PaddlePaddle
diff --git a/‎paddle/phi/infermeta/fusion.cc
Copy file name to clipboard
+82Lines changed: 82 additions & 0 deletions b/‎paddle/phi/infermeta/fusion.cc
Copy file name to clipboard
+82Lines changed: 82 additions & 0 deletions
diff --git a/‎paddle/phi/infermeta/fusion.h
Copy file name to clipboard
+11Lines changed: 11 additions & 0 deletions b/‎paddle/phi/infermeta/fusion.h
Copy file name to clipboard
+11Lines changed: 11 additions & 0 deletions
diff --git a/‎paddle/phi/kernels/CMakeLists.txt
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions b/‎paddle/phi/kernels/CMakeLists.txt
Copy file name to clipboard
+1Lines changed: 1 addition & 0 deletions
diff --git a/‎paddle/phi/kernels/funcs/cublaslt.h
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion b/‎paddle/phi/kernels/funcs/cublaslt.h
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion
@@ -4696,6 +4696,88 @@ void FusedElemwiseActivationGradInferMeta(
   }
 }
 
+void FP8OutHalfGemmFusedInferMeta(
+    const MetaTensor& x,
+    const MetaTensor& y,
+    const MetaTensor& bias,
+    const bool trans_x,
+    const bool trans_y,
+    const float scale,  // only support per-tensor quantization
+    const std::string& output_dtype,
+    const std::string& activation_type,
+    MetaTensor* out) {
+  std::vector<int64_t> dims_x = common::vectorize(x.dims());
+  std::vector<int64_t> dims_y = common::vectorize(y.dims());
+  auto ndims_x = dims_x.size();
+  auto ndims_y = dims_y.size();
+  PADDLE_ENFORCE_GT(ndims_x,
+                    0UL,
+                    phi::errors::InvalidArgument(
+                        "The Input(x) dims size must be greater than 0,"
+                        " but received dims size is 0. "));
+  PADDLE_ENFORCE_GT(ndims_y,
+                    0UL,
+                    phi::errors::InvalidArgument(
+                        "The Input(y) dims size must be greater than 0,"
+                        " but received dims size is 0. "));
+
+  bool x_broadcasted = false, y_broadcasted = false;
+  if (ndims_x == 1) {
+    dims_x.insert(dims_x.begin(), 1);
+    ndims_x = 2;
+    x_broadcasted = true;
+  }
+
+  if (ndims_y == 1) {
+    dims_y.push_back(1);
+    ndims_y = 2;
+    y_broadcasted = true;
+  }
+
+  size_t M = 0, N = 0;
+  if (trans_x) {
+    M = dims_x[ndims_x - 1];
+  } else {
+    M = dims_x[ndims_x - 2];
+  }
+  if (trans_y) {
+    N = dims_y[ndims_y - 2];
+  } else {
+    N = dims_y[ndims_y - 1];
+  }
+
+  std::vector<int64_t> new_dims;
+  if (ndims_x > ndims_y) {
+    new_dims.assign(dims_x.begin(), dims_x.end() - 2);
+  } else if (ndims_x < ndims_y) {
+    new_dims.assign(dims_y.begin(), dims_y.end() - 2);
+  } else {
+    new_dims.reserve(ndims_x);
+    for (size_t i = 0; i < ndims_x - 2; ++i) {
+      new_dims.push_back(std::max(dims_x[i], dims_y[i]));
+    }
+  }
+  if (!x_broadcasted) {
+    new_dims.push_back(M);  // NOLINT
+  }
+  if (!y_broadcasted) {
+    new_dims.push_back(N);  // NOLINT
+  }
+
+  auto ddim_out = common::make_ddim(new_dims);
+
+  out->set_dims(ddim_out);
+  out->set_layout(x.layout());
+  if (output_dtype == "bfloat16") {
+    out->set_dtype(phi::DataType::BFLOAT16);
+  } else if (output_dtype == "float16") {
+    out->set_dtype(phi::DataType::FLOAT16);
+  } else {
+    PADDLE_THROW(phi::errors::Fatal(
+        "fp8_fp8_half_gemm_fused only support bfloat16 and float16 output"));
+  }
+}
+
 void FusedEmbeddingFcLstmInferMeta(const MetaTensor& ids,
                                    const MetaTensor& embeddings,
                                    const MetaTensor& weight_h,
 
@@ -1074,6 +1074,17 @@ void FusedElemwiseActivationGradInferMeta(
     MetaTensor* y_grad,
     MetaConfig config = MetaConfig());
 
+void FP8OutHalfGemmFusedInferMeta(
+    const MetaTensor& x,
+    const MetaTensor& y,
+    const MetaTensor& bias,
+    const bool trans_x,
+    const bool trans_y,
+    const float scale,  // only support per-tensor quantization
+    const std::string& output_dtype,
+    const std::string& activation_type,
+    MetaTensor* out);
+
 void FusedEmbeddingFcLstmInferMeta(const MetaTensor& ids,
                                    const MetaTensor& embeddings,
                                    const MetaTensor& weight_h,
 
@@ -36,6 +36,7 @@ file(
   "selected_rows/gpu/*.cu"
   "sparse/gpu/*.cu"
   "strings/gpu/*.cu"
+  "fusion/fp8_gemm/fp8_gemm_with_cublasLt/*.cu"
   "fusion/gpu/*.cu")
 
 file(
 
@@ -243,7 +243,7 @@ inline cudaDataType_t GetCublasLtDataType<phi::dtype::bfloat16>() {
   return CUDA_R_16BF;
 }
 
-#if CUDA_VERSION >= 12100
+#if CUDA_VERSION >= 12010
 template <typename T>
 void CublasLtMatmulFP8(const phi::GPUContext& dev_ctx,
                        const phi::DenseTensor& mat_a,
Original file line number	Diff line number	Diff line change
`@@ -243,7 +243,7 @@ inline cudaDataType_t GetCublasLtDataType<phi::dtype::bfloat16>() {`
`243`	`243`	`return CUDA_R_16BF;`
`244`	`244`	`}`
`245`	`245`
`246`		`-#if CUDA_VERSION >= 12100`
	`246`	`+#if CUDA_VERSION >= 12010`
`247`	`247`	`template <typename T>`
`248`	`248`	`void CublasLtMatmulFP8(const phi::GPUContext& dev_ctx,`
`249`	`249`	`const phi::DenseTensor& mat_a,`