@bendyline/gezel-inference 1.0.8 → 1.0.9

package/dist/engines/native-manifest.d.ts

@@ -40,12 +40,12 @@
  * but logs that it could not anchor to a baked-in digest.
  */
 /** Native release version this build pins. Placeholder until first public release. */
-export declare const NATIVE_ENGINE_RELEASE = "0.1.57";
+export declare const NATIVE_ENGINE_RELEASE = "0.1.62";
 /** sha256 of the pinned release's `SHA256SUMS` asset. All-zeros = unpinned. */
-export declare const SHA256SUMS_DIGEST = "977b0ad4cdb8983ed44d7a3223267038b1f35b2806d8a1807d7d15c5c7e94f47";
-/** Exact SHA256 values for every engine archive published by native-v0.1.57. */
+export declare const SHA256SUMS_DIGEST = "9c2e726c1cef4152b5ffb8cee8e9970cf24127786d3fafcce3582dcf0df9badb";
+/** Exact SHA256 values for every engine archive published by native-v0.1.62. */
 export declare const NATIVE_ENGINE_ARCHIVE_SHA256: Readonly<Record<string, string>>;
-/** Download size in bytes of every engine archive published by native-v0.1.57. */
+/** Download size in bytes of every engine archive published by native-v0.1.62. */
 export declare const NATIVE_ENGINE_ARCHIVE_BYTES: Readonly<Record<string, number>>;
 /**
  * True only when the pinned release's standalone macOS archives were

package/dist/engines/native-manifest.js

@@ -40,47 +40,51 @@
  * but logs that it could not anchor to a baked-in digest.
  */
 /** Native release version this build pins. Placeholder until first public release. */
-export const NATIVE_ENGINE_RELEASE = '0.1.57';
+export const NATIVE_ENGINE_RELEASE = '0.1.62';
 /** sha256 of the pinned release's `SHA256SUMS` asset. All-zeros = unpinned. */
-export const SHA256SUMS_DIGEST = '977b0ad4cdb8983ed44d7a3223267038b1f35b2806d8a1807d7d15c5c7e94f47';
+export const SHA256SUMS_DIGEST = '9c2e726c1cef4152b5ffb8cee8e9970cf24127786d3fafcce3582dcf0df9badb';
 // BEGIN PINNED NATIVE ARCHIVE HASHES
-/** Exact SHA256 values for every engine archive published by native-v0.1.57. */
+/** Exact SHA256 values for every engine archive published by native-v0.1.62. */
 export const NATIVE_ENGINE_ARCHIVE_SHA256 = Object.freeze({
-    'gezel-native-0.1.57-darwin-arm64-metal.tar.gz': '6b97be648b032980cf73663db4424140f4cc04577972dc283b956af677ce884a',
-    'gezel-native-0.1.57-darwin-arm64.tar.gz': 'e8ca070c4d3961e95946ed32fe7376ddbc97e407dd3f5b78e26cab20c8fb286a',
-    'gezel-native-0.1.57-linux-arm64-cpu.tar.gz': 'cc87822123602a0024c63a0d65789232a665140cb2fe37b106d5bd680d12f795',
-    'gezel-native-0.1.57-linux-arm64-cuda.tar.gz': '315a441e5698800945f7030e9da961e20f3cf379a4b75c98e1d6ba88c8fee1c5',
-    'gezel-native-0.1.57-linux-arm64.tar.gz': '0684b77d23f008f8a2aa0d9d8931437bacd8569fb6c10d3ddab61a5fb25b6a06',
-    'gezel-native-0.1.57-linux-x64-cpu.tar.gz': 'b2100842f51c393b92667535798e49932173d52626f5ae72c40005b3408f1201',
-    'gezel-native-0.1.57-linux-x64-cuda.tar.gz': '2fca7720614687e8491be17366060c604551257199d395f795b27a87216bcde9',
-    'gezel-native-0.1.57-linux-x64-vulkan.tar.gz': 'c1014a1fe1d241ad1ee9abfa5badafcce208cbff8d38b4b4deb4779f31a95bef',
-    'gezel-native-0.1.57-linux-x64.tar.gz': 'a7f3dbe1941c3d03ed121c32be1bdd199aea47080b03d89989590a0f5d13c5ff',
-    'gezel-native-0.1.57-win32-arm64-cpu.zip': '12d66681cec1872f23b015c7d0bd45a25ed262342da6b0ecea8bea303adf2105',
-    'gezel-native-0.1.57-win32-arm64-cuda.zip': 'b437a9f0894673179f91a1af75d5d7d058dafa1b4f48a57eab123596f1c21ca4',
-    'gezel-native-0.1.57-win32-arm64.zip': '21cddb8b733f9a2669d64b1ada18002f60bb2779f2d5e2aacfb88214db00da7e',
-    'gezel-native-0.1.57-win32-x64-cpu.zip': 'a81f7c6754f1954a0ed5de1efc47d6194dc60d912138a9d2b8e4f992168b12c6',
-    'gezel-native-0.1.57-win32-x64-cuda.zip': '9e3f21713fe77c23536dd4124244ac3834f90c82d6e470cb6836a73bb4f7c592',
-    'gezel-native-0.1.57-win32-x64-vulkan.zip': '0d473300317b4f6cb072bdd7875ce1536a85fe50e781c435e1aef1c4cbdb08e5',
-    'gezel-native-0.1.57-win32-x64.zip': '89a073abcfc97e50b9a0c39f56862b51ac21139b7ca4a3e7a75105dca56292a2',
+    'gezel-native-0.1.62-darwin-arm64-metal.tar.gz': '0ac76122fdc7bd33e6774fb5f62dea6ff22220c483c78026667e33ff8014936d',
+    'gezel-native-0.1.62-darwin-arm64.tar.gz': '23607c3e5c3e6ac050f0594bcf4ca997cff91c57a4b65af985511e1d87843764',
+    'gezel-native-0.1.62-linux-arm64-cpu.tar.gz': 'e9a69cdb07f5ad0103da8d99715e4a1bd257219c55457716ff93e34ee4a544ab',
+    'gezel-native-0.1.62-linux-arm64-cuda.tar.gz': '6389e97d85a2d4b53b4a1c6735669e5cd6022f393fdccabe4c18c9957f9073fc',
+    'gezel-native-0.1.62-linux-arm64.tar.gz': 'cd263de007b14be6de5b828985d3066490a339f097ef199086cd51692495b9f0',
+    'gezel-native-0.1.62-linux-x64-cpu.tar.gz': '2f1b382a02082767187340f8ecd037dfa72953004977e5716ca95127429e4f0b',
+    'gezel-native-0.1.62-linux-x64-cuda.tar.gz': '8f28018fdfd5a38e48dc886a92eab392f92193d3a58c322c85dcd16c6f80fd13',
+    'gezel-native-0.1.62-linux-x64-rocm.tar.gz': '37b88b81bdcbe06fa3c5ebe3b07f0d0112b555449cc2d29775d5ad8373150f18',
+    'gezel-native-0.1.62-linux-x64-vulkan.tar.gz': 'c39fff97dceeacb377a1948ebe9f9553a98485cf21cda9492fbfbe18dd279193',
+    'gezel-native-0.1.62-linux-x64.tar.gz': '2d0c927f98379c0cd1dd05d68f2cc4b87178ede077a65e948bf36e124d560993',
+    'gezel-native-0.1.62-win32-arm64-cpu.zip': 'f867d50591df2da2c2d78cbf844a5b87679da9f5abc9a29ed6b06905c5423bed',
+    'gezel-native-0.1.62-win32-arm64-cuda.zip': 'd42c79845f64ee1fb5e97ea45dde44c6b4705ad6bc4cd6ce9360caffc27cf67f',
+    'gezel-native-0.1.62-win32-arm64.zip': 'd4e660181cd387051d4f8728d6ae008004969786f68da15c76e44704db247af0',
+    'gezel-native-0.1.62-win32-x64-cpu.zip': '85a350c01454e77d49e2a7e463089025d71b6f7cd25daa30b5b0fe8498cf3f2a',
+    'gezel-native-0.1.62-win32-x64-cuda.zip': 'edddeeb820f4612453c036bccd4882c507811464520eb772dfa5bf7b4aa2658f',
+    'gezel-native-0.1.62-win32-x64-rocm.zip': '8640280b49fbc4bc6152b734fcb2dd02b0b9c3375797685470da52406a46aa34',
+    'gezel-native-0.1.62-win32-x64-vulkan.zip': 'a57071775f8876800d6e5689d38a3f7362c4b585e8137596ee19e45ee38b6991',
+    'gezel-native-0.1.62-win32-x64.zip': '357e6c79a4d02968a2ba67299e78efed8dc2ab2d4b40a1e8f866073506b6b2b6',
 });
-/** Download size in bytes of every engine archive published by native-v0.1.57. */
+/** Download size in bytes of every engine archive published by native-v0.1.62. */
 export const NATIVE_ENGINE_ARCHIVE_BYTES = Object.freeze({
-    'gezel-native-0.1.57-darwin-arm64-metal.tar.gz': 7387963,
-    'gezel-native-0.1.57-darwin-arm64.tar.gz': 42223786,
-    'gezel-native-0.1.57-linux-arm64-cpu.tar.gz': 9223249,
-    'gezel-native-0.1.57-linux-arm64-cuda.tar.gz': 621626283,
-    'gezel-native-0.1.57-linux-arm64.tar.gz': 42040003,
-    'gezel-native-0.1.57-linux-x64-cpu.tar.gz': 13398337,
-    'gezel-native-0.1.57-linux-x64-cuda.tar.gz': 1000325316,
-    'gezel-native-0.1.57-linux-x64-vulkan.tar.gz': 26290440,
-    'gezel-native-0.1.57-linux-x64.tar.gz': 57190295,
-    'gezel-native-0.1.57-win32-arm64-cpu.zip': 6049234,
-    'gezel-native-0.1.57-win32-arm64-cuda.zip': 243595554,
-    'gezel-native-0.1.57-win32-arm64.zip': 42236484,
-    'gezel-native-0.1.57-win32-x64-cpu.zip': 9131460,
-    'gezel-native-0.1.57-win32-x64-cuda.zip': 756439338,
-    'gezel-native-0.1.57-win32-x64-vulkan.zip': 21673895,
-    'gezel-native-0.1.57-win32-x64.zip': 55668381,
+    'gezel-native-0.1.62-darwin-arm64-metal.tar.gz': 7577614,
+    'gezel-native-0.1.62-darwin-arm64.tar.gz': 42917428,
+    'gezel-native-0.1.62-linux-arm64-cpu.tar.gz': 9518882,
+    'gezel-native-0.1.62-linux-arm64-cuda.tar.gz': 624413294,
+    'gezel-native-0.1.62-linux-arm64.tar.gz': 43076642,
+    'gezel-native-0.1.62-linux-x64-cpu.tar.gz': 13762290,
+    'gezel-native-0.1.62-linux-x64-cuda.tar.gz': 1016218245,
+    'gezel-native-0.1.62-linux-x64-rocm.tar.gz': 313875007,
+    'gezel-native-0.1.62-linux-x64-vulkan.tar.gz': 26687129,
+    'gezel-native-0.1.62-linux-x64.tar.gz': 58389984,
+    'gezel-native-0.1.62-win32-arm64-cpu.zip': 6219263,
+    'gezel-native-0.1.62-win32-arm64-cuda.zip': 245764366,
+    'gezel-native-0.1.62-win32-arm64.zip': 42822090,
+    'gezel-native-0.1.62-win32-x64-cpu.zip': 9357113,
+    'gezel-native-0.1.62-win32-x64-cuda.zip': 761458453,
+    'gezel-native-0.1.62-win32-x64-rocm.zip': 241630451,
+    'gezel-native-0.1.62-win32-x64-vulkan.zip': 21898409,
+    'gezel-native-0.1.62-win32-x64.zip': 56194044,
 });
 // END PINNED NATIVE ARCHIVE HASHES
 /**

