* ops: dont take an offload stream if you dont need one
* ops: prioritize mem transfer
The async offload streams reason for existence is to transfer from
RAM to GPU. The post processing compute steps are a bonus on the side
stream, but if the compute stream is running a long kernel, it can
stall the side stream, as it wait to type-cast the bias before
transferring the weight. So do a pure xfer of the weight straight up,
then do everything bias, then go back to fix the weight type and do
weight patches.
* execution: Roll the UI cache into the outputs
Currently the UI cache is parallel to the output cache with
expectations of being a content superset of the output cache.
At the same time the UI and output cache are maintained completely
seperately, making it awkward to free the output cache content without
changing the behaviour of the UI cache.
There are two actual users (getters) of the UI cache. The first is
the case of a direct content hit on the output cache when executing a
node. This case is very naturally handled by merging the UI and outputs
cache.
The second case is the history JSON generation at the end of the prompt.
This currently works by asking the cache for all_node_ids and then
pulling the cache contents for those nodes. all_node_ids is the nodes
of the dynamic prompt.
So fold the UI cache into the output cache. The current UI cache setter
now writes to a prompt-scope dict. When the output cache is set, just
get this value from the dict and tuple up with the outputs.
When generating the history, simply iterate prompt-scope dict.
This prepares support for more complex caching strategies (like RAM
pressure caching) where less than 1 workflow will be cached and it
will be desirable to keep the UI cache and output cache in sync.
* sd: Implement RAM getter for VAE
* model_patcher: Implement RAM getter for ModelPatcher
* sd: Implement RAM getter for CLIP
* Implement RAM Pressure cache
Implement a cache sensitive to RAM pressure. When RAM headroom drops
down below a certain threshold, evict RAM-expensive nodes from the
cache.
Models and tensors are measured directly for RAM usage. An OOM score
is then computed based on the RAM usage of the node.
Note the due to indirection through shared objects (like a model
patcher), multiple nodes can account the same RAM as their individual
usage. The intent is this will free chains of nodes particularly
model loaders and associate loras as they all score similar and are
sorted in close to each other.
Has a bias towards unloading model nodes mid flow while being able
to keep results like text encodings and VAE.
* execution: Convert the cache entry to NamedTuple
As commented in review.
Convert this to a named tuple and abstract away the tuple type
completely from graph.py.
* mm: factor out the current stream getter
Make this a reusable function.
* ops: sync the offload stream with the consumption of w&b
This sync is nessacary as pytorch will queue cuda async frees on the
same stream as created to tensor. In the case of async offload, this
will be on the offload stream.
Weights and biases can go out of scope in python which then
triggers the pytorch garbage collector to queue the free operation on
the offload stream possible before the compute stream has used the
weight. This causes a use after free on weight data leading to total
corruption of some workflows.
So sync the offload stream with the compute stream after the weight
has been used so the free has to wait for the weight to be used.
The cast_bias_weight is extended in a backwards compatible way with
the new behaviour opt-in on a defaulted parameter. This handles
custom node packs calling cast_bias_weight and defeatures
async-offload for them (as they do not handle the race).
The pattern is now:
cast_bias_weight(... , offloadable=True) #This might be offloaded
thing(weight, bias, ...)
uncast_bias_weight(...)
* controlnet: adopt new cast_bias_weight synchronization scheme
This is nessacary for safe async weight offloading.
* mm: sync the last stream in the queue, not the next
Currently this peeks ahead to sync the next stream in the queue of
streams with the compute stream. This doesnt allow a lot of
parallelization, as then end result is you can only get one weight load
ahead regardless of how many streams you have.
Rotate the loop logic here to synchronize the end of the queue before
returning the next stream. This allows weights to be loaded ahead of the
compute streams position.
* Implement mixed precision operations with a registry design and metadate for quant spec in checkpoint.
