NPUW: Deref #27767
Draft
NPUW: Deref #27767
+199
−84
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
- LazyTensor wasn't prepared to all cases and was crashing in Unpack - Weight bank had a race (device bank was not protected)
github-actions
bot
added
category: Core
dmatveev
commented
Nov 27, 2024
There was a problem hiding this comment.
This will be removed
There was a problem hiding this comment.
This will be removed too
| if (proto_comp_model_desc.device_it + 1 == m_dev_list.end()) { | ||
| LOG_INFO("No fallback expected - clear the OV model for Subgraph[" << idx << "]"); | ||
| proto_comp_model_desc.model.reset(); | ||
| // proto_comp_model_desc.compiled_model = {}; // Shouldn't be here, CPU only |
There was a problem hiding this comment.
NB: It seems CPU's compiled models hold the pointer to the ov::Model - so to the original weights. Shouldn't be the case for NPU
There was a problem hiding this comment.
UPD. Also the case for CPU
Comment on lines
71
to
76
| template<typename F> | ||
| void non_parallel_for(std::size_t count, F &&f) { | ||
| for (std::size_t idx = 0u; idx < count; idx++) { | ||
| f(idx); | ||
| } | ||
| } |
There was a problem hiding this comment.
can be removed (or moved to util as a debug replacement to ov::parallel_for)
Comment on lines
158
to
163
| // FIXME: Uncomment it later (after the CPU copy revert) | ||
| // const auto &device_str = bank.first; | ||
| // if (device_str == "CPU") { | ||
| // // CPU memory is non-detachable | ||
| // continue; | ||
| // } |
There was a problem hiding this comment.
Should be uncommented once tested on device
Comment on lines
113
to
126
| // REVERTME:{{{ | ||
| // Store a copy of the tensor memory even on CPU - to simulate | ||
| // bank load. | ||
|
|
||
| ov::Tensor new_tensor(transformed_tensor.get_element_type(), transformed_tensor.get_shape()); | ||
| dbank.storage[tensor] = new_tensor; | ||
| guard.unlock(); | ||
|
|
||
| transformed_tensor.copy_to(new_tensor); | ||
| return new_tensor; | ||
| // Old code here: | ||
| // m_device_bank[device_for_alloc][tensor] = transformed_tensor; | ||
| // return transformed_tensor; | ||
| // REVERTME:}}} |
There was a problem hiding this comment.
Should be reverted once tested on-device
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Details:
Tickets:
Related: