Add OpenVINO backend for Intel iGPU/dGPU (Arc)

Mirrors the rocm/cuda pattern: ort::ep::OpenVINO with device_type "GPU"
(covers both Intel integrated graphics and Arc discrete GPUs through the
same EP), load-dynamic/dlopen so no OpenVINO toolkit is needed at build
time, and its own --openvino flag / backend = "openvino" config value.
Folded into the `full` feature alongside npu/rocm/cuda.

OpenVINO's provider options go through a generic key/value FFI interface
rather than a fixed C struct (unlike MIGraphX's OrtMIGraphXProviderOptions),
so it should be less exposed to the ABI-version-skew crash MIGraphX hit --
but that's inference from the EP's design, not verified against real
hardware. Like CUDA, this is compile-checked only: no Intel GPU in this
dev environment to runtime-verify against.

Version bump: breadmill 0.2.2 -> 0.2.3.

Cargo.lock

@@ -133,7 +133,7 @@ dependencies = [
 
 [[package]]
 name = "breadmill"
-version = "0.2.2"
+version = "0.2.3"
 dependencies = [
  "breadsearch-shared",
  "hex",

DESIGN.md

@@ -80,7 +80,9 @@ Start from `breadbox/breadbox/src/main.rs`. **Reuse verbatim:** the gtk4-layer-s
 - nomic prefixes + mean-pool + normalize must match between index and query or recall collapses.
 - `ort` linking: prefer the crate's downloaded/bundled ONNX Runtime to avoid version skew with Arch's `onnxruntime`.
 - Office formats (docx/odt) are best-effort in v1; md/txt/org/pdf are the reliable path.
-- GPU EPs (ROCm/CUDA) fail to register silently at the ONNX Runtime level and fall back to CPU — always check
+- GPU EPs (ROCm/CUDA/OpenVINO) fail to register silently at the ONNX Runtime level and fall back to CPU — always
-  startup logs for `Successfully registered` before trusting a GPU build is actually accelerating. See
+  check startup logs for `Successfully registered` before trusting a GPU build is actually accelerating. See
   [README: GPU backend notes](README.md#gpu-backend-notes) for the MIGraphX-vs-ROCMExecutionProvider distinction
   and the per-shape JIT-compile-and-cache behavior that matters for interactive query latency.
+- CUDA and OpenVINO are compile-checked only — no NVIDIA or Intel GPU hardware in this dev environment (AMD-only)
+  to runtime-verify against, unlike ROCm which was confirmed end-to-end on real hardware.

README.md

@@ -21,10 +21,11 @@ Optional features:
 
 | Feature | What it adds |
 |---------|-------------|
-| `npu`   | AMD XDNA NPU via VitisAI ONNX Runtime EP (requires Ryzen AI SDK) |
+| `npu`      | AMD XDNA NPU via VitisAI ONNX Runtime EP (requires Ryzen AI SDK) |
-| `rocm`  | AMD iGPU via the MIGraphX ONNX Runtime EP (ROCm-backed) |
+| `rocm`     | AMD iGPU via the MIGraphX ONNX Runtime EP (ROCm-backed) |
-| `cuda`  | NVIDIA GPU via the CUDA ONNX Runtime EP |
+| `cuda`     | NVIDIA GPU via the CUDA ONNX Runtime EP |
-| `full`  | All three of the above in one binary |
+| `openvino` | Intel iGPU/dGPU (Arc) via the OpenVINO ONNX Runtime EP |
+| `full`     | All four of the above in one binary |
 
 ```
 # NPU build
@@ -36,25 +37,29 @@ cargo build --release -p breadmill --features rocm
 # CUDA (NVIDIA GPU) build
 cargo build --release -p breadmill --features cuda
 
+# OpenVINO (Intel iGPU/dGPU) build
+cargo build --release -p breadmill --features openvino
+
 # All backends in one binary (what the release build ships)
 cargo build --release -p breadmill --features full
 ```
 
-`rocm`/`cuda`/`npu` all use `ort`'s `load-dynamic` mode: at runtime, breadmill
+`rocm`/`cuda`/`npu`/`openvino` all use `ort`'s `load-dynamic` mode: at runtime,
-dlopens whatever `libonnxruntime.so` the dynamic linker resolves (or
+breadmill dlopens whatever `libonnxruntime.so` the dynamic linker resolves (or
 `ORT_DYLIB_PATH` if set). GPU acceleration only works if that ONNX Runtime
 build actually has the matching execution provider compiled in — breadmill
 logs a clear `Successfully registered` / `not enabled in this build` line for
 this at startup (see [GPU backend notes](#gpu-backend-notes) below).
 
