Lightning-AI/lightning-thunder

View on GitHub

PyTorch compiler that accelerates training and inference. Get built-in optimizations for performance, memory, parallelism, and easily write your own.

Stars: 1.5k
Forks: 121
Open beginner issues: 0
Indexed issues: 419

Dominant language: Python
License: Apache-2.0
Last GitHub push: Sep 15, 2026
Latest indexed: Sep 19, 2026
Contributing guide: No contributing guide
Code of conduct: No code of conduct
Beginner labels: help wanted good first issue
PR merge metrics: No merged PRs in 30d

Beginner friendly Everything

Open All

419 open issues indexed

Consider using pre-built Flash Attention kernels via `kernels` Open

Difficulty 5/5 Over a week Newbie friendliness 30/100

Lightning-AI/lightning-thunder#2821 · Opened May 28, 2026
Refactor Triton cross-entropy executor: cleanup | Reduce activation memory in cross-entropy backward Open

Difficulty 3/5 1-2 days Newbie friendliness 58/100

Lightning-AI/lightning-thunder#2818 · 1 comment · 1 reaction · Opened Mar 23, 2026
Remove `xfail`s once PyTorch autograd stabilizes Open

autograd testing

Difficulty 3/5 1-2 days Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2807 · Opened Dec 16, 2025
`args_tensor_mask` has been removed from `torch.ops.higher_order.autograd_function_apply` Open

autograd

Difficulty 2/5 1-3 hours Newbie friendliness 55/100

Lightning-AI/lightning-thunder#2803 · Opened Dec 15, 2025
In-place ops on cloned tensor propagates to the original tensor Open

Difficulty 3/5 1-2 days Newbie friendliness 45/100

Lightning-AI/lightning-thunder#2793 · Opened Dec 10, 2025
Investigate `torch.Tensor.__setitem__` as unsupported in `thunderfx` Open

operators

Lightning-AI/lightning-thunder#2790 · 3 comments · 2 assignees · Opened Dec 3, 2025
[symbolic values] Trace bakes in the initial value of input used for trace construction Open

symbolic values

Lightning-AI/lightning-thunder#2789 · 2 comments · 1 assignee · Opened Dec 3, 2025
`AssertionError` for mutation on intermediates being reordered after `return` Open

Lightning-AI/lightning-thunder#2776 · 1 comment · 1 assignee · Opened Nov 27, 2025
[inference] High thunder.jit/thunderfx overhead for small module Open

Lightning-AI/lightning-thunder#2775 · 4 comments · 1 reaction · 1 assignee · Opened Nov 27, 2025
Too many `update_aliases` after in-place op Open

aliasing in-place

Difficulty 4/5 3-5 days Newbie friendliness 38/100

Lightning-AI/lightning-thunder#2768 · 1 comment · Opened Nov 25, 2025
Errors on in-place ops on tensor aliases unresolved by proxy substitution Open

aliasing in-place

Difficulty 4/5 3-5 days Newbie friendliness 35/100

Lightning-AI/lightning-thunder#2766 · 13 comments · Opened Nov 24, 2025
Replace usage of nvFuser's broadcast_in_dim in executor with broadcast+expand ops Open

enhancement executors nvfuser

Difficulty 3/5 1-2 days Newbie friendliness 45/100

Lightning-AI/lightning-thunder#2765 · Opened Nov 23, 2025
thunderfx: Applying thunderfx should return a nn.Module like object. Open

thunderfx ux

Difficulty 4/5 3-5 days Newbie friendliness 38/100

Lightning-AI/lightning-thunder#2757 · Opened Nov 19, 2025
Integration Test: KV-Cache management with symbolic values Open

enhancement symbolic values

Difficulty 5/5 Over a week Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2742 · 4 comments · Opened Nov 13, 2025
Support symbolic tensor shapes and dynamic dimensions Open

enhancement symbolic values

Difficulty 5/5 Over a week Newbie friendliness 30/100

Lightning-AI/lightning-thunder#2741 · Opened Nov 13, 2025
Support symbolic values in torch.arange and other tensor operations Open

enhancement symbolic values

Difficulty 5/5 Over a week Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2738 · 2 comments · Opened Nov 13, 2025
Support symbolic integer scalars Open

enhancement symbolic values

Lightning-AI/lightning-thunder#2735 · 1 assignee · Opened Nov 13, 2025
[Pattern matching] Replace computation patterns on the FX Graph level Open

pattern matching

Lightning-AI/lightning-thunder#2731 · 1 assignee · Opened Nov 12, 2025
Make prims.shape fusible with nvFuser executor Open

enhancement fusion logic nvfuser primitives

Lightning-AI/lightning-thunder#2730 · 1 assignee · Opened Nov 12, 2025
Using `symbolic values` leads to unfused program. Open

symbolic values

Difficulty 4/5 3-5 days Newbie friendliness 35/100

Lightning-AI/lightning-thunder#2729 · 2 comments · Opened Nov 12, 2025
Tests incompatible with recent transformers Open

Difficulty 3/5 1-2 days Newbie friendliness 42/100

Lightning-AI/lightning-thunder#2726 · 1 comment · Opened Nov 11, 2025
[DTensor] Refactor to re-use prims Open

