2282 Commits

Include renames

Author	SHA1	Message	Date
comfyanonymousandGitHub	4454fab7f0	Remove code to support RMSNorm on old pytorch. (#12499)	2026-02-16 20:09:24 -05:00
comfyanonymousandGitHub	88e6370527	Remove workaround for old pytorch. (#12480)	2026-02-15 20:43:53 -05:00
rattusandGitHub	c0370044cd	MPDynamic: force load flux img_in weight (Fixes flux1 canny+depth lora crash) (#12446) ... * lora: add weight shape calculations. This lets the loader know if a lora will change the shape of a weight so it can take appropriate action. * MPDynamic: force load flux img_in weight This weight is a bit special, in that the lora changes its geometry. This is rather unique, not handled by existing estimate and doesn't work for either offloading or dynamic_vram. Fix for dynamic_vram as a special case. Ideally we can fully precalculate these lora geometry changes at load time, but just get these models working first.	2026-02-15 20:30:09 -05:00
comfyanonymousandGitHub	e1ede29d82	Remove unsafe pickle loading code that was used on pytorch older than 2.4 (#12473) ... ComfyUI hasn't started on pytorch 2.4 since last month.	2026-02-14 22:53:52 -05:00
krigetaandGitHub	dc9822b7df	Add working Qwen 2512 ControlNet (Fun ControlNet) support (#12359)	2026-02-13 22:23:52 -05:00
comfyanonymousandGitHub	712efb466b	Add left padding to LTXAV text encoder. (#12456)	2026-02-13 21:56:54 -05:00
comfyanonymousandGitHub	726af73867	Fix some custom nodes. (#12455)	2026-02-13 20:21:10 -05:00
comfyanonymousandGitHub	831351a29e	Support generating attention masks for left padded text encoders. (#12454)	2026-02-13 20:15:23 -05:00
comfyanonymousandGitHub	e1add563f9	Use torch RMSNorm for flux models and refactor hunyuan video code. (#12432)	2026-02-13 15:35:13 -05:00
rattusandGitHub	8902907d7a	dynamic_vram: Training fixes (#12442)	2026-02-13 15:29:37 -05:00
rattusandGitHub	ae79e33345	llama: use a more efficient rope implementation (#12434) ... Get rid of the cat and unary negation and inplace add-cmul the two halves of the rope. Precompute -sin once at the start of the model rather than every transformer block. This is slightly faster on both GPU and CPU bound setups.	2026-02-12 19:56:42 -05:00
rattusandGitHub	117e214354	ModelPatcherDynamic: force load non leaf weights (#12433) ... The current behaviour of the default ModelPatcher is to .to a model only if its fully loaded, which is how random non-leaf weights get loaded in non-LowVRAM conditions. The however means they never get loaded in dynamic_vram. In the dynamic_vram case, force load them to the GPU.	2026-02-12 19:51:50 -05:00
askmyteapotandGitHub	e5ae670a40	Update ace15.py to allow min_p sampling (#12373)	2026-02-11 20:28:48 -05:00
rattusandGitHub	3fe61cedda	model_patcher: guard against none model_dtype (#12410) ... Handle the case where the _model_dtype exists but is none with the intended fallback.	2026-02-11 14:54:02 -05:00
rattusandGitHub	2a4328d639	ace15: Use dynamic_vram friendly trange (#12409) ... Factor out the ksampler trange and use it in ACE LLM to prevent the silent stall at 0 and rate distortion due to first-step model load.	2026-02-11 14:53:42 -05:00
rattusandGitHub	d297a749a2	dynamic_vram: Fix windows Aimdo crash + Fix LLM performance (#12408) ... * model_management: lazy-cache aimdo_tensor These tensors cosntructed from aimdo-allocations are CPU expensive to make on the pytorch side. Add a cache version that will be valid with signature match to fast path past whatever torch is doing. * dynamic_vram: Minimize fast path CPU work Move as much as possible inside the not resident if block and cache the formed weight and bias rather than the flat intermediates. In extreme layer weight rates this adds up.	2026-02-11 14:50:16 -05:00
