transcribe-cpp-native 0.1.2__tar.gz → 0.1.3__tar.gz

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/CMakeLists.txt

@@ -396,6 +396,22 @@ if(MSVC)
     endif()
 endif()
 
+# Windows MAX_PATH: relocate ExternalProjects (vulkan-shaders-gen) to a flat
+# <build>/e/src/ layout; inherited by add_subdirectory, so the vendored ggml
+# tree stays untouched. Anchored at CMAKE_CURRENT_BINARY_DIR so an embedding
+# project's build root stays clean, and skipped entirely when the embedder
+# already declared its own ExternalProject layout.
+if(WIN32)
+    # The module defines EP_PREFIX/EP_BASE as INHERITED directory properties;
+    # without it the get below cannot see a value set by an embedding project.
+    include(ExternalProject)
+    get_directory_property(_ep_prefix EP_PREFIX)
+    get_directory_property(_ep_base EP_BASE)
+    if(NOT _ep_prefix AND NOT _ep_base)
+        set_property(DIRECTORY PROPERTY EP_PREFIX "${CMAKE_CURRENT_BINARY_DIR}/e")
+    endif()
+endif()
+
 # ggml's CMake declares its own warning flags; let it.
 add_subdirectory(ggml)
 

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/Cargo.lock

@@ -315,7 +315,7 @@ dependencies = [
 
 [[package]]
 name = "transcribe-cpp"
-version = "0.1.2"
+version = "0.1.3"
 dependencies = [
  "hound",
  "log",
@@ -325,7 +325,7 @@ dependencies = [
 
 [[package]]
 name = "transcribe-cpp-sys"
-version = "0.1.2"
+version = "0.1.3"
 dependencies = [
  "cmake",
  "serde_json",

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/Cargo.toml

@@ -11,7 +11,7 @@
 
 [package]
 name = "transcribe-cpp-sys"
-version = "0.1.2"
+version = "0.1.3"
 edition = "2021"
 rust-version = "1.74"
 description = "Native FFI bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml"

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: transcribe-cpp-native
-Version: 0.1.2
+Version: 0.1.3
 Summary: Prebuilt native library for transcribe-cpp (speech-to-text, ggml)
 Keywords: transcription,speech-to-text,asr,ggml,native
 Author: The transcribe.cpp authors

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/README.md

@@ -45,6 +45,10 @@ cmake -B build -DTRANSCRIBE_VULKAN=ON
 cmake --build build
 ```
 
+On Windows, see the [complete build guide](docs/build-windows.md) for Vulkan
+SDK setup, Visual Studio commands, and the short-build-root fallback for
+unusually deep checkouts.
+
 For CUDA (Linux + NVIDIA GPU):
 
 ```bash

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/python/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "hatchling.build"
 
 [project]
 name = "transcribe-cpp"
-version = "0.1.2"
+version = "0.1.3"
 description = "Python bindings for transcribe.cpp"
 readme = "README.md"
 # 3.8 is EOL (2024-10); 3.9 is the floor. The binding is ctypes-only, so there
@@ -34,13 +34,13 @@ classifiers = [
 # packaging fix still resolves); the import-time version/header-hash check in
 # _library.py is the runtime backstop. check_version_sync.py gates this pin
 # against include/transcribe.h.
-dependencies = ["transcribe-cpp-native==0.1.2.*"]
+dependencies = ["transcribe-cpp-native==0.1.3.*"]
 
 [project.optional-dependencies]
 # Opt-in accelerator providers — ADDITIVE: they install alongside the default
 # provider and the best one wins at runtime. Same base-version pin contract
 # as the hard dependency (gated by check_version_sync.py).
-cu12 = ["transcribe-cpp-native-cu12==0.1.2.*"]
+cu12 = ["transcribe-cpp-native-cu12==0.1.3.*"]
 # Test-only deps. Run with: uv run --extra test pytest (from bindings/python).
 # numpy is here so the numpy PCM-input tests run in every lane instead of
 # silently skipping wherever numpy happens to be absent.