DTensor

Difficulty 5/5 Over a week Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2722 · 1 comment · Opened Nov 10, 2025
[benchmark_inference] Investigate why the first and last Dynamo subgraph are not wrapped with cuda graph regions Open

cudagraphs

Difficulty 4/5 3-5 days Newbie friendliness 30/100

Lightning-AI/lightning-thunder#2717 · 1 reaction · Opened Nov 4, 2025
DTensor: nvFP4 grouped_mm and linear support for TensorParallel Open

DTensor enhancement

Lightning-AI/lightning-thunder#2712 · 1 assignee · Opened Nov 3, 2025
custom_op executor AssertionError when running different tests together Open

Lightning-AI/lightning-thunder#2710 · 1 reaction · 2 assignees · Opened Nov 3, 2025
Double first iteration time for `Llama-2-7b-hf` after nvFuser direct bindings Open

nvfuser tracing time

Lightning-AI/lightning-thunder#2700 · 6 comments · 1 assignee · Opened Oct 29, 2025
Symmetric memory support Open

distributed

Difficulty 5/5 Over a week Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2699 · 1 comment · Opened Oct 28, 2025
[Pattern matching] Identify key computation patterns Open

pattern matching

Lightning-AI/lightning-thunder#2693 · 3 comments · 1 assignee · Opened Oct 24, 2025
Long first iteration time for `meta-llama/Llama-4-Maverick-17B-128E` Open

symbolic values thunderfx

Difficulty 4/5 3-5 days Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2687 · 5 comments · Opened Oct 22, 2025
Warn when more than N recompile is triggered. Open

Difficulty 4/5 3-5 days Newbie friendliness 35/100

Lightning-AI/lightning-thunder#2682 · Opened Oct 21, 2025
Milestone 4: Set "symbolic values" as default. Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

Lightning-AI/lightning-thunder#2657 · 1 comment · Opened Oct 15, 2025
Milestone 3: Enable "symbolic values" in benchmarking pipelines Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

Lightning-AI/lightning-thunder#2656 · 1 comment · Opened Oct 15, 2025
Perf is not great on HF Transformers Llama 3.2 1B Open

Difficulty 4/5 3-5 days Newbie friendliness 42/100

Lightning-AI/lightning-thunder#2655 · 2 comments · Opened Oct 15, 2025
Milestone 2: Fix discovered limitations of symbolic values Open

Difficulty 5/5 Over a week Newbie friendliness 15/100

Lightning-AI/lightning-thunder#2654 · Opened Oct 15, 2025
Milestone 1: Understand current limitations of "symbolic values" Open

Difficulty 4/5 3-5 days Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2653 · Opened Oct 15, 2025
Do not perform DTensor isinstance check for all nvFuser inputs at runtime Open

DTensor nvfuser

Lightning-AI/lightning-thunder#2651 · 1 reaction · 1 assignee · Opened Oct 15, 2025
Further improve`test_grad.py::check_vjp_correctness` Open

ci / tests testing

Lightning-AI/lightning-thunder#2645 · 1 assignee · Opened Oct 14, 2025
Write custom `nn.Module`s with `DTensor`s Open

Difficulty 5/5 Over a week Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2621 · 1 comment · Opened Oct 9, 2025
Find and improve performance gaps if any for thunder + nvFP4 Open

Difficulty 5/5 Over a week Newbie friendliness 15/100

Lightning-AI/lightning-thunder#2596 · Opened Oct 3, 2025
Support nvFP4 with thunder (+nvFuser) and enable running the evaluation script with thunder + nvFP4 Open

Difficulty 5/5 Over a week Newbie friendliness 20/100

Lightning-AI/lightning-thunder#2595 · 1 comment · Opened Oct 3, 2025
Setup end-to-end test and benchmark for nvFP4 with SGLang Open

Difficulty 4/5 3-5 days Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2594 · 1 comment · Opened Oct 3, 2025
Support thunder for selected models and enable running the evaluation script with thunder Open

Difficulty 5/5 Over a week Newbie friendliness 35/100

Lightning-AI/lightning-thunder#2593 · 1 reaction · Opened Oct 3, 2025
Determine the configuration and setup end-to-end test and benchmark unmodified SGLang as baseline. Open

Lightning-AI/lightning-thunder#2592 · 1 comment · 1 assignee · Opened Oct 3, 2025
support distributed collective communication operators Open

Lightning-AI/lightning-thunder#2591 · 1 comment · 1 assignee · Opened Oct 3, 2025
Figure out if `DTensor` is involved in any of all the GraphModules Open

Lightning-AI/lightning-thunder#2590 · 1 assignee · Opened Oct 3, 2025
implement missing DTensor operators Open

Lightning-AI/lightning-thunder#2589 · 1 assignee · Opened Oct 3, 2025
think about translating GraphModules into ones with DTensor Open

Difficulty 5/5 Over a week Newbie friendliness 15/100

Lightning-AI/lightning-thunder#2588 · Opened Oct 3, 2025
Collect popular models' `GraphModule`s with distributed collective communication operators in there Open

Lightning-AI/lightning-thunder#2587 · 5 comments · 1 assignee · Opened Oct 3, 2025
Achieve performance parity with AutoDeploy's pipeline Open

enhancement

Difficulty 5/5 Over a week Newbie friendliness 25/100

Lightning-AI/lightning-thunder#2563 · Opened Oct 1, 2025
MoE Model Optimizations Open

performance

Lightning-AI/lightning-thunder#2562 · 1 assignee · Opened Oct 1, 2025

Lightning-AI/lightning-thunder

Get new issues in your inbox