package/dist/engines/resolver.d.ts

@@ -24,7 +24,7 @@ export type EngineResolveEvent = NativeEngineResolveEvent | EngineExtractingEven
 export interface ResolveEngineOptions {
     engine: NativeBinaryName;
     home: string;
-    /** GPU backend for llama-server (`cuda`/`vulkan`/`metal`/`cpu`); omit for single-variant engines. */
+    /** GPU backend for llama-server (`cuda`/`rocm`/`vulkan`/`metal`/`cpu`); omit for single-variant engines. */
     variant?: string;
     /** Native release version; defaults to the source pin / `GEZEL_NATIVE_ENGINE_VERSION`. */
     version?: string;
@@ -100,22 +100,23 @@ export declare function resolveEngine(opts: ResolveEngineOptions): AsyncGenerato
  */
 export declare function resolveEngineToCompletion(opts: ResolveEngineOptions, onEvent?: (e: EngineResolveEvent) => void): Promise<EngineResolveResult>;
 /**
- * The archive variant an engine lives in when the caller names none.
+ * The llama-server build an on-demand install fetches when the probe found
+ * `detected`: the first build down discovery's own fallback chain (ROCm and
+ * CUDA → Vulkan → CPU) that the pinned release publishes for this platform.
+ * Without this, a probe answer the release has no archive for — ROCm before
+ * the release that first carries it, Vulkan on Linux ARM — failed the install
+ * outright instead of taking the next build down.
  *
- * These are facts about *where the archive is*, not backend choices:
- *
- *  - **ds4-server on Linux** ships only in the `-cuda` archive, sharing one
- *    copy of the NVIDIA redistributables with llama-server's CUDA build
- *    instead of shipping them twice (see scripts/native-payload.mjs).
- *  - **llama-server on macOS** ships only in the `-metal` archive; there is no
- *    `darwin-arm64-cpu` asset, and the bare `darwin-arm64` key holds
- *    ds4/sd/whisper/uv but no llama-server. Omitting the variant there fails
- *    the extract with "none of [gezel-llama-server, llama-server] were inside
- *    it" — a loud error, but an avoidable one.
- *
- * Deriving both here is the point: every call site would otherwise have to
- * remember, and one eventually won't. Where the variant is a genuine choice
- * (llama-server on Linux/Windows: cuda / vulkan / cpu) this returns undefined
- * and the decision stays with the caller.
+ * ROCm also has to be worth its several hundred megabytes before anything
+ * downloads: `rocmWorthDownloading` (rocm-devices.ts), which says no on
+ * Windows. An explicit `llamaCppBackendOverride` never comes through here.
  */
+export declare function llamaDownloadVariant(detected: string | undefined, opts?: {
+    platform?: NodeJS.Platform;
+    platformKey?: string | null;
+    version?: string;
+    /** Archive names the release publishes; null when unpinned (every build counts). */
+    archives?: Readonly<Record<string, string>> | null;
+    rocmWorth?: () => boolean;
+}): string | undefined;
 export declare function impliedEngineVariant(engine: NativeBinaryName, platform: NodeJS.Platform): string | undefined;

