PR #16975: Add a few related optimization passes for fp8 gemm custom-…

…calls. Imported from GitHub PR #16975 This caused convergence issue for fp8 training, tested on GPT3 models: Before: ``` NETWORK BACKEND MATH SDPA XLA_EXTRAS GPUs STEPS/SEC LOSS WALLSECS GPT5B XLA fp8 FA 8 1.064 11.019 1571 [PAX STATUS]: Starting training loop. [PAX STATUS] step_i: 100, training loss: 11.015041 [PAX STATUS] step_i: 200, training loss: 11.016165 [PAX STATUS] step_i: 300, training loss: 11.016386 [PAX STATUS] step_i: 400, training loss: 11.014653 [PAX STATUS] step_i: 500, training loss: 11.014734 [PAX STATUS] step_i: 600, training loss: 11.01613 [PAX STATUS] step_i: 700, training loss: 11.009399 [PAX STATUS] step_i: 800, training loss: 11.017071 [PAX STATUS] step_i: 900, training loss: 11.014582 [PAX STATUS] step_i: 1000, training loss: 11.013434 [PAX STATUS] step_i: 1100, training loss: 11.021271 [PAX STATUS] step_i: 1200, training loss: 11.008364 [PAX STATUS] step_i: 1300, training loss: 11.0198145 [PAX STATUS] step_i: 1400, training loss: 11.01253 [PAX STATUS] step_i: 1500, training loss: 11.019016 ``` After: ``` NETWORK BACKEND MATH SDPA GPUs STEPS/SEC LOSS WALLSECS GPT5B XLA fp8 FA 8 1.020 3.797 1647 [PAX STATUS]: Starting training loop. [PAX STATUS] step_i: 100, training loss: 6.150083 [PAX STATUS] step_i: 200, training loss: 5.8871064 [PAX STATUS] step_i: 300, training loss: 5.4491887 [PAX STATUS] step_i: 400, training loss: 5.6384015 [PAX STATUS] step_i: 500, training loss: 5.273538 [PAX STATUS] step_i: 600, training loss: 5.2011905 [PAX STATUS] step_i: 700, training loss: 4.903013 [PAX STATUS] step_i: 800, training loss: 4.62972 [PAX STATUS] step_i: 900, training loss: 4.507727 [PAX STATUS] step_i: 1000, training loss: 4.625259 [PAX STATUS] step_i: 1100, training loss: 4.428066 [PAX STATUS] step_i: 1200, training loss: 4.252451 [PAX STATUS] step_i: 1300, training loss: 3.8448389 [PAX STATUS] step_i: 1400, training loss: 3.8578327 [PAX STATUS] step_i: 1500, training loss: 3.796958 ``` Copybara import of the project: -- 90f5968 by Elfie Guo <[email protected]>: Add a few related optimization pass for fp8 gemm rerwriter. Merging this change closes #16975 FUTURE_COPYBARA_INTEGRATE_REVIEW=#16975 from elfiegg:pass 90f5968 PiperOrigin-RevId: 675755585
openxla · Sep 17, 2024 · 6c1f764 · 6c1f764
1 parent 2b1c609
commit 6c1f764
Show file tree

Hide file tree

Showing 3 changed files with 85 additions and 0 deletions.
diff --git a/xla/service/gpu/gpu_compiler.cc b/xla/service/gpu/gpu_compiler.cc
@@ -1552,6 +1552,12 @@ absl::Status GpuCompiler::OptimizeHloPostLayoutAssignment(
   // Rewrite GEMMs with broadcasted inputs as strided GEMMs.
   pipeline.AddPass<GemmBroadcastFoldingRewriter>();
 
+  pipeline.AddPass<LayoutNormalization>(&NormalizeLayoutForGpuCustomCalls);
+
+  // Layout normalization will create scatters that are not simplified and
+  // also have unsorted update_window_dims.
+  pipeline.AddPass<ScatterSimplifier>();
+
   pipeline.AddPass<HostOffloadLegalize>(
       static_cast<int64_t>(stream_executor::MemoryType::kHost),
       /* after_layout= */ true);

diff --git a/xla/service/gpu/gpu_compiler_test.cc b/xla/service/gpu/gpu_compiler_test.cc
@@ -473,6 +473,62 @@ ENTRY main {
             triton_disabled_module->computation_count());
 }
 
