matteochen
matteochen
Hi guys, Can you please update the Notion template link? Thank you!
## 🐛 Bug Compiling a model with Transformer Engine executor with Cudagraphs enabled is not supported ### To Reproduce #### Code sample ``` import torch import thunder class Module(torch.nn.Module): def...
## 🐛 Bug Pickling a `TraceCtx` currently fails if it contains auto-registered symbols having `.module` pointing to Pytorch. These symbols cannot be looked up leading to a pickle error. ###...
Closes #2752 For reference: https://github.com/NVIDIA/cudnn-frontend/blob/v1.16.0/include/cudnn_frontend/node/scaled_dot_product_flash_attention.h#L816
[This](https://github.com/Lightning-AI/lightning-thunder/blob/93aa254082162605f724c3be7451b641b6669eba/thunder/executors/cudnn_sdpa.py#L337) guard is linked to an old cuDNN version, which shouldn't affect runs with a newer cuDNN version. cc @mruberry
## What does this PR do? Closes #2438. ### Summary - Added ExportStatefulExecutorsTransform (singleton) with a registry of export callbacks. - Executors can register a callback to export runtime state...
From the nsys profiles is visible how 4 dynamo subgraphs are generated but only 2 of them are launched with CUDA Graph: For reference refer to this doc: https://docs.google.com/document/d/1iv8-ujpih7hScQd90nhrl30xJ4D7Q7fRS2hdKuYfu3E/edit?tab=t.0 Repro:...
PR https://github.com/Lightning-AI/lightning-thunder/pull/2618 is not a perfect patch to temporarily speed up CI jobs targeting `check_vjp_correctness` tests. This work maintained this mathematical [identity](https://github.com/Lightning-AI/lightning-thunder/blob/72a9d866b38e8ec0c3a29b1e6544c4787754236c/thunder/tests/test_grad.py#L424) to perform correctness checks; instead of relying on...