Yan Wang issues

Results 27 issues of


                                            Yan Wang

thunder.jit does not work for autocast transform in some cases

## 🐛 Bug In processing #198, thunder.jit fails in `thunder/tests/test_autocast.py` for this case ### To Reproduce ``` import torch import thunder from thunder.core.transforms import autocast def h(a, b, c): return...

bug

help wanted

amp

jit

Fix auto-registered torch.special operators

Before submitting - [ ] Was this discussed/approved via a Github issue? (no need for typos and docs improvements) - [ ] Did you read the [contributor guideline](https://github.com/Lightning-AI/pytorch-lightning/blob/main/.github/CONTRIBUTING.md), Pull Request...

Add the information about the set of auto-registered torch operators in CompileStats

Add functional transformer demo

This notebook demonstrates the acceleration of a transformer model, implemented in a functional style, using Thunder. Key highlights: - Illustrates Thunder-compatible PyTorch code (it doesn't mean that more complicated code...

Different behavior in torch.full: checks if `fill_value` can be cast to `dtype`.

## 🐛 Bug The check function in https://github.com/Lightning-AI/lightning-thunder/blob/019557e1cbd2944a4d8d719954ee8a7c295c539b/thunder/core/prims.py#L2650-L2655 is not the same as the one used in `torch.full` (https://github.com/pytorch/pytorch/blob/f5e704a6f25939478f770f8980c344ab461f0113/c10/util/Half.h#L462) ### To Reproduce An example of inconsistency: ``` import torch import...

operators

Add resnet50 benchmark (#443)

When comparing Thunder Torch Executor to Torch Eager, the ResNet18 gradients are not close for FP32.

*Note*: If you have a model or program that is not supported yet but should be, please use the program coverage template. ## 🐛 Bug ### To Reproduce Steps to...

numerical accuracy

network support

Add graph-by-graph benchmarking of dynamo.ThunderCompiler

As a part of https://github.com/Lightning-AI/lightning-thunder/issues/915. After https://github.com/Lightning-AI/lightning-thunder/pull/947 `torch.compile` can have `thunder.dynamo.ThunderCompiler` as a backend. This PR adds a graph-by-graph benchmarking of different executors(eager, inductor, thunder) working on each `torch.fx.GraphModule` after...

A converter for FXGraph with Torch calls -> FXGraph with Thunder calls

Dynamo represents the `function` argument of `torch.utils.checkpoint.checkpoint(function, args...)` as FX Graph, this FX graph has PyTorch operators in it. This PR creates a converter to replace the torch operators with...