sync: open-core source for 0.5.0

base-convert/Cargo.toml

@@ -12,7 +12,7 @@ members = [
 ]
 
 [workspace.package]
-version = "0.4.7"
+version = "0.5.0"
 edition = "2021"
 license = "Apache-2.0"
 repository = "https://github.com/basecompute/baseRT"

base-convert/crates/base-convert/src/main.rs

@@ -8333,3 +8333,50 @@ mod tied_head_tests {
         );
     }
 }
+
+#[cfg(test)]
+mod row_group_tests {
+    use super::{pack_for_target_rows, TargetScheme};
+    use base_format::TensorDtype;
+
+    // The runtime's base_qN kernels read K / gs scales per row, so a group may
+    // never run across a row end. gemma-4-26B-A4B-it's catalog Q8 was packed
+    // before this choice existed: ffn_down_exps (K = 704) and mlp.down_proj
+    // (K = 2112) at gs 128, with every row decoding against the wrong scales.
+    fn pack(k: usize, rows: usize, target: TargetScheme) -> (base_quant::Packed, TensorDtype) {
+        let w: Vec<f32> = (0..rows * k)
+            .map(|i| ((i % 97) as f32 - 48.0) / 64.0)
+            .collect();
+        pack_for_target_rows(&w, target, Some(k), "layers.0.test.weight").unwrap()
+    }
+
+    #[test]
+    fn q8_rows_that_split_gs128_pack_at_a_dividing_group() {
+        for (k, gs) in [(704, 64), (2112, 64), (2816, 128), (1856, 64), (1888, 32)] {
+            let (p, dt) = pack(k, 3, TargetScheme::BaseQ8);
+            assert_eq!(dt, TensorDtype::BaseQ8, "K = {k}");
+            assert_eq!(p.group_size as usize, gs, "K = {k}");
+            // One f16 scale and one f16 bias per group, groups per row = K / gs.
+            assert_eq!(p.scales.len(), 3 * (k / gs) * 2, "K = {k}");
+            assert_eq!(p.biases.len(), 3 * (k / gs) * 2, "K = {k}");
+        }
+    }
+
+    #[test]
+    fn rows_no_group_divides_stay_unquantized() {
+        // gemma-4's vision ffn.down: 4304 = 16 x 269, below the smallest group.
+        let (p, dt) = pack(4304, 2, TargetScheme::BaseQ8);
+        assert_eq!(dt, TensorDtype::Bf16);
+        assert_eq!(p.group_size, 0);
+        assert!(p.scales.is_empty() && p.biases.is_empty());
+        assert_eq!(p.packed_weights.len(), 2 * 4304 * 2);
+    }
+
+    #[test]
+    fn q4_keeps_its_default_group_where_it_divides() {
+        let (p, dt) = pack(704, 2, TargetScheme::BaseQ4);
+        assert_eq!((dt, p.group_size), (TensorDtype::BaseQ4, 64));
+        let (p, _) = pack(1888, 2, TargetScheme::BaseQ4);
+        assert_eq!(p.group_size, 32);
+    }
+}

base-convert/crates/base-convert/src/selfupdate.rs

@@ -625,9 +625,17 @@ pub fn bundle_dir() -> Result<PathBuf> {
 /// The engine bundle name's platform for this machine:
 /// `basert-engine-<platform>-<v>.tar.gz`.
 pub fn platform() -> Result<&'static str> {
-    match (std::env::consts::OS, std::env::consts::ARCH) {
+    platform_for(std::env::consts::OS, std::env::consts::ARCH)
+}
+
+/// One engine per OS and CPU: Metal on Apple silicon, CUDA on Linux arm64
+/// (GB10), ROCm on Linux x86-64 (AMD Strix Halo and Strix Point; one bundle
+/// carries both chips' kernels).
+fn platform_for(os: &str, arch: &str) -> Result<&'static str> {
+    match (os, arch) {
         ("macos", "aarch64") => Ok("macos-arm64"),
         ("linux", "aarch64") => Ok("linux-arm64-cuda"),
+        ("linux", "x86_64") => Ok("linux-x64-rocm"),
         (os, arch) => bail!("no basert release is published for {os}/{arch}"),
     }
 }
@@ -2307,6 +2315,48 @@ mod tests {
         assert!(releases_in(&json!([pre]), "macos-arm64").is_empty());
     }
 
