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.
* 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
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.
* mp: fix full dynamic unloading
This was not unloading dynamic models when requesting a full unload via
the unpatch() code path.
This was ok, i your workflow was all dynamic models but fails with big
VRAM leaks if you need to fully unload something for a regular ModelPatcher
It also fices the "unload models" button.
* mm: load models outside of Aimdo Mempool
In dynamic_vram mode, escape the Aimdo mempool and load into the regular
mempool. Use a dummy thread to do it.
This function has a dtype argument that allows the caller to set the
dtype in the cast. TIL Some models override this on weight casts, which
means its the highest priority.
Priority scheme is: argument > model dtype > state dict dtype
pinned memory was converted back to pinning the CPU side weight without
any changes. Fix the pinner to use the CPU weight and not the model defined
geometry. This will either save RAM or stop buffer overruns when the types
mismatch.
Fix the model defined weight caster to use the [ s.weight, s.bias ]
interpretation, as xfer_dest might be the flattened pin now. Fix the detection
of needing to cast to not be conditional on !pin.
The code throughout is None safe to just skip the feature cache saving
step if none. Set it none in single frame use so qwen doesn't burn VRAM
on the unused cache.
* ops: introduce autopad for conv3d
This works around pytorch missing ability to causal pad as part of the
kernel and avoids massive weight duplications for padding.
* wan-vae: rework causal padding
This currently uses F.pad which takes a full deep copy and is liable to
be the VRAM peak. Instead, kick spatial padding back to the op and
consolidate the temporal padding with the cat for the cache.
* wan-vae: implement zero pad fast path
The WAN VAE is also QWEN where it is used single-image. These
convolutions are however zero padded 3d convolutions, which means the
VAE is actually just 2D down the last element of the conv weight in
the temporal dimension. Fast path this, to avoid adding zeros that
then just evaporate in convoluton math but cost computation.
* causal_video_ae: Remove attention ResNet
This attention_head_dim argument does not exist on this constructor so
this is dead code. Remove as generic attention mid VAE conflicts with
temporal roll.
* ltx-vae: consoldate causal/non-causal code paths
* ltx-vae: add cache rolling adder
* ltx-vae: use cached adder for resnet
* ltx-vae: Implement rolling VAE
Implement a temporal rolling VAE for the LTX2 VAE.
Usually when doing temporal rolling VAEs you can just chunk on time relying
on causality and cache behind you as you go. The LTX VAE is however
non-causal.
So go whole hog and implement per layer run ahead and backpressure between
the decoder layers using recursive state beween the layers.
Operations are ammended with temporal_cache_state{} which they can use to
hold any state then need for partial execution. Convolutions cache their
inputs behind the up to N-1 frames, and skip connections need to cache the
mismatch between convolution input and output that happens due to missing
future (non-causal) input.
Each call to run_up() processes a layer accross a range on input that
may or may not be complete. It goes depth first to process as much as
possible to try and digest frames to the final output ASAP. If layers run
out of input due to convolution losses, they simply return without action
effectively applying back-pressure to the earlier layers. As the earlier
layers do more work and caller deeper, the partial states are reconciled
and output continues to digest depth first as much as possible.
Chunking is done using a size quota rather than a fixed frame length and
any layer can initiate chunking, and multiple layers can chunk at different
granulatiries. This remove the old limitation of always having to process
1 latent frame to entirety and having to hold 8 full decoded frames as
the VRAM peak.
Use vae.spacial_compression_encode() instead of directly accessing
downscale_ratio to handle both standard VAEs (int) and WAN VAEs (tuple).
Addresses reviewer feedback on PR #11259.
Co-authored-by: ChrisFab16 <christopher@fabritius.dk>
This logic was checking comfy_cast_weights, and going straight to
to the forward_comfy_cast_weights implementation without
attempting to downscale input to fp8 in the event comfy_cast_weights
is set.
The main reason comfy_cast_weights would be set would be for async
offload, which is not a good reason to nix FP8MM.
So instead, and together the underlying exclusions for FP8MM which
are:
* having a weight_function (usually LowVramPatch)
* force_cast_weights (compute dtype override)
* the weight is not Quantized
* the input is already quantized
* the model or layer has MM explictily disabled.
If you get past all of those exclusions, quantize the input tensor.
