[ROCm][Perf] Make AITER FlyDSL GDN prefill default on gfx942/gfx950 (#60645)

Signed-off-by: Mikko Tukiainen <Mikko.Tukiainen@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
Mikko Tukiainen
2026-10-14 10:11:57 +00:00
committed by GitHub
co-authored by Cursor mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 3f466c852c
commit f9b65c30c7
4 changed files with 74 additions and 28 deletions
@@ -178,7 +178,9 @@ def test_rocm_prefill_numerics_and_output_layout(state_dtype, correlated):
model="Qwen/Qwen3.5-0.8B",
revision="2fc06364715b967f1860aea9cf38778875588b17",
max_model_len=1024,
)
),
# auto picks AITER FlyDSL on gfx942/gfx950; this pins the Triton path.
additional_config={"gdn_prefill_backend": "triton"},
)
with set_current_vllm_config(config):
prefill = ChunkGatedDeltaRule()