comfyanonymousandGitHub	76a7fa96db	Make built in lora training work on anima. (#12402)	2026-02-10 22:04:32 -05:00
Kohaku-BlueleafandGitHub	cdcf4119b3	[Trainer] training with proper offloading (#12189) ... * Fix bypass dtype/device moving * Force offloading mode for training * training context var * offloading implementation in training node * fix wrong input type * Support bypass load lora model, correct adapter/offloading handling	2026-02-10 21:45:19 -05:00
rattusandGitHub	123a7874a9	ops: Fix vanilla-fp8 loaded lora quality (#12390) ... This was missing the stochastic rounding required for fp8 downcast to be consistent with model_patcher.patch_weight_to_device. Missed in testing as I spend too much time with quantized tensors and overlooked the simpler ones.	2026-02-10 13:38:28 -05:00
rattusandGitHub	f719f9c062	sd: delay VAE dtype archive until after override (#12388) ... VAEs have host specific dtype logic that should override the dynamic _model_dtype. Defer the archiving of model dtypes until after.	2026-02-10 13:37:46 -05:00
rattusandGitHub	fe053ba5eb	mp: dont deep-clone objects from model_options (#12382) ... If there are non-trivial python objects nested in the model_options, this causes all sorts of issues. Traverse lists and dicts so clones can safely overide settings and BYO objects but stop there on the deepclone.	2026-02-10 13:37:17 -05:00
Tashdid KhanandClaude Opus 4.6	e1ec301db3	Use NORMAL_VRAM instead of SHARED for MPS memory management ... On Apple Silicon, SHARED vram state is treated like HIGH_VRAM in several code paths, which keeps all models loaded on the GPU device. Since MPS uses unified memory (GPU memory IS system memory), this starves the system of memory for inference activations when running large models (e.g. Wan 2.2 14B I2V loads 13.6GB model + 21.6GB text encoder = 35GB, leaving only 13GB for activations on a 48GB system). This caused GPU hangs and kernel panics on macOS when generating video with 14B parameter models. Changes: - Set MPS vram_state to NORMAL_VRAM instead of SHARED, enabling smart memory management to offload unused models during inference - Remove SHARED from unet_inital_load_device HIGH_VRAM fast path, so model loading respects available memory on MPS - Remove MPS shortcut in text_encoder_initial_device that forced the text encoder to stay on GPU regardless of memory pressure On unified memory, "offloading to CPU" is nearly free (same physical memory pool, just a device pointer change), so NORMAL_VRAM behavior is strictly better — it lets ComfyUI intelligently manage what stays on the GPU device based on available memory. Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>	2026-02-09 23:50:46 -05:00
comfyanonymousandGitHub	a4be04c5d7	Ace step prompts match now. (#12376)	2026-02-09 19:45:56 -05:00
bleppingandGitHub	baf8c87455	Iimprovements to ACE-Steps 1.5 text encoding (part 2) (#12350)	2026-02-09 19:41:49 -05:00
rattusandGitHub	62315fbb15	Dynamic VRAM fixes - Ace 1.5 performance + a VRAM leak (#12368) ... * revert threaded model loader change This change was only needed to get around the pytorch 2.7 mempool bugs, and should have been reverted along with #12260. This fixes a different memory leak where pytorch gets confused about cache emptying. * load non comfy weights * MPDynamic: Pre-generate the tensors for vbars Apparently this is an expensive operation that slows down things. * bump to aimdo 1.8 New features: watermark limit feature logging enhancements -O2 build on linux	2026-02-09 16:16:08 -05:00
comfyanonymousandGitHub	f350a84261	Disable prompt weights for ltxv2. (#12354)	2026-02-07 19:16:28 -05:00
comfyanonymousandGitHub	17e7df43d1	Pad ace step 1.5 ref audio if not long enough. (#12341)	2026-02-07 00:02:11 -05:00