-Because all three are dlopen-based, `full` doesn't require the NPU/ROCm/CUDA
+Because all four are dlopen-based, `full` doesn't require the NPU/ROCm/CUDA/
-toolkits to be installed at build time — only at run time, and only for
+OpenVINO toolkits to be installed at build time — only at run time, and only
-whichever single backend you actually select via `--npu`/`--rocm`/`--cuda`
+for whichever single backend you actually select via
-or `backend` in config.toml. The **released binaries are built with
+`--npu`/`--rocm`/`--cuda`/`--openvino` or `backend` in config.toml. The
-`full`**: same binary works CPU-only out of the box, and picks up NPU/ROCm/CUDA
+**released binaries are built with `full`**: same binary works CPU-only out
-acceleration on a machine that has the matching ONNX Runtime available,
+of the box, and picks up NPU/ROCm/CUDA/OpenVINO acceleration on a machine
-without needing a different download. An explicit `--npu`/`--rocm`/`--cuda`
+that has the matching ONNX Runtime available, without needing a different
-flag always overrides `backend` in config.toml, not the other way around.
+download. An explicit `--npu`/`--rocm`/`--cuda`/`--openvino` flag always
+overrides `backend` in config.toml, not the other way around.
 
 ## Setup
 
@@ -117,6 +122,7 @@ breadmill status
 breadmill --npu
 breadmill --rocm
 breadmill --cuda
+breadmill --openvino
 ```
 
 ## Config
@@ -137,7 +143,7 @@ snippet_len  = 200         # max characters in result snippet
 [model]
 name         = "nomic-embed-text-v1.5"
 dim          = 768
-backend      = "cpu"       # "cpu", "npu", "rocm", or "cuda"
+backend      = "cpu"       # "cpu", "npu", "rocm", "cuda", or "openvino"
 ```
 
 `roots` and `excludes` support `~/` expansion. The index respects `.gitignore` files found during the walk.
@@ -154,10 +160,11 @@ Set `backend = "npu"` in config (or pass `--npu`) when running a build compiled
 
 ### GPU backend notes
 
-Both `rocm` and `cuda` need a system ONNX Runtime that was actually built with
+`rocm`, `cuda`, and `openvino` all need a system ONNX Runtime that was
-the matching execution provider — the crate's own downloaded binary is CPU-only.
+actually built with the matching execution provider — the crate's own
-Point `ORT_DYLIB_PATH` at one, or install a distro package that provides
+downloaded binary is CPU-only. Point `ORT_DYLIB_PATH` at one, or install a
-`libonnxruntime.so` with the EP baked in and let the dynamic linker find it.
+distro package that provides `libonnxruntime.so` with the EP baked in and
+let the dynamic linker find it.
 
 **ROCm (`--rocm` / `backend = "rocm"`)** targets ONNX Runtime's **MIGraphX**
 execution provider, not the classic `ROCMExecutionProvider`. Distro
@@ -183,6 +190,18 @@ noticeable for interactive query embedding.
 CUDA/cuDNN install. Unverified on real NVIDIA hardware in this repo — only
 compile-checked, since development happened on an AMD-only machine.
 
+**OpenVINO (`--openvino` / `backend = "openvino"`)** targets
+`OpenVINOExecutionProvider` with `device_type = "GPU"`, covering both Intel
+iGPUs and Arc dGPUs through the same EP (OpenVINO abstracts Intel's whole
+hardware line — CPU/GPU/NPU — behind one provider and a device-type string).
+Needs an OpenVINO-enabled ONNX Runtime and the OpenVINO runtime itself
+installed. Its options go through a generic key/value FFI interface rather
+than a fixed C struct (unlike MIGraphX's `OrtMIGraphXProviderOptions`), so it
+should be less exposed to the kind of ABI-version-skew crash MIGraphX hit —
+but that's inference from reading the EP's design, not something verified
+against real Intel GPU hardware. Also unverified on real hardware in this
+repo — only compile-checked, for the same reason as CUDA.
+
 ## Runtime paths
 
 | Purpose | Path |

breadmill/Cargo.toml

@@ -1,6 +1,6 @@
 [package]
 name = "breadmill"
-version = "0.2.2"
+version = "0.2.3"
 edition = "2021"
 license = "MIT"
 