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/python/src/transcribe_cpp/__init__.py

@@ -46,7 +46,7 @@ from .errors import (
     raise_for_status,
 )
 
-__version__ = "0.1.2"
+__version__ = "0.1.3"
 
 # String-enum types, exported so callers (and type checkers) can name them.
 Backend = Literal["auto", "cpu", "metal", "vulkan", "cpu_accel", "cuda"]
@@ -270,7 +270,8 @@ class BackendDevice:
     # comparable across device kinds.
     memory_free: int
     # Registry index of this device — the value to pass as ``Model(...,
-    # gpu_device=index)`` to select it (0 selects the auto / first device).
+    # gpu_device=index)`` to select it (0 means auto: discrete GPUs are
+    # probed before integrated).
     # None when the device came from Model.device, since the underlying
     # transcribe_model_get_device() does not expose an index; correlate such a
     # device back to backends() by device_id / name instead. The index is

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/rust/sys/README.md

@@ -36,6 +36,31 @@ vcpkg setup is required on any platform. The static link is the default; the
   `libtranscribe`, or `transcribe_init_backends(dir)` for a custom provider
   directory. Implies `shared`.
 
+## Windows Vulkan builds
+
+The `vulkan` feature requires the
+[Vulkan SDK](https://vulkan.lunarg.com/sdk/home#windows) on Windows. Once the
+SDK is installed and a new terminal sees `VULKAN_SDK`, build normally:
+
+```powershell
+cargo build --features vulkan
+```
+
+Windows' legacy path limit can otherwise break ggml's nested Vulkan shader
+build. The build script handles this automatically by compiling through a
+short, per-build NTFS junction under `%LOCALAPPDATA%\tcs`; installed artifacts
+and Cargo metadata still use the durable `OUT_DIR` paths. Junction creation
+does not require administrator rights.
+
+If junction creation is blocked by filesystem or corporate policy, the build
+prints a warning and falls back to the original `OUT_DIR`. Set a short Cargo
+target directory to avoid `MAX_PATH` in that case:
+
+```powershell
+$env:CARGO_TARGET_DIR = "C:\tc-target"
+cargo build --features vulkan
+```
+
 ## Build-flag escape hatch
 
 The features above cover the common, tested configurations. Anything else CMake

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/rust/sys/build.rs

@@ -22,6 +22,9 @@
 //! Escape hatch: anything else CMake accepts can be passed via the
 //! TRANSCRIBE_CMAKE_ARGS (or CMAKE_ARGS) env var — see the passthrough at the
 //! end of main(). This is the "no Cargo feature is a hard ceiling" guarantee.
+//!
+//! Windows: the native build runs through a short NTFS junction to OUT_DIR so
+//! a stock machine builds the Vulkan backend from any checkout depth (MAX_PATH).
 
 use std::env;
 use std::path::{Path, PathBuf};
@@ -196,14 +199,142 @@ fn main() {
         }
     }
 
+    // Windows: build through a short junction to OUT_DIR so the native build
+    // stays under MAX_PATH on stock machines (see windows_short_out_dir).
+    let short = windows_short_out_dir();
+    if let Some(short) = &short {
+        cfg.out_dir(short);
+    }
+
     // Builds + installs into OUT_DIR; the returned path IS the install prefix.
     let prefix = cfg.build();
+    // Emit downstream paths via the durable OUT_DIR, not the junction — cargo
+    // caches them across builds, and the junction may be deleted between runs.
+    let prefix = if short.is_some() {
+        PathBuf::from(env::var_os("OUT_DIR").expect("OUT_DIR"))
+    } else {
+        prefix
+    };
 
     let manifest = find_manifest(&prefix)
         .unwrap_or_else(|| panic!("transcribe-link.json not found under {}", prefix.display()));
     emit_link_lines(&prefix, &manifest);
 }
 
