Revert "Prep for kv-override"

This reverts commit 41e6742.

Author: Nexesenex

expose.h

@@ -41,8 +41,6 @@ struct load_model_inputs
     const char * lora_filename = nullptr;
     const char * lora_base = nullptr;
     const char * mmproj_filename = nullptr;
-    const char * kv_override = nullptr;
-    const char * data, std::vector<llama_model_kv_override> & overrides
     const bool use_mmap = false;
     const bool use_mlock = false;
     const bool use_smartcontext = false;
@@ -196,7 +194,6 @@ extern std::string executable_path;
 extern std::string lora_filename;
 extern std::string lora_base;
 extern std::string mmproj_filename;
-extern std::vector<llama_model_kv_override> kv_overrides;
 extern std::vector<std::string> generated_tokens;
 extern bool generation_finished;
 extern float last_eval_time;

gpttype_adapter.cpp

@@ -52,8 +52,6 @@ std::string executable_path = "";
 std::string lora_filename = "";
 std::string lora_base = "";
 std::string mmproj_filename = "";
-
-std::vector<llama_model_kv_override> kv_overrides;
 bool generation_finished;
 float last_process_time = 0;
 float last_eval_time = 0;
@@ -1691,7 +1689,6 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
     kcpp_data->n_ubatch = GetUBatchSize(inputs.blasubatchsize, in_file_format);
     kcpp_data->flash_attn = inputs.flash_attention;
     kcpp_data->model_filename = inputs.model_filename;
-    kcpp_data->kv_override = inputs.kv_override;
     kcpp_data->use_smartcontext = inputs.use_smartcontext;
     kcpp_data->use_contextshift = inputs.use_contextshift;
     debugmode = inputs.debugmode;
@@ -1901,8 +1898,6 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
         llama_model_params model_params = llama_model_default_params();
         llama_context_params llama_ctx_params = llama_context_default_params();
         llama_ctx_params.n_ctx = clamped_max_context_length;
-        llama_model_kv_override kv_override = imputs.kv_override;
-		
         if(kcpp_data->use_contextshift)
         {
            llama_ctx_params.n_ctx += extra_context_handle_fragmentation;

koboldcpp.py

@@ -130,7 +130,6 @@ class load_model_inputs(ctypes.Structure):
                 ("mmproj_filename", ctypes.c_char_p),
                 ("use_mmap", ctypes.c_bool),
                 ("use_mlock", ctypes.c_bool),
-                ("kv_override", ctypes.c_char_p),
                 ("use_smartcontext", ctypes.c_bool),
                 ("use_contextshift", ctypes.c_bool),
                 ("clblast_info", ctypes.c_int),
@@ -1120,7 +1119,6 @@ def load_model(model_filename):
     inputs.blasthreads = args.blasthreads
     inputs.use_mmap = (not args.nommap)
     inputs.use_mlock = args.usemlock
-    inputs.kv_override = "".encode("UTF-8")
     inputs.lora_filename = "".encode("UTF-8")
     inputs.lora_base = "".encode("UTF-8")
     if args.lora:
@@ -2691,7 +2689,7 @@ def hide_tooltip(event):
 
     tabs = ctk.CTkFrame(root, corner_radius = 0, width=windowwidth, height=windowheight-50)
     tabs.grid(row=0, stick="nsew")
-    tabnames= ["Quick Launch", "Hardware", "Tokens", "GPU AutoLayers", "Model Files", "Network", "Horde Worker","Image Gen","Audio","Extra", "Croco"]
+    tabnames= ["Quick Launch", "Hardware", "Tokens", "GPU AutoLayers", "Model Files", "Network", "Horde Worker","Image Gen","Audio","Extra"]
     navbuttons = {}
     navbuttonframe = ctk.CTkFrame(tabs, width=100, height=int(tabs.cget("height")))
     navbuttonframe.grid(row=0, column=0, padx=2,pady=2)
@@ -2791,8 +2789,6 @@ def hide_tooltip(event):
     mmproj_var = ctk.StringVar()
     nomodel = ctk.IntVar(value=0)
 
-    kv_override_var = ctk.StringVar(value="")
-
     port_var = ctk.StringVar(value=defaultport)
     host_var = ctk.StringVar(value="")
     multiuser_var = ctk.IntVar(value=1)
@@ -3242,8 +3238,6 @@ def changerunmode(a,b,c):
 
     tensor_split_entry,tensor_split_label = makelabelentry(gpu_al_tab, "Tensor Split:", tensor_split_str_vars, 8, 160, tooltip='When using multiple GPUs this option controls how large tensors should be split across all GPUs.\nUses a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order.\nFor example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1.')
 
