A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 10, 19:00 UTC
[ROCM] Reorder arguments and rename parameters for rope_cached_thd_positions_2c_fwd_inplace (#29993)
Signed-off-by: Tres Popp <tres.popp@amd.com>
Tres committed
Jan 5, 2026 at 08:37 UTC
3133c192a3e5b4acfa0a5c1bb6acf90a8da7e23d
1 file changed
+4
-4
vllm/_aiter_ops.py
+4
-4
@@ -1374,14 +1374,14 @@ class rocm_aiter_ops:
1374
key_ = key[..., :rotary_dim]
1375
positions = positions.view(*query.shape[:1])
1376
rope_cached_thd_positions_2c_fwd_inplace(
1377
- positions,
1378
- sin,
1379
- cos,
1377
query_,
1378
key_,
1379
+ cos,
1380
+ sin,
1381
+ positions,
1382
rotate_style,
1383
reuse_freqs_front_part=True,
1384
- is_nope_first=False,
1384
+ nope_first=False,
1385
)
1386
query = query.view(query_shape)
1387
key = key.view(key_shape)