@elizaos/shared 2.0.0-beta.1 → 2.0.3-beta.3

package/local-inference/catalog.js

@@ -0,0 +1,534 @@
+/**
+ * Eliza-curated local model catalog.
+ *
+ * Default local inference is restricted to the active Eliza-1 line:
+ * eliza-1-2b, eliza-1-4b, eliza-1-9b, eliza-1-27b,
+ * and eliza-1-27b-256k.
+ * These ship Gemma 4 bases: E2B/E4B/12B/31B mapped onto the
+ * 2B/4B/9B/27B release tiers (the 2026-06-22 cutover from the legacy
+ * Qwen3.5/3.6 line — see #9033 and
+ * packages/training/scripts/training/model_registry.py for the active
+ * registry). Gemma 4 is a dense SWA + shared-KV + per-layer-embedding
+ * (PLE) + MQA architecture; KV is already minimal so the legacy
+ * QJL/PolarQuant KV kernels are not used (stock KV), while TurboQuant
+ * weight-quant remains active. External Hub search remains custom/opt-in and
+ * never enters first-run or default eligibility.
+ * Separate-drafter MTP is still the required release shape, but runtime
+ * metadata is gated until the Gemma drafter GGUFs are actually hosted.
+ */
+import { resolveHfDownloadBase } from "./hf-proxy.js";
+export const ELIZA_1_HF_REPO = "elizaos/eliza-1";
+export const ELIZA_1_TIER_IDS = [
+    "eliza-1-2b",
+    "eliza-1-4b",
+    "eliza-1-9b",
+    "eliza-1-27b",
+    "eliza-1-27b-256k",
+];
+export const ELIZA_1_RELEASE_TIER_IDS = ELIZA_1_TIER_IDS;
+export const ELIZA_1_VISION_TIER_IDS = [
+    "eliza-1-2b",
+    "eliza-1-4b",
+    "eliza-1-9b",
+    "eliza-1-27b",
+    "eliza-1-27b-256k",
+];
+const _ELIZA_1_VISION_TIER_ID_SET = new Set(ELIZA_1_VISION_TIER_IDS);
+export const ELIZA_1_MTP_TIER_IDS = [
+    "eliza-1-2b",
+    "eliza-1-4b",
+    "eliza-1-9b",
+    "eliza-1-27b",
+    "eliza-1-27b-256k",
+];
+/**
+ * Tiers whose Gemma MTP drafter GGUFs are present at
+ * `bundles/<tier>/mtp/drafter-<tier>.gguf` in the active HF tree.
+ *
+ * Current HF state (2026-06-25): the active bundles only expose legacy
+ * `dflash/` drafter paths, while the Gemma candidate publishes
+ * `candidates/gemma-2b-base-v1/mtp/MISSING.txt`. Keep this empty so the
+ * runtime and downloader do not advertise or fetch missing MTP artifacts.
+ */
+export const ELIZA_1_HOSTED_MTP_TIER_IDS = [];
+function hostedMtpDrafterAvailableForTier(id) {
+    return ELIZA_1_HOSTED_MTP_TIER_IDS.some((mtpId) => mtpId === id);
+}
+/**
+ * On-device (mobile-class) tiers. These are the tiers small enough to run on
+ * a phone, so they advertise the Gemma-4 QAT `Q4_0` quant as the
+ * mobile-preferred variant and ship a LiteRT `.litertlm` bundle for the
+ * on-device LiteRT-LM runtime (NPU/GPU delegate). Mirrors the Kokoro-only
+ * voice policy and the SD-1.5 image-gen tiering (2b/4b).
+ */
+export const ELIZA_1_ON_DEVICE_TIER_IDS = [
+    "eliza-1-2b",
+    "eliza-1-4b",
+];
+const _ELIZA_1_ON_DEVICE_TIER_ID_SET = new Set(ELIZA_1_ON_DEVICE_TIER_IDS);
+export function isOnDeviceTier(id) {
+    return _ELIZA_1_ON_DEVICE_TIER_ID_SET.has(id);
+}
+// The quantized 2B (Gemma 4 E2B) is the shipped first-run default chat model:
+// it is the smallest/entry tier, fits 8 GB-class phones comfortably, downloads
+// fast, and is the model bundled into the AOSP image. Larger tiers (4B/9B/27B)
+// remain available for manual selection on higher-memory hosts.
+export const FIRST_RUN_DEFAULT_MODEL_ID = "eliza-1-2b";
+export const DEFAULT_ELIGIBLE_MODEL_IDS = new Set(ELIZA_1_RELEASE_TIER_IDS);
+export function isDefaultEligibleId(id) {
+    return DEFAULT_ELIGIBLE_MODEL_IDS.has(id);
+}
+/**
+ * Per-tier publish-state hint. Keys are tier ids that are known to have
+ * a pending Hugging Face bundle at the time the catalog snapshot was
+ * cut. Tiers not listed here default to `"published"`. The recommender
+ * consults this map (or a `publishStatus` field on a synthetic
+ * `CatalogModel`) before recommending a first-run default — see
+ * `recommendForFirstRun` and elizaOS/eliza#7629.
+ *
+ * Set the override env var `ELIZA_PUBLISH_STATUS_OVERRIDES` to a JSON
+ * object like `{"eliza-1-2b":"published","eliza-1-9b":"pending"}` to
+ * override at runtime without changing the static map (useful for QA
+ * and for installs that depend on a private HF mirror).
+ *
+ * W3-12 audit (2026-05-14): the following areas require publish attention:
+ *   - 2B vision: enabled in the catalog and canonical vision tier set;
+ *     publish staging must include `vision/mmproj-2b.gguf` or manifest
+ *     validation fails loudly.
+ *   - Voice sub-models (wakeword, turn-detector, speaker-encoder, emotion):
+ *     published under the unified elizaos/eliza-1 `voice/<model-id>/...`
+ *     layout. Per-tier manifests still need to consume these paths directly
+ *     where a bundle wants eager voice downloads.
+ *   - Kokoro same voice preset: `af_same.bin` absent from all
+ *     bundles; I7 eval showed regression. Current bundles ship af_bella
+ *     and standard voices only.
+ */
+export const ELIZA_1_TIER_PUBLISH_STATUS = {};
+export function eliza1TierPublishStatus(id) {
+    const override = readPublishStatusOverride(id);
+    if (override)
+        return override;
+    const hint = ELIZA_1_TIER_PUBLISH_STATUS[id];
+    return hint ?? "published";
+}
+function readPublishStatusOverride(id) {
+    const raw = typeof process !== "undefined"
+        ? process.env.ELIZA_PUBLISH_STATUS_OVERRIDES
+        : undefined;
+    if (!raw)
+        return undefined;
+    try {
+        const parsed = JSON.parse(raw);
+        const value = parsed[id];
+        if (value === "published" || value === "pending")
+            return value;
+    }
+    catch {
+        // Malformed override JSON is non-fatal — fall back to the static
+        // publish-status hint and the catalog's own `publishStatus` field.
+    }
+    return undefined;
+}
+export const ELIZA_1_PLACEHOLDER_IDS = new Set(ELIZA_1_TIER_IDS);
+/**
+ * Per-tier voice backend policy. The FIRST entry is the default backend
+ * for that tier — `runtime-selection.ts` picks it when both backends are
+ * available and no override applies (voice cloning, TTFA target, RTF).
+ * Entries beyond the first are also bundled; tiers that ship only one
+ * backend have a single-element array.
