feat: enhance model loading with prefetching and retry logic

- Introduced `from_pretrained_with_prefetch` to handle racy cache scenarios by re-prefetching and retrying on failures.
- Updated various model pipelines to utilize the new prefetching mechanism, ensuring robust loading of model components.
- Prefetch logic added to multiple models, including HunyuanVideo, LongCatImage, and QwenImage, to mitigate issues with incomplete cache states.
This commit is contained in:
tzhouam committed 2026-06-02 09:01:47 +00:00
1 parent b550709bba
commit 617c6ca55f
20 files changed
+678 -194

No files matched your search

@@ -32,6 +32,7 @@ from vllm_omni.diffusion.data import DiffusionOutput, OmniDiffusionConfig
from vllm_omni.diffusion.distributed.cfg_parallel import CFGParallelMixin
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 from_pretrained_with_prefetch, prefetch_subfolders
from vllm_omni.diffusion.models.interface import SupportsComponentDiscovery
from vllm_omni.diffusion.models.omnigen2.omnigen2_transformer import (
OmniGen2RotaryPosEmbed,
@@ -673,23 +674,40 @@ class OmniGen2Pipeline(CFGParallelMixin, nn.Module, SupportsComponentDiscovery):
)
]
# See ``hub_prefetch.py`` for the transformers v5 multi-worker subfolder
# race; prefetch the whole component set before any from_pretrained.
omnigen2_subfolders = ["scheduler", "vae", "mllm", "processor"]
prefetch_subfolders(model, omnigen2_subfolders, local_files_only=local_files_only)
self.scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
model, subfolder="scheduler", local_files_only=local_files_only
)
self.vae = AutoencoderKL.from_pretrained(model, subfolder="vae", local_files_only=local_files_only).to(
self.device
)
self.vae = from_pretrained_with_prefetch(
AutoencoderKL.from_pretrained,
model,
subfolder="vae",
prefetch_list=omnigen2_subfolders,
local_files_only=local_files_only,
).to(self.device)
transformer_kwargs = get_transformer_config_kwargs(od_config.tf_model_config, OmniGen2Transformer2DModel)
self.transformer = OmniGen2Transformer2DModel(
**transformer_kwargs,
quant_config=od_config.quantization_config,
)
self.mllm = Qwen2_5_VLForConditionalGeneration.from_pretrained(
model, subfolder="mllm", local_files_only=local_files_only
self.mllm = from_pretrained_with_prefetch(
Qwen2_5_VLForConditionalGeneration.from_pretrained,
model,
subfolder="mllm",
prefetch_list=omnigen2_subfolders,
local_files_only=local_files_only,
).to(self.device)
self.processor = Qwen2_5_VLProcessor.from_pretrained(
model, subfolder="processor", local_files_only=local_files_only
self.processor = from_pretrained_with_prefetch(
Qwen2_5_VLProcessor.from_pretrained,
model,
subfolder="processor",
prefetch_list=omnigen2_subfolders,
local_files_only=local_files_only,
)
self.vae_scale_factor = (
2 ** (len(self.vae.config.block_out_channels) - 1) if hasattr(self, "vae") and self.vae is not None else 8