Then hand the new input, quantized or not off to
forward_comfy_cast_weights to handle it. If the weight is offloaded
but input is quantized you will get an offloaded MM8.
If the loader passes 1e32 as the usable memory size, it means force
the full load. This happens with CPU loads and a few other misc cases.
Removing the confusing number and just leave the other details.
Pretty much every error cudaHostRegister can throw also queues the same
error on the async GPU queue. This was fixed for repinning error case,
but there is the bad mmap and just enomem cases that are harder to
detect.
Do some dummy GPU work to clean the error state.
This operation trades in latents which in --gpu-only may be out of the GPU
The two VAE results will follow the --gpu-only defined behaviour so follow
the inpaint image device when calculating the mask in this path.
* make setattr safe for non existent attributes
Handle the case where the attribute doesnt exist by returning a static
sentinel (distinct from None). If the sentinel is passed in as the set
value, del the attr.
* Account for dequantization and type-casts in offload costs
When measuring the cost of offload, identify weights that need a type
change or dequantization and add the size of the conversion result
to the offload cost.
This is mutually exclusive with lowvram patches which already has
a large conservative estimate and wont overlap the dequant cost so\
dont double count.
* Set the compute type on CLIP MPs
So that the loader can know the size of weights for dequant accounting.
In the lowvram case, this now does its math in the model dtype in the
post de-quantization domain. Account for that. The patching was also
put back on the compute stream getting it off-peak so relax the
MATH_FACTOR to only x2 so get out of the worst-case assumption of
everything peaking at once.
slow down the CPU on model load to not run ahead. This fixes a VRAM on
flux 2 load.
I went to try and debug this with the memory trace pickles, which needs
--disable-cuda-malloc which made the bug go away. So I tried this
synchronize and it worked.
The has some very complex interactions with the cuda malloc async and
I dont have solid theory on this one yet.
Still debugging but this gets us over the OOM for the moment.
Im able to push vram above estimate on partial unload. Bump the
estimate. This is experimentally determined with a 720P and 480P
datapoint calibrating for 24GB VRAM total.
TIL that the WAN TE has a 2GB weight followed by 16MB as the next size
down. This means that team 8GB VRAM would fully offload the TE in async
offload mode as it just multiplied this giant size my the num streams.
Do the more complex logic of summing up the upcoming to-load weight
sizes to avoid triple counting this massive weight.
partial unload does the converse of recording the NS most recent
unloads as they go.
* mp: only count the offload cost of math once
This was previously bundling the combined weight storage and computation
cost
* ops: put all post async transfer compute on the main stream
Some models have massive weights that need either complex
dequantization or lora patching. Don't do these patchings on the offload
stream, instead do them on the main stream to syncrhonize the
potentially large vram spikes for these compute processes. This avoids
having to assume a worst case scenario of multiple offload streams
all spiking VRAM is parallel with whatever the main stream is doing.
* hunyuan upsampler: rework imports
Remove the transitive import of VideoConv3d and Resnet and takes these
from actual implementation source.
* model: remove unused give_pre_end
According to git grep, this is not used now, and was not used in the
initial commit that introduced it (see below).
This semantic is difficult to implement temporal roll VAE for (and would
defeat the purpose). Rather than implement the complex if, just delete
the unused feature.
(venv) rattus@rattus-box2:~/ComfyUI$ git log --oneline
220afe33 (HEAD) Initial commit.
(venv) rattus@rattus-box2:~/ComfyUI$ git grep give_pre
comfy/ldm/modules/diffusionmodules/model.py: resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
comfy/ldm/modules/diffusionmodules/model.py: self.give_pre_end = give_pre_end
comfy/ldm/modules/diffusionmodules/model.py: if self.give_pre_end:
(venv) rattus@rattus-box2:~/ComfyUI$ git co origin/master
Previous HEAD position was 220afe33 Initial commit.
HEAD is now at 9d8a8179 Enable async offloading by default on Nvidia. (#10953)
(venv) rattus@rattus-box2:~/ComfyUI$ git grep give_pre
comfy/ldm/modules/diffusionmodules/model.py: resolution, z_channels, give_pre_end=False, tanh_out=False, use_linear_attn=False,
comfy/ldm/modules/diffusionmodules/model.py: self.give_pre_end = give_pre_end
comfy/ldm/modules/diffusionmodules/model.py: if self.give_pre_end:
* move refiner VAE temporal roller to core
Move the carrying conv op to the common VAE code and give it a better
name. Roll the carry implementation logic for Resnet into the base
class and scrap the Hunyuan specific subclass.