-
-
     # load model
     makefileentry(gpu_al_tab, "Model:", "Select GGML Model File", model_var, 40, 576, onchoosefile=on_picked_model_file, filetypes=[("GGML bin or GGUF", ("*.bin","*.gguf"))] ,tooltiptxt="Select a GGUF or GGML model file on disk to be loaded.")
 
@@ -3275,8 +3269,6 @@ def togglerope(a,b,c):
     noqkvlabel.configure(text_color="#ff5555")
     qkvslider,qkvlabel,qkvtitle = makeslider(tokens_tab, "Quantize KV Cache:", quantkv_text, quantkv_var, 0, 22, 30, set=0,tooltip="Enable quantization of KV cache (KVQ). Mode 0 (F16) is default. Modes 1-12 requires FlashAttention and disables ContextShift.\nModes 15-20 work without FA, for incompatible models. 0,13,14 can work with or without.")
 
-
-
     # load model
     makefileentry(tokens_tab, "Model:", "Select GGML or GGML Model File", model_var, 50, 576, onchoosefile=on_picked_model_file, filetypes=[("GGML bin or GGUF", ("*.bin","*.gguf"))] ,tooltiptxt="Select a GGUF or GGML model file on disk to be loaded.")
 
@@ -3432,40 +3424,6 @@ def kcpp_export_template():
     makelabel(extra_tab, "Export as launcher .kcppt template (Expert Only)", 4, 0,tooltiptxt="Creates a KoboldCpp launch template for others to use.\nEmbeds JSON files directly into exported file when saving.\nWhen loaded, forces the backend to be automatically determined.\nWarning! Not recommended for beginners!")
     ctk.CTkButton(extra_tab , text = "Generate LaunchTemplate", command = kcpp_export_template ).grid(row=5,column=0, stick="w", padx= 8, pady=2)
 