comfyanonymousandGitHub	039955c527	Some fixes to previous pr. (#12339)	2026-02-06 20:14:52 -05:00
tdrussellandGitHub	6a26328842	Support fp16 for Cosmos-Predict2 and Anima (#12249)	2026-02-06 20:12:15 -05:00
comfyanonymousandGitHub	204e65b8dc	Fix bug with last pr (#12338)	2026-02-06 19:48:20 -05:00
asagi4andGitHub	a831c19b70	Fix return_word_ids=True with Anima tokenizer (#12328)	2026-02-06 19:38:04 -05:00
comfyanonymousandGitHub	eba6c940fd	Make ace step 1.5 base model work properly with default workflow. (#12337)	2026-02-06 19:14:56 -05:00
comfyanonymousandGitHub	c2d7f07dbf	Fix issue when using disable_unet_model_creation (#12315)	2026-02-05 19:24:09 -05:00
comfyanonymousandGitHub	458292fef0	Fix some lowvram stuff with ace step 1.5 (#12312)	2026-02-05 19:15:04 -05:00
comfyanonymousandGitHub	6555dc65b8	Make ace step 1.5 work without the llm. (#12311)	2026-02-05 16:43:45 -05:00
comfyanonymousandGitHub	35183543e0	Add VAE tiled decode node for audio. (#12299)	2026-02-05 01:12:04 -05:00
bleppingandGitHub	a246cc02b2	Improvements to ACE-Steps 1.5 text encoding (#12283)	2026-02-05 00:17:37 -05:00
comfyanonymousandGitHub	a50c32d63f	Disable sage attention on ace step 1.5 (#12297)	2026-02-04 22:15:30 -05:00
comfyanonymousandGitHub	6125b80979	Add llm sampling options and make reference audio work on ace step 1.5 (#12295)	2026-02-04 21:29:22 -05:00
comfyanonymousandGitHub	c8fcbd66ee	Try to fix ace text encoder slowness on some configs. (#12290)	2026-02-04 19:37:05 -05:00
comfyanonymousandGitHub	26dd7eb421	Fix ace step nan issue on some hardware/pytorch configs. (#12289)	2026-02-04 18:25:06 -05:00
rattusandGitHub	ef73070ea4	mp: Fix checkpoint saving (#12268) ... Fix regression in the recent model saving refactor. Pass the non unet pieces down the layers so that checkpoints are complete.	2026-02-04 02:08:45 -05:00
rattusandGitHub	d30c609f5a	utils: safetensors: dont slice data on torch level (#12266) ... Torch has alignment enforcement when viewing with data type changes but only relative to itself. Do all tensor constructions straight off the memory-view individually so pytorch doesnt see an alignment problem. The is needed for handling misaligned safetensors weights, which are reasonably common in third party models. This limits usage of this safetensors loader to GPU compute only as CPUs kernnel are very likely to bus error. But it works for dynamic_vram, where we really dont want to take a deep copy and we always use GPU copy_ which disentangles the misalignment.	2026-02-04 01:48:47 -05:00
comfyanonymousandGitHub	a31681564d	Fix crash with ace step 1.5 (#12264)	2026-02-04 00:03:21 -05:00
rattusandGitHub	855849c658	mm: Remove Aimdo exemption for empty_cache (#12260) ... Its more important to get the torch caching allocator GC up and running than supporting the pyt2.7 bug. Switch it on. Defeature dynamic_vram + pyt2.7.	2026-02-03 21:39:19 -05:00
comfyanonymousandGitHub	fe2511468d	Support the 4B ace step 1.5 lm model. (#12257) ... Can be used as an alternative to the 1.7B	2026-02-03 19:01:38 -05:00
comfyanonymousandGitHub	b8315e66cb	Fix tiled vae for ace step 1.5 (#12253)	2026-02-03 14:40:45 -05:00
comfyanonymousandGitHub	ab1050bec3	Support ace step 1.5 base model loras. (#12252)	2026-02-03 13:54:23 -05:00
comfyanonymousandGitHub	85fc35e8fa	Fix mac issue. (#12250)	2026-02-03 12:19:39 -05:00
3 people	223364743c	llama: cast logits as a comfy-weight (#12248) ... This is using a different layers weight with .to(). Change it to use the ops caster if the original layer is a comfy weight so that it picks up dynamic_vram and async_offload functionality in full. Co-authored-by: Rattus <rattus128@gmail.com>	2026-02-03 11:31:36 -05:00