README: EPContext verified end-to-end; add runtime wheels (OV 2026.3 + ORT 1.28 OpenVINO EP)

.gitattributes

@@ -94,3 +94,5 @@ openvino/npu/audio_encoder/model.onnx.data filter=lfs diff=lfs merge=lfs -text
 openvino/npu/decoder/model.onnx.data filter=lfs diff=lfs merge=lfs -text
 openvino/npu/embedding/model.onnx.data filter=lfs diff=lfs merge=lfs -text
 openvino/npu/vision_encoder/model.onnx.data filter=lfs diff=lfs merge=lfs -text
+runtime-wheels/openvino-2026.3.0-1-cp312-cp312-manylinux_2_38_x86_64.whl filter=lfs diff=lfs merge=lfs -text
+runtime-wheels/onnxruntime_openvino-1.28.0-cp312-cp312-linux_x86_64.whl filter=lfs diff=lfs merge=lfs -text

README.md

@@ -15,39 +15,48 @@ tags:
 INT4 export of [google/gemma-4-E2B-it](https://huggingface.co/google/gemma-4-E2B-it),
 built with [mobius](https://github.com/onnxruntime/mobius) and quantized with
 [Olive](https://github.com/microsoft/Olive) (K-Quant Q4_K_M + `MatMulNBitsToQDQ`).
-Gemma 4 E2B is an any-to-any multimodal model, so every variant is a
-multi-component package: `decoder`, `embedding`, `vision_encoder`, `audio_encoder`.
+Multi-component: `decoder`, `embedding`, `vision_encoder`, `audio_encoder`.
 
 ## Variants
 
-| Folder | Format | How to load |
+| Folder | Format | Load with |
 |---|---|---|
-| `openvino-epcontext/` | **ONNX `EPContext`** (compiled OV graph wrapped in ONNX + `.xml`/`.bin`) | **ONNX Runtime GenAI + OpenVINO EP** — the precompiled artifact for ORT. |
-| `openvino-ir/` | **OpenVINO IR** (`.xml` + `.bin`) | OpenVINO runtime: `Core().compile_model("…/decoder/openvino_model.xml", "NPU")`. |
-| `openvino/npu/` | Plain ONNX (opset 24) | ORT GenAI + OpenVINO EP (compiled at load). |
+| `openvino-epcontext/` | ONNX `EPContext` (compiled OV graph in ONNX + `.xml`/`.bin`) | **ONNX Runtime + OpenVINO EP** ✅ verified |
+| `openvino-ir/` | OpenVINO IR (`.xml` + `.bin`) | OpenVINO runtime |
+| `openvino/npu/` | Plain ONNX (opset 24) | ORT + OpenVINO EP (compiled at load) |
 
-All three require **OpenVINO ≥ 2026.3** (the release that ships the ONNX-frontend
-translators for opset-24 `RMSNormalization` / `RotaryEmbedding`,
+All require **OpenVINO ≥ 2026.3** (ships the opset-24 `RMSNormalization` /
+`RotaryEmbedding` ONNX-frontend translators,
 [openvinotoolkit/openvino#35545](https://github.com/openvinotoolkit/openvino/pull/35545)).
-Earlier OpenVINO (≤ 2026.2) cannot convert these graphs.
 
-## `openvino-epcontext/` — precompiled OV graph for ONNX Runtime
+## `openvino-epcontext/` — verified precompiled artifact ✅
 
 Each component's `model.onnx` is a single `com.microsoft::EPContext` node
-(`source="OpenVINOExecutionProvider"`, `embed_mode=0`) that points at the
-compiled OpenVINO IR (`model.xml` / `model.bin`) sitting next to it. ONNX
-Runtime's OpenVINO EP loads the precompiled graph directly instead of
-recompiling the ONNX at every session start.
+(`source="OpenVINOExecutionProvider"`) pointing at the compiled OpenVINO IR
+(`model.xml`/`model.bin`) next to it. ONNX Runtime's OpenVINO EP loads the
+precompiled graph directly.
+
+**Verified end-to-end**: the decoder EPContext loads via the OpenVINO EP and a
+forward pass returns `logits (1, seq, 262144)` fp16 (all finite), using ONNX
+Runtime 1.28 built against OpenVINO 2026.3.
 
 ```python
-import onnxruntime_genai as og   # built with the OpenVINO EP (OpenVINO >= 2026.3)
-model = og.Model("openvino-epcontext")
-# ... standard ORT GenAI generation loop ...
+import onnxruntime as ort, numpy as np
+sess = ort.InferenceSession(
+    "openvino-epcontext/decoder/model.onnx",
+    providers=[("OpenVINOExecutionProvider", {"device_type": "NPU"})],  # or GPU / CPU
+)
 ```
 
-Produced with Olive: `OpenVINOConversion` (ONNX → OV IR) → `OpenVINOEncapsulation`
-(IR → `EPContext` ONNX), `DeviceClass=NPU`. To retarget GPU/CPU, regenerate the
-EPContext with the desired `target_device` (the IR itself is device-independent).
+### Runtime (until official releases ship)
+
+The `runtime-wheels/` folder has the two wheels used to verify these models
+(Linux, cp312) — OpenVINO 2026.3 + ONNX Runtime 1.28 with the OpenVINO EP:
+
+```bash
+pip install runtime-wheels/openvino-2026.3.0-*.whl \
+            runtime-wheels/onnxruntime_openvino-1.28.0-*.whl
+```
 
 ## Build pipeline
 
@@ -60,14 +69,6 @@ only uses ops the OpenVINO ONNX frontend supports at opset 24.
 
 Recipe: [microsoft/olive-recipes `google-gemma-4-E2B-it`](https://github.com/microsoft/olive-recipes/tree/main/google-gemma-4-E2B-it).
 
-## Notes / verification
-
-- The OV IR **converts and compiles** with OpenVINO 2026.3 (verified: all 4
-  components convert; decoder `Core().compile_model(...)` on CPU).
-- End-to-end ORT GenAI generation via the OpenVINO EP needs an
-  `onnxruntime-genai` built against OpenVINO ≥ 2026.3; verify on your target
-  once that runtime is available.
-
 ## License
 
 Inherits the [Gemma license](https://ai.google.dev/gemma/terms).

runtime-wheels/onnxruntime_openvino-1.28.0-cp312-cp312-linux_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cb6b7c2afec4e193633c154d4be78f44cbf38527780b23231b3d16d6df7bdee
+size 24341083

runtime-wheels/openvino-2026.3.0-1-cp312-cp312-manylinux_2_38_x86_64.whl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b34951a5d68e34eb3cf259b9e86336cf71715e687c2f2e7204ce0703618e8f0c
+size 35342724