Merge 'origin/master' into hipblas

.devops/tools.sh

@@ -10,13 +10,13 @@ shift
 # Join the remaining arguments into a single string
 arg2="$@"
 
-if [[ $arg1 == '--convert' || $arg1 == '-c' ]]; then
-    python3 ./convert.py $arg2
-elif [[ $arg1 == '--quantize' || $arg1 == '-q' ]]; then
-    ./quantize $arg2
-elif [[ $arg1 == '--run' || $arg1 == '-r' ]]; then
-    ./main $arg2
-elif [[ $arg1 == '--all-in-one' || $arg1 == '-a' ]]; then
+if [[ "$arg1" == '--convert' || "$arg1" == '-c' ]]; then
+    python3 ./convert.py "$arg2"
+elif [[ "$arg1" == '--quantize' || "$arg1" == '-q' ]]; then
+    ./quantize "$arg2"
+elif [[ "$arg1" == '--run' || "$arg1" == '-r' ]]; then
+    ./main "$arg2"
+elif [[ "$arg1" == '--all-in-one' || "$arg1" == '-a' ]]; then
     echo "Converting PTH to GGML..."
     for i in `ls $1/$2/ggml-model-f16.bin*`; do
         if [ -f "${i/f16/q4_0}" ]; then
@@ -26,8 +26,8 @@ elif [[ $arg1 == '--all-in-one' || $arg1 == '-a' ]]; then
             ./quantize "$i" "${i/f16/q4_0}" q4_0
         fi
     done
-elif [[ $arg1 == '--server' || $arg1 == '-s' ]]; then
-    ./server $arg2
+elif [[ "$arg1" == '--server' || "$arg1" == '-s' ]]; then
+    ./server "$arg2"
 else
     echo "Unknown command: $arg1"
     echo "Available commands: "

Makefile

@@ -151,14 +151,11 @@ ifdef LLAMA_MPI
 	CFLAGS += -DGGML_USE_MPI -Wno-cast-qual
 	CXXFLAGS += -DGGML_USE_MPI -Wno-cast-qual
 	OBJS     += ggml-mpi.o
-
-ggml-mpi.o: ggml-mpi.c ggml-mpi.h
-	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_MPI
 
 ifdef LLAMA_OPENBLAS
-	CFLAGS  += -DGGML_USE_OPENBLAS -I/usr/local/include/openblas -I/usr/include/openblas
-	LDFLAGS += -lopenblas
+	CFLAGS  += -DGGML_USE_OPENBLAS $(shell pkg-config --cflags openblas)
+	LDFLAGS += $(shell pkg-config --libs openblas)
 endif # LLAMA_OPENBLAS
 
 ifdef LLAMA_BLIS
@@ -247,9 +244,6 @@ ifdef LLAMA_METAL
 	CXXFLAGS += -DGGML_USE_METAL
 	LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders
 	OBJS     += ggml-metal.o
-
-ggml-metal.o: ggml-metal.m ggml-metal.h
-	$(CC) $(CFLAGS) -c $< -o $@
 endif # LLAMA_METAL
 
 ifneq ($(filter aarch64%,$(UNAME_M)),)
@@ -274,6 +268,16 @@ ifneq ($(filter armv8%,$(UNAME_M)),)
 	CFLAGS += -mfp16-format=ieee -mno-unaligned-access
 endif
 
+ifdef LLAMA_METAL
+ggml-metal.o: ggml-metal.m ggml-metal.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_METAL
+
+ifdef LLAMA_MPI
+ggml-mpi.o: ggml-mpi.c ggml-mpi.h
+	$(CC) $(CFLAGS) -c $< -o $@
+endif # LLAMA_MPI
+
 ifdef LLAMA_NO_K_QUANTS
 k_quants.o: k_quants.c k_quants.h
 	$(CC) $(CFLAGS) -c $< -o $@

README.md

@@ -640,7 +640,7 @@ Please verify the [sha256 checksums](SHA256SUMS) of all downloaded model files t
 
 ```bash
 # run the verification script
-python3 .\scripts\verify-checksum-models.py
+./scripts/verify-checksum-models.py
 ```
 
 - On linux or macOS it is also possible to run the following commands to verify if you have all possible latest files in your self-installed `./models` subdirectory:

build.zig

@@ -1,9 +1,19 @@
 const std = @import("std");
+const commit_hash = @embedFile(".git/refs/heads/master");
 