+/// Windows MAX_PATH mitigation: a short NTFS junction (`%LOCALAPPDATA%\tcs\<hash>`,
+/// no admin needed) resolving to OUT_DIR. The Vulkan ExternalProject nests ~185
+/// chars past OUT_DIR, and MSBuild's FileTracker ignores LongPathsEnabled (FTK1011),
+/// so the build root itself must be short. None = build in OUT_DIR (non-Windows,
+/// or best-effort failure). Idempotent; hash-named per OUT_DIR so checkouts don't collide.
+fn windows_short_out_dir() -> Option<PathBuf> {
+    if !cfg!(windows) {
+        return None;
+    }
+    // Backslash-normalize: mklink rejects forward slashes (MSYS-style CARGO_TARGET_DIR).
+    let out_dir = PathBuf::from(env::var("OUT_DIR").ok()?.replace('/', "\\"));
+    let Some(base) = env::var_os("LOCALAPPDATA")
+        .or_else(|| env::var_os("TEMP"))
+        .map(PathBuf::from)
+    else {
+        println!(
+            "cargo:warning=transcribe-cpp-sys: neither LOCALAPPDATA nor TEMP is set; \
+             building in OUT_DIR (may exceed Windows MAX_PATH in deep checkouts)"
+        );
+        return None;
+    };
+    let base = base.join("tcs");
+
+    let mut hash: u64 = 0xcbf29ce484222325; // FNV-1a: stable across rustc versions
+    for b in out_dir.to_string_lossy().bytes() {
+        hash ^= u64::from(b);
+        hash = hash.wrapping_mul(0x100000001b3);
+    }
+    let link = base.join(format!("{hash:016x}"));
+
+    warn_fallback(
+        std::fs::create_dir_all(&out_dir),
+        "create junction target",
+        &out_dir,
+    )?;
+    // symlink_metadata (not exists()) so a dangling junction is detected and reclaimed.
+    if std::fs::symlink_metadata(&link).is_ok() {
+        match (
+            std::fs::canonicalize(&link),
+            std::fs::canonicalize(&out_dir),
+        ) {
+            (Ok(a), Ok(b)) if a == b => return Some(link),
+            // remove_dir fails if something non-junction squats here (e.g. a
+            // backup tool materialized it as a real tree); the warning names
+            // the path so the user knows what to delete.
+            _ => warn_fallback(std::fs::remove_dir(&link), "remove stale junction", &link)?,
+        }
+    }
+    warn_fallback(
+        std::fs::create_dir_all(&base),
+        "create junction parent",
+        &base,
+    )?;
+
+    // No std API creates junctions; mklink /J needs no extra deps.
+    let output = std::process::Command::new("cmd")
+        .arg("/C")
+        .arg("mklink")
+        .arg("/J")
+        .arg(&link)
+        .arg(&out_dir)
+        .output();
+    let created = output.as_ref().map(|o| o.status.success()).unwrap_or(false);
+    let verified = created
+        && matches!(
+            (std::fs::canonicalize(&link), std::fs::canonicalize(&out_dir)),
+            (Ok(a), Ok(b)) if a == b
+        );
+    if !verified {
+        // Best-effort: fall back to building in the deep OUT_DIR.
+        let detail = output
+            .map(|o| {
+                String::from_utf8_lossy(if o.stderr.is_empty() {
+                    &o.stdout
+                } else {
+                    &o.stderr
+                })
+                .trim()
+                .to_string()
+            })
+            .unwrap_or_else(|e| e.to_string());
+        println!(
+            "cargo:warning=transcribe-cpp-sys: could not create short build junction {} -> {} ({detail}); \
+             building in OUT_DIR (may exceed Windows MAX_PATH in deep checkouts)",
+            link.display(),
+            out_dir.display()
+        );
+        return None;
+    }
+    Some(link)
+}
+
+/// Best-effort junction setup step: on failure, warn like the mklink branch
+/// and bail to the deep-OUT_DIR fallback via `?`. Cargo hides build-script
+/// warnings for registry crates unless the build fails — so this is silent on
+/// success and visible exactly when a deep-path build dies of MAX_PATH.
+fn warn_fallback<T, E: std::fmt::Display>(
+    res: Result<T, E>,
+    action: &str,
+    path: &Path,
+) -> Option<T> {
+    match res {
+        Ok(v) => Some(v),
+        Err(e) => {
+            println!(
+                "cargo:warning=transcribe-cpp-sys: could not {action} {} ({e}); \
+                 building in OUT_DIR (may exceed Windows MAX_PATH in deep checkouts)",
+                path.display()
+            );
+            None
+        }
+    }
+}
+
 /// GNUInstallDirs picks `lib` or `lib64`; find the manifest under either.
 fn find_manifest(prefix: &Path) -> Option<PathBuf> {
     for libdir in ["lib", "lib64"] {

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/rust/transcribe-cpp/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "transcribe-cpp"
-version = "0.1.2"
+version = "0.1.3"
 edition = "2021"
 rust-version = "1.74"
 description = "Rust bindings for transcribe.cpp, a C/C++ speech-to-text library built on ggml"
@@ -40,7 +40,7 @@ dynamic-backends = ["shared", "transcribe-cpp-sys/dynamic-backends"]
 # `transcribe-cpp --no-default-features`, silently re-enabling Metal. The backend
 # features below forward explicitly, so the default (`metal`) still reaches sys
 # via `metal = ["transcribe-cpp-sys/metal"]`.
-transcribe-cpp-sys = { version = "0.1.2", path = "../../..", default-features = false }
+transcribe-cpp-sys = { version = "0.1.3", path = "../../..", default-features = false }
 thiserror = "2"
 log = "0.4"
 

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/rust/transcribe-cpp/README.md

@@ -50,6 +50,11 @@ is the safe wrapper.
 Backends are selected with cargo features forwarded to `transcribe-cpp-sys`:
 `metal` (default on Apple), `vulkan`, `cuda`, and `openmp`.
 
+On Windows, `vulkan` requires the Vulkan SDK. Deep Cargo output paths are
+shortened automatically during the native build; see the
+[Windows Vulkan build notes](https://github.com/handy-computer/transcribe.cpp/blob/main/bindings/rust/sys/README.md#windows-vulkan-builds)
+for prerequisites and the short `CARGO_TARGET_DIR` fallback.
+
 The default link is static and self-contained. Advanced packaging modes are
 available through `shared` and `dynamic-backends`; see the `transcribe-cpp-sys`
 README if you need runtime-loaded backend modules or custom

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/rust/transcribe-cpp/src/backend.rs

@@ -75,11 +75,11 @@ pub struct Device {
     pub memory_free: u64,
     /// Registry index of this device — the value to pass as
     /// [`ModelOptions::gpu_device`](crate::ModelOptions) to select it (0
-    /// selects the auto / first device). `None` when this `Device` came from
-    /// [`crate::Model::device`], since `transcribe_model_get_device` does not
-    /// expose an index; correlate such a device back to [`devices`] by
-    /// `device_id` / `name` instead. Order-dependent and not stable across
-    /// driver updates or hosts.
+    /// means auto: discrete GPUs are probed before integrated). `None` when
+    /// this `Device` came from [`crate::Model::device`], since
+    /// `transcribe_model_get_device` does not expose an index; correlate such
+    /// a device back to [`devices`] by `device_id` / `name` instead.
+    /// Order-dependent and not stable across driver updates or hosts.
     pub index: Option<usize>,
 }
 

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/rust/transcribe-cpp/src/model.rs

@@ -32,7 +32,8 @@ use crate::version;
 pub struct ModelOptions {
     /// Which backend to request. Default [`Backend::Auto`].
     pub backend: Backend,
-    /// GPU device registry index. 0 means auto / first matching device.
+    /// GPU device registry index. 0 means auto: the first device that
+    /// initializes, probing discrete GPUs before integrated.
     pub gpu_device: i32,
 }
 

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/swift/Sources/TranscribeCpp/TranscribeCpp.swift

@@ -17,7 +17,7 @@ public enum Transcribe {
     /// Version this binding was built against. Pinned here for the pre-1.0
     /// base-version load gate; the version-sync milestone will generate it from
     /// `include/transcribe.h` (the single source of truth) rather than hardcode.
-    public static let compiledVersion = "0.1.2"
+    public static let compiledVersion = "0.1.3"
 
     /// `MAJOR.MINOR.PATCH` of the linked native library.
     public static func version() -> String { String(cString: transcribe_version()) }

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/typescript/package-lock.json

@@ -1,12 +1,12 @@
 {
   "name": "transcribe-cpp",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "transcribe-cpp",
-      "version": "0.1.2",
+      "version": "0.1.3",
       "license": "MIT",
       "dependencies": {
         "koffi": "^3.0.2"
@@ -19,11 +19,11 @@
         "node": ">=22"
       },
       "optionalDependencies": {
-        "@transcribe-cpp/darwin-arm64-metal": "0.1.2",
-        "@transcribe-cpp/darwin-x64-cpu": "0.1.2",
-        "@transcribe-cpp/linux-arm64-cpu-vulkan": "0.1.2",
-        "@transcribe-cpp/linux-x64-cpu-vulkan": "0.1.2",
-        "@transcribe-cpp/win32-x64-cpu-vulkan": "0.1.2"
+        "@transcribe-cpp/darwin-arm64-metal": "0.1.3",
+        "@transcribe-cpp/darwin-x64-cpu": "0.1.3",
+        "@transcribe-cpp/linux-arm64-cpu-vulkan": "0.1.3",
+        "@transcribe-cpp/linux-x64-cpu-vulkan": "0.1.3",
+        "@transcribe-cpp/win32-x64-cpu-vulkan": "0.1.3"
       }
     },
     "node_modules/@koromix/koffi-darwin-arm64": {

