From d62ca45eaa05b6a3800d85c1ebff3ce9249287cf Mon Sep 17 00:00:00 2001 From: CalamitousFelicitousness Date: Wed, 12 Aug 2026 03:17:45 +0100 Subject: [PATCH] refactor(video): load the modular denoiser and text encoder through the shared loaders Selected from the specs the pipe already declares, update_components first and load_components for the rest, so only what the workflow declares gets fetched. Weights land in hfcache. Covers the TODO on component names. Encoder sharing left off: shared_te_map matches a substring of the repo name and 4b hits _dynamic_4bit, redirecting Qwen3-VL to another repo. --- modules/video_models/video_modular.py | 64 ++++++++++++++++----------- pipelines/model_minimax.py | 1 + 2 files changed, 39 insertions(+), 26 deletions(-) diff --git a/modules/video_models/video_modular.py b/modules/video_models/video_modular.py index 6a70ebfb4..e272cb3e1 100644 --- a/modules/video_models/video_modular.py +++ b/modules/video_models/video_modular.py @@ -1,7 +1,7 @@ import time import logging import torch -from modules import shared, errors, devices, model_quant +from modules import shared, errors, devices from modules.logger import log @@ -22,30 +22,44 @@ def is_modular(obj) -> bool: return 'Modular' in cls.__name__ -def component_quant_config(pipe) -> dict: - """Per-component quantization config, read off the pipeline's own component specs. +def preload_components(pipe, workflow: str | None, load_config: dict | None = None) -> dict: + """Load the denoiser and text encoder through the shared loaders rather than the pipeline's own. - Component names differ per architecture, so denoisers and text encoders are - recognized by the class each spec declares rather than listed here. The result - carries no default entry, so anything unrecognized loads unquantized. + `load_components` fetches every component into the pipeline's cache directory with no + single-file override, no shared text encoder and no per-component quantization control. + The shared loaders do all three, and everything they need is already on the spec: repo, + subfolder and class. Components differ per architecture, so each is recognized by the + class its spec declares rather than by name. + + Only what the loaded workflow asks for is fetched, so an unused checkpoint partition is + never pulled. `load_components` afterwards loads whatever is still unset, which is the + tokenizer, processors, schedulers and VAEs. """ - model_args = model_quant.create_config(module='Model') - config = {} - for name, spec in getattr(pipe, '_component_specs', {}).items(): # pylint: disable=protected-access - if getattr(spec, 'default_creation_method', None) != 'from_pretrained': + from pipelines import generic + specs = getattr(pipe, '_component_specs', {}) # pylint: disable=protected-access + loaded = {} + for name in missing_components(pipe, workflow): + spec = specs.get(name) + if spec is None or getattr(spec, 'default_creation_method', None) != 'from_pretrained': continue + repo = getattr(spec, 'pretrained_model_name_or_path', None) cls = getattr(spec, 'type_hint', None) + if not repo or cls is None: + continue origin = getattr(cls, '__module__', '') or '' cls_name = getattr(cls, '__name__', '') or '' - if origin.startswith('transformers') and 'text_encoder' in name: - # MiniMax-H3's conditioner keeps its vision tower unquantized, which other arches on the same - # class do not do; a second modular arch would want this passed in rather than assumed here - te_args = model_quant.create_config(module='TE', modules_to_not_convert=['.model.visual']) - if 'quantization_config' in te_args: - config[name] = te_args['quantization_config'] - elif origin.startswith('diffusers') and ('Transformer' in cls_name or 'UNet' in cls_name) and 'quantization_config' in model_args: - config[name] = model_args['quantization_config'] - return config + subfolder = getattr(spec, 'subfolder', None) or name + component = None + if origin.startswith('diffusers') and ('Transformer' in cls_name or 'UNet' in cls_name): + component = generic.load_transformer(repo, cls_name=cls, load_config=load_config, subfolder=subfolder) + elif origin.startswith('transformers') and 'text_encoder' in name: + # sharing stays off: the shared map matches on class plus a substring of the repo name, so a + # quantized repo can be redirected to an unrelated model's encoder. enable it per arch once + # the pipeline's own encoder is known to be interchangeable with the shared one + component = generic.load_text_encoder(repo, cls_name=cls, load_config=load_config, subfolder=subfolder, allow_shared=False) + if component is not None: + loaded[name] = component + return loaded def missing_components(pipe, workflow: str | None) -> list: @@ -65,7 +79,7 @@ def missing_components(pipe, workflow: str | None) -> list: return [name for name in names if getattr(pipe, name, None) is None] -def load_modular_pipe(repo_cls, repo: str, workflow: str | None = None, revision: str | None = None, offline_args: dict | None = None, base: bool = False): +def load_modular_pipe(repo_cls, repo: str, workflow: str | None = None, revision: str | None = None, offline_args: dict | None = None, base: bool = False, load_config: dict | None = None): if repo_cls is None or isinstance(repo_cls, str): log.error(f'Load modular: repo="{repo}" cls="{repo_cls}" pipeline class not found: diffusers too old') return None @@ -81,16 +95,14 @@ def load_modular_pipe(repo_cls, repo: str, workflow: str | None = None, revision **offline_args, ) # the workflow restricts the component fetch only: passing it to from_pretrained instead would prune the blocks tree to one task and disable runtime dispatch between them - load_kwargs = {} - quant_config = component_quant_config(pipe) - if quant_config: - load_kwargs['quantization_config'] = quant_config - log.debug(f'Load modular: quant={next(iter(quant_config.values())).__class__.__name__} modules={list(quant_config)}') + preloaded = preload_components(pipe, workflow, load_config=load_config) + if preloaded: + pipe.update_components(**preloaded) # registered before the rest, which load_components then skips + log.debug(f'Load modular: preloaded={list(preloaded)}') pipe.load_components( workflow=workflow, dtype=devices.dtype, cache_dir=cache_dir, - **load_kwargs, **offline_args, ) loaded = [name for name, component in pipe.components.items() if component is not None] diff --git a/pipelines/model_minimax.py b/pipelines/model_minimax.py index 5ff95bba9..0d13cd065 100644 --- a/pipelines/model_minimax.py +++ b/pipelines/model_minimax.py @@ -20,6 +20,7 @@ def load_minimax(checkpoint_info, diffusers_load_config=None): # pylint: disable workflow=workflow, offline_args=offline_args, base=True, + load_config=diffusers_load_config, ) if pipe is None: return None