package/dist/engines/resolver.js

@@ -3,7 +3,7 @@ import { createReadStream, existsSync, readdirSync } from 'node:fs';
 import { chmod, mkdir, readFile, rename, rm, writeFile } from 'node:fs/promises';
 import { dirname, join } from 'node:path';
 import { HttpStatusError, createLogger, nowIso, resolveDistributionProfile, retryTransient, } from '@bendyline/gezel';
-import { DUCKDB_APPLE_TEAM_ID, DUCKDB_ARCHIVE_SHA256, DUCKDB_ASSET, DUCKDB_BINARY_SHA256, DUCKDB_VERSION, DUCKDB_WINDOWS_PUBLISHER, MXC_EXECUTOR_NAME, MXC_EXECUTOR_SHA256, MXC_LICENSE_SHA256, MXC_TARBALL_INTEGRITY, MXC_TARBALL_LICENSE_ENTRY, MXC_TARBALL_SHA256, MXC_TARBALL_URL, MXC_VERIFIED_SENTINEL, MXC_VERSION, MXC_WINDOWS_PUBLISHER, duckdbAssetUrl, duckdbBinaryName, duckdbInstallDir, duckdbPlatformKey, mxcArch, mxcInstallDir, mxcTarballEntry, resolvePlatformKey, } from '@bendyline/gezel/native';
+import { DUCKDB_APPLE_TEAM_ID, DUCKDB_ARCHIVE_SHA256, DUCKDB_ASSET, DUCKDB_BINARY_SHA256, DUCKDB_VERSION, DUCKDB_WINDOWS_PUBLISHER, MXC_EXECUTOR_NAME, MXC_EXECUTOR_SHA256, MXC_LICENSE_SHA256, MXC_TARBALL_INTEGRITY, MXC_TARBALL_LICENSE_ENTRY, MXC_TARBALL_SHA256, MXC_TARBALL_URL, MXC_VERIFIED_SENTINEL, MXC_VERSION, MXC_WINDOWS_PUBLISHER, duckdbAssetUrl, duckdbBinaryName, duckdbInstallDir, duckdbPlatformKey, mxcArch, mxcInstallDir, mxcTarballEntry, resolveAvailableLlamaBinary, resolvePlatformKey, rocmWorthDownloading, } from '@bendyline/gezel/native';
 import * as tar from 'tar';
 import { downloadWithRetry, friendlyErrors, } from '../utils/download-with-retry.js';
 import { extractZip } from '../utils/extract-zip.js';
@@ -744,6 +744,41 @@ function stripVer(v) {
  * (llama-server on Linux/Windows: cuda / vulkan / cpu) this returns undefined
  * and the decision stays with the caller.
  */
+const LLAMA_DOWNLOAD_BACKENDS = ['cuda', 'rocm', 'vulkan', 'metal', 'cpu'];
+/**
+ * The llama-server build an on-demand install fetches when the probe found
+ * `detected`: the first build down discovery's own fallback chain (ROCm and
+ * CUDA → Vulkan → CPU) that the pinned release publishes for this platform.
+ * Without this, a probe answer the release has no archive for — ROCm before
+ * the release that first carries it, Vulkan on Linux ARM — failed the install
+ * outright instead of taking the next build down.
+ *
+ * ROCm also has to be worth its several hundred megabytes before anything
+ * downloads: `rocmWorthDownloading` (rocm-devices.ts), which says no on
+ * Windows. An explicit `llamaCppBackendOverride` never comes through here.
+ */
+export function llamaDownloadVariant(detected, opts = {}) {
+    const platform = opts.platform ?? process.platform;
+    if (platform === 'darwin')
+        return impliedEngineVariant('llama-server', platform);
+    const preferred = LLAMA_DOWNLOAD_BACKENDS.find((backend) => backend === detected);
+    if (!preferred)
+        return detected;
+    const platformKey = opts.platformKey === undefined ? resolvePlatformKey() : opts.platformKey;
+    const version = stripVer(opts.version ?? effectiveEngineRelease());
+    const ext = platform === 'win32' ? 'zip' : 'tar.gz';
+    const archives = opts.archives === undefined
+        ? isPlaceholderDigest(SHA256SUMS_DIGEST) || process.env.GEZEL_NATIVE_ENGINE_VERSION
+            ? null
+            : NATIVE_ENGINE_ARCHIVE_SHA256
+        : opts.archives;
+    const rocmWorth = opts.rocmWorth ?? (() => rocmWorthDownloading({ platform }));
+    const resolved = resolveAvailableLlamaBinary(preferred, (backend) => {
+        const name = `gezel-native-${version}-${platformKey}-${backend}.${ext}`;
+        return !archives || archives[name] ? name : null;
+    }, true, (backend) => backend !== 'rocm' || rocmWorth());
+    return resolved?.backend ?? preferred;
+}
 export function impliedEngineVariant(engine, platform) {
     if (engine === 'ds4-server' && platform === 'linux')
         return 'cuda';

package/dist/http/model-inventory-cache.d.ts

@@ -4,5 +4,5 @@ export interface CacheEntry {
     at: number;
     models: ModelInfo[];
 }
-export declare const cache: Map<"copilot" | "openai" | "anthropic" | "anthropic-cli" | "codex-cli" | "ollama" | "llama-cpp" | "mlx" | "ds4" | "apple-foundation-models" | "android-mlkit" | "remote", CacheEntry>;
+export declare const cache: Map<"copilot" | "openai" | "anthropic" | "anthropic-cli" | "codex-cli" | "ollama" | "llama-cpp" | "mlx" | "ds4" | "remote" | "apple-foundation-models" | "android-mlkit", CacheEntry>;
 export declare function invalidateModelsCache(provider?: ProviderName): void;

