vLLM's AITER integration gate silently excludes RDNA3 (gfx1100)
[major] config
Fields
id
gotchas/aiter-gate-skips-rdna3-gfx1100Class
config
Severity
major
Symptom
On a gfx1100 Radeon (Radeon PRO W7900 or 7900 XTX) vLLM logs 'AITER is not found or not supported on the current platform, QuarkOCP_MX will fall back to emulation' and then selects TritonInt8ScaledMMLinearKernel instead of AiterInt8ScaledMMLinearKernel. No error is raised; the INT8 model just runs on slower paths.
Root cause
is_aiter_found_and_supported() in vllm/_aiter_ops.py returns get_cdna_version() > 2, which is CDNA3-or-newer only, and the separate RDNA path is_aiter_found_and_supported_on_rdna4() is gated on gfx1201 only. RDNA3 (gfx1100) is in neither branch. AITER upstream lists 'AMD Pro W7900 | gfx1100 (RDNA3) | Experimental', so the library can run its Triton path on gfx1100; vLLM's own gate simply never exposes it. Note the same repo shows AITER CK/ASM kernels are CDNA-only, so gfx1100 needs the Triton WMMA route regardless.
Workaround
Run on the Triton/emulation path deliberately, or apply the four-part patch in issue #51136 (widen the arch gate to gfx1100, allow gfx1100 in aiter_meta/csrc/cpp_itfs/utils.py, add aiter/ops/triton/configs/gemm/gfx1100-GEMM-A8W8.json, and route _rocm_aiter_w8a8_gemm_impl to AITER's Triton gemm_a8w8 WMMA kernel). Otherwise treat W7900/7900 XTX as unsupported for AITER-accelerated formats.
notes
A vLLM maintainer (vllmellm) confirmed RDNA support is not on the vLLM roadmap and redirected to ROCm/aiter#4604. So this gate is unlikely to widen on its own; the issue remains open. Related: gotcha-vllm-rocm-fp8-cold-start-timeout.
Notes
A vLLM maintainer (vllmellm) confirmed RDNA support is not on the vLLM roadmap and redirected to ROCm/aiter#4604. So this gate is unlikely to widen on its own; the issue remains open. Related: gotcha-vllm-rocm-fp8-cold-start-timeout.