[perf] Add fused MLA QKV + strided layernorm (vllm-project#21116)

Signed-off-by: Mickael Seznec <[email protected]>
Co-authored-by: mgoin <[email protected]>
Signed-off-by: avigny <[email protected]>

Author: 2 people

csrc/layernorm_kernels.cu

@@ -15,15 +15,16 @@ namespace vllm {
 // TODO(woosuk): Further optimize this kernel.
 template <typename scalar_t>
 __global__ void rms_norm_kernel(
-    scalar_t* __restrict__ out,           // [..., hidden_size]
-    const scalar_t* __restrict__ input,   // [..., hidden_size]
+    scalar_t* __restrict__ out,          // [..., hidden_size]
+    const scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float)input[blockIdx.x * hidden_size + idx];
+    const float x = (float)input[blockIdx.x * input_stride + idx];
     variance += x * x;
   }
 
@@ -37,7 +38,7 @@ __global__ void rms_norm_kernel(
   __syncthreads();
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * hidden_size + idx];
+    float x = (float)input[blockIdx.x * input_stride + idx];
     out[blockIdx.x * hidden_size + idx] =
         ((scalar_t)(x * s_variance)) * weight[idx];
   }
@@ -50,7 +51,8 @@ __global__ void rms_norm_kernel(
 template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
@@ -59,6 +61,7 @@ fused_add_rms_norm_kernel(
   static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
 
   const int vec_hidden_size = hidden_size / width;
+  const int64_t vec_input_stride = input_stride / width;
   __shared__ float s_variance;
   float variance = 0.0f;
   /* These and the argument pointers are all declared `restrict` as they are
@@ -73,7 +76,8 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> temp = input_v[id];
+    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
+    _f16Vec<scalar_t, width> temp = input_v[strided_id];
     temp += residual_v[id];
     variance += temp.sum_squares();
     residual_v[id] = temp;
@@ -90,10 +94,11 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
     int id = blockIdx.x * vec_hidden_size + idx;
+    int64_t strided_id = blockIdx.x * vec_input_stride + idx;
     _f16Vec<scalar_t, width> temp = residual_v[id];
     temp *= s_variance;
     temp *= weight_v[idx];
-    input_v[id] = temp;
+    input_v[strided_id] = temp;
   }
 }
 
@@ -103,15 +108,16 @@ fused_add_rms_norm_kernel(
 template <typename scalar_t, int width>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_kernel(
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int64_t input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    scalar_t z = input[blockIdx.x * hidden_size + idx];
+    scalar_t z = input[blockIdx.x * input_stride + idx];
     z += residual[blockIdx.x * hidden_size + idx];
     float x = (float)z;
     variance += x * x;
@@ -129,7 +135,7 @@ fused_add_rms_norm_kernel(
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
     float x = (float)residual[blockIdx.x * hidden_size + idx];
-    input[blockIdx.x * hidden_size + idx] =
+    input[blockIdx.x * input_stride + idx] =
         ((scalar_t)(x * s_variance)) * weight[idx];
   }
 }
@@ -141,38 +147,42 @@ void rms_norm(torch::Tensor& out,     // [..., hidden_size]
               torch::Tensor& weight,  // [hidden_size]
               double epsilon) {
   TORCH_CHECK(out.is_contiguous());
-  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(input.stride(-1) == 1);
   TORCH_CHECK(weight.is_contiguous());
 
   int hidden_size = input.size(-1);
   int num_tokens = input.numel() / hidden_size;
+  int64_t input_stride = input.stride(-2);
 
   dim3 grid(num_tokens);
   dim3 block(std::min(hidden_size, 1024));
   const at::cuda::OptionalCUDAGuard device_guard(device_of(input));
   const cudaStream_t stream = at::cuda::getCurrentCUDAStream();
   VLLM_DISPATCH_FLOATING_TYPES(input.scalar_type(), "rms_norm_kernel", [&] {
     vllm::rms_norm_kernel<scalar_t><<<grid, block, 0, stream>>>(
-        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(),
+        out.data_ptr<scalar_t>(), input.data_ptr<scalar_t>(), input_stride,
         weight.data_ptr<scalar_t>(), epsilon, num_tokens, hidden_size);
   });
 }
 
