2254 Commits

Author	SHA1	Message	Date
azazeal04	512e7767e8	Merge branch 'Comfy-Org:master' into master	2026-05-06 15:16:22 +02:00
Talmaj	78b3096bf3	Void model - pass 1 & 2 (CORE-38) (#13403)	2026-05-05 19:59:04 -07:00
drozbay	e5369c0eec	feat: Context windows - add causal_window_fix to improve blending of context windows (CORE-100) (#13563) ... * Context windows: add causal_window_fix toggle * Fix slice_cond to correctly handle causal anchor index for temporal offsets	2026-05-05 16:40:53 -07:00
drozbay	1655f8089a	Add temporal_downscale_ratio to LatentFormat (#13702) ... Co-authored-by: ozbayb <17261091+ozbayb@users.noreply.github.com> Co-authored-by: Alexis Rolland <alexisrolland@hotmail.com> Co-authored-by: Jukka Seppänen <40791699+kijai@users.noreply.github.com> Co-authored-by: Jedrzej Kosinski <kosinkadink1@gmail.com>	2026-05-05 16:30:00 -07:00
azazeal04	fc06fb6cd9	Merge branch 'Comfy-Org:master' into master	2026-05-05 12:23:51 +02:00
Talmaj	fed8d5efa6	feat: Auto-regressive video generation (CORE-25) (#13082)	2026-05-04 21:01:22 -07:00
Jedrzej Kosinski	e758594e3b	Add deploy environment header (Comfy-Env) to partner node API calls (#13425)	2026-05-04 20:17:56 -07:00
Jedrzej Kosinski	ae457da84b	feat: add generic --feature-flag CLI arg and --list-feature-flags registry (#13685)	2026-05-04 19:50:26 -07:00
rattus	1265955b34	ops: handle multi-compute of the same weight (#13705) ... If the same weight is used multiple times within the same prefetch window, it should only apply compute state mutations once. Mark the weight as fully resident on the first pass accordingly.	2026-05-04 16:40:57 -07:00
rattus	1ac78180b3	make control-net load order deterministic (#13701) ... Make this deterministic so speeds dont change base of load order. Load them in reverse order so whatever the caller lists first is the top priority.	2026-05-04 12:58:06 -07:00
azazeal04	3e499174dc	Merge branch 'Comfy-Org:master' into master	2026-05-04 21:56:22 +02:00
rattus	c47633f3be	prefetch: guard against no offload (#13703) ... cast_ will return no stream if there is no work to do. guard against this is the consume logic.	2026-05-04 12:56:05 -07:00
Silver	b138133ffa	Enable triton comfy kitchen via cli-arg (#12730)	2026-05-03 14:07:21 -04:00
Jukka Seppänen	be95871adc	feat: Gemma4 text generation support (CORE-30) (#13376) ... * initial gemma4 support * parity with reference implementation outputs can 100% match transformers with same sdpa flags, checkpoint this and then optimize * Cleanup, video fixes * cleanup, enable fused rms norm by default * update comment * Cleanup * Update sd.py * Various fixes * Add fp8 scaled embedding support * small fixes * Translate think tokens * Fix image encoder attention mask type So it works with basic attention * Handle thinking tokens different only for Gemma4 * Code cleanup * Update nodes_textgen.py * Use embed scale class instead of buffer Slight difference to HF, but technically more accurate and simpler code * Default to fused rms_norm * Update gemma4.py	2026-05-02 22:46:15 -04:00
