diff --git a/csrc/libtorch_stable/minimax_reduce_rms_kernel.cu b/csrc/libtorch_stable/minimax_reduce_rms_kernel.cu index d9af0f5efe..da7fee5670 100644 --- a/csrc/libtorch_stable/minimax_reduce_rms_kernel.cu +++ b/csrc/libtorch_stable/minimax_reduce_rms_kernel.cu @@ -804,35 +804,6 @@ void minimax_reduce_rms_op(MiniMaxReduceRMSParams const& params) { } // namespace tensorrt_llm } // namespace vllm -torch::stable::Tensor minimax_allreduce_rms( - torch::stable::Tensor const& input, - torch::stable::Tensor const& norm_weight, torch::stable::Tensor workspace, - int64_t const rank, int64_t const nranks, double const eps) { - const torch::stable::accelerator::DeviceGuard device_guard( - input.get_device_index()); - auto allreduce_params = vllm::tensorrt_llm::MiniMaxReduceRMSParams(); - - allreduce_params.nranks = static_cast(nranks); - allreduce_params.rank = static_cast(rank); - allreduce_params.dtype = input.scalar_type(); - allreduce_params.size_q = static_cast(input.numel()); - allreduce_params.hidden_dim = static_cast(input.size(-1)); - allreduce_params.stride_q = allreduce_params.hidden_dim; - allreduce_params.workspace = - reinterpret_cast(workspace.mutable_data_ptr()); - allreduce_params.allreduce_in = const_cast(input.const_data_ptr()); - allreduce_params.rms_gamma = const_cast(norm_weight.const_data_ptr()); - allreduce_params.rms_eps = static_cast(eps); - allreduce_params.stream = get_current_cuda_stream(input.get_device_index()); - - torch::stable::Tensor rms_norm_out = torch::stable::empty_like(input); - allreduce_params.rms_norm_out = rms_norm_out.mutable_data_ptr(); - - vllm::tensorrt_llm::minimax_reduce_rms_op(allreduce_params); - - return rms_norm_out; -} - std::tuple minimax_allreduce_rms_qk(torch::stable::Tensor qkv, torch::stable::Tensor const& norm_weight_q, diff --git a/csrc/libtorch_stable/ops.h b/csrc/libtorch_stable/ops.h index d60b68a586..7cf34d8b03 100644 --- a/csrc/libtorch_stable/ops.h +++ b/csrc/libtorch_stable/ops.h @@ -288,10 +288,6 @@ void fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert( int64_t cache_block_size); #ifndef USE_ROCM -torch::stable::Tensor minimax_allreduce_rms( - torch::stable::Tensor const& input, - torch::stable::Tensor const& norm_weight, torch::stable::Tensor workspace, - int64_t const rank, int64_t const nranks, double const eps); std::tuple minimax_allreduce_rms_qk(torch::stable::Tensor qkv, torch::stable::Tensor const& norm_weight_q, diff --git a/csrc/libtorch_stable/torch_bindings.cpp b/csrc/libtorch_stable/torch_bindings.cpp index 1be7217ce7..158999a663 100644 --- a/csrc/libtorch_stable/torch_bindings.cpp +++ b/csrc/libtorch_stable/torch_bindings.cpp @@ -449,10 +449,6 @@ STABLE_TORCH_LIBRARY_FRAGMENT(_C, ops) { "int cache_block_size) -> ()"); #ifndef USE_ROCM - ops.def( - "minimax_allreduce_rms(" - "Tensor input, Tensor norm_weight, Tensor workspace, " - "int rank, int nranks, float eps) -> Tensor"); ops.def( "minimax_allreduce_rms_qk(" "Tensor qkv, Tensor norm_weight_q, Tensor norm_weight_k, " @@ -705,7 +701,6 @@ STABLE_TORCH_LIBRARY_IMPL(_C, CUDA, ops) { "fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert", TORCH_BOX(&fused_deepseek_v4_qnorm_rope_kv_rope_full_cache_fp8_insert)); #ifndef USE_ROCM - ops.impl("minimax_allreduce_rms", TORCH_BOX(&minimax_allreduce_rms)); ops.impl("minimax_allreduce_rms_qk", TORCH_BOX(&minimax_allreduce_rms_qk)); #endif ops.impl("fused_minimax_m3_qknorm_rope_kv_insert", diff --git a/vllm/_custom_ops.py b/vllm/_custom_ops.py index 02404a2f51..7dcb890aec 100644 --- a/vllm/_custom_ops.py +++ b/vllm/_custom_ops.py @@ -3785,20 +3785,6 @@ if hasattr(torch.ops._C, "hadacore_transform"): return torch.empty_like(x) if not inplace else x -if hasattr(torch.ops._C, "minimax_allreduce_rms"): - - @register_fake("_C::minimax_allreduce_rms") - def _minimax_allreduce_rms_fake( - input: torch.Tensor, - norm_weight: torch.Tensor, - workspace: torch.Tensor, - rank: int, - nranks: int, - eps: float, - ) -> torch.Tensor: - return torch.empty_like(input) - - if hasattr(torch.ops._C, "minimax_allreduce_rms_qk"): @register_fake("_C::minimax_allreduce_rms_qk")