Pian Pawakapan

OSError: Could not load shared object file: libllvmlite.so

8

Hi, I've run into an issue while running the command: `python run.py --exp-config vlnce_baselines/config/paper_configs/seq2seq.yaml --run-type train` The error log is as below: `Traceback (most recent call last): File "run.py", line...

Allow non-torch.Tensor inputs in non-strict tracing

3

Previously, torch.export in non-strict mode was failing on non-torch.Tensor inputs while creating fake inputs for tracing (fakify()), and using graph nodes to create constraints. This fixes those 2 stages to...

[export] disable_forced_specializations

8

Summary: By default, some inferred dynamic shapes guards/constraints that are not expressible with the current dynamic shapes language will lead to specialization to the concrete input values provided. If disable_forced_specializations...

[export] add enable MKLDNN in export

3

Adds flag to enable MKLDNN backend, avoiding some decompositions.

[export] handle root swapping in derived dims suggested fixes

3

Summary: The current suggested fixes for derived dims can lead to root dims/derived dims being swapped, e.g. `dy - 1, dy` -> `dx, dx + 1`. This leads to problematic...

[export] handle constant aliasing for export

2

Summary: Currently export will [error out](https://github.com/pytorch/pytorch/blob/2b5ae2611e22d992565f202df9267fe66469efaa/torch/export/_trace.py#L477) if a constant is aliased. This PR supports this by modifying ConstantAttrMap to map constants to a list of FQNs instead of a single...

[export] allow complex guards as runtime asserts

3

With the current state of export's dynamic shapes, we struggle with guards and constraints that are beyond the current dynamic shapes language, expressed with dims and derived dims. While we...

[export] add nonstrict, retrace, serdes patching for _trace._export tests

4

Summary: Not sure if this is a good idea? Handful of new expected failures. But will be useful for a future diff Test Plan: tests Differential Revision: D57638904

[runtime asserts] deduplicate runtime asserts & CSE

13

This PR adds deduplication and CSE for runtime asserts. Existing size computation in the graph is CSE'd along with added runtime asserts, and redundant asserts are removed. Shape calls on...

bug in unbacked_bindings for a*u0

3

Summary: we were storing a*u0 instead of u0 in unbacked_bindings / unbacked_var_to_val Test Plan: - Differential Revision: D64508936