Merge remote-tracking branch 'upstream/concedo'

.editorconfig

@@ -23,3 +23,6 @@ insert_final_newline = unset
 
 [examples/server/public/*]
 indent_size = 2
+
+[examples/llama.swiftui/llama.swiftui.xcodeproj/*]
+indent_style = tab

.github/workflows/kcpp-build-release-win-cuda.yaml

@@ -0,0 +1,36 @@
+name: Koboldcpp Builder Windows CUDA
+
+on: workflow_dispatch
+env:
+  BRANCH_NAME: ${{ github.head_ref || github.ref_name }}
+
+jobs:
+  windows:
+    runs-on: windows-latest
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v3
+        with:
+          ref: concedo_experimental
+
+      - uses: Jimver/cuda-toolkit@v0.2.11
+        id: cuda-toolkit
+        with:
+          cuda: '11.7.1'
+          method: 'network'
+          sub-packages: '["nvcc", "cudart", "cublas", "cublas_dev", "thrust", "visual_studio_integration"]'
+
+      - name: Build
+        id: cmake_build
+        run: |
+          mkdir build
+          cd build
+          cmake .. -DLLAMA_CUBLAS=ON
+          cmake --build . --config Release -j ${env:NUMBER_OF_PROCESSORS}
+
+      - name: Save artifact
+        uses: actions/upload-artifact@v3
+        with:
+          name: kcpp_windows_cuda_binary
+          path: build/bin/Release/

.gitignore

@@ -37,6 +37,7 @@ models-mnt
 /llama-bench
 /llava-cli
 /lookahead
+/lookup
 /main
 /metal
 /perplexity

CMakeLists.txt

@@ -96,7 +96,12 @@ if (LLAMA_CUBLAS)
         add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${LLAMA_CUDA_PEER_MAX_BATCH_SIZE})
 
         if (LLAMA_STATIC)
-            set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            if (WIN32)
+                # As of 12.3.1 CUDA Tookit for Windows does not offer a static cublas library
+                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas CUDA::cublasLt)
+            else ()
+                set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart_static CUDA::cublas_static CUDA::cublasLt_static)
+            endif()
         else()
             set(LLAMA_EXTRA_LIBS ${LLAMA_EXTRA_LIBS} CUDA::cudart CUDA::cublas CUDA::cublasLt)
         endif()

class.py

@@ -9,7 +9,7 @@
 import requests
 import numpy as np
 from typing import List, Optional, Union
-import os
+import os, time
 from . import koboldcpp
 
 import utils
@@ -20,11 +20,9 @@
     InferenceModel,
 )
 
-model_backend_name = "koboldcpp" #specific instead of ggml
+model_backend_name = "KoboldCPP" #specific instead of ggml
 model_backend_type = "ggml" #This should be a generic name in case multiple model backends are compatible (think Hugging Face Custom and Basic Hugging Face)
 
-kcpp_backend_loaded = False
-
 class KoboldCppException(Exception):
     """To be used for errors on cpp side of KoboldCpp."""
 
@@ -35,6 +33,7 @@ def __init__(self, **kwargs):
 class model_backend(InferenceModel):
     def __init__(self) -> None:
         super().__init__()
+        self.kcpp_backend_loaded = False
 
     def is_valid(self, model_name, model_path, menu_path):
 
@@ -257,26 +256,31 @@ def set_input_parameters(self, parameters):
 
     def unload(self):
         print("Attemping to unload library")
-        koboldcpp.unload_libs()
-        global kcpp_backend_loaded
-        kcpp_backend_loaded = False
-        pass
+        self.process.terminate()
+
 