+ *
+ * Policy:
+ *   - Mobile-class tiers (2b / 4b) → Kokoro only. Kokoro is ~82M
+ *     params (a single ~60-80 MB GGUF) and hits ~97ms CPU TTFB, so it is
+ *     both smaller and faster than OmniVoice (~400-625 MB) — the right
+ *     trade for phones. OmniVoice is not shipped in these bundles. On
+ *     mobile, `selectVoiceBackend({ mobile: true })` also forces Kokoro
+ *     regardless of any env override, so the path is Kokoro-exclusive.
+ *   - 9B → OmniVoice first with Kokoro bundled for hosts with enough memory.
+ *   - Large tiers (27b / 27b-256k) → OmniVoice only. The RAM
+ *     and compute budget is large enough that the OmniVoice quality win
+ *     dominates; Kokoro is not shipped in these bundles.
+ */
+export const ELIZA_1_VOICE_BACKENDS = {
+    "eliza-1-2b": ["kokoro"],
+    "eliza-1-4b": ["kokoro"],
+    "eliza-1-9b": ["omnivoice", "kokoro"],
+    "eliza-1-27b": ["omnivoice"],
+    "eliza-1-27b-256k": ["omnivoice"],
+};
+const BASE_REQUIRED_KERNELS = ["turbo3", "turbo4"];
+const TIER_SPECS = {
+    "eliza-1-2b": {
+        id: "eliza-1-2b",
+        params: "2B",
+        sizeGb: 1.4,
+        minRamGb: 4,
+        q4MinRamGb: 4,
+        bucket: "small",
+        contextLength: 131072,
+        textFile: "text/eliza-1-2b-128k.gguf",
+        // WS2: vision enabled — the 2B tier is the standard "small-phone"
+        // default for first-run users, so camera-to-reaction and screen
+        // analysis must work here. The mmproj is ~361 MB Q8_0 (actual:
+        // 361,518,784 bytes, published 2026-05-14); the arbiter owns the
+        // swap with the text weights under pressure.
+        hasVision: true,
+        // WS3: image-gen on the standard small-phone default uses SD 1.5
+        // Q5_0 too; tier-up to Z-Image-Turbo at 9B.
+        hasImageGen: true,
+    },
+    "eliza-1-4b": {
+        id: "eliza-1-4b",
+        params: "4B",
+        sizeGb: 2.6,
+        // 4B is the shipped mid/mobile tier. The 2.6 GB Q4_K_M weights are sized
+        // for a 128k Eliza-1 bundle on the Gemma 4 E4B base. Gemma KV is already
+        // minimal (MQA + windowed-SWA + shared-KV) so the runtime ships stock KV
+        // (f16/q8_0) — the legacy head_dim=128 QJL/Polar kernels do not apply to
+        // Gemma's dual head dims (512 global / 256 swa). The floor stays above the
+        // model size to leave headroom for the OS, app, and KV cache.
+        minRamGb: 6,
+        q4MinRamGb: 6,
+        bucket: "mid",
+        contextLength: 131072,
+        textFile: "text/eliza-1-4b-128k.gguf",
+        hasEmbedding: true,
+        hasVision: true,
+        // WS3: 4B is the last tier that defaults to SD 1.5; flagship-phone
+        // optional path can upgrade to SDXL-Turbo Q4_0.
+        hasImageGen: true,
+    },
+    "eliza-1-9b": {
+        id: "eliza-1-9b",
+        params: "9B",
+        sizeGb: 5.4,
+        minRamGb: 12,
+        q4MinRamGb: 12,
+        bucket: "large",
+        contextLength: 131072,
+        textFile: "text/eliza-1-9b-128k.gguf",
+        gpuProfile: "rtx-3090",
+        hasEmbedding: true,
+        hasVision: true,
+        // WS3: 9B is the boundary tier where Z-Image-Turbo Q4_K_M (~3.4 GB)
+        // becomes the default. FLUX.1 schnell remains opt-in for >=24 GB
+        // shared RAM / >=12 GB VRAM.
+        hasImageGen: true,
+    },
+    "eliza-1-27b": {
+        id: "eliza-1-27b",
+        params: "27B",
+        sizeGb: 16.8,
+        minRamGb: 32,
+        q4MinRamGb: 32,
+        bucket: "large",
+        contextLength: 131072,
+        textFile: "text/eliza-1-27b-128k.gguf",
+        gpuProfile: "rtx-4090",
+        hasEmbedding: true,
+        hasVision: true,
+        hasImageGen: true,
+    },
+    "eliza-1-27b-256k": {
+        id: "eliza-1-27b-256k",
+        params: "27B",
+        parameterLabel: "27B 256k",
+        sizeGb: 16.8,
+        minRamGb: 48,
+        q4MinRamGb: 48,
+        bucket: "large",
+        contextLength: 262144,
+        textFile: "text/eliza-1-27b-256k.gguf",
+        gpuProfile: "h200",
+        hasEmbedding: true,
+        hasVision: true,
+        hasImageGen: true,
+    },
+};
+function tierSlug(id) {
+    return id.slice("eliza-1-".length);
+}
+function tierDisplaySlug(id) {
+    switch (id) {
+        case "eliza-1-2b":
+            return "2B";
+        case "eliza-1-4b":
+            return "4B";
+        case "eliza-1-9b":
+            return "9B";
+        case "eliza-1-27b":
+            return "27B";
+        case "eliza-1-27b-256k":
+            return "27B-256k";
+    }
+    const exhaustive = id;
+    return exhaustive;
+}
+function tierDisplayName(id) {
+    return `eliza-1-${tierDisplaySlug(id)}`;
+}
+function bundleRemotePrefix(id) {
+    return `bundles/${tierSlug(id)}`;
+}
+function bundlePath(_id, rel) {
+    return rel;
+}
+function bundleRemotePath(id, rel) {
+    return `${bundleRemotePrefix(id)}/${rel}`;
+}
+function bundleComponent(id, file) {
+    return { repo: ELIZA_1_HF_REPO, file: bundleRemotePath(id, file) };
+}
+/**
+ * Default OmniVoice K-quant the runtime picks per tier when no
+ * device-class override applies. Mobile-class tiers (2b/4b) default
+ * to Q4_K_M (~4.5 bits/weight, the common sweet spot for llama.cpp /
+ * Ollama / LM Studio). Desktop / workstation tiers default to Q8_0 (≈8
+ * bits/weight, near-bf16 quality) because RAM headroom permits it.
+ */
+function voiceQuantForTier(id) {
+    return id === "eliza-1-2b" || id === "eliza-1-4b" ? "Q4_K_M" : "Q8_0";
+}
+/**
+ * Full K-quant ladder published per tier. The downloader inspects the
+ * device's RAM/SoC class at install time and picks the appropriate level
+ * from this list. The ladder is monotonically decreasing in bits-per-weight
+ * (smallest first): {@link OmniVoiceQuantLevel}.
+ *
+ * Every active tier publishes an OmniVoice ladder. Small tiers keep the
+ * ladder narrow so the installer can stay inside mobile RAM budgets while
+ * still defaulting to the fused OmniVoice path. 9B and 27B-class tiers ship
+ * the full Q3..Q8 ladder so a `--memory-budget okay` host can step down to
+ * Q3_K_M and a `--memory-budget good` host can take Q6_K.