* model: Add temporal roll to main VAE decoder
If there are no attention layers, its a standard resnet and VideoConv3d
is asked for, substitute in the temporal rolloing VAE algorithm. This
reduces VAE usage by the temporal dimension (can be huge VRAM savings).
* model: Add temporal roll to main VAE encoder
If there are no attention layers, its a standard resnet and VideoConv3d
is asked for, substitute in the temporal rolling VAE algorithm. This
reduces VAE usage by the temporal dimension (can be huge VRAM savings).
* mm: default to 0 for NUM_STREAMS
Dont count the compute stream as an offload stream. This makes async
offload accounting easier.
* mm: remove 128MB minimum
This is from a previous offloading system requirement. Remove it to
make behaviour of the loader and partial unloader consistent.
* mp: order the module list by offload expense
Calculate an approximate offloading temporary VRAM cost to offload a
weight and primary order the module load list by that. In the simple
case this is just the same as the module weight, but with Loras, a
weight with a lora consumes considerably more VRAM to do the Lora
application on-the-fly.
This will slightly prioritize lora weights, but is really for
proper VRAM offload accounting.
* mp: Account for the VRAM cost of weight offloading
when checking the VRAM headroom, assume that the weight needs to be
offloaded, and only load if it has space for both the load and offload
* the number of streams.
As the weights are ordered from largest to smallest by offload cost
this is guaranteed to fit in VRAM (tm), as all weights that follow
will be smaller.
Make the partial unload aware of this system as well by saving the
budget for offload VRAM to the model state and accounting accordingly.
Its possible that partial unload increases the size of the largest
offloaded weights, and thus needs to unload a little bit more than
asked to accomodate the bigger temp buffers.
Honor the existing codes floor on model weight loading of 128MB by
having the patcher honor this separately withough regard to offloading.
Otherwise when MM specifies its 128MB minimum, MP will see the biggest
weights, and budget that 128MB to only offload buffer and load nothing
which isnt the intent of these minimums. The same clamp applies in
case of partial offload of the currently loading model.
Clean up a bunch of stacked and no-longer-needed tensors on the QWEN
VRAM peak (currently FFN).
With this I go from OOMing at B=37x1328x1328 to being able to
succesfully run B=47 (RTX5090).
The partial unloader path in model re-use flow skips straight to the
actual unload without any check of the patching UUID. This means that
if you do an upscale flow with a model patch on an existing model, it
will not apply your patchings.
Fix by delaying the partial_unload until after the uuid checks. This
is done by making partial_unload a model of partial_load where extra_mem
is -ve.
* 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.
In the case of --cache-none lazy and subgraph execution can cause
anything to be run multiple times per workflow. If that rerun nodes is
in itself a subgraph generator, this will crash for two reasons.
pending_subgraph_results[] does not cleanup entries after their use.
So when a pending_subgraph_result is consumed, remove it from the list
so that if the corresponding node is fully re-executed this misses
lookup and it fall through to execute the node as it should.
Secondly, theres is an explicit enforcement against dups in the
addition of subgraphs nodes as ephemerals to the dymprompt. Remove this
enforcement as the use case is now valid.
* execution: fold in dependency aware caching
This makes --cache-none compatiable with lazy and expanded
subgraphs.
Currently the --cache-none option is powered by the
DependencyAwareCache. The cache attempts to maintain a parallel
copy of the execution list data structure, however it is only
setup once at the start of execution and does not get meaninigful
updates to the execution list.
This causes multiple problems when --cache-none is used with lazy
and expanded subgraphs as the DAC does not accurately update its
copy of the execution data structure.
DAC has an attempt to handle subgraphs ensure_subcache however
this does not accurately connect to nodes outside the subgraph.
The current semantics of DAC are to free a node ASAP after the
dependent nodes are executed.
This means that if a subgraph refs such a node it will be requed
and re-executed by the execution_list but DAC wont see it in
its to-free lists anymore and leak memory.