rattus	783782d5d7	Implement block prefetch + Lora Async load + and adopt in LTX (Speedup!) (CORE-111) (#13618) ... * mm: Use Aimdo raw allocator for cast buffers pytorch manages allocation of growing buffers on streams poorly. Pyt has no windows support for the expandable segments allocator (which is the right tool for this job), while also segmenting the memory by stream such that it can be generally re-used. So kick the problem to aimdo which can just grow a virtual region thats freed per stream. * plan * ops: move cpu handler up to the caller * ops: split up prefetch from weight prep block prefetching API Split up the casting and weight formating/lora stuff in prep for arbitrary prefetch support. * ops: implement block prefetching API allow a model to construct a prefetch list and operate it for increased async offload. * ltxv2: Implement block prefetching * Implement lora async offload Implement async offload of loras.	2026-05-02 19:23:24 -04:00
Simon Lui	63103d519e	Remove IPEX and clean up checks and add missing synchronize during empty cache. (#13653)	2026-05-01 14:16:41 -07:00
azazeal04	4c654d1d50	Merge branch 'master' into master	2026-05-01 12:27:56 +02:00
Talmaj	cf9cbec596	Reformat models variable into multiline array CORE-59 (#13513) ... Co-authored-by: Talmaj Marinc <talmaj@comfy.org>	2026-05-01 17:20:11 +08:00
Rainer	e9c311b245	OneTainer ERNIE LoRA support (#13640)	2026-04-30 19:33:41 -04:00
blepping	a164c82913	Add high quality preview support for Flux2 latents (#13496)	2026-04-29 19:37:30 -04:00
Talmaj	5eeae3f1d8	Cogvideox (#13402) ... --------- Co-authored-by: kijai <40791699+kijai@users.noreply.github.com> Co-authored-by: Talmaj Marinc <talmaj@comfy.org>	2026-04-29 19:30:08 -04:00
Jukka Seppänen	0e25a6936e	Reduce video tiny VAE peak VRAM and decode time (CORE-127) (#13617) ... * Update taehv.py * Simplify * Simplify pixel_unshuffle dispatch	2026-04-29 12:15:10 -07:00
rattus	fce0398470	dynamicVRAM + --cache-ram 2 (CORE-117) (#13603) ... * pinned_memory: remove JIT RAM pressure release This doesn't work, as freeing intermediates for pins needs to be higher-priority than freeing pins-for-pins if and when you are going to do that. So this is too late as pins-for-pins is model load time and we dont have JIT pins-for-pins. * cacheing: Add a filter to only free intermediates from inactive wfs This is to get priorities in amongst pins straight. * mm: free inactive-ram from RAM cache first Stuff from inactive workflows should be freed before anything else. * caching: purge old ModelPatchers first Dont try and score them, just dump them at the first sign of trouble if they arent part of the workflow.	2026-04-28 19:15:02 -04:00
rattus	b47f15f25a	fix: Handle un-inited meta-tensors in models (fixes a CPU TE crash) (CORE-67) (#13578)	2026-04-27 22:22:31 -04:00
azazeal04	801243b73c	Merge branch 'master' into master	2026-04-25 13:14:31 +02:00
Jukka Seppänen	084e08c6e2	Disable sageattention for SAM3 (#13529) ... Causes Nans	2026-04-23 11:14:42 -07:00
Jukka Seppänen	749d5b4e8d	feat: SAM (segment anything) 3.1 support (CORE-34) (#13408)	2026-04-23 00:07:43 -04:00
rattus	ec4b1659ab	ModelPatcherDynamic: force cast stray weights on comfy layers (#13487) ... the mixed_precision ops can have input_scale parameters that are used in tensor math but arent a weight or bias so dont get proper VRAM management. Treat these as force-castable parameters like the non comfy weight, random params are buffers already are.	2026-04-22 18:13:38 -04:00
blepping	9949c19c63	Derive InterruptProcessingException from BaseException (#13523)	2026-04-22 18:08:19 -04:00