     def _load(self, save_model: bool, initial_load: bool) -> None:
-        global kcpp_backend_loaded
         self.tokenizer = self._get_tokenizer("gpt2")
-        if not kcpp_backend_loaded:
-            kcppargs = KcppArgsObject(model=self.kcpp_filename, model_param=self.kcpp_filename,
-            port=5001, port_param=5001, host='', launch=False, lora=None, threads=self.kcpp_threads, blasthreads=self.kcpp_threads,
-            highpriority=False, contextsize=self.kcpp_ctxsize, blasbatchsize=self.kcpp_blasbatchsize, ropeconfig=[self.kcpp_ropescale, self.kcpp_ropebase],
-            smartcontext=self.kcpp_smartcontext, bantokens=None, forceversion=0, nommap=self.kcpp_nommap,
-            usemlock=False, noavx2=self.kcpp_noavx2, debugmode=self.kcpp_debugmode, skiplauncher=True, hordeconfig=None, noblas=self.kcpp_noblas,
-            useclblast=self.kcpp_useclblast, usecublas=self.kcpp_usecublas, gpulayers=self.kcpp_gpulayers, tensor_split=self.kcpp_tensor_split, config=None,
-            onready='', multiuser=False, foreground=False)
+        kcppargs = KcppArgsObject(model=self.kcpp_filename, model_param=self.kcpp_filename,
+        port=5001, port_param=5001, host='', launch=False, lora=None, threads=self.kcpp_threads, blasthreads=self.kcpp_threads,
+        psutil_set_threads=False, highpriority=False, contextsize=self.kcpp_ctxsize,
+        blasbatchsize=self.kcpp_blasbatchsize, ropeconfig=[self.kcpp_ropescale, self.kcpp_ropebase], stream=False, smartcontext=self.kcpp_smartcontext,
+        unbantokens=False, bantokens=None, usemirostat=None, forceversion=0, nommap=self.kcpp_nommap,
+        usemlock=False, noavx2=self.kcpp_noavx2, debugmode=self.kcpp_debugmode, skiplauncher=True, hordeconfig=None, noblas=self.kcpp_noblas,
+        useclblast=self.kcpp_useclblast, usecublas=self.kcpp_usecublas, gpulayers=self.kcpp_gpulayers, tensor_split=self.kcpp_tensor_split, config=None,
+        onready='', multiuser=False, foreground=False, preloadstory=None, noshift=False, remotetunnel=False)
+
 
-            koboldcpp.main(kcppargs,False) #initialize library without enabling Lite http server
-            kcpp_backend_loaded = True
-        pass
+        #koboldcpp.main(kcppargs,False) #initialize library without enabling Lite http server
+        (self.output_queue, self.input_queue, self.process) = koboldcpp.start_in_seperate_process(kcppargs)
+        while True:
+            data = self.output_queue.get()
+            if data['command'] == 'load status':
+                utils.koboldai_vars.total_layers = data['data']['total']
+                utils.koboldai_vars.loaded_layers = data['data']['loaded']
+            elif data['command'] == 'complete':
+                break
+            time.sleep(0.02)
 
     def _save_settings(self):
         pass
@@ -297,16 +301,31 @@ def _raw_generate(
         # Store context in memory to use it for comparison with generated content
         utils.koboldai_vars.lastctx = decoded_prompt
 
-        genresult = koboldcpp.generate(decoded_prompt,max_new,utils.koboldai_vars.max_length,
-        gen_settings.temp,int(gen_settings.top_k),gen_settings.top_a,gen_settings.top_p,
-        gen_settings.typical,gen_settings.tfs,gen_settings.rep_pen,gen_settings.rep_pen_range,
-        sampler_order=gen_settings.sampler_order,use_default_badwordsids=utils.koboldai_vars.use_default_badwordsids)
+        self.input_queue.put({'command': 'generate', 'data': [(decoded_prompt,max_new,utils.koboldai_vars.max_length,
+                                gen_settings.temp,int(gen_settings.top_k),gen_settings.top_a,gen_settings.top_p,
+                                gen_settings.typical,gen_settings.tfs,gen_settings.rep_pen,gen_settings.rep_pen_range),
+                               {"sampler_order": gen_settings.sampler_order, "use_default_badwordsids": utils.koboldai_vars.use_default_badwordsids}
+                                ]})
+
+        #genresult = koboldcpp.generate(decoded_prompt,max_new,utils.koboldai_vars.max_length,
+        #gen_settings.temp,int(gen_settings.top_k),gen_settings.top_a,gen_settings.top_p,
+        #gen_settings.typical,gen_settings.tfs,gen_settings.rep_pen,gen_settings.rep_pen_range,
+        #sampler_order=gen_settings.sampler_order,use_default_badwordsids=utils.koboldai_vars.use_default_badwordsids)
+
+        genresult = []
+        while True:
+            data = self.output_queue.get()
+            print(data)
+            if data['command'] == 'generated text':
+                genresult.append(data['data'])
+                if self.output_queue.empty():
+                    break
+            time.sleep(0.02)
 
-        outputs = [genresult]
         return GenerationResult(
             model=self,
             out_batches=np.array(
-                [self.tokenizer.encode(x) for x in outputs]
+                [self.tokenizer.encode(x) for x in genresult]
             ),
             prompt=prompt_tokens,
             is_whole_generation=True,

common/common.cpp

@@ -921,7 +921,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  -m FNAME, --model FNAME\n");
     printf("                        model path (default: %s)\n", params.model.c_str());
     printf("  -md FNAME, --model-draft FNAME\n");
-    printf("                        draft model for speculative decoding (default: %s)\n", params.model.c_str());
+    printf("                        draft model for speculative decoding\n");
     printf("  -ld LOGDIR, --logdir LOGDIR\n");
     printf("                        path under which to save YAML logs (no logging if unset)\n");
     printf("  --override-kv KEY=TYPE:VALUE\n");