@@ -14,13 +14,17 @@ npu = ["ort/vitis", "ort/load-dynamic"]
 # libonnxruntime_providers_rocm.so, which most distros don't package.
 rocm = ["ort/migraphx", "ort/load-dynamic"]
 cuda = ["ort/cuda", "ort/load-dynamic"]
+# Intel iGPU/dGPU (Arc) + CPU via OpenVINO. One EP covers Intel's whole
+# hardware line by device_type string ("CPU"/"GPU"/"GPU.0"/"NPU"/"HETERO:...");
+# we always request "GPU" since CPU is already covered by the cpu backend.
+openvino = ["ort/openvino", "ort/load-dynamic"]
 # All backends in one binary. Safe to combine: every backend here uses
 # ort's load-dynamic (dlopen) mode, so none of this links against an actual
-# NPU/ROCm/CUDA toolkit at build time — which ONNX Runtime actually gets
+# NPU/ROCm/CUDA/OpenVINO toolkit at build time — which ONNX Runtime actually
-# loaded (and thus which EPs are really available) is decided at runtime by
+# gets loaded (and thus which EPs are really available) is decided at
-# ORT_DYLIB_PATH / the dynamic linker, per the --npu/--rocm/--cuda flag or
+# runtime by ORT_DYLIB_PATH / the dynamic linker, per the --npu/--rocm/
-# `backend` config value in use for that run.
+# --cuda/--openvino flag or `backend` config value in use for that run.
-full = ["npu", "rocm", "cuda"]
+full = ["npu", "rocm", "cuda", "openvino"]
 
 [[bin]]
 name = "breadmill"

breadmill/src/embed.rs

@@ -27,6 +27,11 @@ pub enum Backend {
     Rocm,
     /// NVIDIA GPU via the CUDA ONNX Runtime execution provider.
     Cuda,
+    /// Intel iGPU/dGPU (Arc) via the OpenVINO ONNX Runtime execution
+    /// provider, requesting device_type "GPU". `cache_dir` stores OpenVINO's
+    /// compiled-model blobs between runs (its own `with_cache_dir`, not an
+    /// env var — unlike MIGraphX this doesn't need a workaround).
+    OpenVino { cache_dir: PathBuf },
 }
 
 pub struct OrtEmbedder {
@@ -165,6 +170,7 @@ fn configure_eps(builder: SessionBuilder, backend: &Backend) -> Result<SessionBu
         Backend::Npu { cache_dir } => npu_session(builder, cache_dir),
         Backend::Rocm => rocm_session(builder),
         Backend::Cuda => cuda_session(builder),
+        Backend::OpenVino { cache_dir } => openvino_session(builder, cache_dir),
     }
 }
 
@@ -253,6 +259,36 @@ fn cuda_session(builder: SessionBuilder) -> Result<SessionBuilder, String> {
     Ok(builder)
 }
 
+// ---- OpenVINO EP (Intel iGPU/dGPU) -------------------------------------------
+
+#[cfg(feature = "openvino")]
+fn openvino_session(builder: SessionBuilder, cache_dir: &Path) -> Result<SessionBuilder, String> {
+    let ov_cache = cache_dir.join("openvino");
+    std::fs::create_dir_all(&ov_cache).map_err(|e| e.to_string())?;
+    eprintln!("breadmill: using OpenVINO execution provider (device_type GPU)");
+    eprintln!("breadmill: OpenVINO model cache: {}", ov_cache.display());
+    eprintln!(
+        "breadmill: note — check the log line above/below for \"Successfully registered \
+         `OpenVINOExecutionProvider`\"; if it's missing, the ONNX Runtime in use wasn't built \
+         with OpenVINO support and inference silently fell back to CPU"
+    );
+    builder
+        .with_execution_providers([
+            ort::ep::OpenVINO::default()
+                .with_device_type("GPU")
+                .with_cache_dir(ov_cache.to_string_lossy())
+                .build(),
+            ort::ep::CPU::default().build(),
+        ])
+        .map_err(|e| e.to_string())
+}
+
+#[cfg(not(feature = "openvino"))]
+fn openvino_session(builder: SessionBuilder, _cache_dir: &Path) -> Result<SessionBuilder, String> {
+    eprintln!("breadmill: OpenVINO backend requested but not compiled in (rebuild with --features openvino); using CPU");
+    Ok(builder)
+}
+
 /// Locate the VitisAI EP config file required by the AMD Ryzen AI SDK.
 ///
 /// Search order:

breadmill/src/main.rs

@@ -39,12 +39,14 @@ fn main() {
     let use_npu = raw_args.iter().any(|a| a == "--npu");
     let use_rocm = raw_args.iter().any(|a| a == "--rocm");
     let use_cuda = raw_args.iter().any(|a| a == "--cuda");
+    let use_openvino = raw_args.iter().any(|a| a == "--openvino");
 
