[ROCm][CI] Update Transformers AR+RMS fusion expectation (#50517)
Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Andreas Karatzas committed
Jul 31, 2026 at 05:16 UTC
c911120a6f499a50f46a484109bf9bb2971ead5f
1 file changed
+3
-3
tests/compile/fusions_e2e/test_tp2_ar_rms.py
+3
-3
@@ -219,9 +219,9 @@ def test_tp2_ar_rms_fusions(
219
220
matches = matches_fn(n_layers)
221
if model_impl == "transformers":
222
- # TODO(BadrBasowid): Match the vLLM backend's fusion count once the
223
- # separate residual add and RMSNorm operations are fused.
224
- matches = matches._replace(aiter_ar_rms_fusion=1)
222
+ # Transformers add+RMSNorm canonicalization exposes every generic
223
+ # AR+RMS fusion site, including the final norm.
224
+ matches = matches._replace(aiter_ar_rms_fusion=matches.ar_rms_fusion)
225
226
# Reduce size of model and skip weight loading time
227
model_kwargs["hf_overrides"] = hf_overrides(n_layers)