Make more options configurable via GpuArtisanConfig.

LiteRT-LM-PiperOrigin-RevId: 916303890

Author: yichunk

runtime/executor/llm_executor_settings.cc

@@ -43,6 +43,10 @@ std::ostream& operator<<(std::ostream& os, const GpuArtisanConfig& config) {
   os << "enable_external_embeddings: " << config.enable_external_embeddings
      << "\n";
   os << "use_submodel: " << config.use_submodel << "\n";
+  os << "prefer_texture_weights: " << config.prefer_texture_weights << "\n";
+  os << "set_enable_host_mapped_pointer: "
+     << config.set_enable_host_mapped_pointer << "\n";
+  os << "disallow_8bit_convs: " << config.disallow_8bit_convs << "\n";
   return os;
 }
 

runtime/executor/llm_executor_settings.h

@@ -75,6 +75,15 @@ struct GpuArtisanConfig {
 
   // Whether the submodel should be used if available.
   bool use_submodel = false;
+
+  // Whether to prefer texture weights over buffers.
+  bool prefer_texture_weights = true;
+
+  // Whether the backend should directly map host memory to the GPU if possible.
+  bool set_enable_host_mapped_pointer = true;
+
+  // Performs f32 convolutions instead of any 8 bit convolutions.
+  bool disallow_8bit_convs = true;
 };
 
 std::ostream& operator<<(std::ostream& os, const GpuArtisanConfig& config);

runtime/executor/llm_executor_settings_test.cc

@@ -200,6 +200,9 @@ max_top_k: 40
 enable_decode_logits: 1
 enable_external_embeddings: 0
 use_submodel: 1
+prefer_texture_weights: 1
+set_enable_host_mapped_pointer: 1
+disallow_8bit_convs: 1
 )";
   EXPECT_EQ(oss.str(), expected_output);
 }
@@ -230,6 +233,9 @@ max_top_k: 40
 enable_decode_logits: 1
 enable_external_embeddings: 0
 use_submodel: 1
+prefer_texture_weights: 1
+set_enable_host_mapped_pointer: 1
+disallow_8bit_convs: 1
 
 max_tokens: 1024
 activation_data_type: FLOAT16
@@ -298,6 +304,9 @@ max_top_k: 40
 enable_decode_logits: 1
 enable_external_embeddings: 0
 use_submodel: 1
+prefer_texture_weights: 1
+set_enable_host_mapped_pointer: 1
+disallow_8bit_convs: 1
 
 max_tokens: 1024
 activation_data_type: FLOAT16