StaticLLMPipeline: Update config (openvinotoolkit#969)

andrei-kochin · Oct 16, 2024 · be23fc6 · be23fc6
1 parent d3bfaa5
commit be23fc6
Showing 1 changed file with 19 additions and 19 deletions.
diff --git a/src/cpp/src/llm_pipeline_static.cpp b/src/cpp/src/llm_pipeline_static.cpp
@@ -79,7 +79,6 @@ void enable_npuw_dq_if_allowed(ov::AnyMap& config,
                                const std::shared_ptr<ov::Model>& model) {
     if (allow_to_enable_npuw_dq(model)) {
         config["NPUW_DQ"] = "YES";
-        pop_option(config, "NPUW_ONLINE_AVOID");
     }
 }
 
@@ -213,33 +212,34 @@ void merge_config_with(ov::AnyMap& lhs, const ov::AnyMap& rhs) {
     }
 }
 
-ov::AnyMap get_default_prefill_config(const std::shared_ptr<ov::Model>& model) {
+ov::AnyMap get_baseline_common_config() {
     ov::AnyMap config = {
-        { "NPU_USE_NPUW", "YES" },
+        { "NPU_COMPILATION_MODE_PARAMS", "compute-layers-with-higher-precision=Sqrt,Power,ReduceMean,Add" },
+        { "NPU_USE_NPUW",  "YES" },
         { "NPUW_FOLD", "YES" },
         { "NPUW_DCOFF_TYPE", "f16" },
-        { "NPUW_DCOFF_SCALE",  "YES" },
-        { "NPUW_WEIGHTS_BANK",  "shared" },
-        { "NPU_COMPILATION_MODE_PARAMS", "compute-layers-with-higher-precision=Sqrt,Power,ReduceMean,Add" },
-        { "NPUW_ONLINE_AVOID", "P:RMSNorm/NPU" }
+        { "NPUW_DCOFF_SCALE", "YES"},
+        { "NPUW_WEIGHTS_BANK", "shared" },
+        { "NPUW_PMM", "NO" }
     };
+    return config;
+}
+
+ov::AnyMap get_default_common_config(const std::shared_ptr<ov::Model>& model) {
+    auto config = get_baseline_common_config();
+    config.emplace("NPUW_WEIGHTS_BANK_ALLOC", "CPU");
     enable_npuw_dq_if_allowed(config, model);
     return config;
 }
 
+ov::AnyMap get_default_prefill_config(const std::shared_ptr<ov::Model>& model) {
+    return get_default_common_config(model);
+}
+
 ov::AnyMap get_default_generate_config(const std::shared_ptr<ov::Model>& model) {
-    ov::AnyMap config = {
-        { "NPU_USE_NPUW", "YES" },
-        { "NPUW_FOLD", "YES" },
-        { "NPUW_DCOFF_TYPE", "f16" },
-        { "NPUW_DCOFF_SCALE", "YES" },
-        { "NPU_COMPILATION_MODE_PARAMS", "compute-layers-with-higher-precision=Sqrt,Power,ReduceMean,Add" },
-        { "NPUW_PARALLEL_COMPILE", "YES" },
-        { "NPUW_FUNCALL_ASYNC", "YES" },
-        { "NPUW_WEIGHTS_BANK",  "shared" },
-        { "NPUW_ONLINE_AVOID", "P:RMSNorm/NPU" }
-    };
-    enable_npuw_dq_if_allowed(config, model);
+    auto config = get_default_common_config(model);
+    config.emplace("NPUW_FUNCALL_ASYNC", "YES");
+    config.emplace("NPUW_PARALLEL_COMPILE", "YES");
     return config;
 }