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/typescript/package.json

@@ -1,6 +1,6 @@
 {
   "name": "transcribe-cpp",
-  "version": "0.1.2",
+  "version": "0.1.3",
   "description": "TypeScript/Node.js bindings for transcribe.cpp — a C/C++ speech-to-text library built on ggml",
   "type": "module",
   "exports": {
@@ -38,11 +38,11 @@
     "koffi": "^3.0.2"
   },
   "optionalDependencies": {
-    "@transcribe-cpp/darwin-arm64-metal": "0.1.2",
-    "@transcribe-cpp/darwin-x64-cpu": "0.1.2",
-    "@transcribe-cpp/linux-x64-cpu-vulkan": "0.1.2",
-    "@transcribe-cpp/linux-arm64-cpu-vulkan": "0.1.2",
-    "@transcribe-cpp/win32-x64-cpu-vulkan": "0.1.2"
+    "@transcribe-cpp/darwin-arm64-metal": "0.1.3",
+    "@transcribe-cpp/darwin-x64-cpu": "0.1.3",
+    "@transcribe-cpp/linux-x64-cpu-vulkan": "0.1.3",
+    "@transcribe-cpp/linux-arm64-cpu-vulkan": "0.1.3",
+    "@transcribe-cpp/win32-x64-cpu-vulkan": "0.1.3"
   },
   "devDependencies": {
     "@types/node": "^22.0.0",

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/bindings/typescript/src/types.ts

@@ -104,18 +104,20 @@ export interface BackendInfo {
    *  to refresh; backend-defined and not comparable across device kinds. */
   memoryFree: number;
   /** Registry index of this device — the value to pass as
-   *  {@link ModelOptions.gpuDevice} to select it (0 selects the auto / first
-   *  device). `null` when this came from `model.device`, since
-   *  `transcribe_model_get_device` does not expose an index; correlate such a
-   *  device back to {@link getAvailableBackends} by `deviceId` / `name`
-   *  instead. Order-dependent and not stable across driver updates or hosts. */
+   *  {@link ModelOptions.gpuDevice} to select it (0 means auto: discrete
+   *  GPUs are probed before integrated). `null` when this came from
+   *  `model.device`, since `transcribe_model_get_device` does not expose an
+   *  index; correlate such a device back to {@link getAvailableBackends} by
+   *  `deviceId` / `name` instead. Order-dependent and not stable across
+   *  driver updates or hosts. */
   index: number | null;
 }
 
 export interface ModelOptions {
   /** "auto" (default), or an explicit backend. */
   backend?: Backend;
-  /** GPU device registry index. 0 means auto / first matching device. */
+  /** GPU device registry index. 0 means auto: the first device that
+   *  initializes, probing discrete GPUs before integrated. */
   gpuDevice?: number;
 }
 

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/docs/build-windows.md

@@ -163,23 +163,12 @@ winget install --id KhronosGroup.VulkanSDK --accept-source-agreements --accept-p
 The SDK sets a machine-wide `VULKAN_SDK` environment variable — **open a new
 terminal** so it's visible.
 
-> ### The one real gotcha: build from a SHORT path
-> ggml builds its `vulkan-shaders-gen` helper as a *nested* ExternalProject,
-> which creates very deep intermediate paths like
-> `...\ggml\src\ggml-vulkan\vulkan-shaders-gen-prefix\src\vulkan-shaders-gen-build\CMakeFiles\CMakeScratch\TryCompile-xxxxx\...`.
-> Under a normal repo path these blow past Windows' 260-char `MAX_PATH`
-> limit, and the build dies early with a misleading
-> `error MSB6003: The specified task executable "link.exe" could not be run`
-> / `DirectoryNotFoundException` on a `.tlog` path — i.e. the compiler check
-> reports "broken." It is **not** a compiler or Vulkan problem.
->
-> The fix is to put the build directory at a short root, e.g. `C:\bv`.
-> (Alternatively, enable Win32 long paths, but a short build dir is the
-> reliable one — MSBuild's file tracker doesn't fully honor long paths.)
+A plain `cmake -B build` from a normal checkout location works — no special
+build root needed:
 
 ```powershell
-cmake -B C:\bv -S . -DTRANSCRIBE_VULKAN=ON
-cmake --build C:\bv --target transcribe-cli --config Release
+cmake -B build -DTRANSCRIBE_VULKAN=ON
+cmake --build build --target transcribe-cli --config Release
 ```
 