* Updated design using Tensor Subclasses
* Fix FP8 MM
* An actually functional POC
* Remove CK reference and ensure correct compute dtype
* Update unit tests
* ruff lint
* Implement mixed precision operations with a registry design and metadate for quant spec in checkpoint.
* Updated design using Tensor Subclasses
* Fix FP8 MM
* An actually functional POC
* Remove CK reference and ensure correct compute dtype
* Update unit tests
* ruff lint
* Fix missing keys
* Rename quant dtype parameter
* Rename quant dtype parameter
* Fix unittests for CPU build
Same change pattern as 7e8dd275c2
applied to WAN2.2
If this suffers an exception (such as a VRAM oom) it will leave the
encode() and decode() methods which skips the cleanup of the WAN
feature cache. The comfy node cache then ultimately keeps a reference
this object which is in turn reffing large tensors from the failed
execution.
The feature cache is currently setup at a class variable on the
encoder/decoder however, the encode and decode functions always clear
it on both entry and exit of normal execution.
Its likely the design intent is this is usable as a streaming encoder
where the input comes in batches, however the functions as they are
today don't support that.
So simplify by bringing the cache back to local variable, so that if
it does VRAM OOM the cache itself is properly garbage when the
encode()/decode() functions dissappear from the stack.
## Summary
Fixed incorrect type hint syntax in `MotionEncoder_tc.__init__()` parameter list.
## Changes
- Line 647: Changed `num_heads=int` to `num_heads: int`
- This corrects the parameter annotation from a default value assignment to proper type hint syntax
## Details
The parameter was using assignment syntax (`=`) instead of type annotation syntax (`:`), which would incorrectly set the default value to the `int` class itself rather than annotating the expected type.
If this suffers an exception (such as a VRAM oom) it will leave the
encode() and decode() methods which skips the cleanup of the WAN
feature cache. The comfy node cache then ultimately keeps a reference
this object which is in turn reffing large tensors from the failed
execution.
The feature cache is currently setup at a class variable on the
encoder/decoder however, the encode and decode functions always clear
it on both entry and exit of normal execution.
Its likely the design intent is this is usable as a streaming encoder
where the input comes in batches, however the functions as they are
today don't support that.
So simplify by bringing the cache back to local variable, so that if
it does VRAM OOM the cache itself is properly garbage when the
encode()/decode() functions dissappear from the stack.
When the VAE catches this VRAM OOM, it launches the fallback logic
straight from the exception context.
Python however refs the entire call stack that caused the exception
including any local variables for the sake of exception report and
debugging. In the case of tensors, this can hold on the references
to GBs of VRAM and inhibit the VRAM allocated from freeing them.
So dump the except context completely before going back to the VAE
via the tiler by getting out of the except block with nothing but
a flag.
The greately increases the reliability of the tiler fallback,
especially on low VRAM cards, as with the bug, if the leak randomly
leaked more than the headroom needed for a single tile, the tiler
would fallback would OOM and fail the flow.
* flux: math: Use _addcmul to avoid expensive VRAM intermediate
The rope process can be the VRAM peak and this intermediate
for the addition result before releasing the original can OOM.
addcmul_ it.
* wan: Delete the self attention before cross attention
This saves VRAM when the cross attention and FFN are in play as the
VRAM peak.
When unloading models in load_models_gpu(), the model finalizer was not
being explicitly detached, leading to a memory leak. This caused
linear memory consumption increase over time as models are repeatedly
loaded and unloaded.
This change prevents orphaned finalizer references from accumulating in
memory during model switching operations.
* flux: Do the xq and xk ropes one at a time
This was doing independendent interleaved tensor math on the q and k
tensors, leading to the holding of more than the minimum intermediates
in VRAM. On a bad day, it would VRAM OOM on xk intermediates.
Do everything q and then everything k, so torch can garbage collect
all of qs intermediates before k allocates its intermediates.
This reduces peak VRAM usage for some WAN2.2 inferences (at least).
