diff --git a/vllm/_aiter_ops.py b/vllm/_aiter_ops.py index 650a229e17..8dbe49f07f 100644 --- a/vllm/_aiter_ops.py +++ b/vllm/_aiter_ops.py @@ -376,6 +376,22 @@ def _rocm_aiter_fused_topk_fake( # Cache whether aiter supports FP8 MLA parameters _AITER_MLA_SUPPORTS_FP8: bool | None = None +_AITER_HAS_FUSED_QK_RMSNORM: bool | None = None + + +def check_aiter_fused_qk_rmsnorm() -> bool: + """Check if aiter provides fused_qk_rmsnorm (requires AITer >= PR #2442).""" + global _AITER_HAS_FUSED_QK_RMSNORM + if _AITER_HAS_FUSED_QK_RMSNORM is None: + try: + from aiter.ops.fused_qk_norm_rope_cache_quant import ( # noqa: F401 + fused_qk_rmsnorm, + ) + + _AITER_HAS_FUSED_QK_RMSNORM = True + except (ImportError, ModuleNotFoundError, AttributeError): + _AITER_HAS_FUSED_QK_RMSNORM = False + return _AITER_HAS_FUSED_QK_RMSNORM def _check_aiter_mla_fp8_support() -> bool: @@ -970,7 +986,14 @@ def _fused_mla_dual_rms_norm_impl( x1_epsilon: float, x2_epsilon: float, ) -> tuple[torch.Tensor, torch.Tensor]: - from aiter.ops.fused_qk_norm_rope_cache_quant import fused_qk_rmsnorm + try: + from aiter.ops.fused_qk_norm_rope_cache_quant import fused_qk_rmsnorm + except (ImportError, ModuleNotFoundError) as exc: + raise ImportError( + "fused_qk_rmsnorm requires a newer AITer version " + "(>= PR #2442). Please upgrade aiter or disable the " + "fuse_mla_dual_rms_norm pass." + ) from exc return fused_qk_rmsnorm( q=x1, diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py index 7ca131b363..2650664256 100644 --- a/vllm/config/vllm.py +++ b/vllm/config/vllm.py @@ -166,10 +166,10 @@ def enable_norm_pad_fusion(cfg: "VllmConfig") -> bool: def enable_mla_dual_rms_norm_fusion(cfg: "VllmConfig") -> bool: - """Enable MLA dual RMS norm fusion when AITer is available.""" - from vllm._aiter_ops import rocm_aiter_ops + """Enable MLA dual RMS norm fusion when AITer has fused_qk_rmsnorm.""" + from vllm._aiter_ops import check_aiter_fused_qk_rmsnorm, rocm_aiter_ops - return rocm_aiter_ops.is_enabled() + return rocm_aiter_ops.is_enabled() and check_aiter_fused_qk_rmsnorm() OPTIMIZATION_LEVEL_00 = {