package/dist/http/routes/engines.js

@@ -40,7 +40,7 @@ export function enginesRoutes(ctx) {
      */
     app.get('/binaries/status', (c) => {
         const backendRaw = process.env.GEZEL_LLAMA_SERVER_BACKEND ?? process.env.GEZEL_LLAMA_DETECTED_BACKEND;
-        const llamaBackend = ['cuda', 'vulkan', 'metal', 'cpu'].find((value) => value === backendRaw);
+        const llamaBackend = ['cuda', 'rocm', 'vulkan', 'metal', 'cpu'].find((value) => value === backendRaw);
         const platformKey = resolvePlatformKey();
         return c.json({
             release: effectiveEngineRelease(),
@@ -64,7 +64,7 @@ export function enginesRoutes(ctx) {
     });
     /**
      * Start (or attach to) a download+verify of `engine` (optionally a GPU
-     * `?variant=cuda|vulkan|metal|cpu`) and stream progress as SSE. The
+     * `?variant=cuda|rocm|vulkan|metal|cpu`) and stream progress as SSE. The
      * manual trigger behind Settings → On-device → "Download engine". The
      * lazy on-device-chat path uses the same registry. Idempotent — a
      * second call while one is in flight attaches to the running job.
@@ -231,7 +231,7 @@ export function enginesRoutes(ctx) {
     app.delete('/llama-cpp/quarantine/:backend', async (c) => {
         const backend = c.req.param('backend');
         if (!isLlamaBackend(backend)) {
-            return c.json({ error: 'backend must be "cuda", "vulkan", "metal", or "cpu"' }, 400);
+            return c.json({ error: 'backend must be "cuda", "rocm", "vulkan", "metal", or "cpu"' }, 400);
         }
         const override = (await ctx.store.readConfig()).llamaCppBackendOverride;
         const { restored } = liftLlamaQuarantine(ctx.home, backend, {

package/dist/models/catalog-drift.d.ts

@@ -36,6 +36,8 @@ export interface EvaluateCatalogDriftOptions {
     installedVersion?: string;
     installed: InstalledPayloadRecord;
     manifest: ChatModelManifest;
+    /** The backend a llama.cpp copy runs on. Defaults to the process's own. */
+    llamaCppBackend?: string | undefined;
     /**
      * False for models resolved out of a read-only overlay (the machine asset
      * store) and while an install is in flight for this id — the verdict still

package/dist/models/catalog-drift.js

@@ -30,7 +30,7 @@ export async function evaluateCatalogDrift(opts) {
     if (!opts.installedVersion || !currentVersion || opts.installedVersion === currentVersion) {
         return { updateAvailable: false };
     }
-    const catalog = describeCatalogPayload(opts.manifest, opts.engine);
+    const catalog = describeCatalogPayload(opts.manifest, opts.engine, opts.llamaCppBackend);
     if (!catalog) {
         // The catalog no longer describes this engine's payload at all (a source
         // block was pulled). Nothing to download, so nothing to offer.

package/dist/models/catalog-payload-identity.d.ts

@@ -30,7 +30,7 @@
  * over what the install already recorded, and its answers steer UI, never
  * trust.
  */
-import type { ChatModelManifest, GezmodelEngine } from '@bendyline/gezel';
+import { type ChatModelManifest, type GezmodelEngine } from '@bendyline/gezel';
 /** One file the catalog expects to land in the model directory. */
 export interface CatalogPayloadFile {
     /** Name as it lands on disk ('/'-joined for nested files, as in `fileSha256`). */
@@ -103,5 +103,5 @@ export declare function catalogPayloadFingerprint(payload: CatalogPayload): stri
  * The single place that knows each source block's shape, so callers compare
  * like for like without re-deriving filenames.
  */
-export declare function describeCatalogPayload(manifest: ChatModelManifest, engine: GezmodelEngine): CatalogPayload | null;
+export declare function describeCatalogPayload(manifest: ChatModelManifest, engine: GezmodelEngine, llamaCppBackend?: string | undefined): CatalogPayload | null;
 export declare function comparePayloadIdentity(opts: ComparePayloadOptions): Promise<PayloadIdentityResult>;

package/dist/models/catalog-payload-identity.js

@@ -33,6 +33,8 @@
 import { createHash } from 'node:crypto';
 import { stat } from 'node:fs/promises';
 import { basename, join } from 'node:path';
+import { llamaCppDraftModelApplies, } from '@bendyline/gezel';
+import { currentLlamaCppBackend } from '../providers/llama-cpp/backend-health.js';
 /**
  * Stable hash of a catalog payload description. Deliberately excludes
  * `revision`: it is a fetch coordinate, and identical hashes mean identical
@@ -53,7 +55,7 @@ export function catalogPayloadFingerprint(payload) {
  * The single place that knows each source block's shape, so callers compare
  * like for like without re-deriving filenames.
  */
-export function describeCatalogPayload(manifest, engine) {
+export function describeCatalogPayload(manifest, engine, llamaCppBackend = currentLlamaCppBackend()) {
     if (engine === 'mlx') {
         const src = manifest.mlx;
         if (!src)
@@ -96,10 +98,13 @@ export function describeCatalogPayload(manifest, engine) {
             sizeBytes: src.approxSizeBytes,
         });
     }
-    // Draft companions exist on both engine sources and are fetched
-    // unconditionally by `buildDownloadPlan`, so read them off the already
-    // engine-selected `src` rather than re-narrowing to one engine.
-    if (src.draftModel) {
+    // Draft companions exist on both engine sources, so read them off the
+    // already engine-selected `src`. A llama.cpp drafter limited to some backends
+    // is no part of this install elsewhere: counting it would offer every other
+    // machine an update for a file it never downloads.
+    const draftApplies = engine !== 'llama-cpp' ||
+        llamaCppDraftModelApplies(manifest.llamaCpp?.draftModel, llamaCppBackend);
+    if (src.draftModel && draftApplies) {
         files.push({
             name: basename(src.draftModel.filename),
             sha256: src.draftModel.sha256,

package/dist/providers/active-install-message.js

@@ -37,7 +37,7 @@ export function notReadyError(message, code) {
  */
 export function noModelYetMessage(engineLabel, active) {
     if (!active) {
-        return `${engineLabel}: no model is installed yet. Download one from Settings → Artificial Intelligence, then try again.`;
+        return `${engineLabel}: no model is installed yet. Download one from Settings → Artificial intelligence, then try again.`;
     }
     if (active.phase !== 'downloading') {
         const verb = active.phase === 'verifying' ? 'being verified' : 'being prepared';

