PR #16841: Delete FP8 Scaling Factors in GEMM Rewriter

Imported from GitHub PR openxla/xla#16841

Removes the scaling factors of C and D (matrix bias and result) from FP8 Custom Calls created in the GEMM rewriter when their data types are not FP8. See openxla/xla#15795.
Copybara import of the project:

--
fd9750fa8474fe72fe641c7b3bc005ff30396e0a by Philipp Hack <[email protected]>:

Removes superfluous FP8 scaling factors in GEMM rewriter.

Merging this change closes #16841

FUTURE_COPYBARA_INTEGRATE_REVIEW=openxla/xla#16841 from philipphack:u_fp8_scales_xla fd9750fa8474fe72fe641c7b3bc005ff30396e0a
PiperOrigin-RevId: 679766659

.bazelrc

@@ -242,6 +242,8 @@ build:cuda_clang --copt=-Qunused-arguments
 # major release. Example: sm_80 kernels can run on sm_89 GPUs but
 # not on sm_90 GPUs. compute_80 kernels though can also run on sm_90 GPUs.
 build:cuda_clang --repo_env=HERMETIC_CUDA_COMPUTE_CAPABILITIES="sm_60,sm_70,sm_80,sm_89,compute_90"
+# Permit newer CUDA versions than Clang is aware of
+build:cuda_clang --copt="-Wno-unknown-cuda-version"
 # Set lld as the linker.
 build:cuda_clang --host_linkopt="-fuse-ld=lld"
 build:cuda_clang --host_linkopt="-lm"

third_party/gpus/cuda/hermetic/cuda_redist_init_repositories.bzl

@@ -219,6 +219,10 @@ def _create_libcuda_symlinks(
         repository_ctx.symlink(nvidia_driver_path, "lib/libcuda.so.1")
         repository_ctx.symlink("lib/libcuda.so.1", "lib/libcuda.so")
 
+def _create_cuda_header_symlinks(repository_ctx):
+    if repository_ctx.name == "cuda_nvcc":
+        repository_ctx.symlink("../cuda_cudart/include/cuda.h", "include/cuda.h")
+
 def use_local_path(repository_ctx, local_path, dirs):
     # buildifier: disable=function-docstring-args
     """Creates repository using local redistribution paths."""
@@ -339,6 +343,7 @@ def _use_downloaded_cuda_redistribution(repository_ctx):
         repository_ctx,
         lib_name_to_version_dict,
     )
+    _create_cuda_header_symlinks(repository_ctx)
     repository_ctx.file("version.txt", major_version)
 
 def _cuda_repo_impl(repository_ctx):

third_party/xla/.bazelrc

@@ -242,6 +242,8 @@ build:cuda_clang --copt=-Qunused-arguments
 # major release. Example: sm_80 kernels can run on sm_89 GPUs but
 # not on sm_90 GPUs. compute_80 kernels though can also run on sm_90 GPUs.
 build:cuda_clang --repo_env=HERMETIC_CUDA_COMPUTE_CAPABILITIES="sm_60,sm_70,sm_80,sm_89,compute_90"
+# Permit newer CUDA versions than Clang is aware of
+build:cuda_clang --copt="-Wno-unknown-cuda-version"
 # Set lld as the linker.
 build:cuda_clang --host_linkopt="-fuse-ld=lld"
 build:cuda_clang --host_linkopt="-lm"

third_party/xla/third_party/tsl/.bazelrc

@@ -242,6 +242,8 @@ build:cuda_clang --copt=-Qunused-arguments
 # major release. Example: sm_80 kernels can run on sm_89 GPUs but
 # not on sm_90 GPUs. compute_80 kernels though can also run on sm_90 GPUs.
 build:cuda_clang --repo_env=HERMETIC_CUDA_COMPUTE_CAPABILITIES="sm_60,sm_70,sm_80,sm_89,compute_90"
+# Permit newer CUDA versions than Clang is aware of
+build:cuda_clang --copt="-Wno-unknown-cuda-version"
 # Set lld as the linker.
 build:cuda_clang --host_linkopt="-fuse-ld=lld"
 build:cuda_clang --host_linkopt="-lm"

third_party/xla/third_party/tsl/third_party/gpus/cuda/hermetic/cuda_redist_init_repositories.bzl

