@elizaos/shared 2.0.0-alpha.538 → 2.0.11-beta.7

package/local-inference/catalog.js

@@ -0,0 +1,547 @@
+/**
+ * Eliza-curated local model catalog.
+ *
+ * Default local inference is restricted to the active Eliza-1 line:
+ * eliza-1-0_8b, eliza-1-2b, eliza-1-4b, eliza-1-9b, eliza-1-27b,
+ * and eliza-1-27b-256k.
+ * These ship Qwen3.5 bases for 0.8B/2B/4B/9B and Qwen3.6 for 27B. The
+ * 2026-05-12 mandate retired the legacy Qwen3 bases; see
+ * packages/training/scripts/training/model_registry.py for the active
+ * registry. External Hub search remains custom/opt-in and never enters
+ * first-run or default eligibility.
+ */
+export const ELIZA_1_HF_REPO = "elizaos/eliza-1";
+export const ELIZA_1_TIER_IDS = [
+    "eliza-1-0_8b",
+    "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-1-0_8b",
+    "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-1-0_8b",
+    "eliza-1-2b",
+    "eliza-1-4b",
+    "eliza-1-9b",
+    "eliza-1-27b",
+    "eliza-1-27b-256k",
+];
+const _ELIZA_1_MTP_TIER_ID_SET = new Set(ELIZA_1_MTP_TIER_IDS);
+function mtpSupportedForTier(id) {
+    return _ELIZA_1_MTP_TIER_ID_SET.has(id);
+}
+// The quantized 4B (Qwen3.5) is the minimum tier that is good enough to ship
+// as the default chat model. The 0.8B/2B tiers remain in the catalog (and as
+// MTP drafter companions) but are no longer first-run defaults — they are too
+// small for a quality conversational experience.
+export const FIRST_RUN_DEFAULT_MODEL_ID = "eliza-1-4b";
+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:
+ *   - 0.8B/2B vision: enabled in the catalog and canonical vision tier set;
+ *     publish staging must include `vision/mmproj-0_8b.gguf` and
+ *     `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 (0_8b / 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-1-0_8b": ["kokoro"],
+    "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",
+    "qjl_full",
+    "polarquant",
+];
+const TIER_SPECS = {
+    "eliza-1-0_8b": {
+        id: "eliza-1-0_8b",
+        params: "0.8B",
+        sizeGb: 0.5,
+        minRamGb: 2,
+        q4MinRamGb: 2,
+        bucket: "small",
+        contextLength: 131072,
+        textFile: "text/eliza-1-0_8b-128k.gguf",
+        // WS2: vision is enabled on the smallest viable tier. The Q4_K_M
+        // mmproj for 0.8B is ~220 MB (see ELIZA_1_BUNDLE_EXTRAS.json), which
+        // fits even on 2 GB-floor devices when the text model is resident.
+        // Camera + screen analysis remain practical on low-tier phones at this
+        // size — the projector cache short-circuits the per-frame cost.
+        hasVision: true,
+        // WS3: image-gen via sd-cpp + SD 1.5 Q5_0 (~1.0 GB). Co-evicts
+        // with vision on the WS1 `vision` resident-role slot; only one of
+        // (VL describe, diffusion generate) is held at a time.
+        hasImageGen: true,
+    },
+    "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 mobile minimum/default. The Q4_K_M weights are 2.6 GB
+        // and the mobile bundle runs a 64k context with compressed KV
+        // (qjl1_256/tbq3_0), so an 8 GB-class phone clears it. 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-1-0_8b":
+            return "0.8B";
+        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 (0_8b/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-1-0_8b" || 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-1-0_8b": ["Q3_K_M", "Q4_K_M", "Q5_K_M"],
+    "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-${voiceQuantForTier(id)}.gguf`;
+    }
+    return "tts/kokoro/kokoro-82m-v1_0-Q4_K_M.gguf";
+}
+function sourceModelForTier(id) {
+    const spec = TIER_SPECS[id];
+    const components = {
+        text: bundleComponent(id, spec.textFile),
+        voice: bundleComponent(id, primaryVoiceFileForTier(id)),
+        asr: bundleComponent(id, "asr/eliza-1-asr.gguf"),
+        vad: bundleComponent(id, "vad/silero-vad-v5.gguf"),
+    };
+    if (spec.hasEmbedding) {
+        components.embedding = bundleComponent(id, "embedding/eliza-1-embedding.gguf");
+    }
+    if (spec.hasVision) {
+        components.vision = bundleComponent(id, `vision/mmproj-${tierSlug(id)}.gguf`);
+    }
+    // Same-file MTP: the NextN head is embedded in the text GGUF
+    // (`qwen35.nextn_predict_layers > 0`), so there is no separate `mtp`
+    // drafter component to download.