package/dist/providers/catalog-model-config.d.ts

@@ -59,5 +59,33 @@ export declare function resolveCatalogLlamaCppEngineConfig(catalog: CatalogServi
         mtp?: boolean | undefined;
         draftModelId?: string | undefined;
         nMax?: number | undefined;
+        sampling?: "greedy" | "probabilistic" | undefined;
+        byBackend?: {
+            cuda?: {
+                type?: "none" | "draft-mtp" | "draft-eagle3" | "draft-dflash" | "draft-simple" | "ngram-mod" | "ngram-simple" | "ngram-map-k" | "ngram-map-k4v" | "ngram-cache" | undefined;
+                nMax?: number | undefined;
+                sampling?: "greedy" | "probabilistic" | undefined;
+            } | undefined;
+            rocm?: {
+                type?: "none" | "draft-mtp" | "draft-eagle3" | "draft-dflash" | "draft-simple" | "ngram-mod" | "ngram-simple" | "ngram-map-k" | "ngram-map-k4v" | "ngram-cache" | undefined;
+                nMax?: number | undefined;
+                sampling?: "greedy" | "probabilistic" | undefined;
+            } | undefined;
+            vulkan?: {
+                type?: "none" | "draft-mtp" | "draft-eagle3" | "draft-dflash" | "draft-simple" | "ngram-mod" | "ngram-simple" | "ngram-map-k" | "ngram-map-k4v" | "ngram-cache" | undefined;
+                nMax?: number | undefined;
+                sampling?: "greedy" | "probabilistic" | undefined;
+            } | undefined;
+            metal?: {
+                type?: "none" | "draft-mtp" | "draft-eagle3" | "draft-dflash" | "draft-simple" | "ngram-mod" | "ngram-simple" | "ngram-map-k" | "ngram-map-k4v" | "ngram-cache" | undefined;
+                nMax?: number | undefined;
+                sampling?: "greedy" | "probabilistic" | undefined;
+            } | undefined;
+            cpu?: {
+                type?: "none" | "draft-mtp" | "draft-eagle3" | "draft-dflash" | "draft-simple" | "ngram-mod" | "ngram-simple" | "ngram-map-k" | "ngram-map-k4v" | "ngram-cache" | undefined;
+                nMax?: number | undefined;
+                sampling?: "greedy" | "probabilistic" | undefined;
+            } | undefined;
+        } | undefined;
     } | undefined;
 } | undefined>;