+    #[test]
+    fn each_os_and_cpu_takes_its_own_engine_bundle() {
+        assert_eq!(platform_for("macos", "aarch64").unwrap(), "macos-arm64");
+        assert_eq!(
+            platform_for("linux", "aarch64").unwrap(),
+            "linux-arm64-cuda"
+        );
+        assert_eq!(platform_for("linux", "x86_64").unwrap(), "linux-x64-rocm");
+        let err = platform_for("macos", "x86_64").unwrap_err().to_string();
+        assert!(
+            err.contains("no basert release is published for macos/x86_64"),
+            "{err}"
+        );
+        assert!(platform_for("windows", "x86_64").is_err());
+
+        // A release with all three bundles: Linux x86-64 takes the ROCm one.
+        let listing = json!([listed(
+            "v0.5.0",
+            "2026-12-10T00:00:00Z",
+            &[
+                "basert-engine-abi-1.txt",
+                "basert-engine-macos-arm64-0.5.0.tar.gz",
+                "basert-engine-macos-arm64-0.5.0.tar.gz.sha256",
+                "basert-engine-linux-arm64-cuda-0.5.0.tar.gz",
+                "basert-engine-linux-arm64-cuda-0.5.0.tar.gz.sha256",
+                "basert-engine-linux-x64-rocm-0.5.0.tar.gz",
+                "basert-engine-linux-x64-rocm-0.5.0.tar.gz.sha256",
+            ]
+        )]);
+        let found = releases_in(&listing, platform_for("linux", "x86_64").unwrap());
+        assert_eq!(found.len(), 1);
+        assert_eq!(
+            found[0].tarball,
+            "basert-engine-linux-x64-rocm-0.5.0.tar.gz"
+        );
+        assert_eq!(
+            found[0].checksum_url,
+            "https://x/v0.5.0/basert-engine-linux-x64-rocm-0.5.0.tar.gz.sha256"
+        );
+        assert_eq!(found[0].abi, Some(1));
+    }
+
     #[test]
     fn how_a_launcher_was_installed_is_read_from_where_it_is() {
         let tmp = TempDir::new().unwrap();

bindings/node/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@baseRT/node",
-  "version": "0.4.7",
+  "version": "0.5.0",
   "private": true,
   "description": "Node.js bindings for BaseRT — LLM inference engine for Apple Silicon (Metal)",
   "main": "dist/index.js",

bindings/python/baseRT/__init__.py

@@ -21,7 +21,7 @@
 from pathlib import Path
 from typing import Any, Callable, Dict, Iterator, List, Optional, Tuple, Union
 
-__version__ = "0.4.7"
+__version__ = "0.5.0"
 
 # ---------------------------------------------------------------------------
 # Library loading

bindings/python/setup.py

@@ -6,7 +6,7 @@
 
 setup(
     name="baseRT",
-    version="0.4.7",
+    version="0.5.0",
     description="Python bindings for the BaseRT LLM inference engine (Apple Silicon / Metal)",
     long_description=long_description,
     long_description_content_type="text/markdown",

bindings/rust/README.md

@@ -57,7 +57,7 @@ Add the dependency from crates.io (install the engine first, e.g.
 
 ```toml
 [dependencies]
-baseRT = "0.4"
+baseRT = "0.5"
 ```
 
 or from a checkout:

bindings/rust/baseRT-sys/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "baseRT-sys"
-version = "0.4.7"
+version = "0.5.0"
 edition = "2021"
 description = "Raw FFI bindings for the BaseRT LLM inference engine"
 license = "Apache-2.0"

bindings/rust/baseRT/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "baseRT"
-version = "0.4.7"
+version = "0.5.0"
 edition = "2021"
 description = "Safe Rust bindings for the BaseRT LLM inference engine (Apple Silicon)"
 license = "Apache-2.0"

bindings/swift/Sources/CBaseRT/include/baseRT.h

@@ -63,8 +63,8 @@ extern "C" {
 // === Versioning ===
 
 #define BASERT_VERSION_MAJOR 0
-#define BASERT_VERSION_MINOR 4
-#define BASERT_VERSION_PATCH 7
+#define BASERT_VERSION_MINOR 5
+#define BASERT_VERSION_PATCH 0
 
 /// Compile-time version, packed as `(MAJOR<<16) | (MINOR<<8) | PATCH`.
 /// Useful for `#if BASERT_VERSION >= 0x000200` feature checks.

include/baseRT/baseRT.h

@@ -63,8 +63,8 @@ extern "C" {
 // === Versioning ===
 
 #define BASERT_VERSION_MAJOR 0
-#define BASERT_VERSION_MINOR 4
-#define BASERT_VERSION_PATCH 7
+#define BASERT_VERSION_MINOR 5
+#define BASERT_VERSION_PATCH 0
 
 /// Compile-time version, packed as `(MAJOR<<16) | (MINOR<<8) | PATCH`.
 /// Useful for `#if BASERT_VERSION >= 0x000200` feature checks.