+ */
+const OMNIVOICE_QUANT_LADDER_BY_TIER = {
+    "eliza-1-2b": ["Q3_K_M", "Q4_K_M", "Q5_K_M"],
+    "eliza-1-4b": ["Q3_K_M", "Q4_K_M", "Q5_K_M"],
+    "eliza-1-9b": ["Q3_K_M", "Q4_K_M", "Q5_K_M", "Q6_K", "Q8_0"],
+    "eliza-1-27b": ["Q3_K_M", "Q4_K_M", "Q5_K_M", "Q6_K", "Q8_0"],
+    "eliza-1-27b-256k": ["Q3_K_M", "Q4_K_M", "Q5_K_M", "Q6_K", "Q8_0"],
+};
+export function voiceQuantLadderForTier(id) {
+    return OMNIVOICE_QUANT_LADDER_BY_TIER[id];
+}
+export function defaultVoiceQuantForTier(id) {
+    return voiceQuantForTier(id);
+}
+function primaryVoiceFileForTier(id) {
+    const defaultBackend = ELIZA_1_VOICE_BACKENDS[id][0];
+    if (defaultBackend === "omnivoice") {
+        return `tts/omnivoice-base-${voiceQuantForTier(id)}.gguf`;
+    }
+    return "tts/kokoro/kokoro-82m-v1_0-Q4_K_M.gguf";
+}
+function sourceModelForTier(id) {
+    const spec = TIER_SPECS[id];
+    const components = {
+        text: bundleComponent(id, spec.textFile),
+        voice: bundleComponent(id, primaryVoiceFileForTier(id)),
+        vad: bundleComponent(id, "vad/silero-vad-v5.gguf"),
+    };
+    // Gemma ASR artifacts are not hosted yet. Do not advertise retired pre-Gemma
+    // ASR files as active tier components. Runtime ASR remains gated by bundle
+    // manifest provenance until Gemma ASR files are published.
+    // LiteRT-LM single-file bundle for the on-device runtime: text + vision +
+    // audio + MTP packed into one QAT (.litertlm) artifact, parallel to the
+    // GGUF `text` component. Only the mobile-class tiers ship it.
+    if (isOnDeviceTier(id)) {
+        components.litert = bundleComponent(id, `text/eliza-1-${tierSlug(id)}.litertlm`);
+    }
+    if (spec.hasEmbedding) {
+        components.embedding = bundleComponent(id, "embedding/eliza-1-embedding.gguf");
+    }
+    if (spec.hasVision) {
+        components.vision = bundleComponent(id, `vision/mmproj-${tierSlug(id)}.gguf`);
+    }
+    // Separate-drafter MTP remains the Gemma release target, but the active HF
+    // tree does not host `mtp/drafter-<tier>.gguf` yet. The available `dflash/`
+    // files are legacy artifacts, so only advertise this component when the
+    // hosted Gemma drafter list is explicitly populated.
+    if (hostedMtpDrafterAvailableForTier(id)) {
+        components.mtp = bundleComponent(id, `mtp/drafter-${tierSlug(id)}.gguf`);
+    }
+    return { finetuned: false, components };
+}
+function runtimeForTier(id, contextLength) {
+    const requiresKernel = contextLength >= 65536
+        ? [...BASE_REQUIRED_KERNELS, "turbo3_tcq"]
+        : BASE_REQUIRED_KERNELS;
+    const runtime = {
+        preferredBackend: "llama-cpp",
+        optimizations: {
+            parallel: 4,
+            flashAttention: true,
+            requiresKernel,
+            // OpenVINO is the right backend for ASR/Whisper on Intel hosts but
+            // never for autoregressive text. The text path uses optimized
+            // llama.cpp kernels plus native MTP heads.
+            unsupportedKernels: ["openvino"],
+            ctxCheckpoints: 4,
+            ctxCheckpointInterval: 4096,
+        },
+    };
+    if (hostedMtpDrafterAvailableForTier(id)) {
+        // Separate-drafter MTP: Gemma 4 ships an official standalone drafter
+        // GGUF, loaded via `-md mtp/drafter-<tier>.gguf --spec-type draft-mtp`.
+        //
+        // Draft window = 1 (single speculative token). The bionic/desktop FFI
+        // MTP engine uses a FIXED window equal to `draftMax` (no adaptive
+        // acceptance schedule; `eliza-inference-ffi.cpp` sets
+        // `sp.draft.n_max = draft_max`), so the catalog value is the live window.
+        // The gemma4-assistant NextN head reliably predicts exactly one token;
+        // its multi-token acceptance collapses past the first, so a larger window
+        // burns draft forwards that get rejected and regresses decode. Measured on
+        // Apple M-series Metal against the eliza-1-2b (Q8) target, greedy, across
+        // 3 prompts:
+        //   draftMax=1 => 1.37-1.66x win | =2 => ~0.90x | =4 => 0.61x | =6 => 0.37x
+        // draftMax=1 is the measured-optimal, never-regress window for this drafter
+        // on every tier; widening it is a per-tier/per-device tuning question that
+        // needs on-hardware measurement before it can beat 1.
+        runtime.mtp = {
+            specType: "draft-mtp",
+            drafterFile: `mtp/drafter-${tierSlug(id)}.gguf`,
+            draftMin: 1,
+            draftMax: 1,
+            gpuLayers: "auto",
+        };
+    }
+    return runtime;
+}
+const QUANT_SUFFIX = {
+    q3_k_m: "q3_k_m",
+    // Google's official Gemma-4 QAT quant. `Q4_0` is the GGUF block format
+    // their QAT checkpoints export to — distinct from the post-training
+    // `q4_k_m` we ship as the desktop default.
+    q4_0: "Q4_0",
+    q4_k_m: "q4_k_m",
+    q5_k_m: "q5_k_m",
+    q6_k: "q6_k",
+    q8_0: "q8_0",
+    // LiteRT-LM mobile bundle suffix. The artifact is a `.litertlm`, not a
+    // `.gguf`; `textLiteRtComponent` overrides the filename, so this suffix
+    // only feeds the non-LiteRT variant-id naming path defensively.
+    wna8o8: "wna8o8",
+};
+function textQuantizationMatrix(args) {
+    const fileBase = args.primaryGgufFile.replace(/\.gguf$/, "");
+    const litertFile = `${fileBase.replace(/-128k$|-256k$/, "")}.litertlm`;
+    const mk = (id, label, scale, minRamScale, status, extra) => ({
+        id,
+        label,
+        ggufFile: extra?.artifactFormat === "litertlm"
+            ? litertFile
+            : id === "q4_k_m"
+                ? args.primaryGgufFile
+                : `${fileBase}-${QUANT_SUFFIX[id]}.gguf`,
+        sizeGb: Number((args.q4SizeGb * scale).toFixed(1)),
+        minRamGb: Math.ceil(args.q4MinRamGb * minRamScale),
+        status,
+        ...extra,
+    });
+    const variants = [
+        mk("q3_k_m", "3-bit", 0.76, 0.85, "planned"),
+        // Gemma-4 QAT Q4_0: same ~4-bit footprint as q4_k_m but the official
+        // quantization-aware-trained checkpoint. Keep this planned until the
+        // tier-specific `*-Q4_0.gguf` artifacts are present in the hosted bundle.
+        mk("q4_0", "4-bit", 0.94, 0.95, "planned", args.onDevice ? { mobilePreferred: true } : undefined),
+        mk("q4_k_m", "4-bit", 1, 1, "published"),
+        mk("q5_k_m", "5-bit", 1.22, 1.18, "planned"),
+        mk("q6_k", "6-bit", 1.45, 1.35, "planned"),
+        mk("q8_0", "8-bit", 1.95, 1.8, "planned"),
+    ];
+    // On-device tiers also advertise the LiteRT-LM `.litertlm` bundle: the
+    // wNa8o8 (4-bit weight / 8-bit activation) mobile schema run by the
+    // LiteRT-LM runtime (NPU/GPU delegate), not llama.cpp. It carries the same
+    // ~4-bit footprint as the QAT Q4_0 GGUF.
