diff --git a/vllm/_aiter_ops.py b/vllm/_aiter_ops.py index 8ab06d446b..4a915c0dd6 100644 --- a/vllm/_aiter_ops.py +++ b/vllm/_aiter_ops.py @@ -409,37 +409,6 @@ 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. - - Supports both the new private name ``_fused_qk_rmsnorm`` - (AITER >= PR #2958) and the old public name ``fused_qk_rmsnorm`` - (AITER >= PR #2442). - - TODO(rbrugaro-amd): remove the legacy fused_qk_rmsnorm path once - AITER stabilizes the API (https://github.com/ROCm/aiter/issues/3207). - """ - 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): - 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: @@ -1267,43 +1236,17 @@ def _fused_mla_dual_rms_norm_impl( x1_epsilon: float, x2_epsilon: float, ) -> tuple[torch.Tensor, torch.Tensor]: - try: - import aiter.ops.fused_qk_norm_rope_cache_quant as aiter_ops - except (ImportError, ModuleNotFoundError, AttributeError) as exc: - raise ImportError( - "fused_qk_rmsnorm requires AITer >= PR #2442. " - "Please upgrade aiter or disable the " - "fuse_mla_dual_rms_norm pass." - ) from exc + from aiter.ops.fused_qk_norm_rope_cache_quant import _fused_qk_rmsnorm - if hasattr(aiter_ops, "_fused_qk_rmsnorm"): - return aiter_ops._fused_qk_rmsnorm( - q_out=None, - q=x1, - q_weight=x1_weight, - q_eps=x1_epsilon, - k_out=None, - k=x2, - k_weight=x2_weight, - k_eps=x2_epsilon, - ) - - # TODO(rbrugaro-amd): remove the legacy fused_qk_rmsnorm path once - # AITER stabilizes the API (https://github.com/ROCm/aiter/issues/3207). - if hasattr(aiter_ops, "fused_qk_rmsnorm"): - return aiter_ops.fused_qk_rmsnorm( - q=x1, - q_weight=x1_weight, - q_eps=x1_epsilon, - k=x2, - k_weight=x2_weight, - k_eps=x2_epsilon, - ) - - raise ImportError( - "fused_qk_rmsnorm requires AITer >= PR #2442. " - "Please upgrade aiter or disable the " - "fuse_mla_dual_rms_norm pass." + return _fused_qk_rmsnorm( + q_out=None, + q=x1, + q_weight=x1_weight, + q_eps=x1_epsilon, + k_out=None, + k=x2, + k_weight=x2_weight, + k_eps=x2_epsilon, ) diff --git a/vllm/compilation/passes/pass_manager.py b/vllm/compilation/passes/pass_manager.py index 67a3e4cbae..2fe863189c 100644 --- a/vllm/compilation/passes/pass_manager.py +++ b/vllm/compilation/passes/pass_manager.py @@ -7,7 +7,7 @@ from typing import Any, ParamSpec, TypeVar from torch import fx as fx from vllm import envs -from vllm._aiter_ops import check_aiter_fused_qk_rmsnorm, rocm_aiter_ops +from vllm._aiter_ops import rocm_aiter_ops from vllm.compilation.passes.utility.post_cleanup import PostCleanupPass from vllm.config import VllmConfig, set_current_vllm_config from vllm.logger import init_logger @@ -177,11 +177,7 @@ class PostGradPassManager(CustomGraphPass): # type: ignore[misc] self.passes += [ScatterSplitReplacementPass(config)] self.passes += [QkNormRopeKvCacheFusionPass(config)] - if ( - self.pass_config.fuse_mla_dual_rms_norm - and rocm_aiter_ops.is_enabled() - and check_aiter_fused_qk_rmsnorm() - ): + if self.pass_config.fuse_mla_dual_rms_norm and rocm_aiter_ops.is_enabled(): self.passes += [MLADualRMSNormFusionPass(config)] if self.pass_config.fuse_rope_kvcache: diff --git a/vllm/config/vllm.py b/vllm/config/vllm.py index 4b4a97f41b..bda5013a5f 100644 --- a/vllm/config/vllm.py +++ b/vllm/config/vllm.py @@ -187,10 +187,10 @@ 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 has fused_qk_rmsnorm.""" - from vllm._aiter_ops import check_aiter_fused_qk_rmsnorm, rocm_aiter_ops + """Enable MLA dual RMS norm fusion on ROCm with AITER.""" + from vllm._aiter_ops import rocm_aiter_ops - return rocm_aiter_ops.is_enabled() and check_aiter_fused_qk_rmsnorm() + return rocm_aiter_ops.is_enabled() def enable_qk_norm_rope_kvcache(cfg: "VllmConfig") -> bool: