## Summary
Changes:
- Deprecate `ram` and `vram` configs. If these are set in invokeai.yaml,
they will be ignored.
- Create new `max_cache_ram_gb` and `max_cache_vram_gb` configs with the
same definitions as the old configs.
The main motivation of this change is to make the migration path
smoother for users who had previously added `ram` /`vram` to their
config files. Now, these users will be automatically migrated into the
new dynamic limit behavior (which is better in most cases). These users
will have to manually re-add `max_cache_ram_gb` and `max_cache_vram_gb`
to their configs if they wish to go back to specifying manual limits.
## Related Issues / Discussions
See the release notes for RC v5.6.0rc1 for the old migration behavior
that we are trying to improve:
https://github.com/invoke-ai/InvokeAI/releases/tag/v5.6.0rc1
## QA Instructions
- [x] Test that if `ram` or `vram` are present in a user's
`invokeai.yaml`, these values are ignored.
- [x] Test that `max_cache_ram_gb` and `max_cache_vram_gb` are applied,
if set.
## Merge Plan
- Don't forget to update the RC release notes accordingly.
## Checklist
- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [x] _Tests added / updated (if applicable)_
- [x] _Documentation added / updated (if applicable)_
- [ ] _Updated `What's New` copy (if doing a release after this PR)_
## Summary
This PR contains a bugfix for an edge case with model unloading (from
VRAM to RAM). Thanks to @JPPhoto for finding it.
The bug was triggered under the following conditions:
- A GGML-quantized model is loaded in VRAM
- We run a Spandrel image-to-image invocation (which is wrapped in a
`torch.inference_mode()` context manager.
- The model cache attempts to unload the GGML-quantized model from VRAM
to RAM.
- Doing this inside of the `torch.inference_mode()` cm results in the
following error:
```
[2025-01-07 15:48:17,744]::[InvokeAI]::ERROR --> Error while invoking session 98a07259-0c03-4111-a8d8-107041cb86f9, invocation d8daa90b-7e4c-4fc4-807c-50ba9be1a4ed (spandrel_image_to_image): Cannot set version_counter for inference tensor
[2025-01-07 15:48:17,744]::[InvokeAI]::ERROR --> Traceback (most recent call last):
File "/home/ryan/src/InvokeAI/invokeai/app/services/session_processor/session_processor_default.py", line 129, in run_node
output = invocation.invoke_internal(context=context, services=self._services)
File "/home/ryan/src/InvokeAI/invokeai/app/invocations/baseinvocation.py", line 300, in invoke_internal
output = self.invoke(context)
File "/home/ryan/.pyenv/versions/3.10.14/envs/InvokeAI_3.10.14/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/home/ryan/src/InvokeAI/invokeai/app/invocations/spandrel_image_to_image.py", line 167, in invoke
with context.models.load(self.image_to_image_model) as spandrel_model:
File "/home/ryan/src/InvokeAI/invokeai/backend/model_manager/load/load_base.py", line 60, in __enter__
self._cache.lock(self._cache_record, None)
File "/home/ryan/src/InvokeAI/invokeai/backend/model_manager/load/model_cache/model_cache.py", line 224, in lock
self._load_locked_model(cache_entry, working_mem_bytes)
File "/home/ryan/src/InvokeAI/invokeai/backend/model_manager/load/model_cache/model_cache.py", line 272, in _load_locked_model
vram_bytes_freed = self._offload_unlocked_models(model_vram_needed, working_mem_bytes)
File "/home/ryan/src/InvokeAI/invokeai/backend/model_manager/load/model_cache/model_cache.py", line 458, in _offload_unlocked_models
cache_entry_bytes_freed = self._move_model_to_ram(cache_entry, vram_bytes_to_free)
File "/home/ryan/src/InvokeAI/invokeai/backend/model_manager/load/model_cache/model_cache.py", line 330, in _move_model_to_ram
return cache_entry.cached_model.partial_unload_from_vram(
File "/home/ryan/.pyenv/versions/3.10.14/envs/InvokeAI_3.10.14/lib/python3.10/site-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/home/ryan/src/InvokeAI/invokeai/backend/model_manager/load/model_cache/cached_model/cached_model_with_partial_load.py", line 182, in partial_unload_from_vram
cur_state_dict = self._model.state_dict()
File "/home/ryan/.pyenv/versions/3.10.14/envs/InvokeAI_3.10.14/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1939, in state_dict
module.state_dict(destination=destination, prefix=prefix + name + '.', keep_vars=keep_vars)
File "/home/ryan/.pyenv/versions/3.10.14/envs/InvokeAI_3.10.14/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1936, in state_dict
self._save_to_state_dict(destination, prefix, keep_vars)
File "/home/ryan/.pyenv/versions/3.10.14/envs/InvokeAI_3.10.14/lib/python3.10/site-packages/torch/nn/modules/module.py", line 1843, in _save_to_state_dict
destination[prefix + name] = param if keep_vars else param.detach()
RuntimeError: Cannot set version_counter for inference tensor
```
### Explanation
From the `torch.inference_mode()` docs:
> Code run under this mode gets better performance by disabling view
tracking and version counter bumps.