Octopus	cc6f9500a1	fix: use Parameter assignment for Stable_Zero123 cc_projection weights (fixes #13492) (#13518) ... On Windows with aimdo enabled, disable_weight_init.Linear uses lazy initialization that sets weight and bias to None to avoid unnecessary memory allocation. This caused a crash when copy_() was called on the None weight attribute in Stable_Zero123.__init__. Replace copy_() with direct torch.nn.Parameter assignment, which works correctly on both Windows (aimdo enabled) and other platforms.	2026-04-22 15:05:43 -07:00
Jukka Seppänen	eb22225387	Support standalone LTXV audio VAEs (#13499)	2026-04-21 10:46:37 -07:00
comfyanonymous	ad94d47221	Make the ltx audio vae more native. (#13486)	2026-04-21 11:02:42 -04:00
azazeal04	c232e313f8	Merge branch 'master' into master	2026-04-21 07:41:31 +02:00
comfyanonymous	3d816db07f	Some optimizations to make Ernie inference a bit faster. (#13472)	2026-04-18 23:02:29 -04:00
Jukka Seppänen	b9dedea57d	feat: SUPIR model support (CORE-17) (#13250)	2026-04-18 23:02:01 -04:00
Bedovyy	b41ab53b6f	Use ErnieTEModel_ not ErnieTEModel. (#13431)	2026-04-16 10:11:58 -04:00
Jun Yamog	1de83f91c3	Fix OOM regression in _apply() for quantized models during inference (#13372) ... Skip unnecessary clone of inference-mode tensors when already inside torch.inference_mode(), matching the existing guard in set_attr_param. The unconditional clone introduced in 20561aa9 caused transient VRAM doubling during model movement for FP8/quantized models.	2026-04-15 02:10:36 -07:00
comfyanonymous	cb0bbde402	Fix ernie on devices that don't support fp64. (#13414)	2026-04-14 22:54:47 -04:00
comfyanonymous	722bc73319	Make text generation work with ministral model. (#13395) ... Needs template before it works properly.	2026-04-13 20:43:57 -04:00
comfyanonymous	402ff1cdb7	Fix issue with ernie image. (#13393)	2026-04-13 16:38:42 -04:00
comfyanonymous	c2657d5fb9	Fix typo. (#13382)	2026-04-12 23:37:13 -04:00
comfyanonymous	31283d2892	Implement Ernie Image model. (#13369)	2026-04-11 22:29:31 -04:00
comfyanonymous	55ebd287ee	Add a supports_fp64 function. (#13368)	2026-04-11 21:06:36 -04:00
Jukka Seppänen	a134423890	SDPose: resize input always (#13349)	2026-04-10 11:26:55 -10:00
huemin	b615af1c65	Add support for small flux.2 decoder (#13314)	2026-04-07 03:44:18 -04:00
comfyanonymous	40862c0776	Support Ace Step 1.5 XL model. (#13317)	2026-04-07 03:13:47 -04:00
azazeal04	49fef1697c	Replace empty tensor initialization with zeros	2026-04-04 20:21:27 +02:00
azazeal04	5d119f0532	Implement allow_fp16 configuration in model detection ... Add logic to set allow_fp16 based on weight standard deviation	2026-04-04 20:09:10 +02:00
azazeal04	b2e08bb963	Avoid overwriting existing keys in state_dict ... Prevent overwriting existing keys when mapping t_embedder_2 keys to t_embedder.	2026-04-04 20:05:17 +02:00
azazeal04	6965b83a2a	Fix condition and update transformer_options handling	2026-04-04 19:57:37 +02:00
azazeal04	479536e34d	Prevent overwriting existing keys in state_dict ... Ensure new key is added only if it doesn't exist.	2026-04-04 19:45:05 +02:00
azazeal04	a73ac772b5	Fix key mapping for t_embedder compatibility ... Ensure compatibility by mapping keys for t_embedder.	2026-04-04 19:42:55 +02:00
azazeal04	5f041f4e5e	Refactor time embedding and projection layers	2026-04-04 19:38:55 +02:00
azazeal04	de2ff57f3c	Implement _compute_twinflow_adaln method ... Added a method to compute TwinFlow adaLN input with delta-time conditioning.	2026-04-04 18:49:19 +02:00
azazeal04	4fe75dcce8	Add configuration for TwinFlow-Z-Image model	2026-04-04 17:23:46 +02:00
