more consistency fixes

koboldcpp.py

@@ -1898,7 +1898,7 @@ def embeddings_load_model(model_filename):
     inputs = embeddings_load_model_inputs()
     inputs.model_filename = model_filename.encode("UTF-8")
     inputs.gpulayers = (999 if args.embeddingsgpu else 0)
-    inputs.flash_attention = False
+    inputs.flash_attention = args.flashattention
     inputs.threads = args.threads
     inputs.use_mmap = args.usemmap
     inputs.embeddingsmaxctx = (args.embeddingsmaxctx if args.embeddingsmaxctx else args.contextsize) # for us to clamp to contextsize if embeddingsmaxctx unspecified

otherarch/embeddings_adapter.cpp

@@ -120,6 +120,8 @@ bool embeddingstype_load_model(const embeddings_load_model_inputs inputs)
     model_params.use_mmap = inputs.use_mmap;
     model_params.use_mlock = false;
     model_params.n_gpu_layers = inputs.gpulayers; //offload if possible
+    int kcpp_parseinfo_maindevice = inputs.kcpp_main_gpu<=0?0:inputs.kcpp_main_gpu;
+    model_params.main_gpu = kcpp_parseinfo_maindevice;
     model_params.split_mode = llama_split_mode::LLAMA_SPLIT_MODE_LAYER;
 
     llama_model * embeddingsmodel = llama_model_load_from_file(modelfile.c_str(), model_params);

otherarch/tts_adapter.cpp

@@ -532,6 +532,8 @@ bool ttstype_load_model(const tts_load_model_inputs inputs)
     tts_model_params.use_mlock = false;
     tts_model_params.n_gpu_layers = inputs.gpulayers; //offload if possible
     tts_model_params.split_mode = llama_split_mode::LLAMA_SPLIT_MODE_LAYER;
+    int kcpp_parseinfo_maindevice = inputs.kcpp_main_gpu<=0?0:inputs.kcpp_main_gpu;
+    tts_model_params.main_gpu = kcpp_parseinfo_maindevice;
     tts_ctx_params.n_ctx = 8192;
     tts_ctx_params.offload_kqv = true;
     tts_ctx_params.n_batch = 8192;