+    if (args.onDevice) {
+        variants.push(mk("wna8o8", "4-bit", 0.94, 0.95, "planned", {
+            artifactFormat: "litertlm",
+        }));
+    }
+    return { defaultVariantId: "q4_k_m", variants };
+}
+function blurbForTier(id) {
+    const displayName = tierDisplayName(id);
+    switch (id) {
+        case "eliza-1-2b":
+            return `${displayName} - smallest/entry local tier for low-memory phones and CPU fallback.`;
+        case "eliza-1-4b":
+            return `${displayName} - balanced local tier for modern laptops and desktops.`;
+        case "eliza-1-9b":
+            return `${displayName} - workstation local tier for stronger reasoning.`;
+        case "eliza-1-27b":
+            return `${displayName} - high-quality local tier for GPU workstations.`;
+        case "eliza-1-27b-256k":
+            return `${displayName} - long-context local tier for high-memory GPU workstations.`;
+    }
+    const exhaustive = id;
+    return exhaustive;
+}
+function chatTier(id) {
+    const spec = TIER_SPECS[id];
+    return {
+        id,
+        displayName: tierDisplayName(id),
+        hfRepo: ELIZA_1_HF_REPO,
+        hfPathPrefix: bundleRemotePrefix(id),
+        ggufFile: bundlePath(id, spec.textFile),
+        bundleManifestFile: bundlePath(id, "eliza-1.manifest.json"),
+        params: spec.params,
+        parameterLabel: spec.parameterLabel,
+        quant: "Eliza-1 optimized local runtime",
+        sizeGb: spec.sizeGb,
+        minRamGb: spec.minRamGb,
+        category: "chat",
+        bucket: spec.bucket,
+        contextLength: spec.contextLength,
+        tokenizerFamily: "gemma4",
+        runtimeClass: "fused-eliza1",
+        sourceModel: sourceModelForTier(id),
+        voiceBackends: ELIZA_1_VOICE_BACKENDS[id],
+        runtime: runtimeForTier(id, spec.contextLength),
+        gpuProfile: spec.gpuProfile,
+        quantization: textQuantizationMatrix({
+            primaryGgufFile: bundlePath(id, spec.textFile),
+            q4SizeGb: spec.sizeGb,
+            q4MinRamGb: spec.q4MinRamGb,
+            onDevice: isOnDeviceTier(id),
+        }),
+        blurb: blurbForTier(id),
+        publishStatus: eliza1TierPublishStatus(id),
+    };
+}
+export const MODEL_CATALOG = ELIZA_1_TIER_IDS.map((id) => chatTier(id));
+export function findCatalogModel(id) {
+    return MODEL_CATALOG.find((m) => m.id === id);
+}
+export function buildHuggingFaceResolveUrlForPath(model, filePath) {
+    const cleanFilePath = filePath.replace(/^\/+/, "");
+    const cleanPrefix = model.hfPathPrefix?.replace(/^\/+|\/+$/g, "");
+    const pathWithPrefix = cleanPrefix &&
+        cleanFilePath !== cleanPrefix &&
+        !cleanFilePath.startsWith(`${cleanPrefix}/`)
+        ? `${cleanPrefix}/${cleanFilePath}`
+        : cleanFilePath;
+    if (model.hub === "modelscope") {
+        const base = process.env.ELIZA_MODELSCOPE_BASE_URL?.trim().replace(/\/+$/, "") ||
+            "https://www.modelscope.cn";
+        const encodedPath = pathWithPrefix
+            .split("/")
+            .map((segment) => encodeURIComponent(segment))
+            .join("/");
+        return `${base}/models/${model.hfRepo}/resolve/master/${encodedPath}`;
+    }
+    const { base } = resolveHfDownloadBase();
+    const encodedPath = pathWithPrefix
+        .split("/")
+        .map((segment) => encodeURIComponent(segment))
+        .join("/");
+    return `${base}/${model.hfRepo}/resolve/main/${encodedPath}?download=true`;
+}
+export function buildHuggingFaceResolveUrl(model) {
+    return buildHuggingFaceResolveUrlForPath(model, model.ggufFile);
+}
+//# sourceMappingURL=catalog.js.map

