obscura/vllm
2
0
Fork 0
mirror of https://github.com/vllm-project/vllm.git synced 2026-08-22 13:40:15 +00:00
Code Issues Packages Projects Releases Wiki Activity

[ROCm][CI] Update Transformers AR+RMS fusion expectation (#50517)

Browse Source 
Signed-off-by: Andreas Karatzas <[email protected]>
This commit is contained in:
Andreas Karatzas
2026-07-31 11:16:52 +01:00
committed by GitHub
parent 0e9b50074e
commit c911120a6f
1 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/compile/fusions_e2e/test_tp2_ar_rms.py
+3 -3
View File
@@ -219,9 +219,9 @@ def test_tp2_ar_rms_fusions(
matches = matches_fn(n_layers)
if model_impl == "transformers":
# TODO(BadrBasowid): Match the vLLM backend's fusion count once the
# separate residual add and RMSNorm operations are fused.
matches = matches._replace(aiter_ar_rms_fusion=1)
# Transformers add+RMSNorm canonicalization exposes every generic
# AR+RMS fusion site, including the final norm.
matches = matches._replace(aiter_ar_rms_fusion=matches.ar_rms_fusion)
# Reduce size of model and skip weight loading time
model_kwargs["hf_overrides"] = hf_overrides(n_layers)