[Inference] Use cse pass (#64523)

yuanlehome · web-flow · commit fe30f9f36fa3 · 2024-05-28T23:07:52.000-07:00
* use cse pass

* fix
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -114,6 +114,7 @@
 
 #include "paddle/common/flags.h"
 #include "paddle/fluid/ir_adaptor/translator/translate.h"
+#include "paddle/fluid/pir/transforms/general/common_subexpression_elimination_pass.h"
 #include "paddle/fluid/pir/transforms/general/constant_folding_pass.h"
 #include "paddle/fluid/pir/transforms/general/dead_code_elimination_pass.h"
 #include "paddle/fluid/pir/transforms/general/inplace_pass.h"
@@ -906,7 +907,7 @@ bool AnalysisPredictor::PrepareExecutor() {
         ctx->GetOrRegisterDialect<cinn::dialect::OperatorDialect>();
         ctx->GetOrRegisterDialect<pir::shape::ShapeDialect>();
         auto pass_manager = std::make_shared<::pir::PassManager>(
-            ::pir::IrContext::Instance(), 2);
+            ::pir::IrContext::Instance(), config_.pm_opt_level_);
         if (!config_.glog_info_disabled()) {
           pass_manager->EnablePrintStatistics();
         }
@@ -999,7 +1000,7 @@ bool AnalysisPredictor::PrepareExecutor() {
       // Apply some basic passes required by the framework
       ::pir::PassManager basic_pass_pm(::pir::IrContext::Instance(),
                                        config_.pm_opt_level_);
-
+      basic_pass_pm.AddPass(::pir::CreateCommonSubexpressionEliminationPass());
       auto params_sync_among_devices_pass =
           ::pir::CreateParamsSyncAmongDevicesPass();
       params_sync_among_devices_pass->SetNotOwned(pir::Pass::kPlaceAttr,
diff --git a/paddle/fluid/inference/api/paddle_pass_builder.cc b/paddle/fluid/inference/api/paddle_pass_builder.cc
@@ -610,6 +610,7 @@ const std::vector<std::string> kPirGpuPasses{
     "embedding_eltwise_layernorm_fuse_pass",
     "fused_flash_attn_pass",
     "multihead_matmul_fuse_pass",
+    "fused_weight_only_linear_pass",
     "matmul_add_act_fuse_pass",
     "fc_elementwise_layernorm_fuse_pass",
     "matmul_scale_fuse_pass",
diff --git a/paddle/fluid/pir/transforms/gpu/fused_weight_only_linear_pass.cc b/paddle/fluid/pir/transforms/gpu/fused_weight_only_linear_pass.cc
@@ -32,9 +32,6 @@ int getSMVersion() {
 #if defined(PADDLE_WITH_CUDA) && defined(PADDLE_WITH_CUTLASS)
   sm_version = paddle::platform::GetGPUComputeCapability(
       paddle::platform::GetCurrentDeviceId());
-#else
-  PADDLE_THROW(common::errors::Unavailable(
-      "fused_weight_only_linear_pass needs paddle compiled with CUDA."));
 #endif
   return sm_version;
 }
@@ -280,7 +277,7 @@ class FusedWeightOnlyLinearPass : public pir::PatternRewritePass {
         sm_version_(getSMVersion()) {}
 
   pir::RewritePatternSet InitializePatterns(pir::IrContext *context) override {
-    std::string algo = "weight_only_int4";
+    std::string algo = "weight_only_int8";
     if (Has("weight_only_algo")) {
       algo = Get<std::string>("weight_only_algo");
     }