-// Zig Version: 0.11.0-dev.3379+629f0d23b
+// Zig Version: 0.11.0-dev.3986+e05c242cd
 pub fn build(b: *std.build.Builder) void {
     const target = b.standardTargetOptions(.{});
     const optimize = b.standardOptimizeOption(.{});
+
+    const config_header = b.addConfigHeader(
+        .{ .style = .blank, .include_path = "build-info.h" },
+        .{
+            .BUILD_NUMBER = 0,
+            .BUILD_COMMIT = commit_hash[0 .. commit_hash.len - 1], // omit newline
+        },
+    );
+
     const lib = b.addStaticLibrary(.{
         .name = "llama",
         .target = target,
@@ -13,24 +23,21 @@ pub fn build(b: *std.build.Builder) void {
     lib.linkLibCpp();
     lib.addIncludePath(".");
     lib.addIncludePath("./examples");
-    lib.addCSourceFiles(&.{
-        "ggml.c",
-    }, &.{"-std=c11"});
-    lib.addCSourceFiles(&.{
-        "llama.cpp",
-    }, &.{"-std=c++11"});
+    lib.addConfigHeader(config_header);
+    lib.addCSourceFiles(&.{"ggml.c"}, &.{"-std=c11"});
+    lib.addCSourceFiles(&.{"llama.cpp"}, &.{"-std=c++11"});
     b.installArtifact(lib);
 
     const examples = .{
         "main",
         "baby-llama",
         "embedding",
-        // "metal",
+        "metal",
         "perplexity",
         "quantize",
         "quantize-stats",
         "save-load-state",
-        // "server",
+        "server",
         "simple",
         "train-text-from-scratch",
     };
@@ -43,16 +50,19 @@ pub fn build(b: *std.build.Builder) void {
         });
         exe.addIncludePath(".");
         exe.addIncludePath("./examples");
+        exe.addConfigHeader(config_header);
         exe.addCSourceFiles(&.{
-            std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{example_name, example_name}),
+            std.fmt.comptimePrint("examples/{s}/{s}.cpp", .{ example_name, example_name }),
             "examples/common.cpp",
         }, &.{"-std=c++11"});
         exe.linkLibrary(lib);
         b.installArtifact(exe);
+
         const run_cmd = b.addRunArtifact(exe);
         run_cmd.step.dependOn(b.getInstallStep());
         if (b.args) |args| run_cmd.addArgs(args);
-        const run_step = b.step("run_" ++ example_name, "Run the app");
+
+        const run_step = b.step("run-" ++ example_name, "Run the app");
         run_step.dependOn(&run_cmd.step);
     }
 }

examples/common.cpp

@@ -168,6 +168,18 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--rope-freq-base") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_base = std::stof(argv[i]);
+        } else if (arg == "--rope-freq-scale") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = std::stof(argv[i]);
         } else if (arg == "--memory-f32") {
             params.memory_f16 = false;
         } else if (arg == "--top-p") {
@@ -285,6 +297,7 @@ bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.lora_adapter = argv[i];
+            params.use_mmap = false;
         } else if (arg == "--lora-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -492,6 +505,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --cfg-scale N         strength of guidance (default: %f, 1.0 = disable)\n", params.cfg_scale);
     fprintf(stderr, "  --cfg-smooth-factor N smooth factor between old and new logits (default: %f, 1.0 = no smoothing)\n", params.cfg_smooth_factor);
     fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
     fprintf(stderr, "  --ignore-eos          ignore end of stream token and continue generating (implies --logit-bias 2-inf)\n");
     fprintf(stderr, "  --no-penalize-nl      do not penalize newline token\n");
     fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
@@ -520,7 +535,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     fprintf(stderr, "  --mtest               compute maximum memory usage\n");
     fprintf(stderr, "  --export              export the computation graph to 'llama.ggml'\n");
     fprintf(stderr, "  --verbose-prompt      print prompt before generation\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  -m FNAME, --model FNAME\n");
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
@@ -572,6 +587,8 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
     lparams.use_mlock    = params.use_mlock;
     lparams.logits_all   = params.perplexity;
     lparams.embedding    = params.embedding;
+    lparams.rope_freq_base  = params.rope_freq_base;
+    lparams.rope_freq_scale = params.rope_freq_scale;
 
     return lparams;
 }

examples/common.h

@@ -32,6 +32,8 @@ struct gpt_params {
     int32_t main_gpu                        = 0;   // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0}; // how split tensors should be distributed across GPUs
     int32_t n_probs                         = 0;   // if greater than 0, output the probabilities of top n_probs tokens.
+    float   rope_freq_base                  = 10000.0f; // RoPE base frequency
+    float   rope_freq_scale                 = 1.0f;     // RoPE frequency scaling factor
 