-#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                       \
-  VLLM_DISPATCH_FLOATING_TYPES(                                                \
-      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {                  \
-        vllm::fused_add_rms_norm_kernel<scalar_t, width>                       \
-            <<<grid, block, 0, stream>>>(input.data_ptr<scalar_t>(),           \
-                                         residual.data_ptr<scalar_t>(),        \
-                                         weight.data_ptr<scalar_t>(), epsilon, \
-                                         num_tokens, hidden_size);             \
+#define LAUNCH_FUSED_ADD_RMS_NORM(width)                                    \
+  VLLM_DISPATCH_FLOATING_TYPES(                                             \
+      input.scalar_type(), "fused_add_rms_norm_kernel", [&] {               \
+        vllm::fused_add_rms_norm_kernel<scalar_t, width>                    \
+            <<<grid, block, 0, stream>>>(                                   \
+                input.data_ptr<scalar_t>(), input_stride,                   \
+                residual.data_ptr<scalar_t>(), weight.data_ptr<scalar_t>(), \
+                epsilon, num_tokens, hidden_size);                          \
       });
 
 void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
                         torch::Tensor& residual,  // [..., hidden_size]
                         torch::Tensor& weight,    // [hidden_size]
                         double epsilon) {
+  TORCH_CHECK(residual.is_contiguous());
+  TORCH_CHECK(weight.is_contiguous());
   int hidden_size = input.size(-1);
+  int64_t input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -194,9 +204,16 @@ void fused_add_rms_norm(torch::Tensor& input,     // [..., hidden_size]
   auto inp_ptr = reinterpret_cast<std::uintptr_t>(input.data_ptr());
   auto res_ptr = reinterpret_cast<std::uintptr_t>(residual.data_ptr());
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
-  bool ptrs_are_aligned =
-      inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+  constexpr int vector_width = 8;
+  constexpr int req_alignment_bytes =
+      vector_width * 2;  // vector_width * sizeof(bfloat16 or float16) (float32
+                         // falls back to non-vectorized version anyway)
+  bool ptrs_are_aligned = inp_ptr % req_alignment_bytes == 0 &&
+                          res_ptr % req_alignment_bytes == 0 &&
+                          wt_ptr % req_alignment_bytes == 0;
+  bool offsets_are_multiple_of_vector_width =
+      hidden_size % vector_width == 0 && input_stride % vector_width == 0;
+  if (ptrs_are_aligned && offsets_are_multiple_of_vector_width) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/layernorm_quant_kernels.cu

@@ -23,16 +23,17 @@ namespace vllm {
 // TODO(woosuk): Further optimize this kernel.
 template <typename scalar_t, typename fp8_type>
 __global__ void rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,           // [..., hidden_size]
-    const scalar_t* __restrict__ input,   // [..., hidden_size]
+    fp8_type* __restrict__ out,          // [..., hidden_size]
+    const scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int input_stride,
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
     const float epsilon, const int num_tokens, const int hidden_size) {
   __shared__ float s_variance;
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    const float x = (float)input[blockIdx.x * hidden_size + idx];
+    const float x = (float)input[blockIdx.x * input_stride + idx];
     variance += x * x;
   }
 
@@ -49,7 +50,7 @@ __global__ void rms_norm_static_fp8_quant_kernel(
   float const scale_inv = 1.0f / *scale;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    float x = (float)input[blockIdx.x * hidden_size + idx];
+    float x = (float)input[blockIdx.x * input_stride + idx];
     float const out_norm = ((scalar_t)(x * s_variance)) * weight[idx];
     out[blockIdx.x * hidden_size + idx] =
         scaled_fp8_conversion<true, fp8_type>(out_norm, scale_inv);
@@ -63,8 +64,9 @@ __global__ void rms_norm_static_fp8_quant_kernel(
 template <typename scalar_t, int width, typename fp8_type>
 __global__ std::enable_if_t<(width > 0) && _typeConvert<scalar_t>::exists>
 fused_add_rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,           // [..., hidden_size]
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    fp8_type* __restrict__ out,    // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
@@ -74,6 +76,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
   static_assert(sizeof(_f16Vec<scalar_t, width>) == sizeof(scalar_t) * width);
 
   const int vec_hidden_size = hidden_size / width;
+  const int vec_input_stride = input_stride / width;
   __shared__ float s_variance;
   float variance = 0.0f;
   /* These and the argument pointers are all declared `restrict` as they are
@@ -87,8 +90,9 @@ fused_add_rms_norm_static_fp8_quant_kernel(
       reinterpret_cast<const _f16Vec<scalar_t, width>*>(weight);
 