package/local-inference/catalog.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"catalog.js","sourceRoot":"","sources":["../../src/local-inference/catalog.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;GAiBG;AAEH,OAAO,EAAE,qBAAqB,EAAE,MAAM,eAAe,CAAC;AAQtD,MAAM,CAAC,MAAM,eAAe,GAAG,iBAA0B,CAAC;AAE1D,MAAM,CAAC,MAAM,gBAAgB,GAAG;IAC9B,YAAY;IACZ,YAAY;IACZ,YAAY;IACZ,aAAa;IACb,kBAAkB;CACV,CAAC;AAIX,MAAM,CAAC,MAAM,wBAAwB,GACnC,gBAAsD,CAAC;AAEzD,MAAM,CAAC,MAAM,uBAAuB,GAAG;IACrC,YAAY;IACZ,YAAY;IACZ,YAAY;IACZ,aAAa;IACb,kBAAkB;CAC4B,CAAC;AAEjD,MAAM,2BAA2B,GAA8B,IAAI,GAAG,CACpE,uBAAuB,CACxB,CAAC;AAEF,MAAM,CAAC,MAAM,oBAAoB,GAAG;IAClC,YAAY;IACZ,YAAY;IACZ,YAAY;IACZ,aAAa;IACb,kBAAkB;CAC4B,CAAC;AAEjD;;;;;;;;GAQG;AACH,MAAM,CAAC,MAAM,2BAA2B,GACtC,EAAiD,CAAC;AAEpD,SAAS,gCAAgC,CAAC,EAAgB;IACxD,OAAO,2BAA2B,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,KAAK,EAAE,CAAC,CAAC;AACnE,CAAC;AAED;;;;;;GAMG;AACH,MAAM,CAAC,MAAM,0BAA0B,GAAG;IACxC,YAAY;IACZ,YAAY;CACkC,CAAC;AAEjD,MAAM,8BAA8B,GAA8B,IAAI,GAAG,CACvE,0BAA0B,CAC3B,CAAC;AAEF,MAAM,UAAU,cAAc,CAAC,EAAgB;IAC7C,OAAO,8BAA8B,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;AAChD,CAAC;AAED,8EAA8E;AAC9E,+EAA+E;AAC/E,+EAA+E;AAC/E,gEAAgE;AAChE,MAAM,CAAC,MAAM,0BAA0B,GAAiB,YAAY,CAAC;AAErE,MAAM,CAAC,MAAM,0BAA0B,GAAwB,IAAI,GAAG,CACpE,wBAAwB,CACzB,CAAC;AAEF,MAAM,UAAU,mBAAmB,CAAC,EAAU;IAC5C,OAAO,0BAA0B,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;AAC5C,CAAC;AAED;;;;;;;;;;;;;;;;;;;;;;;;GAwBG;AACH,MAAM,CAAC,MAAM,2BAA2B,GAEpC,EAAE,CAAC;AAEP,MAAM,UAAU,uBAAuB,CACrC,EAAyB;IAEzB,MAAM,QAAQ,GAAG,yBAAyB,CAAC,EAAE,CAAC,CAAC;IAC/C,IAAI,QAAQ;QAAE,OAAO,QAAQ,CAAC;IAC9B,MAAM,IAAI,GACR,2BAID,CAAC,EAAE,CAAC,CAAC;IACN,OAAO,IAAI,IAAI,WAAW,CAAC;AAC7B,CAAC;AAED,SAAS,yBAAyB,CAChC,EAAU;IAEV,MAAM,GAAG,GACP,OAAO,OAAO,KAAK,WAAW;QAC5B,CAAC,CAAC,OAAO,CAAC,GAAG,CAAC,8BAA8B;QAC5C,CAAC,CAAC,SAAS,CAAC;IAChB,IAAI,CAAC,GAAG;QAAE,OAAO,SAAS,CAAC;IAC3B,IAAI,CAAC;QACH,MAAM,MAAM,GAAG,IAAI,CAAC,KAAK,CAAC,GAAG,CAA4B,CAAC;QAC1D,MAAM,KAAK,GAAG,MAAM,CAAC,EAAE,CAAC,CAAC;QACzB,IAAI,KAAK,KAAK,WAAW,IAAI,KAAK,KAAK,SAAS;YAAE,OAAO,KAAK,CAAC;IACjE,CAAC;IAAC,MAAM,CAAC;QACP,iEAAiE;QACjE,mEAAmE;IACrE,CAAC;IACD,OAAO,SAAS,CAAC;AACnB,CAAC;AAED,MAAM,CAAC,MAAM,uBAAuB,GAAwB,IAAI,GAAG,CACjE,gBAAgB,CACjB,CAAC;AAIF;;;;;;;;;;;;;;;;;;GAkBG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAG/B;IACF,YAAY,EAAE,CAAC,QAAQ,CAAC;IACxB,YAAY,EAAE,CAAC,QAAQ,CAAC;IACxB,YAAY,EAAE,CAAC,WAAW,EAAE,QAAQ,CAAC;IACrC,aAAa,EAAE,CAAC,WAAW,CAAC;IAC5B,kBAAkB,EAAE,CAAC,WAAW,CAAC;CAClC,CAAC;AAEF,MAAM,qBAAqB,GAAyB,CAAC,QAAQ,EAAE,QAAQ,CAAC,CAAC;AA0BzE,MAAM,UAAU,GAA6C;IAC3D,YAAY,EAAE;QACZ,EAAE,EAAE,YAAY;QAChB,MAAM,EAAE,IAAI;QACZ,MAAM,EAAE,GAAG;QACX,QAAQ,EAAE,CAAC;QACX,UAAU,EAAE,CAAC;QACb,MAAM,EAAE,OAAO;QACf,aAAa,EAAE,MAAM;QACrB,QAAQ,EAAE,2BAA2B;QACrC,kEAAkE;QAClE,gEAAgE;QAChE,+DAA+D;QAC/D,iEAAiE;QACjE,6CAA6C;QAC7C,SAAS,EAAE,IAAI;QACf,iEAAiE;QACjE,4CAA4C;QAC5C,WAAW,EAAE,IAAI;KAClB;IACD,YAAY,EAAE;QACZ,EAAE,EAAE,YAAY;QAChB,MAAM,EAAE,IAAI;QACZ,MAAM,EAAE,GAAG;QACX,yEAAyE;QACzE,yEAAyE;QACzE,yEAAyE;QACzE,yEAAyE;QACzE,2EAA2E;QAC3E,8DAA8D;QAC9D,QAAQ,EAAE,CAAC;QACX,UAAU,EAAE,CAAC;QACb,MAAM,EAAE,KAAK;QACb,aAAa,EAAE,MAAM;QACrB,QAAQ,EAAE,2BAA2B;QACrC,YAAY,EAAE,IAAI;QAClB,SAAS,EAAE,IAAI;QACf,mEAAmE;QACnE,gDAAgD;QAChD,WAAW,EAAE,IAAI;KAClB;IACD,YAAY,EAAE;QACZ,EAAE,EAAE,YAAY;QAChB,MAAM,EAAE,IAAI;QACZ,MAAM,EAAE,GAAG;QACX,QAAQ,EAAE,EAAE;QACZ,UAAU,EAAE,EAAE;QACd,MAAM,EAAE,OAAO;QACf,aAAa,EAAE,MAAM;QACrB,QAAQ,EAAE,2BAA2B;QACrC,UAAU,EAAE,UAAU;QACtB,YAAY,EAAE,IAAI;QAClB,SAAS,EAAE,IAAI;QACf,oEAAoE;QACpE,iEAAiE;QACjE,6BAA6B;QAC7B,WAAW,EAAE,IAAI;KAClB;IACD,aAAa,EAAE;QACb,EAAE,EAAE,aAAa;QACjB,MAAM,EAAE,KAAK;QACb,MAAM,EAAE,IAAI;QACZ,QAAQ,EAAE,EAAE;QACZ,UAAU,EAAE,EAAE;QACd,MAAM,EAAE,OAAO;QACf,aAAa,EAAE,MAAM;QACrB,QAAQ,EAAE,4BAA4B;QACtC,UAAU,EAAE,UAAU;QACtB,YAAY,EAAE,IAAI;QAClB,SAAS,EAAE,IAAI;QACf,WAAW,EAAE,IAAI;KAClB;IACD,kBAAkB,EAAE;QAClB,EAAE,EAAE,kBAAkB;QACtB,MAAM,EAAE,KAAK;QACb,cAAc,EAAE,UAAU;QAC1B,MAAM,EAAE,IAAI;QACZ,QAAQ,EAAE,EAAE;QACZ,UAAU,EAAE,EAAE;QACd,MAAM,EAAE,OAAO;QACf,aAAa,EAAE,MAAM;QACrB,QAAQ,EAAE,4BAA4B;QACtC,UAAU,EAAE,MAAM;QAClB,YAAY,EAAE,IAAI;QAClB,SAAS,EAAE,IAAI;QACf,WAAW,EAAE,IAAI;KAClB;CACF,CAAC;AAEF,SAAS,QAAQ,CAAC,EAAgB;IAChC,OAAO,EAAE,CAAC,KAAK,CAAC,UAAU,CAAC,MAAM,CAAC,CAAC;AACrC,CAAC;AAED,SAAS,eAAe,CAAC,EAAgB;IACvC,QAAQ,EAAE,EAAE,CAAC;QACX,KAAK,YAAY;YACf,OAAO,IAAI,CAAC;QACd,KAAK,YAAY;YACf,OAAO,IAAI,CAAC;QACd,KAAK,YAAY;YACf,OAAO,IAAI,CAAC;QACd,KAAK,aAAa;YAChB,OAAO,KAAK,CAAC;QACf,KAAK,kBAAkB;YACrB,OAAO,UAAU,CAAC;IACtB,CAAC;IACD,MAAM,UAAU,GAAU,EAAE,CAAC;IAC7B,OAAO,UAAU,CAAC;AACpB,CAAC;AAED,SAAS,eAAe,CAAC,EAAgB;IACvC,OAAO,WAAW,eAAe,CAAC,EAAE,CAAC,EAAE,CAAC;AAC1C,CAAC;AAED,SAAS,kBAAkB,CAAC,EAAgB;IAC1C,OAAO,WAAW,QAAQ,CAAC,EAAE,CAAC,EAAE,CAAC;AACnC,CAAC;AAED,SAAS,UAAU,CAAC,GAAiB,EAAE,GAAW;IAChD,OAAO,GAAG,CAAC;AACb,CAAC;AAED,SAAS,gBAAgB,CAAC,EAAgB,EAAE,GAAW;IACrD,OAAO,GAAG,kBAAkB,CAAC,EAAE,CAAC,IAAI,GAAG,EAAE,CAAC;AAC5C,CAAC;AAMD,SAAS,eAAe,CACtB,EAAgB,EAChB,IAAY;IAEZ,OAAO,EAAE,IAAI,EAAE,eAAe,EAAE,IAAI,EAAE,gBAAgB,CAAC,EAAE,EAAE,IAAI,CAAC,EAAE,CAAC;AACrE,CAAC;AA2BD;;;;;;GAMG;AACH,SAAS,iBAAiB,CAAC,EAAgB;IACzC,OAAO,EAAE,KAAK,YAAY,IAAI,EAAE,KAAK,YAAY,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,CAAC,MAAM,CAAC;AACxE,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,8BAA8B,GAEhC;IACF,YAAY,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC;IAC5C,YAAY,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC;IAC5C,YAAY,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,CAAC;IAC5D,aAAa,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,CAAC;IAC7D,kBAAkB,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,MAAM,EAAE,MAAM,CAAC;CACnE,CAAC;AAEF,MAAM,UAAU,uBAAuB,CACrC,EAAgB;IAEhB,OAAO,8BAA8B,CAAC,EAAE,CAAC,CAAC;AAC5C,CAAC;AAED,MAAM,UAAU,wBAAwB,CACtC,EAAgB;IAEhB,OAAO,iBAAiB,CAAC,EAAE,CAAC,CAAC;AAC/B,CAAC;AAED,SAAS,uBAAuB,CAAC,EAAgB;IAC/C,MAAM,cAAc,GAAG,sBAAsB,CAAC,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;IACrD,IAAI,cAAc,KAAK,WAAW,EAAE,CAAC;QACnC,OAAO,sBAAsB,iBAAiB,CAAC,EAAE,CAAC,OAAO,CAAC;IAC5D,CAAC;IACD,OAAO,wCAAwC,CAAC;AAClD,CAAC;AAED,SAAS,kBAAkB,CAAC,EAAgB;IAC1C,MAAM,IAAI,GAAG,UAAU,CAAC,EAAE,CAAC,CAAC;IAC5B,MAAM,UAAU,GAAuB;QACrC,IAAI,EAAE,eAAe,CAAC,EAAE,EAAE,IAAI,CAAC,QAAQ,CAAC;QACxC,KAAK,EAAE,eAAe,CAAC,EAAE,EAAE,uBAAuB,CAAC,EAAE,CAAC,CAAC;QACvD,GAAG,EAAE,eAAe,CAAC,EAAE,EAAE,wBAAwB,CAAC;KACnD,CAAC;IAEF,6EAA6E;IAC7E,2EAA2E;IAC3E,2DAA2D;IAE3D,0EAA0E;IAC1E,wEAAwE;IACxE,8DAA8D;IAC9D,IAAI,cAAc,CAAC,EAAE,CAAC,EAAE,CAAC;QACvB,UAAU,CAAC,MAAM,GAAG,eAAe,CACjC,EAAE,EACF,gBAAgB,QAAQ,CAAC,EAAE,CAAC,WAAW,CACxC,CAAC;IACJ,CAAC;IAED,IAAI,IAAI,CAAC,YAAY,EAAE,CAAC;QACtB,UAAU,CAAC,SAAS,GAAG,eAAe,CACpC,EAAE,EACF,kCAAkC,CACnC,CAAC;IACJ,CAAC;IACD,IAAI,IAAI,CAAC,SAAS,EAAE,CAAC;QACnB,UAAU,CAAC,MAAM,GAAG,eAAe,CACjC,EAAE,EACF,iBAAiB,QAAQ,CAAC,EAAE,CAAC,OAAO,CACrC,CAAC;IACJ,CAAC;IACD,2EAA2E;IAC3E,4EAA4E;IAC5E,wEAAwE;IACxE,qDAAqD;IACrD,IAAI,gCAAgC,CAAC,EAAE,CAAC,EAAE,CAAC;QACzC,UAAU,CAAC,GAAG,GAAG,eAAe,CAAC,EAAE,EAAE,eAAe,QAAQ,CAAC,EAAE,CAAC,OAAO,CAAC,CAAC;IAC3E,CAAC;IAED,OAAO,EAAE,SAAS,EAAE,KAAK,EAAE,UAAU,EAAE,CAAC;AAC1C,CAAC;AAED,SAAS,cAAc,CACrB,EAAgB,EAChB,aAAqB;IAErB,MAAM,cAAc,GAClB,aAAa,IAAI,KAAK;QACpB,CAAC,CAAC,CAAC,GAAG,qBAAqB,EAAE,YAAY,CAAC;QAC1C,CAAC,CAAC,qBAAqB,CAAC;IAC5B,MAAM,OAAO,GAA4B;QACvC,gBAAgB,EAAE,WAAW;QAC7B,aAAa,EAAE;YACb,QAAQ,EAAE,CAAC;YACX,cAAc,EAAE,IAAI;YACpB,cAAc;YACd,mEAAmE;YACnE,8DAA8D;YAC9D,2CAA2C;YAC3C,kBAAkB,EAAE,CAAC,UAAU,CAAC;YAChC,cAAc,EAAE,CAAC;YACjB,qBAAqB,EAAE,IAAI;SAC5B;KACF,CAAC;IAEF,IAAI,gCAAgC,CAAC,EAAE,CAAC,EAAE,CAAC;QACzC,qEAAqE;QACrE,wEAAwE;QACxE,EAAE;QACF,sEAAsE;QACtE,kEAAkE;QAClE,sDAAsD;QACtD,0EAA0E;QAC1E,uEAAuE;QACvE,0EAA0E;QAC1E,2EAA2E;QAC3E,0EAA0E;QAC1E,aAAa;QACb,4EAA4E;QAC5E,4EAA4E;QAC5E,2EAA2E;QAC3E,sDAAsD;QACtD,OAAO,CAAC,GAAG,GAAG;YACZ,QAAQ,EAAE,WAAW;YACrB,WAAW,EAAE,eAAe,QAAQ,CAAC,EAAE,CAAC,OAAO;YAC/C,QAAQ,EAAE,CAAC;YACX,QAAQ,EAAE,CAAC;YACX,SAAS,EAAE,MAAM;SAClB,CAAC;IACJ,CAAC;IAED,OAAO,OAAO,CAAC;AACjB,CAAC;AAED,MAAM,YAAY,GAA0C;IAC1D,MAAM,EAAE,QAAQ;IAChB,uEAAuE;IACvE,oEAAoE;IACpE,2CAA2C;IAC3C,IAAI,EAAE,MAAM;IACZ,MAAM,EAAE,QAAQ;IAChB,MAAM,EAAE,QAAQ;IAChB,IAAI,EAAE,MAAM;IACZ,IAAI,EAAE,MAAM;IACZ,uEAAuE;IACvE,wEAAwE;IACxE,gEAAgE;IAChE,MAAM,EAAE,QAAQ;CACjB,CAAC;AAEF,SAAS,sBAAsB,CAAC,IAU/B;IACC,MAAM,QAAQ,GAAG,IAAI,CAAC,eAAe,CAAC,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;IAC7D,MAAM,UAAU,GAAG,GAAG,QAAQ,CAAC,OAAO,CAAC,eAAe,EAAE,EAAE,CAAC,WAAW,CAAC;IACvE,MAAM,EAAE,GAAG,CACT,EAAyB,EACzB,KAA0C,EAC1C,KAAa,EACb,WAAmB,EACnB,MAA4C,EAC5C,KAGC,EAC2B,EAAE,CAAC,CAAC;QAChC,EAAE;QACF,KAAK;QACL,QAAQ,EACN,KAAK,EAAE,cAAc,KAAK,UAAU;YAClC,CAAC,CAAC,UAAU;YACZ,CAAC,CAAC,EAAE,KAAK,QAAQ;gBACf,CAAC,CAAC,IAAI,CAAC,eAAe;gBACtB,CAAC,CAAC,GAAG,QAAQ,IAAI,YAAY,CAAC,EAAE,CAAC,OAAO;QAC9C,MAAM,EAAE,MAAM,CAAC,CAAC,IAAI,CAAC,QAAQ,GAAG,KAAK,CAAC,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC;QAClD,QAAQ,EAAE,IAAI,CAAC,IAAI,CAAC,IAAI,CAAC,UAAU,GAAG,WAAW,CAAC;QAClD,MAAM;QACN,GAAG,KAAK;KACT,CAAC,CAAC;IAEH,MAAM,QAAQ,GAAiC;QAC7C,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC;QAC5C,qEAAqE;QACrE,qEAAqE;QACrE,0EAA0E;QAC1E,EAAE,CACA,MAAM,EACN,OAAO,EACP,IAAI,EACJ,IAAI,EACJ,SAAS,EACT,IAAI,CAAC,QAAQ,CAAC,CAAC,CAAC,EAAE,eAAe,EAAE,IAAI,EAAE,CAAC,CAAC,CAAC,SAAS,CACtD;QACD,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC,EAAE,WAAW,CAAC;QACxC,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC;QAC5C,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC;QAC1C,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,SAAS,CAAC;KAC1C,CAAC;IAEF,uEAAuE;IACvE,oEAAoE;IACpE,2EAA2E;IAC3E,yCAAyC;IACzC,IAAI,IAAI,CAAC,QAAQ,EAAE,CAAC;QAClB,QAAQ,CAAC,IAAI,CACX,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,EAAE;YAC3C,cAAc,EAAE,UAAU;SAC3B,CAAC,CACH,CAAC;IACJ,CAAC;IAED,OAAO,EAAE,gBAAgB,EAAE,QAAQ,EAAE,QAAQ,EAAE,CAAC;AAClD,CAAC;AAED,SAAS,YAAY,CAAC,EAAgB;IACpC,MAAM,WAAW,GAAG,eAAe,CAAC,EAAE,CAAC,CAAC;IACxC,QAAQ,EAAE,EAAE,CAAC;QACX,KAAK,YAAY;YACf,OAAO,GAAG,WAAW,sEAAsE,CAAC;QAC9F,KAAK,YAAY;YACf,OAAO,GAAG,WAAW,yDAAyD,CAAC;QACjF,KAAK,YAAY;YACf,OAAO,GAAG,WAAW,mDAAmD,CAAC;QAC3E,KAAK,aAAa;YAChB,OAAO,GAAG,WAAW,kDAAkD,CAAC;QAC1E,KAAK,kBAAkB;YACrB,OAAO,GAAG,WAAW,8DAA8D,CAAC;IACxF,CAAC;IACD,MAAM,UAAU,GAAU,EAAE,CAAC;IAC7B,OAAO,UAAU,CAAC;AACpB,CAAC;AAED,SAAS,QAAQ,CAAC,EAAgB;IAChC,MAAM,IAAI,GAAG,UAAU,CAAC,EAAE,CAAC,CAAC;IAC5B,OAAO;QACL,EAAE;QACF,WAAW,EAAE,eAAe,CAAC,EAAE,CAAC;QAChC,MAAM,EAAE,eAAe;QACvB,YAAY,EAAE,kBAAkB,CAAC,EAAE,CAAC;QACpC,QAAQ,EAAE,UAAU,CAAC,EAAE,EAAE,IAAI,CAAC,QAAQ,CAAC;QACvC,kBAAkB,EAAE,UAAU,CAAC,EAAE,EAAE,uBAAuB,CAAC;QAC3D,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,cAAc,EAAE,IAAI,CAAC,cAAc;QACnC,KAAK,EAAE,iCAAiC;QACxC,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,QAAQ,EAAE,IAAI,CAAC,QAAQ;QACvB,QAAQ,EAAE,MAAM;QAChB,MAAM,EAAE,IAAI,CAAC,MAAM;QACnB,aAAa,EAAE,IAAI,CAAC,aAAa;QACjC,eAAe,EAAE,QAAQ;QACzB,YAAY,EAAE,cAAc;QAC5B,WAAW,EAAE,kBAAkB,CAAC,EAAE,CAAC;QACnC,aAAa,EAAE,sBAAsB,CAAC,EAAE,CAAC;QACzC,OAAO,EAAE,cAAc,CAAC,EAAE,EAAE,IAAI,CAAC,aAAa,CAAC;QAC/C,UAAU,EAAE,IAAI,CAAC,UAAU;QAC3B,YAAY,EAAE,sBAAsB,CAAC;YACnC,eAAe,EAAE,UAAU,CAAC,EAAE,EAAE,IAAI,CAAC,QAAQ,CAAC;YAC9C,QAAQ,EAAE,IAAI,CAAC,MAAM;YACrB,UAAU,EAAE,IAAI,CAAC,UAAU;YAC3B,QAAQ,EAAE,cAAc,CAAC,EAAE,CAAC;SAC7B,CAAC;QACF,KAAK,EAAE,YAAY,CAAC,EAAE,CAAC;QACvB,aAAa,EAAE,uBAAuB,CAAC,EAAE,CAAC;KAC3C,CAAC;AACJ,CAAC;AAED,MAAM,CAAC,MAAM,aAAa,GAAmB,gBAAgB,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,EAAE,CACvE,QAAQ,CAAC,EAAE,CAAC,CACb,CAAC;AAEF,MAAM,UAAU,gBAAgB,CAAC,EAAU;IACzC,OAAO,aAAa,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAAC;AAChD,CAAC;AAED,MAAM,UAAU,iCAAiC,CAC/C,KAAmB,EACnB,QAAgB;IAEhB,MAAM,aAAa,GAAG,QAAQ,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC,CAAC;IACnD,MAAM,WAAW,GAAG,KAAK,CAAC,YAAY,EAAE,OAAO,CAAC,YAAY,EAAE,EAAE,CAAC,CAAC;IAClE,MAAM,cAAc,GAClB,WAAW;QACX,aAAa,KAAK,WAAW;QAC7B,CAAC,aAAa,CAAC,UAAU,CAAC,GAAG,WAAW,GAAG,CAAC;QAC1C,CAAC,CAAC,GAAG,WAAW,IAAI,aAAa,EAAE;QACnC,CAAC,CAAC,aAAa,CAAC;IACpB,IAAI,KAAK,CAAC,GAAG,KAAK,YAAY,EAAE,CAAC;QAC/B,MAAM,IAAI,GACR,OAAO,CAAC,GAAG,CAAC,yBAAyB,EAAE,IAAI,EAAE,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;YACjE,2BAA2B,CAAC;QAC9B,MAAM,WAAW,GAAG,cAAc;aAC/B,KAAK,CAAC,GAAG,CAAC;aACV,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,kBAAkB,CAAC,OAAO,CAAC,CAAC;aAC7C,IAAI,CAAC,GAAG,CAAC,CAAC;QACb,OAAO,GAAG,IAAI,WAAW,KAAK,CAAC,MAAM,mBAAmB,WAAW,EAAE,CAAC;IACxE,CAAC;IACD,MAAM,EAAE,IAAI,EAAE,GAAG,qBAAqB,EAAE,CAAC;IACzC,MAAM,WAAW,GAAG,cAAc;SAC/B,KAAK,CAAC,GAAG,CAAC;SACV,GAAG,CAAC,CAAC,OAAO,EAAE,EAAE,CAAC,kBAAkB,CAAC,OAAO,CAAC,CAAC;SAC7C,IAAI,CAAC,GAAG,CAAC,CAAC;IACb,OAAO,GAAG,IAAI,IAAI,KAAK,CAAC,MAAM,iBAAiB,WAAW,gBAAgB,CAAC;AAC7E,CAAC;AAED,MAAM,UAAU,0BAA0B,CAAC,KAAmB;IAC5D,OAAO,iCAAiC,CAAC,KAAK,EAAE,KAAK,CAAC,QAAQ,CAAC,CAAC;AAClE,CAAC"}

package/local-inference/device-fit.d.ts

@@ -0,0 +1,73 @@
+/**
+ * Device-fit selector for the one forced local model (eliza-1).