     // sampling parameters
     std::unordered_map<llama_token, float> logit_bias; // logit bias for specific tokens

examples/embd-input/README.md

@@ -17,7 +17,7 @@ make
 import torch
 
 bin_path = "../LLaVA-13b-delta-v1-1/pytorch_model-00003-of-00003.bin"
-pth_path = "./examples/embd_input/llava_projection.pth"
+pth_path = "./examples/embd-input/llava_projection.pth"
 
 dic = torch.load(bin_path)
 used_key = ["model.mm_projector.weight","model.mm_projector.bias"]

examples/embd-input/llava.py

@@ -59,7 +59,7 @@ def chat_with_image(self, image, question):
     # Also here can use pytorch_model-00003-of-00003.bin directly.
     a.load_projection(os.path.join(
         os.path.dirname(__file__) ,
-        "llava_projetion.pth"))
+        "llava_projection.pth"))
     respose = a.chat_with_image(
         Image.open("./media/llama1-logo.png").convert('RGB'),
         "what is the text in the picture?")

examples/main/README.md

@@ -293,5 +293,5 @@ These options provide extra functionality and customization when running the LLa
 -   `-mg i, --main-gpu i`: When using multiple GPUs this option controls which GPU is used for small tensors for which the overhead of splitting the computation across all GPUs is not worthwhile. The GPU in question will use slightly more VRAM to store a scratch buffer for temporary results. By default GPU 0 is used. Requires cuBLAS.
 -   `-ts SPLIT, --tensor-split SPLIT`: When using multiple GPUs this option controls how large tensors should be split across all GPUs. `SPLIT` is a comma-separated list of non-negative values that assigns the proportion of data that each GPU should get in order. For example, "3,2" will assign 60% of the data to GPU 0 and 40% to GPU 1. By default the data is split in proportion to VRAM but this may not be optimal for performance. Requires cuBLAS.
 -   `-lv, --low-vram`: Do not allocate a VRAM scratch buffer for holding temporary results. Reduces VRAM usage at the cost of performance, particularly prompt processing speed. Requires cuBLAS.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model. This allows you to adapt the pretrained model to specific tasks or domains.
+-   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.

examples/main/main.cpp

@@ -84,9 +84,17 @@ int main(int argc, char ** argv) {
         return 0;
     }
 
+    if (params.rope_freq_base != 10000.0) {
+        fprintf(stderr, "%s: warning: changing RoPE frequency base to %g (default 10000.0)\n", __func__, params.rope_freq_base);
+    }
+
+    if (params.rope_freq_scale != 1.0) {
+        fprintf(stderr, "%s: warning: scaling RoPE frequency by %g (default 1.0)\n", __func__, params.rope_freq_scale);
+    }
+
     if (params.n_ctx > 2048) {
-        fprintf(stderr, "%s: warning: model might not support context sizes greater than 2048 tokens (%d specified);"
-                "expect poor results\n", __func__, params.n_ctx);
+        fprintf(stderr, "%s: warning: base model only supports context sizes no greater than 2048 tokens (%d specified);"
+                " you are on your own\n", __func__, params.n_ctx);
     } else if (params.n_ctx < 8) {
         fprintf(stderr, "%s: warning: minimum context size is 8, using minimum size.\n", __func__);
         params.n_ctx = 8;