     // Build a view of argv without backend flags for command matching.
+    let backend_flags = ["--npu", "--rocm", "--cuda", "--openvino"];
     let args: Vec<&str> = raw_args
         .iter()
         .skip(1)
-        .filter(|a| a.as_str() != "--npu" && a.as_str() != "--rocm" && a.as_str() != "--cuda")
+        .filter(|a| !backend_flags.contains(&a.as_str()))
         .map(|s| s.as_str())
         .collect();
 
@@ -59,7 +61,7 @@ fn main() {
             }
         }
         Some("--reindex") | Some("reindex") => {
-            if let Err(e) = run_daemon(true, use_npu, use_rocm, use_cuda) {
+            if let Err(e) = run_daemon(true, use_npu, use_rocm, use_cuda, use_openvino) {
                 eprintln!("breadmill: {}", e);
                 std::process::exit(1);
             }
@@ -76,7 +78,7 @@ fn main() {
             cli_status();
         }
         None | Some("serve") | Some("--serve") => {
-            if let Err(e) = run_daemon(false, use_npu, use_rocm, use_cuda) {
+            if let Err(e) = run_daemon(false, use_npu, use_rocm, use_cuda, use_openvino) {
                 eprintln!("breadmill: {}", e);
                 std::process::exit(1);
             }
@@ -84,7 +86,7 @@ fn main() {
         Some(cmd) => {
             eprintln!("breadmill: unknown command: {}", cmd);
             eprintln!(
-                "usage: breadmill [serve|reindex|fetch-model|query <text>|status] [--npu|--rocm|--cuda] [--version]"
+                "usage: breadmill [serve|reindex|fetch-model|query <text>|status] [--npu|--rocm|--cuda|--openvino] [--version]"
             );
             std::process::exit(1);
         }
@@ -93,7 +95,13 @@ fn main() {
 
 // ---- Daemon -----------------------------------------------------------------
 
-fn run_daemon(force_reindex: bool, use_npu: bool, use_rocm: bool, use_cuda: bool) -> Result<(), String> {
+fn run_daemon(
+    force_reindex: bool,
+    use_npu: bool,
+    use_rocm: bool,
+    use_cuda: bool,
+    use_openvino: bool,
+) -> Result<(), String> {
     let config = breadsearch_shared::Config::load();
     let state_dir = breadsearch_shared::state_dir();
     let cache_dir = breadsearch_shared::cache_dir();
@@ -115,6 +123,8 @@ fn run_daemon(force_reindex: bool, use_npu: bool, use_rocm: bool, use_cuda: bool
         "rocm"
     } else if use_cuda {
         "cuda"
+    } else if use_openvino {
+        "openvino"
     } else {
         config.model.backend.as_str()
     };
@@ -132,6 +142,10 @@ fn run_daemon(force_reindex: bool, use_npu: bool, use_rocm: bool, use_cuda: bool
             eprintln!("breadmill: CUDA backend selected");
             Backend::Cuda
         }
+        "openvino" => {
+            eprintln!("breadmill: OpenVINO backend selected");
+            Backend::OpenVino { cache_dir: cache_dir.clone() }
+        }
         _ => Backend::Cpu,
     };
 

breadsearch-shared/src/lib.rs

@@ -129,7 +129,8 @@ pub struct ModelConfig {
     pub name: String,
     #[serde(default = "default_dim")]
     pub dim: usize,
-    /// Compute backend: "cpu", "npu" (VitisAI/XDNA), "rocm" (MIGraphX/AMD GPU), or "cuda" (NVIDIA GPU).
+    /// Compute backend: "cpu", "npu" (VitisAI/XDNA), "rocm" (MIGraphX/AMD GPU),
+    /// "cuda" (NVIDIA GPU), or "openvino" (Intel iGPU/dGPU).
     #[serde(default = "default_backend")]
     pub backend: String,
 }

packaging/breadmill.service

@@ -8,7 +8,8 @@ Type=simple
 # Required (not just a perf tweak) when backend = "rocm": without a valid
 # cache dir, ONNX Runtime's MIGraphX EP reads an uninitialized cache path and
 # crashes on the first query instead of just recompiling every restart. Inert
-# for cpu/npu/cuda backends. See README.md#gpu-backend-notes.
+# for cpu/npu/cuda/openvino backends (openvino's cache dir is set in code,
+# not via env var). See README.md#gpu-backend-notes.
 Environment=ORT_MIGRAPHX_MODEL_CACHE_PATH=%h/.cache/breadsearch/migraphx-cache
 ExecStart=%h/.cargo/bin/breadmill
 Restart=on-failure