Revert "[Quantization] Enable FP8 online quantization for Z-image text encode…"

This reverts commit ac66282746.
This commit is contained in:
Hongsheng Liu authored and GitHub committed 2026-04-30 15:05:26 +08:00
1 parent fa945b2626
commit cf301fe75b
5 files changed
+19 -188

No files matched your search

@@ -28,7 +28,7 @@ from diffusers.image_processor import PipelineImageInput, VaeImageProcessor
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
from diffusers.utils import logging
from diffusers.utils.torch_utils import randn_tensor
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer
from transformers import AutoModel, AutoTokenizer
from vllm.model_executor.models.utils import AutoWeightsLoader
from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
@@ -36,7 +36,6 @@ from vllm_omni.diffusion.distributed.autoencoders.autoencoder_kl import Distribu
from vllm_omni.diffusion.distributed.utils import get_local_device
from vllm_omni.diffusion.model_loader.diffusers_loader import DiffusersPipelineLoader
from vllm_omni.diffusion.model_loader.hub_prefetch import prefetch_subfolders
from vllm_omni.diffusion.models.utils import create_transformers_model
from vllm_omni.diffusion.models.z_image.z_image_transformer import (
ZImageTransformer2DModel,
)
@@ -170,25 +169,13 @@ class ZImagePipeline(nn.Module, DiffusionPipelineProfilerMixin):
super().__init__()
self.od_config = od_config
self.weights_sources = [
DiffusersPipelineLoader.ComponentSource(
model_or_path=od_config.model,
subfolder="text_encoder",
revision=od_config.revision,
prefix="text_encoder.",
),
DiffusersPipelineLoader.ComponentSource(
model_or_path=od_config.model,
subfolder="transformer",
revision=od_config.revision,
revision=None,
prefix="transformer.",
fall_back_to_pt=True,
),
DiffusersPipelineLoader.ComponentSource(
model_or_path=od_config.model,
subfolder="vae",
revision=od_config.revision,
prefix="vae.",
),
)
]
self._execution_device = get_local_device()
model = od_config.model
@@ -205,19 +192,12 @@ class ZImagePipeline(nn.Module, DiffusionPipelineProfilerMixin):
model, subfolder="scheduler", local_files_only=local_files_only
)
text_encoder_config = AutoConfig.from_pretrained(
self.text_encoder = AutoModel.from_pretrained(
model, subfolder="text_encoder", local_files_only=local_files_only
)
self.text_encoder = create_transformers_model(
AutoModelForCausalLM,
od_config,
hf_config=text_encoder_config,
).to(self._execution_device)
if text_encoder_config.tie_word_embeddings:
self.text_encoder.lm_head.weight = self.text_encoder.get_input_embeddings().weight
vae_config = DistributedAutoencoderKL.load_config(model, subfolder="vae", local_files_only=local_files_only)
self.vae = DistributedAutoencoderKL.from_config(vae_config).to(self._execution_device)
self.vae = DistributedAutoencoderKL.from_pretrained(
model, subfolder="vae", local_files_only=local_files_only
).to(self._execution_device)
self.transformer = ZImageTransformer2DModel(quant_config=od_config.quantization_config)
self.tokenizer = AutoTokenizer.from_pretrained(model, subfolder="tokenizer", local_files_only=local_files_only)