   for (int idx = threadIdx.x; idx < vec_hidden_size; idx += blockDim.x) {
+    int stride_id = blockIdx.x * vec_input_stride + idx;
     int id = blockIdx.x * vec_hidden_size + idx;
-    _f16Vec<scalar_t, width> temp = input_v[id];
+    _f16Vec<scalar_t, width> temp = input_v[stride_id];
     temp += residual_v[id];
     variance += temp.sum_squares();
     residual_v[id] = temp;
@@ -125,8 +129,9 @@ fused_add_rms_norm_static_fp8_quant_kernel(
 template <typename scalar_t, int width, typename fp8_type>
 __global__ std::enable_if_t<(width == 0) || !_typeConvert<scalar_t>::exists>
 fused_add_rms_norm_static_fp8_quant_kernel(
-    fp8_type* __restrict__ out,           // [..., hidden_size]
-    scalar_t* __restrict__ input,         // [..., hidden_size]
+    fp8_type* __restrict__ out,    // [..., hidden_size]
+    scalar_t* __restrict__ input,  // [..., hidden_size]
+    const int input_stride,
     scalar_t* __restrict__ residual,      // [..., hidden_size]
     const scalar_t* __restrict__ weight,  // [hidden_size]
     const float* __restrict__ scale,      // [1]
@@ -135,7 +140,7 @@ fused_add_rms_norm_static_fp8_quant_kernel(
   float variance = 0.0f;
 
   for (int idx = threadIdx.x; idx < hidden_size; idx += blockDim.x) {
-    scalar_t z = input[blockIdx.x * hidden_size + idx];
+    scalar_t z = input[blockIdx.x * input_stride + idx];
     z += residual[blockIdx.x * hidden_size + idx];
     float x = (float)z;
     variance += x * x;
@@ -169,7 +174,9 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
                                torch::Tensor& weight,  // [hidden_size]
                                torch::Tensor& scale,   // [1]
                                double epsilon) {
+  TORCH_CHECK(out.is_contiguous());
   int hidden_size = input.size(-1);
+  int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -183,8 +190,9 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
               vllm::rms_norm_static_fp8_quant_kernel<scalar_t, fp8_t>
                   <<<grid, block, 0, stream>>>(
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),
-                      weight.data_ptr<scalar_t>(), scale.data_ptr<float>(),
-                      epsilon, num_tokens, hidden_size);
+                      input_stride, weight.data_ptr<scalar_t>(),
+                      scale.data_ptr<float>(), epsilon, num_tokens,
+                      hidden_size);
             });
       });
 }
@@ -198,7 +206,7 @@ void rms_norm_static_fp8_quant(torch::Tensor& out,     // [..., hidden_size]
                                                                width, fp8_t> \
                   <<<grid, block, 0, stream>>>(                              \
                       out.data_ptr<fp8_t>(), input.data_ptr<scalar_t>(),     \
-                      residual.data_ptr<scalar_t>(),                         \
+                      input_stride, residual.data_ptr<scalar_t>(),           \
                       weight.data_ptr<scalar_t>(), scale.data_ptr<float>(),  \
                       epsilon, num_tokens, hidden_size);                     \
             });                                                              \
@@ -210,7 +218,10 @@ void fused_add_rms_norm_static_fp8_quant(
     torch::Tensor& weight,    // [hidden_size]
     torch::Tensor& scale,     // [1]
     double epsilon) {
+  TORCH_CHECK(out.is_contiguous());
+  TORCH_CHECK(residual.is_contiguous());
   int hidden_size = input.size(-1);
+  int input_stride = input.stride(-2);
   int num_tokens = input.numel() / hidden_size;
 
   dim3 grid(num_tokens);
@@ -234,7 +245,7 @@ void fused_add_rms_norm_static_fp8_quant(
   auto wt_ptr = reinterpret_cast<std::uintptr_t>(weight.data_ptr());
   bool ptrs_are_aligned =
       inp_ptr % 16 == 0 && res_ptr % 16 == 0 && wt_ptr % 16 == 0;
-  if (ptrs_are_aligned && hidden_size % 8 == 0) {
+  if (ptrs_are_aligned && hidden_size % 8 == 0 && input_stride % 8 == 0) {
     LAUNCH_FUSED_ADD_RMS_NORM(8);
   } else {
     LAUNCH_FUSED_ADD_RMS_NORM(0);

csrc/quantization/fp8/common.cu

@@ -88,6 +88,8 @@ void static_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
                              torch::Tensor const& input,  // [..., d]
                              torch::Tensor const& scale)  // [1]
 {
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
   int const block_size = 256;
   int const num_tokens = input.numel() / input.size(-1);
   int const num_elems = input.numel();
@@ -111,6 +113,8 @@ void dynamic_scaled_fp8_quant(torch::Tensor& out,          // [..., d]
                               torch::Tensor const& input,  // [..., d]
                               torch::Tensor& scale)        // [1]
 {
+  TORCH_CHECK(input.is_contiguous());
+  TORCH_CHECK(out.is_contiguous());
   int const block_size = 256;
   int const num_tokens = input.numel() / input.size(-1);
   int const num_elems = input.numel();