|
| 1 | +/// <reference path="./cross-origin-storage.d.ts" /> |
| 2 | + |
| 3 | +const HASH_ALGORITHM = 'SHA-256'; |
| 4 | + |
| 5 | +/** |
| 6 | + * Name of the Cache API bucket used to persist the url→hash mapping. |
| 7 | + */ |
| 8 | +const HASH_CACHE_NAME = 'experimental_transformers-hash-cache'; |
| 9 | + |
| 10 | +/** |
| 11 | + * Builds the hash descriptor object expected by the cross-origin storage API. |
| 12 | + * |
| 13 | + * @param {string} value Hex-encoded SHA-256 hash. |
| 14 | + * @returns {{ algorithm: string, value: string }} |
| 15 | + */ |
| 16 | +const makeHashDescriptor = (value) => ({ algorithm: HASH_ALGORITHM, value }); |
| 17 | + |
| 18 | +/** |
| 19 | + * A cache implementation backed by the experimental `navigator.crossOriginStorage` API, |
| 20 | + * which allows sharing cached files (identified by content hash) across origins. |
| 21 | + * |
| 22 | + * Implements {@link import('../cache.js').CacheInterface}. |
| 23 | + * |
| 24 | + * @see https://github.com/explainers-by-googlers/cross-origin-storage |
| 25 | + */ |
| 26 | +export class CrossOriginStorage { |
| 27 | + /** @type {Promise<Cache> | null} */ |
| 28 | + #hashCache = null; |
| 29 | + |
| 30 | + /** |
| 31 | + * Returns (and lazily opens) the hash cache, reusing the same promise across concurrent callers. |
| 32 | + * @returns {Promise<Cache>} |
| 33 | + */ |
| 34 | + _getHashCache = () => { |
| 35 | + this.#hashCache ??= caches.open(HASH_CACHE_NAME); |
| 36 | + return this.#hashCache; |
| 37 | + }; |
| 38 | + |
| 39 | + /** |
| 40 | + * Returns whether the `navigator.crossOriginStorage` API is available in the current environment. |
| 41 | + * @returns {boolean} |
| 42 | + */ |
| 43 | + static isAvailable = () => typeof navigator !== 'undefined' && 'crossOriginStorage' in navigator; |
| 44 | + |
| 45 | + /** |
| 46 | + * Looks up a cached response for the given URL by resolving its SHA-256 hash and requesting |
| 47 | + * the corresponding file handle from cross-origin storage. |
| 48 | + * |
| 49 | + * Implements `CacheInterface.match`. |
| 50 | + * |
| 51 | + * @param {string} request The URL of the resource to look up. |
| 52 | + * @returns {Promise<Response|undefined>} The cached `Response`, or `undefined` if not found. |
| 53 | + */ |
| 54 | + match = async (request) => { |
| 55 | + const hashValue = await this._getFileHash(request); |
| 56 | + if (!hashValue) { |
| 57 | + return undefined; |
| 58 | + } |
| 59 | + try { |
| 60 | + const [handle] = await navigator.crossOriginStorage.requestFileHandles([makeHashDescriptor(hashValue)]); |
| 61 | + const blob = await handle.getFile(); |
| 62 | + return new Response(blob); |
| 63 | + } catch { |
| 64 | + return undefined; |
| 65 | + } |
| 66 | + }; |
| 67 | + |
| 68 | + /** |
| 69 | + * Stores a response in cross-origin storage, keyed by its SHA-256 hash. |
| 70 | + * |
| 71 | + * For LFS-backed URLs the hash is resolved cheaply via `_getFileHash` (which checks |
| 72 | + * `HASH_CACHE_NAME` first, then falls back to fetching the Git LFS pointer file) |
| 73 | + * without reading the response body a second time. |
| 74 | + * |
| 75 | + * For non-LFS resources the hash is unknown upfront. In that case the body is consumed |
| 76 | + * in the background: the stream is read to compute the content hash, the file is written |
| 77 | + * into cross-origin storage, and the computed hash is persisted to `HASH_CACHE_NAME` |
| 78 | + * so that future `match` calls can resolve the file without a network round-trip. |
| 79 | + * |
| 80 | + * Implements `CacheInterface.put`. |
| 81 | + * |
| 82 | + * @param {string} request The URL of the resource (used as the hash-cache key). |