Rather than try and cover all the cases where the execution list
changes from inside the cache, move the while problem to the
executor which maintains an always up-to-date copy of the wanted
data-structure.
The executor now has a fast-moving run-local cache of its own.
Each _to node has its own mini cache, and the cache is unconditionally
primed at the time of add_strong_link.
add_strong_link is called for all of static workflows, lazy links
and expanded subgraphs so its the singular source of truth for
output dependendencies.
In the case of a cache-hit, the executor cache will hold the non-none
value (it will respect updates if they happen somehow as well).
In the case of a cache-miss, the executor caches a None and will
wait for a notification to update the value when the node completes.
When a node completes execution, it simply releases its mini-cache
and in turn its strong refs on its direct anscestor outputs, allowing
for ASAP freeing (same as the DependencyAwareCache but a little more
automatic).
This now allows for re-implementation of --cache-none with no cache
at all. The dependency aware cache was also observing the dependency
sematics for the objects and UI cache which is not accurate (this
entire logic was always outputs specific).
This also prepares for more complex caching strategies (such as RAM
pressure based caching), where a cache can implement any freeing
strategy completely independently of the DepedancyAwareness
requirement.
* main: re-implement --cache-none as no cache at all
The execution list now tracks the dependency aware caching more
correctly that the DependancyAwareCache.
Change it to a cache that does nothing.
* test_execution: add --cache-none to the test suite
--cache-none is now expected to work universally. Run it through the
full unit test suite. Propagate the server parameterization for whether
or not the server is capabale of caching, so that the minority of tests
that specifically check for cache hits can if else. Hard assert NOT
caching in the else to give some coverage of --cache-none expected
behaviour to not acutally cache.
* execution: fold in dependency aware caching
This makes --cache-none compatiable with lazy and expanded
subgraphs.
Currently the --cache-none option is powered by the
DependencyAwareCache. The cache attempts to maintain a parallel
copy of the execution list data structure, however it is only
setup once at the start of execution and does not get meaninigful
updates to the execution list.
This causes multiple problems when --cache-none is used with lazy
and expanded subgraphs as the DAC does not accurately update its
copy of the execution data structure.
DAC has an attempt to handle subgraphs ensure_subcache however
this does not accurately connect to nodes outside the subgraph.
The current semantics of DAC are to free a node ASAP after the
dependent nodes are executed.
This means that if a subgraph refs such a node it will be requed
and re-executed by the execution_list but DAC wont see it in
its to-free lists anymore and leak memory.
Rather than try and cover all the cases where the execution list
changes from inside the cache, move the while problem to the
executor which maintains an always up-to-date copy of the wanted
data-structure.
The executor now has a fast-moving run-local cache of its own.
Each _to node has its own mini cache, and the cache is unconditionally
primed at the time of add_strong_link.
add_strong_link is called for all of static workflows, lazy links
and expanded subgraphs so its the singular source of truth for
output dependendencies.
In the case of a cache-hit, the executor cache will hold the non-none
value (it will respect updates if they happen somehow as well).
In the case of a cache-miss, the executor caches a None and will
wait for a notification to update the value when the node completes.
When a node completes execution, it simply releases its mini-cache
and in turn its strong refs on its direct anscestor outputs, allowing
for ASAP freeing (same as the DependencyAwareCache but a little more
automatic).
This now allows for re-implementation of --cache-none with no cache
at all. The dependency aware cache was also observing the dependency
sematics for the objects and UI cache which is not accurate (this
entire logic was always outputs specific).
This also prepares for more complex caching strategies (such as RAM
pressure based caching), where a cache can implement any freeing
strategy completely independently of the DepedancyAwareness
requirement.
* main: re-implement --cache-none as no cache at all
The execution list now tracks the dependency aware caching more
correctly that the DependancyAwareCache.
Change it to a cache that does nothing.
* test_execution: add --cache-none to the test suite
--cache-none is now expected to work universally. Run it through the
full unit test suite. Propagate the server parameterization for whether
or not the server is capabale of caching, so that the minority of tests
that specifically check for cache hits can if else. Hard assert NOT
caching in the else to give some coverage of --cache-none expected
behaviour to not acutally cache.
Same change pattern as 7e8dd275c243ad460ed5015d2e13611d81d2a569
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.
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.
* 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.
rattusandGitHub