common/common.h

@@ -45,7 +45,7 @@ struct gpt_params {
     int32_t n_ctx                           = 512;   // context size
     int32_t n_batch                         = 512;   // batch size for prompt processing (must be >=32 to use BLAS)
     int32_t n_keep                          = 0;     // number of tokens to keep from initial prompt
-    int32_t n_draft                         = 16;    // number of tokens to draft during speculative decoding
+    int32_t n_draft                         = 8;     // number of tokens to draft during speculative decoding
     int32_t n_chunks                        = -1;    // max number of chunks to process (-1 = unlimited)
     int32_t n_parallel                      = 1;     // number of parallel sequences to decode
     int32_t n_sequences                     = 1;     // number of sequences to decode
@@ -248,3 +248,4 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
 void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
+

convert-hf-to-gguf.py

@@ -182,6 +182,8 @@ def from_model_architecture(model_architecture):
             return QwenModel
         if model_architecture == "MixtralForCausalLM":
             return MixtralModel
+        if model_architecture == "PhiForCausalLM":
+            return Phi2Model
         return Model
 
     def _is_model_safetensors(self) -> bool:
@@ -221,6 +223,8 @@ def _get_model_architecture(self) -> gguf.MODEL_ARCH:
             return gguf.MODEL_ARCH.QWEN
         if arch == "MixtralForCausalLM":
             return gguf.MODEL_ARCH.LLAMA
+        if arch == "PhiForCausalLM":
+            return gguf.MODEL_ARCH.PHI2
 
         raise NotImplementedError(f'Architecture "{arch}" not supported!')
 
@@ -980,6 +984,24 @@ def write_tensors(self):
             print(f"{new_name}, n_dims = {n_dims}, {old_dtype} --> {data.dtype}")
             self.gguf_writer.add_tensor(new_name, data)
 
+
+class Phi2Model(Model):
+    def set_gguf_parameters(self):
+        block_count = self.hparams["n_layer"]
+
+        self.gguf_writer.add_name("Phi2")
+        self.gguf_writer.add_context_length(self.hparams["n_positions"])
+        self.gguf_writer.add_embedding_length(self.hparams["n_embd"])
+        self.gguf_writer.add_feed_forward_length(4 * self.hparams["n_embd"])
+        self.gguf_writer.add_block_count(block_count)
+        self.gguf_writer.add_head_count(self.hparams["n_head"])
+        self.gguf_writer.add_head_count_kv(self.hparams["n_head"])
+        self.gguf_writer.add_layer_norm_eps(self.hparams["layer_norm_epsilon"])
+        self.gguf_writer.add_rope_dimension_count(self.hparams["rotary_dim"])
+        self.gguf_writer.add_file_type(self.ftype)
+        self.gguf_writer.add_add_bos_token(False)
+
+
 ###### CONVERSION LOGIC ######
 
 

examples/CMakeLists.txt

@@ -33,6 +33,7 @@ else()
     add_subdirectory(simple)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
+    add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)

examples/baby-llama/baby-llama.cpp

@@ -575,10 +575,7 @@ static struct ggml_tensor * forward(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]
@@ -844,10 +841,7 @@ static struct ggml_tensor * forward_batch(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, n_batch]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
             assert_shape_4d(KQ_scaled, n_past + N, N, n_head, n_batch);
 
             // KQ_masked = mask_past(KQ_scaled)
@@ -1131,10 +1125,7 @@ static struct ggml_tensor * forward_lora(
 
             // KQ_scaled = KQ / sqrt(n_embd/n_head)
             // KQ_scaled shape [n_past + N, N, n_head, 1]
-            struct ggml_tensor * KQ_scaled =
-                ggml_scale(ctx0,
-                        KQ,
-                        ggml_new_f32(ctx0, 1.0f/sqrtf(float(n_embd)/n_head)));
+            struct ggml_tensor * KQ_scaled = ggml_scale(ctx0, KQ, 1.0f/sqrtf(float(n_embd)/n_head));
 
             // KQ_masked = mask_past(KQ_scaled)
             // KQ_masked shape [n_past + N, N, n_head, 1]

examples/export-lora/export-lora.cpp

@@ -309,7 +309,7 @@ static struct ggml_cgraph * build_graph_lora(
 ) {
     struct ggml_tensor * ab = ggml_mul_mat(ctx, lora_a, lora_b);
     if (scaling != 1.0f) {
-        ab = ggml_scale(ctx, ab, ggml_new_f32(ctx, scaling));
+        ab = ggml_scale(ctx, ab, scaling);
     }
     struct ggml_tensor * res = ggml_add_inplace(ctx, tensor, ab);