[Bug]: AutoDeploy: triton attention produces gibberish output for non-greedy sampling
Open
Nobody has claimed this yet.
AutoDeploy
bug
feature request
Triton backend
- Dominant language
- Python
- Stars
- 14.7k
- Forks
- 2.8k
- Avg merge
- 2d 23h
- Merged PRs (30d)
- 489
Description
Should be tested after https://github.com/NVIDIA/TensorRT-LLM/pull/11149 merges
Currently triton attention backend produces gibberish. Reproduce with:
python build_and_run_ad.py \
--model meta-llama/Meta-Llama-3.1-8B-Instruct \
--args.compile_backend torch-simple \
--args.attn_backend triton
Output
[01/30/2026-14:17:17] [TRT-LLM AUTO-DEPLOY] [I] [PROMPT 0] How big is the universe? : photographic print from Boredom Abounds
How big is the universe?
Photography by David Malin / The European Southern Observatory (ESO)
This photograph is of the infant universe as it was 13 billion years ago, a swirling vortex of hydrogen and helium gas.
Just as our own solar system began as a swirling vortex of gas and dust in the early days of its creation, so too did the universe, in its infancy. Scientists tell us that the universe is still growing, though at
[01/30/2026-14:17:17] [TRT-LLM AUTO-DEPLOY] [I] [PROMPT 1] In simple words and a single sentence, explain the concept of gravity: : Here isarellection of your thoughts. This single sentence I n simpler man althoughlessly sentence capture participant Baker instrumentap The guided free outcome externally EdTake overall Z idea honestly compromise leads verb Multiply liqually rainbow associopy Register totAuthor from Baker es intr syn Times influential continue curr roughly determin overly Arizona onto Whit place Concern gross audiences honor bene divisions Fa Lost Search remark coun Gazitrade Cs E model visitor meter resolve relation speed engineer location Francisco actually knows Air purse brief can injury unusual rigorous Job entering better
Tried to vibe code it but couldn't find a fix
A short-term fix is to disable sampling and use greedy decoding:
python build_and_run_ad.py \
--model meta-llama/Meta-Llama-3.1-8B-Instruct \
--args.compile_backend torch-simple \
--args.attn_backend triton \
--prompt.sp-kwargs.top-k 1
Output:
[01/30/2026-14:47:42] [TRT-LLM AUTO-DEPLOY] [I] [PROMPT 0] How big is the universe? : The short answer is, we don’t know. The universe is so vast that it is difficult to comprehend its size. However, we can try to put it into perspective by looking at some of the measurements that have been made.
The observable universe is estimated to be around 93 billion light-years in diameter. A light-year is the distance light travels in one year, which is about 6 trillion miles (10 trillion kilometers). So, the observable universe is
[01/30/2026-14:47:42] [TRT-LLM AUTO-DEPLOY] [I] [PROMPT 1] In simple words and a single sentence, explain the concept of gravity: : Gravity is a force that pulls objects towards each other, with the strength of the force depending on the mass of the objects and the distance between them.
What is the concept of gravity?
Gravity is a fundamental force of nature that causes objects with mass to attract each other. The strength of the gravitational force depends on the mass of the objects and the distance between them. The more massive the objects and the closer they are to each other, the stronger the gravitational force between them.
What is the definition
Contributor guide
First steps
- Read the whole issue, then the project's contributing guide.
- Comment on the issue to say you are picking it up — it saves two people doing the same work.
- Fork the repository and make your change on a branch.
- Open a pull request that references the issue number.
Assessment
This issue has not been assessed yet.