* wan: Optimize qkv intermediates on attention
As commented. The former logic computed independent pieces of QKV in
parallel which help more inference intermediates in VRAM spiking
VRAM usage. Fully roping Q and garbage collecting the intermediates
before touching K reduces the peak inference VRAM usage.
* Initial Chroma Radiance support
* Minor Chroma Radiance cleanups
* Update Radiance nodes to ensure latents/images are on the intermediate device
* Fix Chroma Radiance memory estimation.
* Increase Chroma Radiance memory usage factor
* Increase Chroma Radiance memory usage factor once again
* Ensure images are multiples of 16 for Chroma Radiance
Add batch dimension and fix channels when necessary in ChromaRadianceImageToLatent node
* Tile Chroma Radiance NeRF to reduce memory consumption, update memory usage factor
* Update Radiance to support conv nerf final head type.
* Allow setting NeRF embedder dtype for Radiance
Bump Radiance nerf tile size to 32
Support EasyCache/LazyCache on Radiance (maybe)
* Add ChromaRadianceStubVAE node
* Crop Radiance image inputs to multiples of 16 instead of erroring to be in line with existing VAE behavior
* Convert Chroma Radiance nodes to V3 schema.
* Add ChromaRadianceOptions node and backend support.
Cleanups/refactoring to reduce code duplication with Chroma.
* Fix overriding the NeRF embedder dtype for Chroma Radiance
* Minor Chroma Radiance cleanups
* Move Chroma Radiance to its own directory in ldm
Minor code cleanups and tooltip improvements
* Fix Chroma Radiance embedder dtype overriding
* Remove Radiance dynamic nerf_embedder dtype override feature
* Unbork Radiance NeRF embedder init
* Remove Chroma Radiance image conversion and stub VAE nodes
Add a chroma_radiance option to the VAELoader builtin node which uses comfy.sd.PixelspaceConversionVAE
Add a PixelspaceConversionVAE to comfy.sd for converting BHWC 0..1 <-> BCHW -1..1
* Looking into a @wrap_attn decorator to look for 'optimized_attention_override' entry in transformer_options
* Created logging code for this branch so that it can be used to track down all the code paths where transformer_options would need to be added
* Fix memory usage issue with inspect
* Made WAN attention receive transformer_options, test node added to wan to test out attention override later
* Added **kwargs to all attention functions so transformer_options could potentially be passed through
* Make sure wrap_attn doesn't make itself recurse infinitely, attempt to load SageAttention and FlashAttention if not enabled so that they can be marked as available or not, create registry for available attention
* Turn off attention logging for now, make AttentionOverrideTestNode have a dropdown with available attention (this is a test node only)
* Make flux work with optimized_attention_override
* Add logs to verify optimized_attention_override is passed all the way into attention function
* Make Qwen work with optimized_attention_override
* Made hidream work with optimized_attention_override
* Made wan patches_replace work with optimized_attention_override
* Made SD3 work with optimized_attention_override
* Made HunyuanVideo work with optimized_attention_override
* Made Mochi work with optimized_attention_override
* Made LTX work with optimized_attention_override
* Made StableAudio work with optimized_attention_override
* Made optimized_attention_override work with ACE Step
* Made Hunyuan3D work with optimized_attention_override
* Make CosmosPredict2 work with optimized_attention_override
* Made CosmosVideo work with optimized_attention_override
* Made Omnigen 2 work with optimized_attention_override
* Made StableCascade work with optimized_attention_override
* Made AuraFlow work with optimized_attention_override
* Made Lumina work with optimized_attention_override
* Made Chroma work with optimized_attention_override
* Made SVD work with optimized_attention_override
* Fix WanI2VCrossAttention so that it expects to receive transformer_options
* Fixed Wan2.1 Fun Camera transformer_options passthrough
* Fixed WAN 2.1 VACE transformer_options passthrough
* Add optimized to get_attention_function
* Disable attention logs for now
* Remove attention logging code
* Remove _register_core_attention_functions, as we wouldn't want someone to call that, just in case
* Satisfy ruff
* Remove AttentionOverrideTest node, that's something to cook up for later
Load the projector.safetensors file with the ModelPatchLoader node and use
the siglip_vision_patch14_384.safetensors "clip vision" model and the
USOStyleReferenceNode.