package/dist/providers/ds4/build-provider.js

@@ -1,5 +1,5 @@
 import { basename, dirname, join } from 'node:path';
-import { DEFAULT_LOCAL_ENGINE_IDLE_TIMEOUT_MS, createLogger, } from '@bendyline/gezel';
+import { DEFAULT_LOCAL_ENGINE_IDLE_TIMEOUT_MS, createLogger, ds4SsdStreamingSupported, } from '@bendyline/gezel';
 import { gezelPaths } from '@bendyline/gezel/paths';
 import { LlamaCppProvider } from '../llama-cpp/index.js';
 import { minViableLocalContextTokens, pressureIdleGraceMs } from '../native/capacity-broker.js';
@@ -282,7 +282,7 @@ export async function buildDs4Provider(opts) {
     // build). `shouldUseDs4SsdStreaming` still refuses a model that cannot fit,
     // which is what keeps the flip safe: overriding that on a 90.89 GiB build
     // OOM-killed the engine mid-load on a 121.63 GiB host.
-    const { ds4VisionResidentBytes, planDs4ExpertCache, shouldUseDs4SsdStreaming } = await import('./residency.js');
+    const { ds4VisionResidentBytes, planDs4ExpertCache, planDs4MeasuredLaunch, shouldUseDs4SsdStreaming, } = await import('./residency.js');
     let visionEncoderSizeBytes = installedModel?.visionEncoderSizeBytes;
     if (vision.enabled && vision.visionEncoderPath && !visionEncoderSizeBytes) {
         const { stat: statVisionEncoder } = await import('node:fs/promises');
@@ -291,11 +291,10 @@ export async function buildDs4Provider(opts) {
             .catch(() => undefined);
     }
     const visionResidentBytes = ds4VisionResidentBytes(vision.enabled ? visionEncoderSizeBytes : undefined);
+    const streamingSupported = ds4SsdStreamingSupported(ds4Source, process.platform);
     const ssdStreaming = shouldUseDs4SsdStreaming({
         configured: config.ds4SsdStreaming,
-        ...(ds4Source?.ssdStreamingSupported !== undefined
-            ? { ssdStreamingSupported: ds4Source.ssdStreamingSupported }
-            : {}),
+        ...(streamingSupported !== undefined ? { ssdStreamingSupported: streamingSupported } : {}),
         modelSizeBytes: residentWeightBytes,
         totalRamBytes,
         ...(visionResidentBytes > 0 ? { companionBytes: visionResidentBytes } : {}),
@@ -304,15 +303,30 @@ export async function buildDs4Provider(opts) {
         log.warn(`[ds4] ignored unsafe full-residency override for ${effectiveModelId ?? modelPath}; ` +
             `model=${modelSizeBytes ?? 'unknown'} bytes, system=${Math.round(totalRamGb)} GiB`);
     }
-    if (config.ds4SsdStreaming === true && ds4Source?.ssdStreamingSupported === false) {
-        log.warn(`[ds4] ignored SSD-streaming override for ${effectiveModelId ?? modelPath}; this model keeps its lookup tables on disk but does not implement routed-expert streaming`);
+    if (config.ds4SsdStreaming === true && streamingSupported === false) {
+        log.warn(`[ds4] ignored SSD-streaming override for ${effectiveModelId ?? modelPath}; ds4 cannot stream this model's routed experts on this backend`);
     }
-    const cachePlanWith = (extraResidentBytes) => planDs4ExpertCache({
-        configuredGb: config.ds4CacheExpertsGb,
-        catalogCacheBytes: ds4Source?.cacheExpertsBytes,
-        catalogResidentBytes: (ds4Source?.residentBytes ?? 0) + visionResidentBytes + extraResidentBytes,
+    // A measured footprint prices the launch for its mode; a measured streamed
+    // footprint also sizes the expert cache against the local-model budget.
+    const modelBudgetBytes = ssdStreaming && ds4Source?.streamingMemory
+        ? await (await import('../native/device-capacity.js')).deviceModelBudgetBytes(config)
+        : undefined;
+    const measuredLaunchWith = (extraResidentBytes) => planDs4MeasuredLaunch({
+        source: ds4Source,
+        numCtx,
+        ssdStreaming,
         totalRamBytes: totalRamGb * 1024 ** 3,
+        budgetBytes: modelBudgetBytes,
+        sidecarBytes: visionResidentBytes + extraResidentBytes,
+        configuredCacheGb: config.ds4CacheExpertsGb,
     });
+    const cachePlanWith = (extraResidentBytes) => measuredLaunchWith(extraResidentBytes)?.cache ??
+        planDs4ExpertCache({
+            configuredGb: config.ds4CacheExpertsGb,
+            catalogCacheBytes: ds4Source?.cacheExpertsBytes,
+            catalogResidentBytes: (ds4Source?.residentBytes ?? 0) + visionResidentBytes + extraResidentBytes,
+            totalRamBytes: totalRamGb * 1024 ** 3,
+        });
     // DSpark speculative decoding. Resolved here, beside the residency decision
     // it depends on: ds4 aborts at startup if an external `--mtp-model` is
     // combined with `--ssd-streaming` for any family but DeepSeek V4.1.
@@ -386,7 +400,9 @@ export async function buildDs4Provider(opts) {
         throw err;
     }
     if (cachePlan.clamped) {
-        log.warn(`[ds4] clamped expert cache ${cachePlan.requestedGb} → ${cachePlan.cacheGb} GiB to preserve system headroom`);
+        log.warn(`[ds4] clamped expert cache ${cachePlan.requestedGb} → ${cachePlan.cacheGb} GiB to ${cachePlan.limit === 'model-budget'
+            ? "fit this device's local-model budget"
+            : 'preserve system headroom'}`);
     }
     const cacheExpertsGb = cachePlan.cacheGb;
     const kvDir = join(home, 'engines', 'ds4', 'kv');
@@ -433,6 +449,11 @@ export async function buildDs4Provider(opts) {
                 ? 'interactive'
                 : 'background',
             requirement: async () => {
+                // A measured launch is reserved at exactly its working set: no
+                // full-residency floor, and the expert cache this launch was given.
+                const measured = measuredLaunchWith(dspark.enabled ? dsparkCompanionBytes : 0);
+                if (measured)
+                    return { bytes: measured.workingSetBytes, exclusive: !ssdStreaming };
                 const { ds4AdmissionGpuBytes, ds4BaseResidentBytes, ds4ResidentLine, ds4ProjectedResidentBytes, ds4ResidentBytesForMode, } = await import('./residency.js');
                 const line = ds4ResidentLine(ds4Source ?? {});
                 const projected = line ? ds4ProjectedResidentBytes(line, numCtx) : ds4Source?.residentBytes;
@@ -444,7 +465,7 @@ export async function buildDs4Provider(opts) {
                         modelSizeBytes: residentWeightBytes,
                     });
                 const workingSetBytes = base + visionResidentBytes + (dspark.enabled ? dsparkCompanionBytes : 0);
-                const bytes = ds4ResidentBytesForMode(workingSetBytes, ssdStreaming, ds4Source?.ssdStreamingSupported);
+                const bytes = ds4ResidentBytesForMode(workingSetBytes, ssdStreaming, streamingSupported);
                 const gpuBytes = ds4AdmissionGpuBytes({
                     workingSetBytes,
                     reservationBytes: bytes,
@@ -528,6 +549,7 @@ export async function buildDs4Provider(opts) {
                 `backend=${backendFlag.slice(2)} ctx=${numCtx} ` +
                 `prefillChunk=${ds4Source?.prefillChunk ?? 'engine-default'} ` +
                 `ssdStreaming=${ssdStreaming} cacheExpertsGiB=${ssdStreaming ? cacheExpertsGb : 0} ` +
+                `memory=${measuredLaunchWith(0) ? 'measured' : 'catalog'} ` +
                 `vision=${vision.enabled} mtp=${mtp.enabled} exactSampling=${mtp.exactSampling} (${mtp.reason}) ` +
                 `dspark=${dspark.enabled} (${dspark.reason})`);
             return { command: binary, args, baseUrl: `http://127.0.0.1:${port}`, cwd: ds4BundleDir };

