support lite xpu choose device id (#36610)

paddle/fluid/inference/analysis/argument.h

@@ -238,6 +238,7 @@ struct Argument {
   DECL_ARGUMENT_FIELD(xpu_autotune_file, XpuAutotuneFile, std::string);
   DECL_ARGUMENT_FIELD(xpu_precision, XpuPrecision, std::string);
   DECL_ARGUMENT_FIELD(xpu_adaptive_seqlen, XpuAdaptiveSeqlen, bool);
+  DECL_ARGUMENT_FIELD(xpu_device_id, XpuDeviceId, int);
 
   DECL_ARGUMENT_FIELD(use_nnadapter, UseNNAdapter, bool);
   DECL_ARGUMENT_FIELD(nnadapter_model_cache_dir, NNAdapterModelCacheDir,

paddle/fluid/inference/analysis/ir_pass_manager.cc

@@ -202,6 +202,7 @@ void IRPassManager::CreatePasses(Argument *argument,
                 new std::string(argument->xpu_autotune_file()));
       pass->Set("precision", new std::string(argument->xpu_precision()));
       pass->Set("adaptive_seqlen", new bool(argument->xpu_adaptive_seqlen()));
+      pass->Set("xpu_device_id", new int(argument->xpu_device_id()));
       // NNAdapter Related
       pass->Set("use_nnadapter", new bool(argument->use_nnadapter()));
       pass->Set("nnadapter_model_cache_dir",

paddle/fluid/inference/analysis/ir_passes/lite_subgraph_pass.cc

@@ -243,6 +243,7 @@ void LiteSubgraphPass::SetUpEngine(
   bool use_gpu = Get<bool>("use_gpu");
   bool enable_int8 = Get<bool>("enable_int8");
   bool use_xpu = Get<bool>("use_xpu");
+  int xpu_device_id = Get<int>("xpu_device_id");
   int xpu_l3_workspace_size = Get<int>("xpu_l3_workspace_size");
   int cpu_math_library_num_threads = Get<int>("cpu_math_library_num_threads");
   bool locked = Get<bool>("locked");
@@ -305,6 +306,7 @@ void LiteSubgraphPass::SetUpEngine(
   };
   config.cpu_math_library_num_threads = cpu_math_library_num_threads;
   config.xpu_l3_workspace_size = xpu_l3_workspace_size;
+  config.device_id = xpu_device_id;
   config.locked = locked;
   config.autotune = autotune;
   config.autotune_file = autotune_file;

paddle/fluid/inference/api/analysis_predictor.cc

@@ -619,6 +619,7 @@ void AnalysisPredictor::PrepareArgument() {
     argument_.SetXpuAutotuneFile(config_.xpu_autotune_file_);
     argument_.SetXpuPrecision(config_.xpu_precision_);
     argument_.SetXpuAdaptiveSeqlen(config_.xpu_adaptive_seqlen_);
+    argument_.SetXpuDeviceId(config_.xpu_device_id_);
     // NNAdapter related
     argument_.SetUseNNAdapter(config_.NNAdapter().use_nnadapter);
     argument_.SetNNAdapterDeviceNames(

paddle/fluid/inference/lite/engine.cc

@@ -67,6 +67,7 @@ paddle::lite_api::PaddlePredictor* EngineManager::Create(
   lite_cxx_config.set_xpu_conv_autotune(cfg.autotune, cfg.autotune_file);
   lite_cxx_config.set_xpu_multi_encoder_method(cfg.precision,
                                                cfg.adaptive_seqlen);
+  lite_cxx_config.set_xpu_dev_per_thread(cfg.device_id);
 #endif
 
 #ifdef LITE_SUBGRAPH_WITH_NPU

paddle/fluid/inference/lite/engine.h

@@ -39,6 +39,9 @@ struct EngineConfig {
   std::vector<std::string> neglected_passes;
   lite_api::LiteModelType model_type{lite_api::LiteModelType::kProtobuf};
   bool model_from_memory{true};
+  // TODO(wilber): now only works for xpu, lite gpu can support device_id or
+  // not?
+  int device_id = 0;
 
   // for xpu
   size_t xpu_l3_workspace_size;