 A successful configure prints `Found Vulkan: ... found components: glslc`
@@ -187,12 +176,26 @@ and `Including Vulkan backend`. Run it the same way; the CLI auto-selects the
 GPU:
 
 ```powershell
-C:\bv\bin\Release\transcribe-cli.exe `
+build\bin\Release\transcribe-cli.exe `
   -m models\parakeet-tdt-0.6b-v3\parakeet-tdt-0.6b-v3-Q8_0.gguf `
   samples\jfk.wav
 # backend: Vulkan0  (e.g. "Intel(R) Iris(R) Xe Graphics")
 ```
 
+> ### Only for very deep checkouts: MAX_PATH
+> ggml builds its `vulkan-shaders-gen` helper as a nested ExternalProject;
+> transcribe.cpp flattens it to `<build>\e\src\` on Windows so normal paths
+> stay under Windows' 260-char `MAX_PATH` limit. But if your build directory
+> path is itself very long (roughly 120+ characters), the intermediate paths
+> can still overflow, and the build dies early with a misleading
+> `error MSB6003: The specified task executable "link.exe" could not be run`
+> / `DirectoryNotFoundException` on a `.tlog` path — i.e. the compiler check
+> reports "broken." It is **not** a compiler or Vulkan problem.
+>
+> The fix is to put the build directory at a short root instead, e.g.
+> `cmake -B C:\bv -S . -DTRANSCRIBE_VULKAN=ON`. (Enabling Win32 long paths
+> does not reliably help — MSBuild's file tracker doesn't fully honor them.)
+
 > **Performance note — measure *warm*, not the first run.** A single
 > `transcribe-cli` invocation pays a large one-time cost on Vulkan: the
 > backend compiles its compute pipelines (SPIR-V → device shaders) and
@@ -214,4 +217,4 @@ C:\bv\bin\Release\transcribe-cli.exe `
 | winget install fails with `1602` | UAC / elevation prompt was declined | rerun and accept the prompt |
 | `git` / `cmake` not found after install | shell has stale `PATH` | open a new terminal |
 | build is Debug / slow | VS generator is multi-config | add `--config Release` |
-| Vulkan build: `MSB6003 ... link.exe could not be run` / `DirectoryNotFoundException` on a `.tlog` | `MAX_PATH` (260) exceeded by ggml's nested shader-gen ExternalProject | build from a short root, e.g. `cmake -B C:\bv ...` |
+| Vulkan build: `MSB6003 ... link.exe could not be run` / `DirectoryNotFoundException` on a `.tlog` | `MAX_PATH` (260) exceeded — build dir path too deep even for the flattened `e\src` shader-gen layout | build from a short root, e.g. `cmake -B C:\bv ...` |

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/include/transcribe.h

@@ -153,7 +153,7 @@
  */
 #define TRANSCRIBE_VERSION_MAJOR 0
 #define TRANSCRIBE_VERSION_MINOR 1
-#define TRANSCRIBE_VERSION_PATCH 2
+#define TRANSCRIBE_VERSION_PATCH 3
 
