failing to link dur to "mach-o"

Author: MarcusDunn

llama-cpp-2/Cargo.toml

@@ -26,6 +26,10 @@ anyhow = "1.0.79"
 name = "grammar_bias"
 harness = false
 
+[[bench]]
+name = "generate"
+harness = false
+
 [features]
 cublas = ["llama-cpp-sys-2/cublas"]
 

llama-cpp-2/benches/generate.rs

@@ -0,0 +1,61 @@
+use anyhow::Context;
+use criterion::{Criterion, criterion_group, criterion_main};
+use pprof::criterion::{Output, PProfProfiler};
+use llama_cpp_2::context::params::LlamaContextParams;
+use llama_cpp_2::llama_backend::LlamaBackend;
+use llama_cpp_2::llama_batch::LlamaBatch;
+use llama_cpp_2::model::{AddBos, LlamaModel};
+use llama_cpp_2::model::params::LlamaModelParams;
+use llama_cpp_2::token::data_array::LlamaTokenDataArray;
+
+fn generate(c: &mut Criterion) {
+    let api = hf_hub::api::sync::ApiBuilder::new()
+        .with_progress(true)
+        .build()
+        .unwrap();
+    let file = api
+        .model("TheBloke/Llama-2-7B-Chat-GGUF".to_string())
+        .get("llama-2-7b-chat.Q4_K_M.gguf")
+        .unwrap();
+    let backend = LlamaBackend::init().unwrap();
+    let model_params = LlamaModelParams::default();
+    let model = LlamaModel::load_from_file(&backend, &file, &model_params).unwrap();
+    let mut ctx = model
+        .new_context(&backend, LlamaContextParams::default())
+        .unwrap();
+
+    c.bench_function("generate 50 tokens", |b| {
+        b.iter(|| {
+            let tokens_list = model.str_to_token("Hello, my name is", AddBos::Always).unwrap();
+            let mut n_ctx = tokens_list.len() as i32;
+            let mut batch = LlamaBatch::new(512, 1);
+            let last_index: i32 = (tokens_list.len() - 1) as i32;
+            for (i, token) in (0_i32..).zip(tokens_list.into_iter()) {
+                let is_last = i == last_index;
+                batch.add(token, i, &[0], is_last).unwrap();
+            }
+            ctx.decode(&mut batch).unwrap();
+
+            for _ in 0..50 {
+                let candidates = ctx.candidates_ith(batch.n_tokens() - 1);
+                let candidates_p = LlamaTokenDataArray::from_iter(candidates, false);
+                let new_token_id = ctx.sample_token_greedy(candidates_p);
+                if new_token_id == model.token_eos() {
+                    break;
+                }
+                batch.clear();
+                batch.add(new_token_id, n_ctx, &[0], true).unwrap();
+                n_ctx += 1;
+                ctx.decode(&mut batch).unwrap();
+            }
+            ctx.clear_kv_cache_seq(0, None, None)
+        });
+    });
+}
+
+criterion_group!(
+    name = benches;
+    config = Criterion::default().with_profiler(PProfProfiler::new(100, Output::Flamegraph(None)));
+    targets = generate
+);
+criterion_main!(benches);

llama-cpp-2/benches/grammar_bias.rs

@@ -32,7 +32,7 @@ fn criterion_benchmark(c: &mut Criterion) {
     let model_params = LlamaModelParams::default();
     let model = LlamaModel::load_from_file(&backend, &file, &model_params).unwrap();
     let mut ctx = model
-        .new_context(&backend, &LlamaContextParams::default())
+        .new_context(&backend, LlamaContextParams::default())
         .unwrap();
     let grammar = LlamaGrammar::from_str(include_str!("../src/grammar/json.gbnf")).unwrap();
 

llama-cpp-sys-2/build.rs

@@ -13,6 +13,7 @@ fn main() {
 
     let mut ggml = cc::Build::new();
     let mut ggml_cuda = if cublas_enabled { Some(cc::Build::new()) } else { None };
+    let mut ggml_metal= if cfg!(target_os = "macos") { Some(cc::Build::new()) } else { None };
     let mut llama_cpp = cc::Build::new();
 
     ggml.cpp(false);
@@ -60,20 +61,22 @@ fn main() {
         llama_cpp.define("GGML_USE_ACCELERATE", None);
         llama_cpp.define("ACCELERATE_NEW_LAPACK", None);
         llama_cpp.define("ACCELERATE_LAPACK_ILP64", None);
-        println!("cargo:rustc-link-lib=framework=Accelerate");
+        println!("cargo:rustc-link-arg=framework=Accelerate");
 