Disabling version counter bumps results in the aforementioned error when
saving `GGMLTensor`s to a state_dict.
This incompatibility between `GGMLTensors` and `torch.inference_mode()`
is likely caused by the custom tensor type implementation. There may
very well be a way to get these to cooperate, but for now it is much
simpler to remove the `torch.inference_mode()` contexts.
Note that there are several other uses of `torch.inference_mode()` in
the Invoke codebase, but they are all tight wrappers around the
inference forward pass and do not contain the model load/unload process.
## Related Issues / Discussions
Original discussion:
https://discord.com/channels/1020123559063990373/1149506274971631688/1326180753159094303
## QA Instructions
Find a sequence of operations that triggers the condition. For me, this
was:
- Reserve VRAM in a separate process so that there was ~12GB left.
- Fresh start of Invoke
- Run FLUX inference with a GGML 8K model
- Run Spandrel upscaling
Tests:
- [x] Confirmed that I can reproduce the error and that it is no longer
hit after the change
- [x] Confirm that there is no speed regression from switching from
`torch.inference_mode()` to `torch.no_grad()`.
- Before: `50.354s`, After: `51.536s`
## Checklist
- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [x] _Tests added / updated (if applicable)_
- [x] _Documentation added / updated (if applicable)_
- [ ] _Updated `What's New` copy (if doing a release after this PR)_
Currently translated at 16.5% (273 of 1645 strings)
translationBot(ui): update translation (Polish)
Currently translated at 15.4% (254 of 1645 strings)
translationBot(ui): update translation (Polish)
Currently translated at 10.8% (178 of 1645 strings)
Co-authored-by: Nik Nikovsky <zejdzztegomaila@gmail.com>
Translate-URL: https://hosted.weblate.org/projects/invokeai/web-ui/pl/
Translation: InvokeAI/Web UI
Currently translated at 100.0% (1649 of 1649 strings)
translationBot(ui): update translation (Vietnamese)
Currently translated at 100.0% (1645 of 1645 strings)
translationBot(ui): update translation (Vietnamese)
Currently translated at 100.0% (1645 of 1645 strings)
translationBot(ui): update translation (Vietnamese)
Currently translated at 100.0% (1645 of 1645 strings)
Co-authored-by: Linos <linos.coding@gmail.com>
Translate-URL: https://hosted.weblate.org/projects/invokeai/web-ui/vi/
Translation: InvokeAI/Web UI
## Summary
This PR enables RAM/VRAM cache size limits to be determined dynamically
based on availability.
**Config Changes**
This PR modifies the app configs in the following ways:
- A new `device_working_mem_gb` config was added. This is the amount of
non-model working memory to keep available on the execution device (i.e.
GPU) when using dynamic cache limits. It default to 3GB.
- The `ram` and `vram` configs now default to `None`. If these configs
are set, they will take precedence over the dynamic limits. **Note: Some
users may have previously overriden the `ram` and `vram` values in their
`invokeai.yaml`. They will need to remove these configs to enable the
new dynamic limit feature.**
**Working Memory**
In addition to the new `device_working_mem_gb` config described above,
memory-intensive operations can estimate the amount of working memory
that they will need and request it from the model cache. This is
currently applied to the VAE decoding step for all models. In the
future, we may apply this to other operations as we work out which ops
tend to exceed the default working memory reservation.
**Mitigations for https://github.com/invoke-ai/InvokeAI/issues/7513**
This PR includes some mitigations for the issue described in
https://github.com/invoke-ai/InvokeAI/issues/7513. Without these
mitigations, it would occur with higher frequency when dynamic RAM
limits are used and the RAM is close to maxed-out.
## Limitations / Future Work
- Only _models_ can be offloaded to RAM to conserve VRAM. I.e. if VAE
decoding requires more working VRAM than available, the best we can do
is keep the full model on the CPU, but we will still hit an OOM error.
In the future, we could detect this ahead of time and switch to running
inference on the CPU for those ops.
