mirror of
https://github.com/comfyanonymous/ComfyUI.git
synced 2026-06-17 13:30:16 +08:00
bece6b2aec
- ModelPatcher.deepclone_multigpu: remove copy.deepcopy fallback. Require cached_patcher_init (raise a descriptive RuntimeError if missing) and always go through clone(model_override=...) with empty backup containers so the per-device clone owns a pristine, unpatched module instead of a deepcopy of an already-loaded/already-patched one. Also call register_load_device on the new patcher so ModelPatcherDynamic per-device bookkeeping (e.g. dynamic_pins) is populated for the requested load device. - comfy/sd.py: register cached_patcher_init on the CLIP and VAE patchers returned by load_checkpoint_guess_config, and on the patcher returned by load_diffusion_model's companion paths. Add load_checkpoint_clip_patcher, load_checkpoint_vae_patcher, and load_vae_patcher reload helpers so the same loader context can be reused to produce per-device clones. - nodes.py: VAELoader registers cached_patcher_init on the produced VAE's patcher when there is a single backing file (skip for pixel_space and composite image-TAESDs which aren't addressable by a single path). - comfy_extras/nodes_multigpu.py: SelectModelDevice / SelectCLIPDevice / SelectVAEDevice now retarget via deepclone_multigpu when the requested device differs from the current load_device, so the consumed model is not just relabeled but actually rehomed onto the chosen device. Verified on runner-2 (2x RTX 4090, comfy-aimdo 0.4.4): - 10/10 focused unit tests (deepclone behavior, missing-factory error path, Select*Device behavior). - Device-switch-after-consumption end-to-end (SD1.5) produces bit-identical PNGs on cuda:0 and cuda:1. - Z Image multigpu CFG split: ~1.90x speedup (10.5s vs 19.9s steady). - Qwen Image multigpu CFG split (real text negative, cfg=4): ~1.69x speedup (32.5s vs 54.8s steady) -- matches pre-refactor numbers. - Baseline (patch stashed) and patched produce identical timings on both models, so the refactor is performance-neutral. Amp-Thread-ID: https://ampcode.com/threads/T-019e5783-b810-74b1-8ca9-09d675de1479 Co-authored-by: Amp <amp@ampcode.com> |
||
|---|---|---|
| .. | ||
| audio_encoders | ||
| background_removal | ||
| cldm | ||
| comfy_types | ||
| extra_samplers | ||
| image_encoders | ||
| k_diffusion | ||
| ldm | ||
| sd1_tokenizer | ||
| t2i_adapter | ||
| taesd | ||
| text_encoders | ||
| weight_adapter | ||
| bg_removal_model.py | ||
| cli_args.py | ||
| clip_config_bigg.json | ||
| clip_model.py | ||
| clip_vision_config_g.json | ||
| clip_vision_config_h.json | ||
| clip_vision_config_vitl_336_llava.json | ||
| clip_vision_config_vitl_336.json | ||
| clip_vision_config_vitl.json | ||
| clip_vision_siglip2_base_naflex.json | ||
| clip_vision_siglip_384.json | ||
| clip_vision_siglip_512.json | ||
| clip_vision.py | ||
| conds.py | ||
| context_windows.py | ||
| controlnet.py | ||
| deploy_environment.py | ||
| diffusers_convert.py | ||
| diffusers_load.py | ||
| float.py | ||
| gligen.py | ||
| hooks.py | ||
| latent_formats.py | ||
| lora_convert.py | ||
| lora.py | ||
| memory_management.py | ||
| model_base.py | ||
| model_detection.py | ||
| model_management.py | ||
| model_patcher.py | ||
| model_prefetch.py | ||
| model_sampling.py | ||
| multigpu.py | ||
| nested_tensor.py | ||
| ops.py | ||
| options.py | ||
| patcher_extension.py | ||
| pinned_memory.py | ||
| pixel_space_convert.py | ||
| quant_ops.py | ||
| rmsnorm.py | ||
| sample.py | ||
| sampler_helpers.py | ||
| samplers.py | ||
| sd1_clip_config.json | ||
| sd1_clip.py | ||
| sd.py | ||
| sdxl_clip.py | ||
| supported_models_base.py | ||
| supported_models.py | ||
| utils.py | ||