
20212021
},
2022-
"expected_e2e_ms": 1742.7,
2022+
"expected_e2e_ms": 2103.05,
20232023
"expected_avg_denoise_ms": 173.83,
20242024
"expected_median_denoise_ms": 178.08
20252025
}

0 commit comments

Comments
 (0)