- There is often a non-negligible amount of VRAM 'reserved' by the torch
CUDA allocator, but not used by any allocated tensors. We may be able to
tune the torch CUDA allocator to work better for our use case.
Reference:
https://pytorch.org/docs/stable/notes/cuda.html#optimizing-memory-usage-with-pytorch-cuda-alloc-conf
- There may be some ops that require high working memory that haven't
been updated to request extra memory yet. We will update these as we
uncover them.
- If a model is 'locked' in VRAM, it won't be partially unloaded if a
later model load requests extra working memory. This should be uncommon,
but I can think of cases where it would matter.
## Related Issues / Discussions
- #7492
- #7494
- #7500
- #7505
## QA Instructions
Run a variety of models near the cache limits to ensure that model
switching works properly for the following configurations:
- [x] CUDA, `enable_partial_loading=true`, all other configs default
(i.e. dynamic memory limits)
- [x] CUDA, `enable_partial_loading=true`, CPU and CUDA memory reserved
in another process so there is limited RAM/VRAM remaining, all other
configs default (i.e. dynamic memory limits)
- [x] CUDA, `enable_partial_loading=false`, all other configs default
(i.e. dynamic memory limits)
- [x] CUDA, ram/vram limits set (these should take precedence over the
dynamic limits)
- [x] MPS, all other default (i.e. dynamic memory limits)
- [x] CPU, all other default (i.e. dynamic memory limits)
## Merge Plan
- [x] Merge #7505 first and change target branch to main
## Checklist
- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [x] _Tests added / updated (if applicable)_
- [x] _Documentation added / updated (if applicable)_
- [ ] _Updated `What's New` copy (if doing a release after this PR)_
## Summary
This PR adds support for partial loading of models onto the GPU. This
enables models to run with much lower peak VRAM requirements (e.g. full
FLUX dev with 8GB of VRAM).
The partial loading feature is enabled behind a new config flag:
`enable_partial_loading=True`. This flag defaults to `False`.
**Note about performance:**
The `ram` and `vram` config limits are still applied when
`enable_partial_loading=True` is set. This can result in significant
slowdowns compared to the 'old' behaviour. Consider the case where the
VRAM limit is set to `vram=0.75` (GB) and we are trying to run an 8GB
model. When `enable_partial_loading=False`, we attempt to load the
entire model into VRAM, and if it fits (no OOM error) then it will run
at full speed. When `enable_partial_loading=True`, since we have the
option to partially load the model we will only load 0.75 GB into VRAM
and leave the remaining 7.25 GB in RAM. This will cause inference to be
much slower than before. To workaround this, it is important that your
`ram` and `vram` configs are carefully tuned. In a future PR, we will
add the ability to dynamically set the RAM/VRAM limits based on the
available memory / VRAM.
## Related Issues / Discussions
- #7492
- #7494
- #7500
## QA Instructions
Tests with `enable_partial_loading=True`, `vram=2`, on CUDA device:
For all tests, we expect model memory to stay below 2 GB. Peak working
memory will be higher.
- [x] SD1 inference
- [x] SDXL inference
- [x] FLUX non-quantized inference
- [x] FLUX GGML-quantized inference
- [x] FLUX BnB quantized inference
- [x] Variety of ControlNet / IP-Adapter / LoRA smoke tests
Tests with `enable_partial_loading=True`, and hack to force all models
to load 10%, on CUDA device:
- [x] SD1 inference
- [x] SDXL inference
- [x] FLUX non-quantized inference
- [x] FLUX GGML-quantized inference
- [x] FLUX BnB quantized inference
- [x] Variety of ControlNet / IP-Adapter / LoRA smoke tests
Tests with `enable_partial_loading=False`, `vram=30`:
We expect no change in behaviour when `enable_partial_loading=False`.
- [x] SD1 inference
- [x] SDXL inference
- [x] FLUX non-quantized inference
- [x] FLUX GGML-quantized inference
- [x] FLUX BnB quantized inference
- [x] Variety of ControlNet / IP-Adapter / LoRA smoke tests
Other platforms:
- [x] No change in behavior on MPS, even if
`enable_partial_loading=True`.
- [x] No change in behavior on CPU-only systems, even if
`enable_partial_loading=True`.
## Merge Plan
- [x] Merge #7500 first, and change the target branch to main
## Checklist
- [x] _The PR has a short but descriptive title, suitable for a
changelog_
- [x] _Tests added / updated (if applicable)_
- [x] _Documentation added / updated (if applicable)_
- [ ] _Updated `What's New` copy (if doing a release after this PR)_