azazeal04	fec629ba3d	Implement twinflow_z_image function	2026-04-04 16:09:22 +02:00
azazeal04	e978c61a79	Implement twinflow_z_image_key_mapping function ... Add twinflow_z_image_key_mapping function for key mapping.	2026-04-04 16:06:41 +02:00
azazeal04	b76510b549	Add TwinFlow_Z_Image model to supported models	2026-04-04 16:03:52 +02:00
azazeal04	dc413071ff	Add functions to check and extract TwinFlow-Z-Image config	2026-04-04 15:59:50 +02:00
azazeal04	84c983a779	Add detection for TwinFlow-Z-Image checkpoints	2026-04-04 15:57:19 +02:00
azazeal04	21ed4a2242	Add TwinFlow_Z_Image class to model_base.py ... Added TwinFlow_Z_Image class for new model integration.	2026-04-04 15:55:09 +02:00
azazeal04	02b5222dcc	Add twinflow_z_image_lora_to_diffusers function	2026-04-04 15:51:35 +02:00
azazeal04	00c4e72a97	Implement TwinFlow-Z-Image LoRA compatibility ... Add compatibility for TwinFlow-Z-Image LoRAs to target t_embedder_2.* keys.	2026-04-04 15:49:32 +02:00
comfyanonymous	0c63b4f6e3	Remove dead code. (#13251)	2026-04-01 20:22:06 -04:00
comfyanonymous	e2ddf28d78	Fix some fp8 scaled checkpoints no longer working. (#13239)	2026-03-31 14:27:17 -07:00
rattus	8d723d2caa	Fix/tweak pinned memory accounting (#13221) ... * mm: Lower windows pin threshold Some workflows have more extranous use of shared GPU memory than is accounted for in the 5% pin headroom. Lower this for safety. * mm: Remove pin count clearing threshold. TOTAL_PINNED_MEMORY is shared between the legacy and aimdo pinning systems, however this catch-all assumes only the legacy system exists. Remove the catch-all as the PINNED_MEMORY buffer is coherent already.	2026-03-29 16:43:24 -07:00
Jukka Seppänen	a500f1edac	CORE-13 feat: Support RT-DETRv4 detection model (#12748)	2026-03-28 23:34:10 -04:00
comfyanonymous	3f77450ef1	Fix #13214 (#13216)	2026-03-28 22:35:59 -04:00
rattus	b353a7c863	Integrate RAM cache with model RAM management (#13173)	2026-03-27 21:34:16 -04:00
comfyanonymous	3a56201da5	Allow flux conditioning without a pooled output. (#13198)	2026-03-27 20:36:26 -04:00
Jukka Seppänen	b0fd65e884	fix: regression in text generate with LTXAV model (#13170)	2026-03-26 09:55:05 -07:00
comfyanonymous	2a1f402601	Make Qwen 8B work with TextGenerate node. (#13160)	2026-03-25 23:21:44 -04:00
Jukka Seppänen	404d7b9978	feat: Support Qwen3.5 text generation models (#12771)	2026-03-25 22:48:28 -04:00
Kohaku-Blueleaf	5ebb0c2e0b	FP8 bwd training (#13121)	2026-03-24 20:39:04 -04:00
Jukka Seppänen	e87858e974	feat: LTX2: Support reference audio (ID-LoRA) (#13111)	2026-03-23 18:22:24 -04:00
Talmaj	d49420b3c7	LongCat-Image edit (#13003)	2026-03-21 23:51:05 -04:00
rattus	25b6d1d629	wan: vae: Fix light/color change (#13101) ... There was an issue where the resample split was too early and dropped one of the rolling convolutions a frame early. This is most noticable as a lighting/color change between pixel frames 5->6 (latent 2->3), or as a lighting change between the first and last frame in an FLF wan flow.	2026-03-21 18:44:35 -04:00
comfyanonymous	11c15d8832	Fix fp16 intermediates giving different results. (#13100)	2026-03-21 17:53:25 -04:00
comfyanonymous	b5d32e6ad2	Fix sampling issue with fp16 intermediates. (#13099)	2026-03-21 17:47:42 -04:00
Jedrzej Kosinski	87cda1fc25	Move inline comfy.context_windows imports to top-level in model_base.py (#13083) ... The recent PR that added resize_cond_for_context_window methods to model classes used inline 'import comfy.context_windows' in each method body. This moves that import to the top-level import section, replacing 4 duplicate inline imports with a single top-level one.	2026-03-20 20:03:42 -04:00