+    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,
+        },
+        kvCache: {
+            typeK: "qjl1_256",
+            typeV: "tbq3_0",
+            requiresFork: "buun-llama-cpp",
+        },
+    };
+    if (mtpSupportedForTier(id)) {
+        // Same-file MTP: no separate `drafterFile`. The NextN head lives in
+        // the text GGUF and is activated by `--spec-type draft-mtp` with no
+        // `-md`. These tiers carry a single NextN head
+        // (`nextn_predict_layers = 1`); benchmarks show `draft-n-max 2` is the
+        // throughput peak (a single head autoregressed past 2 collapses
+        // acceptance), so we do not scale the draft window with context length.
+        runtime.mtp = {
+            specType: "draft-mtp",
+            draftMin: 1,
+            draftMax: 2,
+            gpuLayers: "auto",
+        };
+    }
+    return runtime;
+}
+const QUANT_SUFFIX = {
+    q3_k_m: "q3_k_m",
+    q4_0: "Q4_0",
+    q4_k_m: "q4_k_m",
+    q5_k_m: "q5_k_m",
+    q6_k: "q6_k",
+    q8_0: "q8_0",
+};
+function textQuantizationMatrix(args) {
+    const fileBase = args.primaryGgufFile.replace(/\.gguf$/, "");
+    const mk = (id, label, scale, minRamScale, status) => ({
+        id,
+        label,
+        ggufFile: id === "q4_k_m"
+            ? args.primaryGgufFile
+            : `${fileBase}-${QUANT_SUFFIX[id]}.gguf`,
+        sizeGb: Number((args.q4SizeGb * scale).toFixed(1)),
+        minRamGb: Math.ceil(args.q4MinRamGb * minRamScale),
+        status,
+    });
+    return {
+        defaultVariantId: "q4_k_m",
+        variants: [
+            mk("q3_k_m", "3-bit", 0.76, 0.85, "planned"),
+            mk("q4_k_m", "4-bit", 1, 1, "published"),
+            mk("q5_k_m", "5-bit", 1.22, 1.18, "planned"),
+            mk("q6_k", "6-bit", 1.45, 1.35, "planned"),
+            mk("q8_0", "8-bit", 1.95, 1.8, "planned"),
+        ],
+    };
+}
+function blurbForTier(id) {
+    const displayName = tierDisplayName(id);
+    switch (id) {
+        case "eliza-1-0_8b":
+            return `${displayName} - smallest local tier for low-memory phones and CPU fallback.`;
+        case "eliza-1-2b":
+            return `${displayName} - recommended first-run local tier for responsive text and voice.`;
+        case "eliza-1-4b":
+            return `${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 `${displayName} - long-context local tier for high-memory GPU workstations.`;
+    }
+    const exhaustive = id;
+    return exhaustive;
+}
+const drafterId = (id) => `${id}-drafter`;
+// DFlash speculative-decoding draft companions published under each tier's
+// HuggingFace bundle (elizaos/eliza-1 -> bundles/<slug>/dflash/drafter-<slug>.gguf).
+// Sizes are the published gguf byte sizes verified against the HF repo (2026-05);
+// params/minRamGb are companion metadata (these are hidden, runtimeRole-gated).
