gfx1100 (RDNA3, W7900) works via Triton path but vLLM integration never enables it — confirm roadmap status

Summary

On Radeon PRO W7900 (gfx1100, RDNA3) with vLLM, AITER is never enabled: is_aiter_found_and_supported() in vLLM only covers CDNA3+ (get_cdna_version() > 2) and a separate path covers RDNA4 (gfx1201). RDNA3 (gfx1100) falls in the gap. We verified this with a 4-step patch that enables AITER on gfx1100 end-to-end (Triton WMMA GEMM + GDN decode + causal_conv1d + sampler) — so the library itself works here, but vLLM's integration layer has no path for it.

We raised this on vLLM; maintainers replied that RDNA support on AITER is not part of the vLLM roadmap and pointed here. Opening this to (a) confirm whether gfx1100/RDNA3 support is on AITER's roadmap, and (b) share our working patch as a reference.

Environment

• GPU: AMD Radeon PRO W7900 (gfx1100, RDNA3, 48GB)
• ROCm: 7.2.4
• vLLM: 0.26.0+rocm723 (also verified against latest main)
• aiter: 0.1.16.post3
• Model: Qwen3.6-27B-Quark-W8A8-INT8 (W8A8)

AITER upstream status (from README)

So the Triton path is usable on gfx1100, but nothing in the vLLM/AITER integration exposes it.

Our 4-step patch (reference)

1. vllm/_aiter_ops.py: is_aiter_found_and_supported() → on_mi3xx() or on_gfx1100()
2. aiter_meta/csrc/cpp_itfs/utils.py: add gfx1100 (and RDNA archs) to the compile allow-list
3. Add aiter/ops/triton/configs/gemm/gfx1100-GEMM-A8W8.json (M-band tuning config for RDNA3)
4. vllm/_aiter_ops.py: in _rocm_aiter_w8a8_gemm_impl, route gfx1100 to AITER's Triton gemm_a8w8 (CK gemm_a8w8_CK is XDL-only, cannot compile on RDNA3)

After patch: INFO ... Selected AiterInt8ScaledMMLinearKernel for QuarkW8A8Int8

Questions

1. Is gfx1100 (RDNA3) AITER support on the roadmap, or Experimental-only indefinitely?
2. Would vLLM/AITER welcome a PR that adds a gfx1100 path to is_aiter_found_and_supported() (mirroring the RDNA4 path)?
3. Any guidance on the Triton GEMM tuning config for RDNA3 (BLOCK sizes / warps / stages) beyond what we derived empirically?

Reference

• vLLM issue: https://github.com/vllm-project/vllm/issues/51136 (maintainer: "RDNA support on AITER is not part of vLLM roadmap")
• Our patch + tuning config: https://github.com/okone1995/Radeon-hackathon-2026-07/tree/main/claim-agent/bench-results/vllm/aiter-gfx1100

Start by reviewing the README's gfx1100 support status and the referenced vLLM issue, then inspect vllm/_aiter_ops.py and aiter_meta/csrc/cpp_itfs/utils.py alongside the supplied patch. Compare the existing CDNA3 and RDNA4 paths with the proposed gfx1100 Triton configuration. Done means maintainers confirm roadmap status and whether a PR using this approach is wanted.