ROCm / ROCm/aiter

[Issue]: Crash with Qwen3.5 with tool calling enabled

Open
#4,142 1 comment 0 reactions 1 assignee View on GitHub

@ChuanLi1101 is already working on this.

Since Jul 12, 2026.

Dominant language
Python
Stars
565
Forks
585
Avg merge
3d 4h
Merged PRs (30d)
366

Description

Problem Description

Serving Qwen3.5 with tool calling enabled causes a crash and a gpu core dump during the serving of a chat request.
I found this while running one of SWEBench workloads.

Operating System

Centos Stream Release 9

CPU

AMD EPYC 9655

GPU

AMD Instinct MI355X

ROCm Version

7.14

ROCm Component

No response

Steps to Reproduce

Files:

Serve command:

HSA_COREDUMP_PATTERN=/tmp/gpucore.%p
HSA_ENABLE_DEBUG=1
AMD_SERIALIZE_KERNEL=3
VLLM_ROCM_USE_AITER=1
/usr/local/bin/vllm serve Qwen/Qwen3.5-35B-A3B --port "30000" --tensor-parallel-size "8" --enable-auto-tool-choice --tool-call-parser "qwen3_xml"

Client command:

bash request.sh
(Optional for Linux users) Output of /opt/rocm/bin/rocminfo --support

No response

Additional Information

Unrelated to #2187, as setting VLLM_ROCM_USE_AITER_MHA=0 did not make any difference.

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.