+ *
+ * The product rule (see `packages/ui/docs/local-model-simplification/README.md`):
+ * **always run the biggest eliza-1 tier we can, balanced against a 128k context
+ * target, with every memory optimization applied** so the floors are as low as
+ * physics allows:
+ *
+ *  - **TurboQuant weights** — the catalog `sizeGb` per tier already reflects the
+ *    TurboQuant-compressed GGUF (e.g. a 2B model is 1.4 GB, not ~4 GB bf16).
+ *  - **Stock q8_0 KV-cache** — Gemma 4's KV is already minimal by construction
+ *    (MQA = 1 KV head, windowed sliding-window attention on most layers, and
+ *    shared-KV layers reusing earlier KV), so a 128k window fits in a fraction
+ *    of a GB at stock q8_0 without the legacy QJL/Polar kernels. The head_dim=128
+ *    QJL kernel does not apply to Gemma's dual head dims (512 global / 256 swa)
+ *    and is not used. KV quant is forced, never a user option.
+ *
+ * `minRamGb` per tier is therefore "TurboQuant weights + a 128k q8_0 KV cache +
+ * runtime/OS overhead". Picking the largest tier whose `minRamGb` fits free RAM
+ * is exactly "biggest model that still gets a 128k window".
+ *
+ * **Dynamic fit-to-RAM context (#8809 M10b).** Picking the largest tier whose
+ * static `minRamGb` floor fits is the coarse gate. On top of that we compute the
+ * *largest context window that actually fits the device's free RAM* for the
+ * chosen tier, from the per-token q8_0 KV rate (`kvBytesPerTokenForTier`) rather
+ * than only honouring the catalog ceiling. On a roomy host that is the tier's
+ * native window; on a host that clears the weights+overhead floor but not a full
+ * 128k cache, the window shrinks to the largest 4k-stepped size that fits. The
+ * tier choice never drops below 2B (0.8B is gone); when even a minimum window
+ * (`ELIZA_1_MIN_LOCAL_CONTEXT`) cannot fit any tier, we return `null`, which the
+ * caller reads as "this modality should route to Cloud" (the AUTO policy).
+ */
+import type { Eliza1TierId } from "./catalog.js";
+/** The KV-cache quantization eliza-1 always uses on-device. Gemma 4's KV is
+ * already minimal (MQA + windowed-SWA + shared-KV), so stock q8_0 is sufficient;
+ * the legacy head_dim=128 QJL kernel is incompatible with Gemma's dual head dims
+ * (512 global / 256 swa) and is not used. */
+export declare const ELIZA_1_KV_QUANT: "q8_0";
+/** The consumer context target. We never advertise less than this if it fits. */
+export declare const ELIZA_1_CONTEXT_TARGET = 131072;
+/** Floor below which a cramped local window is worse than routing to Cloud. */
+export declare const ELIZA_1_MIN_LOCAL_CONTEXT = 8192;
+export interface Eliza1Fit {
+    /** The chosen tier (always the largest that fits the policy). */
+    tierId: Eliza1TierId;
+    /** The context window to load — native target when it fits, else downscaled. */
+    contextLength: number;
+    /** The KV quant to load with (stock q8_0 on-device — Gemma KV is already minimal). */
+    kvQuant: typeof ELIZA_1_KV_QUANT;
+    /** True when context was reduced below the tier's native window to fit RAM. */
+    contextDownscaled: boolean;
+    /** Per-token q8_0 KV footprint used to compute the window (bytes). */
+    kvBytesPerToken: number;
+    /**
+     * The largest 4k-stepped window that fits free RAM for this tier at the q8_0
+     * KV rate, before clamping to the tier's native ceiling. Equals
+     * `contextLength` once clamped on roomy hosts; larger than `contextLength`
+     * only when free RAM could hold more than the tier's native window.
+     */
+    maxFittingContext: number;
+    /** Why this tier — surfaced in diagnostics, never as a user control. */
+    reason: "native-fit" | "context-downscaled";
+}
+/**
+ * Pick the best on-device eliza-1 configuration for a device with `freeRamGb` of
+ * usable RAM (Apple-silicon unified RAM, discrete-GPU VRAM, or CPU RAM — the
+ * caller normalizes this, e.g. via the device-tier classifier).
+ *
+ * Returns `null` when nothing acceptable fits locally → the caller routes this
+ * modality to Cloud (AUTO). Never returns a tier smaller than 2B (0.8B is gone).
+ */
+export declare function selectBestEliza1Fit(freeRamGb: number): Eliza1Fit | null;
+//# sourceMappingURL=device-fit.d.ts.map

package/local-inference/device-fit.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"device-fit.d.ts","sourceRoot":"","sources":["../../src/local-inference/device-fit.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;GA+BG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,MAAM,WAAW,CAAC;AAI9C;;;6CAG6C;AAC7C,eAAO,MAAM,gBAAgB,EAAG,MAAe,CAAC;AAEhD,iFAAiF;AACjF,eAAO,MAAM,sBAAsB,SAAS,CAAC;AAE7C,+EAA+E;AAC/E,eAAO,MAAM,yBAAyB,OAAO,CAAC;AAuC9C,MAAM,WAAW,SAAS;IACxB,iEAAiE;IACjE,MAAM,EAAE,YAAY,CAAC;IACrB,gFAAgF;IAChF,aAAa,EAAE,MAAM,CAAC;IACtB,sFAAsF;IACtF,OAAO,EAAE,OAAO,gBAAgB,CAAC;IACjC,+EAA+E;IAC/E,iBAAiB,EAAE,OAAO,CAAC;IAC3B,sEAAsE;IACtE,eAAe,EAAE,MAAM,CAAC;IACxB;;;;;OAKG;IACH,iBAAiB,EAAE,MAAM,CAAC;IAC1B,wEAAwE;IACxE,MAAM,EAAE,YAAY,GAAG,oBAAoB,CAAC;CAC7C;AAqBD;;;;;;;GAOG;AACH,wBAAgB,mBAAmB,CAAC,SAAS,EAAE,MAAM,GAAG,SAAS,GAAG,IAAI,CAuDvE"}