@@ -219,6 +219,10 @@ def _create_libcuda_symlinks(
         repository_ctx.symlink(nvidia_driver_path, "lib/libcuda.so.1")
         repository_ctx.symlink("lib/libcuda.so.1", "lib/libcuda.so")
 
+def _create_cuda_header_symlinks(repository_ctx):
+    if repository_ctx.name == "cuda_nvcc":
+        repository_ctx.symlink("../cuda_cudart/include/cuda.h", "include/cuda.h")
+
 def use_local_path(repository_ctx, local_path, dirs):
     # buildifier: disable=function-docstring-args
     """Creates repository using local redistribution paths."""
@@ -339,6 +343,7 @@ def _use_downloaded_cuda_redistribution(repository_ctx):
         repository_ctx,
         lib_name_to_version_dict,
     )
+    _create_cuda_header_symlinks(repository_ctx)
     repository_ctx.file("version.txt", major_version)
 
 def _cuda_repo_impl(repository_ctx):

third_party/xla/xla/service/gpu/ir_emitter_unnested.cc

@@ -740,8 +740,7 @@ absl::Status IrEmitterUnnested::EmitCublasLtMatmulThunk(
 
 absl::Status IrEmitterUnnested::EmitCublasLtMatmulThunkF8(
     const HloCustomCallInstruction* instr) {
-  TF_RET_CHECK(instr->operand_count() == 6 || instr->operand_count() == 7 ||
-               instr->operand_count() == 8);
+  TF_RET_CHECK(instr->operand_count() > 3 && instr->operand_count() < 8);
   TF_ASSIGN_OR_RETURN(const auto gpu_config,
                       instr->backend_config<xla::gpu::GpuBackendConfig>());
   const xla::gpu::GemmBackendConfig& config = gpu_config.gemm_backend_config();
@@ -777,22 +776,22 @@ absl::Status IrEmitterUnnested::EmitCublasLtMatmulThunkF8(
   TF_ASSIGN_OR_RETURN(
       BufferAllocation::Slice b_scale,
       GetAllocationSliceForHlo(instr->operand(a_scale_index + 1)));
+
+  // cublasLT requires c_scale/d_scale to be null when C/D is not FP8.
+  // Currently, C cannot be FP8.
+  BufferAllocation::Slice c_scale, d_scale;
 #if GOOGLE_CUDA
-  TF_ASSIGN_OR_RETURN(
-      BufferAllocation::Slice c_scale,
-      GetAllocationSliceForHlo(instr->operand(a_scale_index + 2)));
-  TF_ASSIGN_OR_RETURN(
-      BufferAllocation::Slice d_scale,
-      GetAllocationSliceForHlo(instr->operand(a_scale_index + 3)));
-#else  // TENSORFLOW_USE_ROCM
-  BufferAllocation::Slice c_scale;
-  BufferAllocation::Slice d_scale;
+  if (instr->shape().tuple_shapes(0).element_type() == F8E4M3FN ||
+      instr->shape().tuple_shapes(0).element_type() == F8E5M2) {
+    TF_ASSIGN_OR_RETURN(d_scale,
+                        GetAllocationSliceForHlo(instr->operands().back()));
+  }
 #endif
 
   BufferAllocation::Slice bias;
   if (has_vector_bias) {
     TF_ASSIGN_OR_RETURN(
-        bias, GetAllocationSliceForHlo(instr->operand(a_scale_index + 4)));
+        bias, GetAllocationSliceForHlo(instr->operand(a_scale_index + 2)));
   }
 