         // 	MK_LDFLAGS  += -framework Foundation -framework Metal -framework MetalKit
         // https://github.com/ggerganov/llama.cpp/blob/3c0d25c4756742ebf15ad44700fabc0700c638bd/Makefile#L509-L511
-        println!("cargo:rustc-link-lib=framework Foundation");
-        println!("cargo:rustc-link-lib=framework Metal");
-        println!("cargo:rustc-link-lib=framework MetalKit");
-
+        println!("cargo:rustc-link-arg=framework=Foundation");
+        println!("cargo:rustc-link-arg=framework=Metal");
+        println!("cargo:rustc-link-arg=framework=MetalKit");
+    }
 
-        // https://github.com/ggerganov/llama.cpp/blob/3c0d25c4756742ebf15ad44700fabc0700c638bd/Makefile#L517-L520
-        ggml
-            .file("llama.cpp/ggml-metal.m")
-            .file("llama.cpp/ggml-metal.h");
+    if let Some(ggml_metal) = &mut ggml_metal {
+        metal_hack(ggml_metal);
+        ggml_metal
+            .file("llama.cpp/ggml-metal")
+            .include("llama.cpp");
     }
+
     if cfg!(target_os = "dragonfly") {
         llama_cpp.define("__BSD_VISIBLE", None);
     }
@@ -83,6 +86,12 @@ fn main() {
         ggml_cuda.compile("ggml-cuda");
     }
 
+
+    if let Some(ggml_metal) = ggml_metal {
+        println!("compiling ggml-metal");
+        ggml_metal.compile("ggml-metal")
+    }
+
     if cfg!(target_os = "linux") {
         ggml.define("_GNU_SOURCE", None);
     }
@@ -97,6 +106,7 @@ fn main() {
 
     llama_cpp
         .define("_XOPEN_SOURCE", Some("600"))
+        .include("llama.cpp")
         .std("c++17")
         .file("llama.cpp/llama.cpp");
 
@@ -124,3 +134,43 @@ fn main() {
         .write_to_file(out_path.join("bindings.rs"))
         .expect("failed to write bindings to file");
 }
+
+
+// courtesy of https://github.com/rustformers/llm
+fn metal_hack(build: &mut cc::Build) {
+    const GGML_METAL_METAL_PATH: &str = "llama.cpp/ggml-metal.metal";
+    const GGML_METAL_PATH: &str = "llama.cpp/ggml-metal.m";
+
+    let out_dir = PathBuf::from(env::var("OUT_DIR").expect("OUT_DIR is not defined"));
+
+    let ggml_metal_path = {
+        let ggml_metal_metal = std::fs::read_to_string(GGML_METAL_METAL_PATH)
+            .expect("Could not read ggml-metal.metal")
+            .replace('\\', "\\\\")
+            .replace('\n', "\\n")
+            .replace('\r', "\\r")
+            .replace('\"', "\\\"");
+
+        let ggml_metal =
+            std::fs::read_to_string(GGML_METAL_PATH).expect("Could not read ggml-metal.m");
+
+        let needle = r#"NSString * src = [NSString stringWithContentsOfFile:sourcePath encoding:NSUTF8StringEncoding error:&error];"#;
+        if !ggml_metal.contains(needle) {
+            panic!("ggml-metal.m does not contain the needle to be replaced; the patching logic needs to be reinvestigated. Contact a `llama-cpp-sys-2` developer!");
+        }
+
+        // Replace the runtime read of the file with a compile-time string
+        let ggml_metal = ggml_metal.replace(
+            needle,
+            &format!(r#"NSString * src  = @"{ggml_metal_metal}";"#),
+        );
+
+        let patched_ggml_metal_path = out_dir.join("ggml-metal.m");
+        std::fs::write(&patched_ggml_metal_path, ggml_metal)
+            .expect("Could not write temporary patched ggml-metal.m");
+
+        patched_ggml_metal_path
+    };
+
+    build.file(ggml_metal_path);
+}