+TEST_F(GpuCompilerTest,
+       CublasF8NumericallySameWithTritonFallbackAndWithoutTriton) {
+  auto cc = backend()
+                .default_stream_executor()
+                ->GetDeviceDescription()
+                .cuda_compute_capability();
+  if (!cc.IsAtLeastAmpere()) {
+    GTEST_SKIP() << "Autotuning results have only been generated for Ampere "
+                 << "and Hopper GPUs";
+  }
+  const absl::string_view hlo_string = R"(
+HloModule test 
+
+ENTRY main {
+  p0 = f8e4m3fn[12288,4096]{0,1} parameter(0)
+  p1 = f8e4m3fn[4096,16384]{0,1} parameter(1)
+  dot = bf16[12288,16384]{1,0} dot(p0, p1), lhs_contracting_dims={1}, rhs_contracting_dims={0}
+  bitcast = bf16[] constant(0.956)
+  broadcast = bf16[12288,16384]{1,0} broadcast(bitcast), dimensions={}
+  ROOT multiply = bf16[12288,16384]{1,0} multiply(dot, broadcast)
+  })";
+
+  HloModuleConfig config;
+  DebugOptions triton_enabled_debug_options = GetDebugOptionsForTest();
+  triton_enabled_debug_options
+      .set_xla_gpu_require_complete_aot_autotune_results(true);
+  config.set_debug_options(triton_enabled_debug_options);
+
+  // Load autotuning DB. We shouldn't depend on actual execution times in a unit
+  // test.
+  std::string path =
+      tsl::io::JoinPath(tsl::testing::XlaSrcRoot(), "service", "gpu",
+                        "gpu_compiler_test_autotune_db.textproto");
+  TF_EXPECT_OK(AutotunerUtil::LoadAutotuneResultsFromFile(path));
+
+  TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr<HloModule> module,
+                          ParseAndReturnVerifiedModule(hlo_string, config));
+  TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr<HloModule> triton_enabled_module,
+                          GetOptimizedModule(std::move(module)));
+
+  AutotunerUtil::ClearAutotuneResults();
+  DebugOptions triton_disabled_debug_options = GetDebugOptionsForTest();
+  triton_disabled_debug_options.set_xla_gpu_enable_triton_gemm(false);
+  triton_disabled_debug_options.set_xla_gpu_cublas_fallback(true);
+  config.set_debug_options(triton_disabled_debug_options);
+
+  TF_ASSERT_OK_AND_ASSIGN(module,
+                          ParseAndReturnVerifiedModule(hlo_string, config));
+  TF_ASSERT_OK_AND_ASSIGN(std::unique_ptr<HloModule> triton_disabled_module,
+                          GetOptimizedModule(std::move(module)));
+
+  EXPECT_TRUE(RunAndCompareTwoModules(std::move(triton_enabled_module),
+                                      std::move(triton_disabled_module),
+                                      ErrorSpec{1e-6, 1e-6}, false));
+}
+
 class FloatNormalizationTest : public GpuCompilerTest,
                                public ::testing::WithParamInterface<
                                    std::pair<PrimitiveType, PrimitiveType>> {};

diff --git a/xla/service/gpu/gpu_compiler_test_autotune_db.textproto b/xla/service/gpu/gpu_compiler_test_autotune_db.textproto
@@ -71,3 +71,26 @@ results {
     }
   }
 }
+results {
+  device: "CUDA: 9.0, Cores: 114, GPU clock: 1.755 GHz, Memory bandwidth: 2039 GB/s, L2 cache: 50 MB"
+  hlo: "(bf16[12288,16384]{1,0}, s8[33554432]{0}) custom-call(f8e4m3fn[4096,12288]{0,1}, f8e4m3fn[4096,16384]{0,1}, f32[], f32[], f32[], f32[]), custom_call_target=\"__cublas$lt$matmul$f8\", backend_config={\"force_earliest_schedule\":false,\"gemm_backend_config\":{\"alpha_imag\":0,\"alpha_real\":0.95703125,\"beta\":0,\"damax_output\":false,\"dot_dimension_numbers\":{\"lhs_batch_dimensions\":[],\"lhs_contracting_dimensions\":[\"0\"],\"rhs_batch_dimensions\":[],\"rhs_contracting_dimensions\":[\"0\"]},\"epilogue\":\"DEFAULT\",\"grad_x\":false,\"grad_y\":false,\"lhs_stride\":\"50331648\",\"precision_config\":{\"algorithm\":\"ALG_UNSET\",\"operand_precision\":[\"DEFAULT\",\"DEFAULT\"]},\"rhs_stride\":\"67108864\"},\"operation_queue_id\":\"0\",\"wait_on_operation_queues\":[]}"
+  result {
+    gemm {
+    }
+    run_time {
+      nanos: 1
+    }
+  }
+}
+results {
+  device: "CUDA: 9.0, Cores: 114, GPU clock: 1.755 GHz, Memory bandwidth: 2039 GB/s, L2 cache: 50 MB"
+  hlo: "{\n  tmp_0 = f8e4m3fn[12288,4096]{0,1} parameter(0)\n  tmp_1 = f8e4m3fn[4096,16384]{0,1} parameter(1)\n  tmp_2 = bf16[12288,16384]{1,0} dot(f8e4m3fn[12288,4096]{0,1} tmp_0, f8e4m3fn[4096,16384]{0,1} tmp_1), lhs_contracting_dims={1}, rhs_contracting_dims={0}\n  tmp_3 = bf16[] constant({...})\n  tmp_4 = bf16[12288,16384]{1,0} broadcast(bf16[] tmp_3), dimensions={}\n  ROOT tmp_5 = bf16[12288,16384]{1,0} multiply(bf16[12288,16384]{1,0} tmp_2, bf16[12288,16384]{1,0} tmp_4)\n}"
+  result {
+    gemm {
+      algorithm: -1
+    }
+    run_time {
+      nanos: 1
+    }
+  }
+}