examples/server/README.md

@@ -16,7 +16,7 @@ Command line options:
 -   `--memory-f32`: Use 32-bit floats instead of 16-bit floats for memory key+value. Not recommended.
 -   `--mlock`: Lock the model in memory, preventing it from being swapped out when memory-mapped.
 -   `--no-mmap`: Do not memory-map the model. By default, models are mapped into memory, which allows the system to load only the necessary parts of the model as needed.
--   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model. This allows you to adapt the pretrained model to specific tasks or domains.
+-   `--lora FNAME`: Apply a LoRA (Low-Rank Adaptation) adapter to the model (implies --no-mmap). This allows you to adapt the pretrained model to specific tasks or domains.
 -   `--lora-base FNAME`: Optional model to use as a base for the layers modified by the LoRA adapter. This flag is used in conjunction with the `--lora` flag, and specifies the base model for the adaptation.
 -   `-to N`, `--timeout N`: Server read/write timeout in seconds. Default `600`.
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
@@ -66,6 +66,7 @@ Using [curl](https://curl.se/). On Windows `curl.exe` should be available in the
 ```sh
 curl --request POST \
     --url http://localhost:8080/completion \
+    --header "Content-Type: application/json" \
     --data '{"prompt": "Building a website can be done in 10 simple steps:","n_predict": 128}'
 ```
 

examples/server/chat.sh

@@ -32,6 +32,7 @@ tokenize() {
         --silent \
         --request POST \
         --url "${API_URL}/tokenize" \
+        --header "Content-Type: application/json" \
         --data-raw "$(jq -ns --arg content "$1" '{content:$content}')" \
     | jq '.tokens[]'
 }
@@ -64,6 +65,7 @@ chat_completion() {
         --no-buffer \
         --request POST \
         --url "${API_URL}/completion" \
+        --header "Content-Type: application/json" \
         --data-raw "${DATA}")
 
     printf "\n"

examples/server/server.cpp

@@ -608,6 +608,8 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stderr, "  -v, --verbose         verbose output (default: %s)\n", server_verbose ? "enabled" : "disabled");
     fprintf(stderr, "  -t N, --threads N     number of threads to use during computation (default: %d)\n", params.n_threads);
     fprintf(stderr, "  -c N, --ctx-size N    size of the prompt context (default: %d)\n", params.n_ctx);
+    fprintf(stderr, "  --rope-freq-base N    RoPE base frequency (default: %.1f)\n", params.rope_freq_base);
+    fprintf(stderr, "  --rope-freq-scale N   RoPE frequency scaling factor (default: %g)\n", params.rope_freq_scale);
     fprintf(stderr, "  -b N, --batch-size N  batch size for prompt processing (default: %d)\n", params.n_batch);
     fprintf(stderr, "  --memory-f32          use f32 instead of f16 for memory key+value (default: disabled)\n");
     fprintf(stderr, "                        not recommended: doubles context memory required and no measurable increase in quality\n");
@@ -632,7 +634,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     fprintf(stderr, "                        model path (default: %s)\n", params.model.c_str());
     fprintf(stderr, "  -a ALIAS, --alias ALIAS\n");
     fprintf(stderr, "                        set an alias for the model, will be added as `model` field in completion response\n");
-    fprintf(stderr, "  --lora FNAME          apply LoRA adapter\n");
+    fprintf(stderr, "  --lora FNAME          apply LoRA adapter (implies --no-mmap)\n");
     fprintf(stderr, "  --lora-base FNAME     optional model to use as a base for the layers modified by the LoRA adapter\n");
     fprintf(stderr, "  --host                ip address to listen (default  (default: %s)\n", sparams.hostname.c_str());
     fprintf(stderr, "  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
@@ -722,6 +724,22 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
             }
             params.n_ctx = std::stoi(argv[i]);
         }
+        else if (arg == "--rope-freq-base")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_base = std::stof(argv[i]);
+        }
+        else if (arg == "--rope-freq-scale")
+        {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.rope_freq_scale = std::stof(argv[i]);
+        }
         else if (arg == "--memory-f32" || arg == "--memory_f32")
         {
             params.memory_f16 = false;
@@ -820,6 +838,7 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 break;
             }
             params.lora_adapter = argv[i];
+            params.use_mmap = false;
         }
         else if (arg == "--lora-base")
         {

flake.nix

@@ -43,6 +43,8 @@
             "-DLLAMA_METAL=ON"
           ]);
           installPhase = ''
+            runHook preInstall
+
             mkdir -p $out/bin
             mv bin/* $out/bin/
             mv $out/bin/main $out/bin/llama
@@ -51,6 +53,8 @@
             echo "#!${llama-python}/bin/python" > $out/bin/convert.py
             cat ${./convert.py} >> $out/bin/convert.py
             chmod +x $out/bin/convert.py
+
+            runHook postInstall
           '';
           meta.mainProgram = "llama";
         };