drozbay	589228e671	Add slice_cond and per-model context window cond resizing (#12645) ... * Add slice_cond and per-model context window cond resizing * Fix cond_value.size() call in context window cond resizing * Expose additional advanced inputs for ContextWindowsManualNode Necessary for WanAnimate context windows workflow, which needs cond_retain_index_list = 0 to work properly with its reference input. ---------	2026-03-19 20:42:42 -07:00
rattus	f49856af57	ltx: vae: Fix missing init variable (#13074) ... Forgot to push this ammendment. Previous test results apply to this.	2026-03-19 22:34:58 -04:00
rattus	82b868a45a	Fix VRAM leak in tiler fallback in video VAEs (#13073) ... * sd: soft_empty_cache on tiler fallback This doesnt cost a lot and creates the expected VRAM reduction in resource monitors when you fallback to tiler. * wan: vae: Don't recursion in local fns (move run_up) Moved Decoder3d’s recursive run_up out of forward into a class method to avoid nested closure self-reference cycles. This avoids cyclic garbage that delays garbage of tensors which in turn delays VRAM release before tiled fallback. * ltx: vae: Don't recursion in local fns (move run_up) Mov the recursive run_up out of forward into a class method to avoid nested closure self-reference cycles. This avoids cyclic garbage that delays garbage of tensors which in turn delays VRAM release before tiled fallback.	2026-03-19 22:30:27 -04:00
comfyanonymous	8458ae2686	Revert "fix: run text encoders on MPS GPU instead of CPU for Apple Silicon (#…" (#13070) ... This reverts commit b941913f1d.	2026-03-19 15:27:55 -04:00
Jukka Seppänen	fd0261d2bc	Reduce tiled decode peak memory (#13050)	2026-03-19 13:29:34 -04:00
rattus	ab14541ef7	memory: Add more exclusion criteria to pinned read (#13067)	2026-03-19 10:03:20 -07:00
rattus	fabed694a2	ltx: vae: implement chunked encoder + CPU IO chunking (Big VRAM reductions) (#13062) ... * ltx: vae: add cache state to downsample block * ltx: vae: Add time stride awareness to causal_conv_3d * ltx: vae: Automate truncation for encoder Other VAEs just truncate without error. Do the same. * sd/ltx: Make chunked_io a flag in its own right Taking this bi-direcitonal, so make it a for-purpose named flag. * ltx: vae: implement chunked encoder + CPU IO chunking People are doing things with big frame counts in LTX including V2V flows. Implement the time-chunked encoder to keep the VRAM down, with the converse of the new CPU pre-allocation technique, where the chunks are brought from the CPU JIT. * ltx: vae-encode: round chunk sizes more strictly Only powers of 2 and multiple of 8 are valid due to cache slicing.	2026-03-19 09:58:47 -07:00
comfyanonymous	f6b869d7d3	fp16 intermediates doen't work for some text enc models. (#13056)	2026-03-18 19:42:28 -04:00
comfyanonymous	56ff88f951	Fix regression. (#13053)	2026-03-18 18:35:25 -04:00
Jukka Seppänen	9fff091f35	Further Reduce LTX VAE decode peak RAM usage (#13052)	2026-03-18 18:32:26 -04:00
comfyanonymous	dcd659590f	Make more intermediate values follow the intermediate dtype. (#13051)	2026-03-18 18:14:18 -04:00
