From bb49f2b14505c7541bb3c82159c40f4ebe27ccf0 Mon Sep 17 00:00:00 2001
From: Benjamin Chetioui <bchetioui@google.com>
Date: Thu, 19 Dec 2024 05:12:45 -0800
Subject: [PATCH] [XLA:GPU] Remove the call to `FloatNormalization` preceding
 the normalization rewriter.

It should no longer be necessary. This is part of a series of changes aimed
at decreasing the slight added complexity introduced by the normalization
rewriter.

PiperOrigin-RevId: 707877388
---
 xla/service/gpu/gpu_compiler.cc | 3 ---
 1 file changed, 3 deletions(-)

diff --git a/xla/service/gpu/gpu_compiler.cc b/xla/service/gpu/gpu_compiler.cc
index f4e87d417eec5..faeaa7a6c4667 100755
--- a/xla/service/gpu/gpu_compiler.cc
+++ b/xla/service/gpu/gpu_compiler.cc
@@ -1584,9 +1584,6 @@ absl::Status GpuCompiler::OptimizeHloPostLayoutAssignment(
     if ((cuda_cc != nullptr &&
          cuda_cc->IsAtLeast(se::CudaComputeCapability::AMPERE)) ||
         rocm_cc != nullptr) {
-      // Triton compilation needs normalized operations on bf16 (i.e. converted
-      // to f32).
-      add_float_normalization(pipeline);
       pipeline.AddPass<HloPassFix<GpuAlgebraicSimplifier>>(simplifier_options,
                                                            gpu_version);
       pipeline.AddPass<HloCSE>(/*is_layout_sensitive=*/true);