| 83 | + * @param {Response} response The response whose body will be written to the cache. |
| 84 | + * @returns {Promise<void>} |
| 85 | + */ |
| 86 | + put = async (request, response) => { |
| 87 | + const hashValue = await this._getFileHash(request); |
| 88 | + |
| 89 | + if (hashValue) { |
| 90 | + // Fast path: LFS hash already known. Consume the body and store directly. |
| 91 | + const blob = await response.blob(); |
| 92 | + await this._storeBlobInCOS(blob, hashValue); |
| 93 | + } else { |
| 94 | + // Slow path: hash unknown. Process in the background so put() returns promptly. |
| 95 | + // The caller already holds a reference to the original response; we receive it |
| 96 | + // here only to buffer and hash its body. |
| 97 | + this._processAndStore(request, response.body); |
| 98 | + } |
| 99 | + }; |
| 100 | + |
| 101 | + /** |
| 102 | + * Writes a blob into cross-origin storage using the given pre-computed hex hash string. |
| 103 | + * |
| 104 | + * @param {Blob} blob |
| 105 | + * @param {string} hashHex Hex-encoded SHA-256 hash of `blob`. |
| 106 | + * @returns {Promise<void>} |
| 107 | + */ |
| 108 | + _storeBlobInCOS = async (blob, hashHex) => { |
| 109 | + const [handle] = await navigator.crossOriginStorage.requestFileHandles([makeHashDescriptor(hashHex)], { |
| 110 | + create: true, |
| 111 | + }); |
| 112 | + const writableStream = await handle.createWritable(); |
| 113 | + await writableStream.write(blob); |
| 114 | + await writableStream.close(); |
| 115 | + }; |
| 116 | + |
| 117 | + /** |
| 118 | + * Background task for non-LFS resources: consumes `stream`, computes the SHA-256 hash |
| 119 | + * of the resulting blob, stores it in cross-origin storage, and persists the computed |
| 120 | + * hash to `HASH_CACHE_NAME` keyed by `request` so future `match` calls can resolve the |
| 121 | + * file without a network round-trip. |
| 122 | + * |
| 123 | + * Called fire-and-forget from `put` — errors are swallowed so failures never surface to |
| 124 | + * the caller. |
| 125 | + * |
| 126 | + * @param {string} request The original resource URL. |
| 127 | + * @param {ReadableStream} stream The response body stream to consume. |
| 128 | + * @returns {Promise<void>} |
| 129 | + */ |
| 130 | + _processAndStore = async (request, stream) => { |
| 131 | + try { |
| 132 | + const chunks = []; |
| 133 | + for await (const chunk of stream) { |
| 134 | + chunks.push(chunk); |
| 135 | + } |
| 136 | + const blob = new Blob(chunks); |
| 137 | + const hashHex = await this._getBlobHash(blob); |
| 138 | + |
| 139 | + await this._storeBlobInCOS(blob, hashHex); |
| 140 | + |
| 141 | + // Persist the computed hash so future match() calls resolve without the network. |
| 142 | + try { |
| 143 | + const hashCache = await this._getHashCache(); |
| 144 | + await hashCache.put(request, new Response(hashHex)); |
| 145 | + } catch { |
| 146 | + // Cache API unavailable (e.g. non-secure context): COS entry still written. |
| 147 | + } |
| 148 | + } catch { |
| 149 | + // Non-fatal: background store failure must not affect the caller. |
| 150 | + } |
| 151 | + }; |
| 152 | + |
| 153 | + /** |
| 154 | + * Deletes the cache entry for the given request. |
| 155 | + * |
| 156 | + * Removes the hash entry from `HASH_CACHE_NAME`. Note: cross-origin storage itself does not |
| 157 | + * expose a delete API, so only the local hash mapping is removed. For non-LFS URLs this |
| 158 | + * permanently prevents `match` from resolving the file. For LFS-backed URLs, `match` will |
| 159 | + * re-fetch the LFS pointer file on the next call and repopulate the hash cache automatically. |
