[Bug][ROCm][RDNA3] vLLM 0.23 RDNA image: Qwen3 answers garbage after the first request on RX 7900 XTX (fix in PR #1403) #1404

Image: rocm/vllm:rocm7.14.1_rdna_ubuntu24.04_py3.14_pytorch_2.11_vllm_0.23.0 (vLLM 0.23.1.dev1+g9ddef7117, branch release/0.23.0)
GPU: AMD Radeon RX 7900 XTX (gfx1100), ROCm 7.14.1, torch 2.11.0+rocm7.14.1

Problem

Qwen3 models (1.7B and 8B checked) answer garbage from the first or second request onward. The same prompt at temperature 0 gives ': 1.0.0.0…', then '1111…' or ' of of of…'. Qwen2.5 1.5B/7B are correct on the same card.

Cause

One non_blocking host-to-device copy from pageable memory: async_copy_to_gpu(query_start_loc_np, out=self.input_buffers.query_start_loc) in vllm/v1/worker/gpu/model_runner.py. Its first elements intermittently stay stale even after a full device synchronize, so attention reads the KV cache at wrong offsets. Upstream main stages this copy through pinned memory since vllm-project#56888 (included in vLLM v0.30.0+), but that change is not in the 0.23.0 RDNA image.

Ask

Either of these resolves it:

1. Merge the backport PR [Bugfix][ROCm] Stage pageable H2D copies in async_copy_to_gpu through pinned memory #1403 (one function, validated on an RX 7900 XTX: 0/12 → 12/12 correct), and rebuild the RDNA image; or
2. Publish an RDNA image built from a vLLM release that contains [MRV2] Buffer util simplifications vllm-project/vllm#56888 (v0.30.0 or later).

The repro script, the before/after table and how the copy was isolated are in the PR.

Contact: Hadi Honarvar, hadi@opod.io or hadi.work.ca@gmail.com (https://opod.io)