Lightning-AI / Lightning-AI/lightning-thunder

Double first iteration time for `Llama-2-7b-hf` after nvFuser direct bindings

Open
#2,700 6 comments 0 reactions 1 assignee View on GitHub

@rdspring1 is already working on this.

Since Oct 29, 2025.

nvfuser tracing time
Dominant language
Python
Stars
1.5k
Forks
121
PR merge metrics
No merged PRs in 30d

Description

## 🐛 Bug

As per title, the following command takes double the time for the first iteration at commit 8b542cf16e12cf04f6375691621bb3adb0c4acea :

```
python thunder/benchmarks/benchmark_litgpt.py --model_name Llama-2-7b-hf --compile thunder_inductor_cat --checkpoint_activations False --low_precision_mode none --micro_batch_size 1 --global_batch_size 64 --use_sdpa False --block_size 4096 --max_iters 10 --warmup_iters 5
```

compared to before #2502 :

```python
# Before
iter 0: loss 0.1650, iter time: 105627.25ms, t: 4096
# After
iter 0: loss 0.1650, iter time: 215099.65ms, t: 4096
```

Tested on the latest container on B200

cc @rdspring1 @kshitij12345 @crcrpar

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Assessment

This issue has not been assessed yet.

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.