| 160 | + * |
| 161 | + * Implements `CacheInterface.delete`. |
| 162 | + * |
| 163 | + * @param {string} request |
| 164 | + * @returns {Promise<boolean>} Resolves to `true` if the hash entry was deleted, `false` otherwise. |
| 165 | + */ |
| 166 | + delete = async (request) => { |
| 167 | + try { |
| 168 | + const hashCache = await this._getHashCache(); |
| 169 | + return await hashCache.delete(request); |
| 170 | + } catch { |
| 171 | + return false; |
| 172 | + } |
| 173 | + }; |
| 174 | + |
| 175 | + /** |
| 176 | + * Resolves the SHA-256 hash for a given URL. |
| 177 | + * |
| 178 | + * Returns the cached hash immediately if one has been persisted to `HASH_CACHE_NAME`. |
| 179 | + * Otherwise falls back to `_getLfsFileHash` to retrieve the hash from the Hugging Face |
| 180 | + * LFS pointer file, persisting the result to `HASH_CACHE_NAME` for future lookups. |
| 181 | + * |
| 182 | + * Returns `null` if the hash cannot be determined (e.g. non-LFS URL with no cached entry). |
| 183 | + * |
| 184 | + * @param {string} url The resource URL to resolve a hash for. |
| 185 | + * @returns {Promise<string|null>} The hex-encoded SHA-256 hash, or `null` if unavailable. |
| 186 | + */ |
| 187 | + _getFileHash = async (url) => { |
| 188 | + try { |
| 189 | + const hashCache = await this._getHashCache(); |
| 190 | + const cached = await hashCache.match(url); |
| 191 | + if (cached) { |
| 192 | + return cached.text(); |
| 193 | + } |
| 194 | + |
| 195 | + const hash = await this._getLfsFileHash(url); |
| 196 | + if (hash) { |
| 197 | + await hashCache.put(url, new Response(hash)); |
| 198 | + return hash; |
| 199 | + } |
| 200 | + |
| 201 | + return null; |
| 202 | + } catch { |
| 203 | + return null; |
| 204 | + } |
| 205 | + }; |
| 206 | + |
| 207 | + /** |
| 208 | + * Attempts to retrieve the SHA-256 hash for a Hugging Face resource URL from its raw |
| 209 | + * Git LFS pointer file. |
| 210 | + * |
| 211 | + * Only applicable to URLs containing `/resolve/` (i.e. Hugging Face resolved file URLs). |
| 212 | + * The `/resolve/` segment is rewritten to `/raw/` to fetch the LFS pointer directly. |
| 213 | + * Returns `null` for non-LFS URLs or when the network request fails. |
| 214 | + * |
| 215 | + * @see https://huggingface.co/docs/hub/en/storage-backends#xet |
| 216 | + * @param {string} url The resolved Hugging Face URL of the resource. |
| 217 | + * @returns {Promise<string|null>} The hex-encoded SHA-256 hash, or `null` if unavailable. |
| 218 | + */ |
| 219 | + _getLfsFileHash = async (url) => { |
| 220 | + if (!url.includes('/resolve/')) { |
| 221 | + return null; |
| 222 | + } |
| 223 | + |
| 224 | + const rawUrl = url.replace('/resolve/', '/raw/'); |
| 225 | + |
| 226 | + try { |
| 227 | + const text = await fetch(rawUrl).then((r) => r.text()); |
| 228 | + const match = text.match(/^oid sha256:([0-9a-f]+)$/m); |
| 229 | + return match ? match[1] : null; |
| 230 | + } catch { |
| 231 | + return null; |
| 232 | + } |
| 233 | + }; |
| 234 | + |
| 235 | + /** |
| 236 | + * Computes the SHA-256 hash of a `Blob`'s contents. |
| 237 | + * |
| 238 | + * @param {Blob} blob The blob to hash. |
| 239 | + * @returns {Promise<string>} The lowercase hex-encoded SHA-256 hash. |
| 240 | + */ |
| 241 | + _getBlobHash = async (blob) => { |
| 242 | + const arrayBuffer = await blob.arrayBuffer(); |
| 243 | + const hashBuffer = await crypto.subtle.digest(HASH_ALGORITHM, arrayBuffer); |
| 244 | + const hashArray = Array.from(new Uint8Array(hashBuffer)); |
| 245 | + return hashArray.map((byte) => byte.toString(16).padStart(2, '0')).join(''); |
| 246 | + }; |
| 247 | +} |
0 commit comments