Merge branch 'develop' into andreybest-main

Cargo.toml

@@ -6,7 +6,7 @@ members = [
     "crates/llm",
     "crates/llm-base",
     "crates/models/*",
-    "binaries/*"
+    "binaries/*",
 ]
 resolver = "2"
 default-members = ["binaries/llm-cli", "crates/llm"]
@@ -27,12 +27,12 @@ anyhow = "1.0"
 rustyline = { version = "11.0.0", features = ["derive"] }
 serde = { version = "1.0", features = ["derive"] }
 serde_json = { version = "1.0" }
-spinoff = { version = "0.7.0", default-features = false, features = ["dots2"] }
+spinoff = { version = "0.8.0", default-features = false, features = ["dots2"] }
 clap = { version = "4.1.8", features = ["derive"] }
 memmap2 = "0.5.10"
 tracing-subscriber = { version = "0.3", features = ["env-filter"] }
 tracing = { version = "0.1", features = ["log"] }
-llm-samplers = "=0.0.6"
+llm-samplers = "=0.0.7"
 
 # Config for 'cargo dist'
 [workspace.metadata.dist]
@@ -45,7 +45,12 @@ ci = ["github"]
 # The installers to generate for each app
 installers = ["shell", "powershell"]
 # Target platforms to build apps for (Rust target-triple syntax)
-targets = ["x86_64-unknown-linux-gnu", "x86_64-apple-darwin", "x86_64-pc-windows-msvc", "aarch64-apple-darwin"]
+targets = [
+    "x86_64-unknown-linux-gnu",
+    "x86_64-apple-darwin",
+    "x86_64-pc-windows-msvc",
+    "aarch64-apple-darwin",
+]
 
 # The profile that 'cargo dist' will build with
 [profile.dist]

README.md

@@ -287,6 +287,7 @@ Absolutely! Please see the [contributing guide](./doc/CONTRIBUTING.md).
   inference API on your local machine using `llm`.
 - [secondbrain](https://github.com/juliooa/secondbrain): Desktop app to download and run LLMs locally in your computer using `llm`.
 - [floneum](https://floneum.com/): A graph editor for local AI workflows.
+- [poly](https://github.com/pixelspark/poly): A versatile LLM serving back-end with tasks, streaming completion, memory retrieval, and more.
 
 #### Libraries
 

binaries/generate-ggml-bindings/src/main.rs

@@ -27,6 +27,8 @@ fn generate_main(ggml_path: &Path, src_path: &Path) {
         .allowlist_file(r".*ggml.h")
         .header(ggml_path.join("k_quants.h").to_string_lossy())
         .allowlist_file(r".*k_quants.h")
+        .header(ggml_path.join("ggml-alloc.h").to_string_lossy())
+        .allowlist_file(r".*ggml-alloc.h")
         // Suppress some warnings
         .raw_line("#![allow(non_upper_case_globals)]")
         .raw_line("#![allow(non_camel_case_types)]")
@@ -88,6 +90,9 @@ fn generate_metal(ggml_path: &Path, src_path: &Path) {
     generate_extra("metal", ggml_path, src_path, |b| {
         b.header(ggml_path.join("ggml-metal.h").to_string_lossy())
             .allowlist_file(r".*ggml-metal\.h")
+            .raw_line("use super::ggml_tensor;")
+            .raw_line("use super::ggml_log_callback;")
+            .raw_line("use super::ggml_cgraph;")
     });
 }
 

binaries/llm-cli/src/cli_args.rs

@@ -290,6 +290,15 @@ pub struct Generate {
     /// top_p - The probability for the top tokens are added until the result is greater or equal to P and at least min_keep tokens have been seen.
     ///   p(0.95): The cumulative probability after which no more tokens are kept for sampling.
     ///   min_keep(1): Minimum tokens to keep. Setting this to 0 is not recommended.
+    ///
+    /// top_a (default: disabled) - This sampler prunes tokens that don't meet a threshold based on the most probable token. The formula is `a1 * pow(max_prob, a2)`. See https://github.com/BlinkDL/RWKV-LM#the-top-a-sampling-method for more information.
+    ///   a1(0.0): Threshold scale. A reasonable value is 0.2. Setting either a1 or a2 to 0 disables the sampler.
+    ///   a2(0.0): Threshold power. A reasonable value is 2.
+    ///   min_keep(1): Minimum tokens to keep. Setting this to 0 is not recommended.
+    ///
+    /// min_p (default: disabled) - This sampler prunes tokens that don't meet a certain percentage of the most probable token. For example if `p` is `0.05` then after `min_keep` is satisfied, other tokens must be at least 5% of the most probable token. See https://github.com/ggerganov/llama.cpp/issues/3483 for more information.
+    ///   p(0.0): Probability threshold. 0.05 to 0.2 are good starting values to try. Setting this to 0 disables the sampler.
+    ///   min_keep(1): Minimum tokens to keep. Setting this to 0 is not recommended.
     #[arg(long = "sampler", short = 's', verbatim_doc_comment)]
     pub sampler_options: Vec<String>,
 
@@ -533,7 +542,7 @@ impl ModelLoad {
         let tokenizer_source = match self.model_and_tokenizer.to_source() {
             Ok(vs) => vs,
             Err(err) => {
-                if let Some(sp) = sp.take() {
+                if let Some(mut sp) = sp.take() {
                     sp.fail(&format!("Failed to load tokenizer: {}", err));
                 }
                 return Err(err);
@@ -586,7 +595,7 @@ impl ModelLoad {
                     file_size,
                     tensor_count,
                 } => {
-                    if let Some(sp) = sp.take() {
+                    if let Some(mut sp) = sp.take() {
                         sp.success(&format!(
                             "Loaded {tensor_count} tensors ({}) after {}ms",
                             bytesize::to_string(file_size, false),
@@ -601,7 +610,7 @@ impl ModelLoad {
         if model.is_err() {
             // If we've failed at loading the model, we probably haven't stopped the spinner yet.
             // Cancel it now if needed.
-            if let Some(sp) = sp {
+            if let Some(mut sp) = sp {
                 sp.fail("Failed to load model")
             }
         }

binaries/llm-cli/src/interactive.rs

@@ -141,7 +141,7 @@ fn feed_prompt_with_spinner(
         prompt.insert(0, '\n');
     }
 
-    let sp = spinoff::Spinner::new(spinoff::spinners::Dots2, "".to_string(), None);
+    let mut sp = spinoff::Spinner::new(spinoff::spinners::Dots2, "".to_string(), None);
     let result = session.feed_prompt(
         model,
         &prompt,
@@ -165,8 +165,7 @@ fn session_ends_with_newline(session: &llm::InferenceSession) -> bool {
     session
         .decoded_tokens()
         .last()
-        .map(|t| *t == b'\n')
-        .unwrap_or(true)
+        .map_or(true, |t| *t == b'\n')
 }
 
 fn readline_loop(mut body: impl FnMut(String) -> eyre::Result<()>) -> eyre::Result<()> {