Hacktoberfest 2026: the issues maintainers tagged for October, open and beginner-friendly. Browse Hacktoberfest issues

Grout prefill is much slower than paper artifact on RTX 5070 Ti, and referenced commit 4631fb01629fc3bc36ec4302b0bf3636f23e1453 is not found

Open
  • Is something wrong?
#1,404 0 comments 0 reactions 0 assignees View on GitHub

Maintainers usually reply within 2 days

Nobody has claimed this yet.

Assessment

Difficulty
4/5
Estimated time
3-5 days
Newbie friendliness
48/100
Issue type
Bug
Clarity
Mostly clear
Activity status
Active
Tech stack
rust
Domain
machine-learning, performance

Research direction

Start by checking the paper artifact’s Grout commit reference against the public repository history, then reproduce the reported benchmark using the command and environment in the issue. Compare prefill timings and the listed GatherRow, RmsNorm, and Rope profiles with the artifact results. Done means identifying the exact artifact version and whether the reported performance difference is reproducible; the issue does not name specific source files or tests.

Written by the indexing model from the issue text.

Description

cutile-rs perf

Migrated from NVlabs/cutile-rs#171, opened by @kitty-eu-org on 2026-06-23. 9 comments remain on the original.

Hi, thanks for releasing cutile-rs and Grout.

I am trying to reproduce the Grout end-to-end inference benchmark mentioned in the cutile-rs README / paper artifacts, but I am seeing much slower prefill performance on my machine.

Environment

  • GPU: NVIDIA GeForce RTX 5070 Ti, 16GB
  • Driver: 610.62
  • CUDA: 13.3
  • OS: Ubuntu 22.04 under WSL2
  • Model: Qwen3-4B-Instruct-2507
  • Grout repo: https://github.com/huggingface/grout
  • Grout commit tested: 487f076b90ff4fe0834effa3be58fc2f67cf236e
  • Branch: main

CUDA tools:

nvcc: release 13.3, V13.3.33
ptxas: release 13.3, V13.3.33

Issue

The cutile-rs paper artifact appears to reference this Grout commit:

4631fb01629fc3bc36ec4302b0bf3636f23e1453

However, I cannot find this commit in the public huggingface/grout repository. The current public main commit I tested is:

487f076 chore: adapt grout to cutile 0.0.2 changes (#2)

Could you clarify which Grout commit / branch was used for the paper artifact results, and whether that exact version is publicly available?

Benchmark command

I used:

export CUDA_TOOLKIT_PATH=/usr/local/cuda-13.3
export PATH=/usr/local/cuda-13.3/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-13.3/lib64:/usr/local/cuda-13.3/targets/x86_64-linux/lib:${LD_LIBRARY_PATH:-}

GROUT_CUDA_GRAPH_DECODE=1 \
target/release/grout \
  --model /home/hezhaozhao/models/Qwen3-4B-Instruct-2507 \
  --prompt "请用中文简单介绍一下大型语言模型。" \
  --max-new-tokens 256 \
  --device-argmax

Results with current public Grout main

With cutile = 0.0.2 / current Cargo.toml:

t/s: 40.47 prompt, 74.93 decode, 51.72 end-to-end
prompt_tokens=21
generated_tokens=114
prompt_s=0.519
decode_s=1.521
total_s=2.204

With --profile and GROUT_PROFILE_OPS=1, the apparent hotspots were:

RmsNorm        total_ms=241.128 calls=145
GatherRow      total_ms=229.666 calls=1
Rope           total_ms=117.887 calls=72
Attention      total_ms=0.407 calls=36

So attention does not seem to be the bottleneck. GatherRow, RmsNorm, and Rope look suspicious for prefill latency.

Longer prompt test

For a longer prompt:

prompt_tokens=1292
generated_tokens=0
prompt_s=3.049
prompt t/s=423.78
Dominant language
Rust
Stars
3.7k
Forks
312
Avg merge
6d 5h
Merged PRs (30d)
68

Getting set up

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.

More from NVIDIA/cuda-rust

All issues in NVIDIA/cuda-rust

Similar issues

More Rust issues

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.