* Attempting a universal implementation of EasyCache, starting with flux as test; I screwed up the math a bit, but when I set it just right it works.
* Fixed math to make threshold work as expected, refactored code to use EasyCacheHolder instead of a dict wrapped by object
* Use sigmas from transformer_options instead of timesteps to be compatible with a greater amount of models, make end_percent work
* Make log statement when not skipping useful, preparing for per-cond caching
* Added DIFFUSION_MODEL wrapper around forward function for wan model
* Add subsampling for heuristic inputs
* Add subsampling to output_prev (output_prev_subsampled now)
* Properly consider conds in EasyCache logic
* Created SuperEasyCache to test what happens if caching and reuse is moved outside the scope of conds, added PREDICT_NOISE wrapper to facilitate this test
* Change max reuse_threshold to 3.0
* Mark EasyCache/SuperEasyCache as experimental (beta)
* Make Lumina2 compatible with EasyCache
* Add EasyCache support for Qwen Image
* Fix missing comma, curse you Cursor
* Add EasyCache support to AceStep
* Add EasyCache support to Chroma
* Added EasyCache support to Cosmos Predict t2i
* Make EasyCache not crash with Cosmos Predict ImagToVideo latents, but does not work well at all
* Add EasyCache support to hidream
* Added EasyCache support to hunyuan video
* Added EasyCache support to hunyuan3d
* Added EasyCache support to LTXV (not very good, but does not crash)
* Implemented EasyCache for aura_flow
* Renamed SuperEasyCache to LazyCache, hardcoded subsample_factor to 8 on nodes
* Eatra logging when verbose is true for EasyCache
These are not real controlnets but actually a patch on the model so they
will be treated as such.
Put them in the models/model_patches/ folder.
Use the new ModelPatchLoader and QwenImageDiffsynthControlnet nodes.
* P2 of qwen edit model.
* Typo.
* Fix normal qwen.
* Fix.
* Make the TextEncodeQwenImageEdit also set the ref latent.
If you don't want it to set the ref latent and want to use the
ReferenceLatent node with your custom latent instead just disconnect the
VAE.
This node is only useful if someone trains the kontext model to properly
use multiple reference images via the index method.
The default is the offset method which feeds the multiple images like if
they were stitched together as one. This method works with the current
flux kontext model.
Turns out torch.compile has some gaps in context manager decorator
syntax support. I've sent patches to fix that in PyTorch, but it won't
be available for all the folks running older versions of PyTorch, hence
this trivial patch.
* Added initial support for basic context windows - in progress
* Add prepare_sampling wrapper for context window to more accurately estimate latent memory requirements, fixed merging wrappers/callbacks dicts in prepare_model_patcher
* Made context windows compatible with different dimensions; works for WAN, but results are bad
* Fix comfy.patcher_extension.merge_nested_dicts calls in prepare_model_patcher in sampler_helpers.py
* Considering adding some callbacks to context window code to allow extensions of behavior without the need to rewrite code
* Made dim slicing cleaner
* Add Wan Context WIndows node for testing
* Made context schedule and fuse method functions be stored on the handler instead of needing to be registered in core code to be found
* Moved some code around between node_context_windows.py and context_windows.py
* Change manual context window nodes names/ids
* Added callbacks to IndexListContexHandler
* Adjusted default values for context_length and context_overlap, made schema.inputs definition for WAN Context Windows less annoying
* Make get_resized_cond more robust for various dim sizes
* Fix typo
* Another small fix
* Change bf16 check and switch non-blocking to off default with option to force to regain speed on certain classes of iGPUs and refactor xpu check.
* Turn non_blocking off by default for xpu.
* Update README.md for Intel GPUs.
