📖 [Story] Improve the torch_tensorrt.dynamo.compile path

[gs-olive]

TL;DR

Improve the performance acceleration, robustness, and overall quality of the torch_tensorrt.dynamo.compile path. Simultaneously use these improvements to improve the quality of the torch_tensorrt.dynamo.fx_ts_compat path.

Goal(s)

By adding lowering passes, improving converter coverage, and building features to accelerate the performance of our torch.compile backend, we can improve the quality of the path. Additionally, in conjunction with #1940, many of these upgrades can be ported over to our torch._dynamo.export path.

Implementation Phases

1. Module-Level Acceleration ✨[Feature] Upgrade Prototype System for Module-Level Acceleration in Dynamo Path #1894
2. Improved Converter Coverage (↔ [Converter] Add support for assorted operators in the FX aten path #1769, ↔ [Converter] Add operations to accelerate Transformer encoder #1754)
3. Support for truncate_long_and_double (fix: Add support for truncate_long_and_double in FX #1865)
4. Engine caching + cross-session state capture
5. Dynamic shape support + avoiding unnecessary recompilation

Additional context

Related to #1940

[github-actions]

This issue has not seen activity for 90 days, Remove stale label or comment or this will be closed in 10 days