   BufferAllocation::Slice d_amax;

third_party/xla/xla/service/gpu/matmul_utils.cc

@@ -484,8 +484,8 @@ bool IsTf32Allowed(PrecisionConfig::Algorithm algorithm,
   if (has_vector_bias) {
     int vector_bias_index = has_matrix_bias ? 3 : 2;
     if (primitive_util::IsF8Type(lhs_shape.element_type())) {
-      // FP8 gemms have 4 scales as inputs which come before the vector bias.
-      vector_bias_index += 4;
+      // FP8 gemms have 2 scales as inputs which come before the vector bias.
+      vector_bias_index += 2;
     }
     vector_bias_shape = gemm->operand(vector_bias_index)->shape();
   }

third_party/xla/xla/service/gpu/transforms/gemm_rewriter.cc

@@ -1083,12 +1083,18 @@ class GemmRewriterVisitor : public DfsHloRewriteVisitor {
 
     // cuBLASLt FP8 GEMM kernels require the scaling factors to be in F32
     // format. Set the factors to one when no scaling factors were captured.
-    Literal one_literal = LiteralUtil::One(F32);
-    HloInstruction *one = instr->AddInstruction(
-        HloInstruction::CreateConstant(one_literal.Clone()));
     std::array<bool, 2> mult_scale{a.mult_scale, b.mult_scale};
     std::array<HloInstruction *, 2> scales{a.scale, b.scale}, inv_scales,
         scales_f32;
+    HloInstruction *one_constant = nullptr;
+    auto one = [&one_constant, instr]() -> HloInstruction * {
+      if (!one_constant) {
+        one_constant = instr->AddInstruction(
+            HloInstruction::CreateConstant(LiteralUtil::One(F32)));
+      }
+      return one_constant;
+    };
+
     for (int i = 0; i < scales.size(); ++i) {
       if (scales[i]) {
         if (!ShapeUtil::IsScalar(scales[i]->shape())) {
@@ -1099,15 +1105,15 @@ class GemmRewriterVisitor : public DfsHloRewriteVisitor {
         }
         if (!mult_scale[i]) {
           inv_scales[i] = instr->AddInstruction(HloInstruction::CreateBinary(
-              scales[i]->shape(), HloOpcode::kDivide, one, scales[i]));
+              scales[i]->shape(), HloOpcode::kDivide, one(), scales[i]));
         }
         scales_f32[i] = mult_scale[i] ? scales[i] : inv_scales[i];
         if (scales_f32[i]->shape().element_type() != F32) {
           scales_f32[i] = instr->AddInstruction(HloInstruction::CreateConvert(
               ShapeUtil::MakeScalarShape(F32), scales_f32[i]));
         }
       } else {
-        scales_f32[i] = one;
+        scales_f32[i] = one();
       }
     }
 
@@ -1249,7 +1255,7 @@ class GemmRewriterVisitor : public DfsHloRewriteVisitor {
         PadShapeToMultipleOf16(instr->shape(), out_batch_dims);
 
     std::vector<HloInstruction *> operands_list = {
-        a.fp8_input, b.fp8_input, scales_f32[0], scales_f32[1], one, one};
+        a.fp8_input, b.fp8_input, scales_f32[0], scales_f32[1]};
 
     HloInstruction *new_custom_call =
         instr->AddInstruction(HloInstruction::CreateCustomCall(
@@ -1415,13 +1421,16 @@ class GemmRewriterVisitor : public DfsHloRewriteVisitor {
       }
     }
 
-    // If necessary, invert the scaling factor of D and convert to F32.
+    // If necessary, invert the scaling factor of D and convert to F32. When no
+    // scaling factor was captured, set the factor to one.
     if (d_scale) {
       TF_ASSIGN_OR_RETURN(d_scale,
                           InvertAndConvertScalar(d_scale, !mult_scale));
-      TF_RETURN_IF_ERROR(existing_gemm->ReplaceOperandWith(
-          gemm_backend_config.beta() == 0.0 ? 5 : 6, d_scale));
+    } else {
+      d_scale = instr->AddInstruction(
+          HloInstruction::CreateConstant(LiteralUtil::One(F32)));
     }
+    existing_gemm->AppendOperand(d_scale);
 
     // If present, elide the calculation of the maximum of the absolute values
     // of the result of the GEMM.