package/dist/providers/ds4/residency.d.ts

@@ -4,6 +4,10 @@ export { DS4_FULL_RESIDENCY_HEADROOM_BYTES };
  * Conservative floor for the full-residency working set. Exclusivity is an
  * explicit device lease at process startup, not a consequence of this number:
  * the automatic model budget can exceed 96 GiB on larger machines.
+ *
+ * Only for a launch nobody measured: an entry with `fullResidencyMemory` is
+ * reserved at that working set ({@link planDs4MeasuredLaunch}). Flooring a
+ * 45 GiB Qwen3.8 launch here put it over a 96 GiB Mac's 80 GiB budget.
  */
 export declare const DS4_FULL_RESIDENCY_RESERVATION_BYTES: number;
 /**
@@ -73,12 +77,27 @@ export interface Ds4ExpertCachePlanOptions {
     catalogCacheBytes?: number;
     catalogResidentBytes?: number;
     totalRamBytes?: number;
+    /**
+     * A measured streamed footprint at the launch window, sidecars included
+     * ({@link ds4StreamingFixedBytes}). With {@link budgetBytes} it replaces the
+     * catalog subtraction and the generic 24 GiB headroom.
+     */
+    measuredFixedBytes?: number;
+    /** This device's local-model budget, the ceiling the capacity broker admits under. */
+    budgetBytes?: number;
+    /** Smallest cache the entry says is worth streaming with. */
+    minCacheBytes?: number;
 }
 export interface Ds4ExpertCachePlan {
     cacheGb: number;
     requestedGb: number;
     clamped: boolean;
     safe: boolean;
+    /**
+     * What bounded the cache: the device's local-model budget beside a measured
+     * footprint, or the generic {@link DS4_FULL_RESIDENCY_HEADROOM_BYTES}.
+     */
+    limit: 'model-budget' | 'headroom';
 }
 /**
  * Pick a bounded routed-expert cache for streaming mode. Catalog guidance is
@@ -86,8 +105,66 @@ export interface Ds4ExpertCachePlan {
  * the old RAM-tier heuristic which incorrectly gave both models 64 GiB on a
  * 128 GiB Mac. Manual values are still accepted, but clamped so fixed model
  * state + the cache cannot consume the runtime/OS reserve.
+ *
+ * The 24 GiB headroom guards a footprint guessed as `residentBytes −
+ * cacheExpertsBytes`, which omits graph and staging memory ds4 also holds.
+ * An entry with a measured streamed footprint has no such gap, so its cache
+ * may use whatever this device's local-model budget leaves beside it. A fixed
+ * 24 GiB refused Qwen3.8 on 32 GiB Macs whose budget held it with room to spare.
  */
 export declare function planDs4ExpertCache(opts: Ds4ExpertCachePlanOptions): Ds4ExpertCachePlan;
+/** A measured ds4 memory line from the catalog: bytes at a window, plus a per-token slope. */
+export interface Ds4MeasuredLine {
+    bytes: number;
+    ctxTokens: number;
+    bytesPerToken: number;
+}
+/** Evaluate a measured line at a window, never below its context-free part. */
+export declare function ds4MeasuredBytesAt(line: Ds4MeasuredLine, ctxTokens: number): number;
+/** The memory fields of a catalog `ds4` block that price a launch. */
+export interface Ds4MemorySource {
+    cacheExpertsBytes?: number | undefined;
+    fullResidencyMemory?: {
+        workingSetBytes: number;
+        ctxTokens: number;
+        bytesPerToken: number;
+    } | undefined;
+    streamingMemory?: {
+        fixedBytes: number;
+        ctxTokens: number;
+        bytesPerToken: number;
+        minCacheBytes?: number;
+    } | undefined;
+}
+/** Measured streamed footprint without the expert cache, at a window; undefined when unmeasured. */
+export declare function ds4StreamingFixedBytes(source: Ds4MemorySource | undefined, ctxTokens: number): number | undefined;
+export interface Ds4MeasuredLaunchOptions {
+    source: Ds4MemorySource | undefined;
+    numCtx: number;
+    ssdStreaming: boolean;
+    totalRamBytes: number;
+    /** Local-model budget, needed to size a streamed launch's cache. */
+    budgetBytes?: number | undefined;
+    /** Resident memory beside the model: a vision encoder, a DSpark companion. */
+    sidecarBytes?: number;
+    configuredCacheGb?: number | undefined;
+}
+export interface Ds4MeasuredLaunch {
+    /** What the launch holds, sidecars included. The broker reserves exactly this. */
+    workingSetBytes: number;
+    /** The part of {@link workingSetBytes} the window does not move. */
+    contextFreeBytes: number;
+    bytesPerToken: number;
+    /** Streamed launches only. */
+    cache?: Ds4ExpertCachePlan;
+}
+/**
+ * Price a launch from the entry's measurement for the chosen mode. Undefined
+ * when that mode is unmeasured (or a streamed launch has no budget to size
+ * against), so callers keep their catalog-`residentBytes` path, including the
+ * 96 GiB full-residency floor that stands in for a missing measurement.
+ */
+export declare function planDs4MeasuredLaunch(opts: Ds4MeasuredLaunchOptions): Ds4MeasuredLaunch | undefined;
 /** Effective capacity-broker reservation for the chosen DS4 residency mode. */
 export declare function ds4ResidentBytesForMode(catalogResidentBytes: number, ssdStreaming: boolean, ssdStreamingSupported?: boolean): number;
 /** Allowance for ds4's compute and staging buffers beyond weights and KV rows. */