Better FX offload #817
Merged
Better FX offload #817
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
nickfraser
pushed a commit
that referenced
this pull request
Feb 1, 2024
Giuseppe5
added a commit
that referenced
this pull request
Feb 6, 2024
Giuseppe5
added a commit
that referenced
this pull request
Feb 6, 2024
* optimum: initial optimum integration * Refined solution for offloading * Fix (optimum): clean-up (#802) * Fix (optimum): dataloader and forward cleanup (#807) * Fix (optimum): forward pass + fx (#808) * FX forward, GPTQ, Export (#809) * Forward pass with fx and pkv * Restore eval * Restore quantization * Experimental export * Fix GPTQ + Export * Fix 2GB ONNX export error * Fix gptq + speedup * Feat (offload/fx): better buffer/params + call_functional (#816) * Fix: typo to setting weight handlers * Feat (optimum): better call_function FX offload (#817) * Refactored per row quantization. JIT not working (#818) * Better structure for QDQ weights (#822) * Fix (export): flag for torch qcdq export (#823) * Setup: remove optimum folder (#825) * Add/fix comments * Fix llm example * Misc: pre-commit fix * Fix (graph/equalize): new transpose interface * Fix (examples/llm): no constant folding for group quant --------- Co-authored-by: Nick Fraser <[email protected]>
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Outdated suggestions cannot be applied.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.
A bit more interesting mechanism to offload call_function with FX graph.
We check the device of all inputs, and if the "main device" (e.g., any accelerator that is not cpu or disk) is uniform, nothing happens.
Otherwise we align the input on the main device.
After the computation, the original devices are restored to avoid memory leaks.