Skip to content

armgit / ente

Public Imported from GitHub
0

πŸ’š End-to-end encrypted cloud for everything.

Only do model hash check on download

laurenspriemc143ce675d0b
diff --git a/desktop/src/main/services/ml-worker.ts b/desktop/src/main/services/ml-worker.ts
index 5def938fe5..e718c91361 100644
--- a/desktop/src/main/services/ml-worker.ts
+++ b/desktop/src/main/services/ml-worker.ts
@@ -151,53 +151,68 @@ const logMLRuntimeEvents = (native: MLNative) => {
/**
* The ML models (and the CLIP tokenizer vocabulary) used by the Rust
- * pipeline, identified by their name in the models.ente.com bucket and pinned
- * by their SHA-256. These are the same assets that the mobile apps use (see
- * ml_model_assets.dart).
+ * pipeline, identified by their name in the models.ente.com bucket. These are
+ * the same assets that the mobile apps use (see ml_model_assets.dart).
+ *
+ * Fresh downloads are verified against the pinned SHA-256; copies already on
+ * disk are revalidated with only a cheap size check (a corrupt model that
+ * nevertheless retains the expected size gets caught by the Rust side, which
+ * reports it back to us for eviction and redownload).
*/
const modelSpecs = {
faceDetection: {
name: "yolov5s_face_640_640_static_b1.onnx",
+ byteSize: 32355091,
sha256: "e047647409403d52696035ecd445792173e50d7fbdcccac97b958a585db9aa3d",
},
faceEmbedding: {
name: "mobilefacenet_portable_static_b1.onnx",
+ byteSize: 5278803,
sha256: "0763fc33f54e138476194da95987e133b3e976075a6b1d3e1b2caedb251b1a36",
},
clipImage: {
name: "mobileclip_s2_image_gelu_opset20.onnx",
+ byteSize: 143057352,
sha256: "205a430af825e501c5138e5bb9abea942482a7a4fd4a680e98e47cf0830dce7e",
},
clipText: {
name: "mobileclip_s2_text_opset18_quant.onnx",
+ byteSize: 67144712,
sha256: "d92f33dfcff83077fc2e0d3414250710efbb51795dfd89767bdbefb5fdc47322",
},
clipTextVocab: {
name: "bpe_simple_vocab_16e6.txt",
+ byteSize: 3194984,
sha256: "67603cfda2e032ad77b5f8808af37789d590db664b26df8705d2bf8b3c553fc8",
},
petFaceDetection: {
name: "yolov5s_pet_face_fp16_V2.onnx",
+ byteSize: 14758020,
sha256: "7876d97992eeb5f3a9f3b35eff5e0e133012928172a8b005093108d8c3ad2d1c",
},
petFaceEmbeddingDog: {
name: "dog_face_embedding128.onnx",
+ byteSize: 4141071,
sha256: "fb04d781eb1f7adf6ce3432dc0c5873f16cc051b5c98c14c754afb39e2b92462",
},
petFaceEmbeddingCat: {
name: "cat_face_embedding128.onnx",
+ byteSize: 4141071,
sha256: "32b10694a27f6404d2beaddbd64f07ad555f72dccb12ee60a7afe5dcf6aad6cd",
},
petBodyDetection: {
name: "yolov5s_object_fp16.onnx",
+ byteSize: 14987107,
sha256: "113f0c18632eb2c4f6deebcd40eb01c676492e9b43923c2d336e1b4012fce9ef",
},
petBodyEmbeddingDog: {
name: "dog_body_embedding192.onnx",
+ byteSize: 4569361,
sha256: "1d85aa20358137e30f11c2d0baa9a2248b9997928d501fe15365d1fc57522770",
},
petBodyEmbeddingCat: {
name: "cat_body_embedding192.onnx",
+ byteSize: 4569361,
sha256: "62fb5891e61be69a96510d8ec56e7525a9541b0283e54574d27c86c9b4a26ddf",
},
} as const;
@@ -210,7 +225,7 @@ const modelSavePath = (modelName: string) =>
/**
* Cached promises to the on-disk paths of verified models, so that each model
- * is downloaded and hash-verified at most once per process. Multiple parallel
+ * is downloaded and verified at most once per process. Multiple parallel
* calls for the same model await the same promise.
*/
const _verifiedModelPaths = new Map<string, Promise<string>>();
@@ -245,7 +260,9 @@ const evictModel = async (modelPath: string) => {
/**
* Download the model described by {@link spec} if we don't already have it.
*
- * Also verify the SHA-256 of the on-disk model, redownloading it on mismatch.
+ * An existing on-disk copy is revalidated with only a cheap size check so
+ * that resolving already downloaded models stays fast; fresh downloads are
+ * verified against the pinned SHA-256.
*
* @returns the path to the model on the local machine.
*/
@@ -254,27 +271,45 @@ const modelPathDownloadingIfNeeded = async (spec: ModelSpec) => {
await cleanupOldModelsIfNeeded();
- if (!existsSync(savePath)) {
- log.info(`ML model ${spec.name} not found, downloading`);
- await downloadModel(savePath, spec.name);
- }
-
- let sha256 = await computeFileSHA256(savePath);
- if (sha256 != spec.sha256) {
+ if (existsSync(savePath)) {
+ const size = (await fs.stat(savePath)).size;
+ if (size == spec.byteSize) return savePath;
log.error(
- `The SHA-256 ${sha256} of model ${spec.name} does not match the expected value, downloading again`,
+ `The size ${size} of model ${spec.name} does not match the expected size, downloading again`,
);
- await downloadModel(savePath, spec.name);
- sha256 = await computeFileSHA256(savePath);
- if (sha256 != spec.sha256)
- throw new Error(
- `The SHA-256 of the downloaded model ${spec.name} does not match the expected value`,
- );
+ } else {
+ log.info(`ML model ${spec.name} not found, downloading`);
}
+ await downloadModelVerifyingSHA256(savePath, spec);
return savePath;
};
+/**
+ * Download the model described by {@link spec} to {@link savePath}, verifying
+ * the SHA-256 of the downloaded contents and retrying once on mismatch.
+ */
+const downloadModelVerifyingSHA256 = async (
+ savePath: string,
+ spec: ModelSpec,
+) => {
+ await downloadModel(savePath, spec.name);
+ if ((await computeFileSHA256(savePath)) == spec.sha256) return;
+
+ log.error(
+ `The SHA-256 of the downloaded model ${spec.name} does not match the expected value, downloading again`,
+ );
+ await downloadModel(savePath, spec.name);
+ if ((await computeFileSHA256(savePath)) == spec.sha256) return;
+
+ // Don't leave the bad download behind: were it to happen to pass the size
+ // check, a later call would use it without noticing.
+ await fs.rm(savePath, { force: true });
+ throw new Error(
+ `The SHA-256 of the downloaded model ${spec.name} does not match the expected value`,
+ );
+};
+
const computeFileSHA256 = (filePath: string) =>
new Promise<string>((resolve, reject) => {
const hash = createHash("sha256");