Anton Bukov	b941913f1d	fix: run text encoders on MPS GPU instead of CPU for Apple Silicon (#12809) ... On Apple Silicon, `vram_state` is set to `VRAMState.SHARED` because CPU and GPU share unified memory. However, `text_encoder_device()` only checked for `HIGH_VRAM` and `NORMAL_VRAM`, causing all text encoders to fall back to CPU on MPS devices. Adding `VRAMState.SHARED` to the condition allows non-quantized text encoders (e.g. bf16 Gemma 3 12B) to run on the MPS GPU, providing significant speedup for text encoding and prompt generation. Note: quantized models (fp4/fp8) that use float8_e4m3fn internally will still fall back to CPU via the `supports_cast()` check in `CLIP.__init__()`, since MPS does not support fp8 dtypes.	2026-03-17 21:21:32 -04:00
rattus	cad24ce262	cascade: remove dead weight init code (#13026) ... This weight init process is fully shadowed be the weight load and doesnt work in dynamic_vram were the weight allocation is deferred.	2026-03-17 20:59:10 -04:00
comfyanonymous	68d542cc06	Fix case where pixel space VAE could cause issues. (#13030)	2026-03-17 20:46:22 -04:00
Jukka Seppänen	735a0465e5	Inplace VAE output processing to reduce peak RAM consumption. (#13028)	2026-03-17 20:20:49 -04:00
rattus	035414ede4	Reduce WAN VAE VRAM, Save use cases for OOM/Tiler (#13014) ... * wan: vae: encoder: Add feature cache layer that corks singles If a downsample only gives you a single frame, save it to the feature cache and return nothing to the top level. This increases the efficiency of cacheability, but also prepares support for going two by two rather than four by four on the frames. * wan: remove all concatentation with the feature cache The loopers are now responsible for ensuring that non-final frames are processes at least two-by-two, elimiating the need for this cat case. * wan: vae: recurse and chunk for 2+2 frames on decode Avoid having to clone off slices of 4 frame chunks and reduce the size of the big 6 frame convolutions down to 4. Save the VRAMs. * wan: encode frames 2x2. Reduce VRAM usage greatly by encoding frames 2 at a time rather than 4. * wan: vae: remove cloning The loopers now control the chunking such there is noever more than 2 frames, so just cache these slices directly and avoid the clone allocations completely. * wan: vae: free consumer caller tensors on recursion * wan: vae: restyle a little to match LTX	2026-03-17 17:34:39 -04:00
rattus	1a157e1f97	Reduce LTX VAE VRAM usage and save use cases from OOMs/Tiler (#13013) ... * ltx: vae: scale the chunk size with the users VRAM Scale this linearly down for users with low VRAM. * ltx: vae: free non-chunking recursive intermediates * ltx: vae: cleanup some intermediates The conv layer can be the VRAM peak and it does a torch.cat. So cleanup the pieces of the cat. Also clear our the cache ASAP as each layer detect its end as this VAE surges in VRAM at the end due to the ended padding increasing the size of the final frame convolutions off-the-books to the chunker. So if all the earlier layers free up their cache it can offset that surge. Its a fragmentation nightmare, and the chance of it having to recache the pyt allocator is very high, but you wont OOM.	2026-03-17 17:32:43 -04:00
Paulo Muggler Moreira	8cc746a864	fix: disable SageAttention for Hunyuan3D v2.1 DiT (#12772)	2026-03-16 22:27:27 -04:00
comfyanonymous	ca17fc8355	Fix potential issue. (#13009)	2026-03-16 21:38:40 -04:00
Kohaku-Blueleaf	20561aa919	[Trainer] FP4, 8, 16 training by native dtype support and quant linear autograd function (#12681)	2026-03-16 21:31:50 -04:00