From 7c0541b3859a4e3618f0379dc2aa882b445903bb Mon Sep 17 00:00:00 2001
From: Lianmin Zheng <lianminzheng@gmail.com>
Date: Sun, 9 Mar 2025 22:41:13 -0700
Subject: [PATCH] Move activation.cu to sgl-kernel/elementwise (#4250)

---
 sgl-kernel/csrc/elementwise/activation.cu | 82 +++++++++++++++++++++++
 sgl-kernel/setup.py                       |  2 +-
 2 files changed, 83 insertions(+), 1 deletion(-)
 create mode 100644 sgl-kernel/csrc/elementwise/activation.cu

diff --git a/sgl-kernel/csrc/elementwise/activation.cu b/sgl-kernel/csrc/elementwise/activation.cu
new file mode 100644
index 000000000..242281fd9
--- /dev/null
+++ b/sgl-kernel/csrc/elementwise/activation.cu
@@ -0,0 +1,82 @@
+/*
+ * Copyright (c) 2024 by FlashInfer team.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *   http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+#include <flashinfer/activation.cuh>
+
+#include "pytorch_extension_utils.h"
+
+using namespace flashinfer;
+
+__device__ __forceinline__ float silu(const float& val) {
+  return val / (1.0f + __expf(-val));
+}
+
+__device__ __forceinline__ float gelu(const float& val) {
+  constexpr float kAlpha = M_SQRT1_2;
+  return val * 0.5f * (1.0f + ::erf(val * kAlpha));
+}
+
+__device__ __forceinline__ float gelu_tanh(const float& val) {
+  const float cdf = 0.5f * (1.0f + math::tanh((0.7978845608028654f * (val + 0.044715f * val * val * val))));
+  return val * cdf;
+}
+
+void silu_and_mul(at::Tensor& out, at::Tensor& input, int64_t cuda_stream) {
+  int d = input.size(-1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+  dim3 grid(num_tokens);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input.scalar_type(), c_type, [&] {
+    uint32_t vec_size = 16 / sizeof(c_type);
+    dim3 block(std::min(d / vec_size, 1024U));
+    flashinfer::activation::act_and_mul_kernel<c_type, silu>
+        <<<grid, block, 0, stream>>>(static_cast<c_type*>(out.data_ptr()), static_cast<c_type*>(input.data_ptr()), d);
+
+    return true;
+  });
+}
+
+void gelu_tanh_and_mul(at::Tensor& out, at::Tensor& input, int64_t cuda_stream) {
+  int d = input.size(-1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+  dim3 grid(num_tokens);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input.scalar_type(), c_type, [&] {
+    uint32_t vec_size = 16 / sizeof(c_type);
+    dim3 block(std::min(d / vec_size, 1024U));
+    flashinfer::activation::act_and_mul_kernel<c_type, gelu_tanh>
+        <<<grid, block, 0, stream>>>(static_cast<c_type*>(out.data_ptr()), static_cast<c_type*>(input.data_ptr()), d);
+
+    return true;
+  });
+}
+
+void gelu_and_mul(at::Tensor& out, at::Tensor& input, int64_t cuda_stream) {
+  int d = input.size(-1) / 2;
+  int64_t num_tokens = input.numel() / input.size(-1);
+  dim3 grid(num_tokens);
+
+  cudaStream_t stream = reinterpret_cast<cudaStream_t>(cuda_stream);
+  DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(input.scalar_type(), c_type, [&] {
+    uint32_t vec_size = 16 / sizeof(c_type);
+    dim3 block(std::min(d / vec_size, 1024U));
+    flashinfer::activation::act_and_mul_kernel<c_type, gelu>
+        <<<grid, block, 0, stream>>>(static_cast<c_type*>(out.data_ptr()), static_cast<c_type*>(input.data_ptr()), d);
+
+    return true;
+  });
+}
diff --git a/sgl-kernel/setup.py b/sgl-kernel/setup.py
index 3fcce9474..47c1531a4 100644
--- a/sgl-kernel/setup.py
+++ b/sgl-kernel/setup.py
@@ -97,6 +97,7 @@ sources = [
     "csrc/allreduce/trt_reduce_internal.cu",
     "csrc/allreduce/trt_reduce_kernel.cu",
     "csrc/attention/lightning_attention_decode_kernel.cu",
+    "csrc/elementwise/activation.cu",
     "csrc/elementwise/fused_add_rms_norm_kernel.cu",
     "csrc/elementwise/rope.cu",
     "csrc/gemm/bmm_fp8.cu",
@@ -111,7 +112,6 @@ sources = [
     "csrc/speculative/eagle_utils.cu",
     "csrc/speculative/speculative_sampling.cu",
     "csrc/torch_extension.cc",
-    "3rdparty/flashinfer/csrc/activation.cu",
     "3rdparty/flashinfer/csrc/norm.cu",
     "3rdparty/flashinfer/csrc/renorm.cu",
     "3rdparty/flashinfer/csrc/sampling.cu",