From 4b2ca1169c8a3fb0891e0807030327db60093b15 Mon Sep 17 00:00:00 2001
From: Concedo <39025047+LostRuins@users.noreply.github.com>
Date: Wed, 13 Aug 2025 19:28:53 +0800
Subject: [PATCH] more consistency fixes

---
 koboldcpp.py                     | 2 +-
 otherarch/embeddings_adapter.cpp | 2 ++
 otherarch/tts_adapter.cpp        | 2 ++
 3 files changed, 5 insertions(+), 1 deletion(-)

diff --git a/koboldcpp.py b/koboldcpp.py
index 79d9547e2..d63bfe0eb 100644
--- a/koboldcpp.py
+++ b/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
diff --git a/otherarch/embeddings_adapter.cpp b/otherarch/embeddings_adapter.cpp
index daad79ab6..e9a57e031 100644
--- a/otherarch/embeddings_adapter.cpp
+++ b/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);
diff --git a/otherarch/tts_adapter.cpp b/otherarch/tts_adapter.cpp
index 56aa605b6..86caf67e5 100644
--- a/otherarch/tts_adapter.cpp
+++ b/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;