 #define TRANSCRIBE_VERSION_STRINGIZE_(x) #x
 #define TRANSCRIBE_VERSION_STRINGIZE(x)  TRANSCRIBE_VERSION_STRINGIZE_(x)
@@ -634,15 +634,17 @@ TRANSCRIBE_API bool transcribe_model_accepts_ext_kind(const struct transcribe_mo
 /*
  * Backend request.
  *
- * AUTO    Pick the best available backend. Takes the first GPU/IGPU
- *         device that successfully initializes in ggml's device
- *         registry order — which is build-time prioritized (Metal on
- *         Apple, Vulkan / CUDA / SYCL on Linux, …). Host-memory
- *         accelerators (BLAS, AMX, …) are additionally layered onto
- *         the scheduler when present — they run on the same memory
- *         as the CPU backend and are orthogonal to the GPU/CPU split.
- *         Always succeeds: CPU is the final fallback when no GPU
- *         initializes.
+ * AUTO    Pick the best available backend. Takes the first GPU device
+ *         that successfully initializes, probing every discrete GPU
+ *         before any integrated GPU; within a tier, devices are tried
+ *         in ggml's device registry order — which is build-time
+ *         prioritized (Metal on Apple, Vulkan / CUDA / SYCL on
+ *         Linux, …). An integrated GPU is selected only when no
+ *         discrete GPU initializes. Host-memory accelerators (BLAS,
+ *         AMX, …) are additionally layered onto the scheduler when
+ *         present — they run on the same memory as the CPU backend
+ *         and are orthogonal to the GPU/CPU split. Always succeeds:
+ *         CPU is the final fallback when no GPU initializes.
  *
  * CPU     Strict CPU only. No GPU, no IGPU, and no host-memory
  *         accelerators (BLAS/AMX). This is the right choice for
@@ -875,9 +877,11 @@ TRANSCRIBE_API transcribe_status transcribe_model_get_device(const struct transc
  *             for the semantics of each value. Default is AUTO.
  *
  * gpu_device: Multi-GPU selector. 0 (the default) means "auto / the first
- *             device of the chosen kind": AUTO picks the first GPU/IGPU in
- *             ggml's registry order, and explicit METAL/VULKAN/CUDA requests
- *             pick the first matching device, as before.
+ *             device of the chosen kind": AUTO picks the first GPU that
+ *             initializes, and explicit METAL/VULKAN/CUDA requests pick the
+ *             first matching device — in both cases probing every discrete
+ *             GPU before any integrated GPU, in ggml's registry order
+ *             within each tier.
  *
  *             A value > 0 selects the GPU/IGPU device at that global ggml
  *             registry index — the same index space transcribe_get_backend_device()
@@ -895,8 +899,9 @@ TRANSCRIBE_API transcribe_status transcribe_model_get_device(const struct transc
  *             device whose vendor doesn't match an explicit GPU request, or
  *             is non-zero alongside a CPU / CPU_ACCEL request (there is no
  *             GPU to select). Note there is no way to explicitly select the
- *             device at registry index 0; that is exactly what AUTO /
- *             first-of-kind already picks.
+ *             device at registry index 0 — 0 is the auto sentinel. An
+ *             integrated GPU sitting at index 0 is therefore reachable only
+ *             via the probe order, when no discrete GPU initializes.
  */
 struct transcribe_model_load_params {
     uint64_t                   struct_size;

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/src/transcribe-backend.cpp

@@ -86,6 +86,22 @@ BackendKind classify_device(ggml_backend_dev_t dev) {
     return classify_backend_type(dev_type, reg_name);
 }
 
+std::vector<size_t> gpu_probe_order(const std::vector<enum ggml_backend_dev_type> & dev_types) {
+    std::vector<size_t> order;
+    order.reserve(dev_types.size());
+    for (size_t i = 0; i < dev_types.size(); ++i) {
+        if (dev_types[i] == GGML_BACKEND_DEVICE_TYPE_GPU) {
+            order.push_back(i);
+        }
+    }
+    for (size_t i = 0; i < dev_types.size(); ++i) {
+        if (dev_types[i] == GGML_BACKEND_DEVICE_TYPE_IGPU) {
+            order.push_back(i);
+        }
+    }
+    return order;
+}
+
 namespace {
 
 // Shared body for safe_* teardown wrappers. The test hook fires after the

{transcribe_cpp_native-0.1.2 → transcribe_cpp_native-0.1.3}/src/transcribe-backend.h

@@ -52,6 +52,15 @@ BackendKind classify_backend_type(enum ggml_backend_dev_type dev_type, const cha
 // vendor. Never returns Unknown for a valid device pointer.
 BackendKind classify_device(ggml_backend_dev_t dev);
 
+// Probe order for GPU device selection. `dev_types` holds the ggml
+// device type of every registry device, indexed by registry position;
+// the returned indices are the candidates to try: every discrete GPU
+// (registry order) before any integrated GPU (registry order), non-GPU
+// devices excluded. Discrete-first matters because Vulkan enumeration
+// on hybrid-graphics machines often lists the display iGPU before the
+// dGPU. Unit-testable core for try_init_kind.
+std::vector<size_t> gpu_probe_order(const std::vector<enum ggml_backend_dev_type> & dev_types);
+
 // A resolved backend plan. Produced by load_common::init_backends
 // from a transcribe_backend_request and consumed by every helper
 // that needs to know where the graph will run.