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 7, 09:00 UTC
[ROCm][Bugfix]: Disable AITER Triton ROPE by default (#35601)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Rohan Potdar committed
Mar 3, 2026 at 13:43 UTC
3a8eef5869b8997af22f7b204eba56f9e654875e
2 files changed
+5
-6
vllm/envs.py
+3
-3
@@ -106,7 +106,7 @@ if TYPE_CHECKING:
106
VLLM_ROCM_USE_AITER_MLA: bool = True
107
VLLM_ROCM_USE_AITER_MHA: bool = True
108
VLLM_ROCM_USE_AITER_FP4_ASM_GEMM: bool = False
109
- VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = True
109
+ VLLM_ROCM_USE_AITER_TRITON_ROPE: bool = False
110
VLLM_ROCM_USE_AITER_FP8BMM: bool = True
111
VLLM_ROCM_USE_AITER_FP4BMM: bool = True
112
VLLM_ROCM_USE_AITER_UNIFIED_ATTENTION: bool = False
@@ -949,9 +949,9 @@ environment_variables: dict[str, Callable[[], Any]] = {
949
os.getenv("VLLM_ROCM_USE_AITER_FP4_ASM_GEMM", "False").lower() in ("true", "1")
950
),
951
# Whether to use aiter rope.
952
- # By default is enabled.
952
+ # By default is disabled.
953
"VLLM_ROCM_USE_AITER_TRITON_ROPE": lambda: (
954
- os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "True").lower() in ("true", "1")
954
+ os.getenv("VLLM_ROCM_USE_AITER_TRITON_ROPE", "False").lower() in ("true", "1")
955
),
956
# Whether to use aiter triton fp8 bmm kernel
957
# By default is enabled.
vllm/platforms/rocm.py
+2
-3
@@ -592,7 +592,6 @@ class RocmPlatform(Platform):
592
use_aiter_rms_norm = rocm_aiter_ops.is_rmsnorm_enabled()
593
use_aiter_fp8_linear = rocm_aiter_ops.is_linear_fp8_enabled()
594
use_aiter_fused_se = rocm_aiter_ops.is_fusion_moe_shared_experts_enabled()
595
- use_aiter_triton_rope = rocm_aiter_ops.is_triton_rotary_embed_enabled()
595
# Aiter rms norm perform best when CUDA Graph capture is enabled.
596
if (
597
use_aiter_rms_norm
@@ -619,9 +618,9 @@ class RocmPlatform(Platform):
618
and "-grouped_topk" not in compilation_config.custom_ops
619
):
620
compilation_config.custom_ops.append("+grouped_topk")
622
- # Enable rotary embedding when using AITER if its not disabled by user
621
+ # Enable rotary embedding customop when using AITER if not disabled by user
622
if (
624
- use_aiter_triton_rope
623
+ rocm_aiter_ops.is_enabled()
624
and "+rotary_embedding" not in compilation_config.custom_ops
625
and "-rotary_embedding" not in compilation_config.custom_ops
626
):