-    # croco tab
-    croco_tab = tabcontent["Croco"]
-
-    # makelabelentry(croco_tab, "Context Size:" , context_var, 2, 160,tooltip="How many threads to use during BLAS processing.\nIf left blank, uses same value as regular thread count.")
-
-    makelabelentry(croco_tab, "Threads:" , threads_var, 4, 80,tooltip="How many threads to use.\nRecommended value is your CPU core count, defaults are usually OK.")
-
-    makelabelentry(croco_tab, "BLAS threads:" , blas_threads_var, 6, 80,tooltip="How many threads to use during BLAS processing.\nIf left blank, uses same value as regular thread count.")
-    
-    # makelabelentry(croco_tab, "Logical Blas Batch Size:" , blas_size_var, 8, 160,tooltip="How many tokens to process at once per batch.\nLarger values use more memory unless Physical Batch supersedes it.")
-    
-    # makelabelentry(croco_tab, "Physical Blas Batch Size:" , blasubatchsize_var, 10, 160,tooltip="How many tokens to process at once per batch.\nLarger values use more memory.")
-
-    makelabelentry(croco_tab, "GPU Layers:", gpulayers_var, 12, 80,tooltip="How many layers to offload onto the GPU.\nVRAM intensive, usage increases with model and context size.\nRequires some trial and error to find the best fit value.\n\nCommon values for total layers, accuracy not guaranteed.\n\nLlama/Mistral 7b/8b: 33\nSolar 10.7b/11b: 49\nLlama 13b: 41\nLlama 20b(stack): 63\nLlama/Yi 34b: 61\nMixtral 8x7b: 33\nLlama 70b: 81")
-
-    makelabelentry(croco_tab, "Positive Layer offset:", poslayeroffset_var, 14, 80, tooltip="Adds layers to the GPU layers autoloader calculation in case of under-exploitation of your GPU(s)..")
-
-    makelabelentry(croco_tab, "Negative Layer Offset:", neglayeroffset_var, 16, 80, tooltip="Removes layers to the GPU layers autoloader calculation in case of Out of Memory (OOM) error..")
-
-    makelabelentry(croco_tab, "Tensor Split:", tensor_split_str_vars, 18, 280, tooltip='When using multiple GPUs this option controls how large tensors should be split across all GPUs.\nUses a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order.\nFor example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1.')
-
-    makelabelentry(croco_tab, "RoPE Scale:", customrope_scale, 20, 80, tooltip="For Linear RoPE scaling. RoPE frequency scale.")
-
-    makelabelentry(croco_tab, "RoPE Base:", customrope_base, 22, 160, tooltip="For NTK Aware Scaling. RoPE frequency base.")
-
-    makelabelentry(croco_tab, "Quantize KV Cache:", quantkv_var, 24, 80, tooltip="Enable quantization of KV cache (KVQ). Mode 0 (F16) is default. Modes 1-12 requires FlashAttention and disables ContextShift.\nModes 15-20 work without FA, for incompatible models. 0,13,14 can work with or without.")
-
-    makelabelentry(croco_tab, "Opt. model metadata KV override:", kv_override_var, 26, 420, tooltip="Supersede metadata of a model, like Epislon _ e.g : llama.attention.layer_norm_rms_epsilon=float:1e5, 1.25e5, 3e6, etc.")
-
-    makefileentry(croco_tab, "Model:", "Select GGML or GGML Model File", model_var, 28, 576, onchoosefile=on_picked_model_file, filetypes=[("GGML bin or GGUF", ("*.bin","*.gguf"))] ,tooltiptxt="Select a GGUF or GGML model file on disk to be loaded.")
-    model_var.trace("w", gui_changed_modelfile)
-
-    ctk.CTkButton(croco_tab, text = "Run Benchmark", command = guibench ).grid(row=32, stick="se", padx= 0, pady=0)
-
     # launch
     def guilaunch():
         if model_var.get() == "" and sd_model_var.get() == "" and whisper_model_var.get() == "" and nomodel.get()!=1:
@@ -3592,8 +3550,6 @@ def export_vars():
             pass
         args.mmproj = None if mmproj_var.get() == "" else mmproj_var.get()
 
-        args.kv_override = None if kv_override_var.get() == "" else kv_override_var.get()
-
         args.ssl = None if (ssl_cert_var.get() == "" or ssl_key_var.get() == "") else ([ssl_cert_var.get(), ssl_key_var.get()])
         args.password = None if (password_var.get() == "") else (password_var.get())
 
@@ -3754,8 +3710,6 @@ def import_vars(dict):
 
         mmproj_var.set(dict["mmproj"] if ("mmproj" in dict and dict["mmproj"]) else "")
 
-        kv_override_var.set(dict["kv_override"] if ("kv_override" in dict and dict["kv_override"]) else "")
-
         ssl_cert_var.set("")
         ssl_key_var.set("")
         if "ssl" in dict and dict["ssl"]:
@@ -5060,9 +5014,6 @@ def range_checker(arg: str):
     advparser.add_argument("--ssl", help="Allows all content to be served over SSL instead. A valid UNENCRYPTED SSL cert and key .pem files must be provided", metavar=('[cert_pem]', '[key_pem]'), nargs='+')
     advparser.add_argument("--nocertify", help="Allows insecure SSL connections. Use this if you have cert errors and need to bypass certificate restrictions.", action='store_true')
     advparser.add_argument("--mmproj", help="Select a multimodal projector file for LLaVA.", default="")
-
-    advparser.add_argument("--kv_override", help="Supersede metadata of a model, like Epislon (e.g : llama.attention.layer_norm_rms_epsilon=float:1e5, 1.25e5, 3e6, etc)", metavar=('[kv_override]'), nargs='+')
-
     advparser.add_argument("--password", help="Enter a password required to use this instance. This key will be required for all text endpoints. Image endpoints are not secured.", default=None)
     advparser.add_argument("--ignoremissing", help="Ignores all missing non-essential files, just skipping them instead.", action='store_true')
     advparser.add_argument("--chatcompletionsadapter", help="Select an optional ChatCompletions Adapter JSON file to force custom instruct tags.", default="")