+const TIER_DRAFTERS = {
+    "eliza-1-0_8b": {
+        ggufRel: "dflash/drafter-0_8b.gguf",
+        params: "0.5B",
+        sizeGb: 0.24,
+        minRamGb: 2,
+        bucket: "small",
+    },
+    "eliza-1-2b": {
+        ggufRel: "dflash/drafter-2b.gguf",
+        params: "0.8B",
+        sizeGb: 0.68,
+        minRamGb: 4,
+        bucket: "small",
+    },
+};
+function drafterCompanion(id) {
+    const drafter = TIER_DRAFTERS[id];
+    if (!drafter) {
+        throw new Error(`No DFlash drafter spec for tier ${id}`);
+    }
+    return {
+        id: drafterId(id),
+        displayName: `${tierDisplayName(id)} drafter`,
+        hfRepo: ELIZA_1_HF_REPO,
+        hfPathPrefix: bundleRemotePrefix(id),
+        ggufFile: bundlePath(id, drafter.ggufRel),
+        params: drafter.params,
+        quant: "Eliza-1 DFlash drafter companion",
+        sizeGb: drafter.sizeGb,
+        minRamGb: drafter.minRamGb,
+        category: "chat",
+        bucket: drafter.bucket,
+        tokenizerFamily: "qwen35",
+        hiddenFromCatalog: true,
+        runtimeRole: "mtp-drafter",
+        companionForModelId: id,
+        blurb: "DFlash speculative-decoding draft companion.",
+    };
+}
+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: "qwen35",
+        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,
+        }),
+        blurb: blurbForTier(id),
+        publishStatus: eliza1TierPublishStatus(id),
+    };
+}
+export const MODEL_CATALOG = ELIZA_1_TIER_IDS.map((id) => chatTier(id));
+// DFlash speculative-decoding drafter companions. Findable by id (installer
+// companion download + registry self-heal via findCatalogModel) but
+// intentionally NOT part of the published hub catalog (MODEL_CATALOG) — they are
+// hidden, mtp-drafter companion entries keyed to their parent tier.
+const DRAFTER_COMPANIONS = ELIZA_1_TIER_IDS.filter((id) => TIER_DRAFTERS[id]).map((id) => drafterCompanion(id));
+export function findCatalogModel(id) {
+    return (MODEL_CATALOG.find((m) => m.id === id) ??
+        DRAFTER_COMPANIONS.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 = process.env.ELIZA_HF_BASE_URL?.trim().replace(/\/+$/, "") ||
+        "https://huggingface.co";
+    const encodedPath = pathWithPrefix
+        .split("/")
+        .map((segment) => encodeURIComponent(segment))
+        .join("/");
+    return `${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;;;;;;;;;;;GAWG;AASH,MAAM,CAAC,MAAM,eAAe,GAAG,iBAA0B,CAAC;AAE1D,MAAM,CAAC,MAAM,gBAAgB,GAAG;IAC9B,cAAc;IACd,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,cAAc;IACd,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,cAAc;IACd,YAAY;IACZ,YAAY;IACZ,YAAY;IACZ,aAAa;IACb,kBAAkB;CAC4B,CAAC;AAEjD,MAAM,wBAAwB,GAA8B,IAAI,GAAG,CACjE,oBAAoB,CACrB,CAAC;AAEF,SAAS,mBAAmB,CAAC,EAAgB;IAC3C,OAAO,wBAAwB,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC;AAC1C,CAAC;AAED,6EAA6E;AAC7E,6EAA6E;AAC7E,8EAA8E;AAC9E,iDAAiD;AACjD,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,cAAc,EAAE,CAAC,QAAQ,CAAC;IAC1B,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;IAClD,QAAQ;IACR,QAAQ;IACR,UAAU;IACV,YAAY;CACb,CAAC;AA0BF,MAAM,UAAU,GAA6C;IAC3D,cAAc,EAAE;QACd,EAAE,EAAE,cAAc;QAClB,MAAM,EAAE,MAAM;QACd,MAAM,EAAE,GAAG;QACX,QAAQ,EAAE,CAAC;QACX,UAAU,EAAE,CAAC;QACb,MAAM,EAAE,OAAO;QACf,aAAa,EAAE,MAAM;QACrB,QAAQ,EAAE,6BAA6B;QACvC,iEAAiE;QACjE,qEAAqE;QACrE,mEAAmE;QACnE,uEAAuE;QACvE,gEAAgE;QAChE,SAAS,EAAE,IAAI;QACf,+DAA+D;QAC/D,kEAAkE;QAClE,uDAAuD;QACvD,WAAW,EAAE,IAAI;KAClB;IACD,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,0EAA0E;QAC1E,8DAA8D;QAC9D,uEAAuE;QACvE,wEAAwE;QACxE,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,cAAc;YACjB,OAAO,MAAM,CAAC;QAChB,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,cAAc,IAAI,EAAE,KAAK,YAAY,IAAI,EAAE,KAAK,YAAY;QACxE,CAAC,CAAC,QAAQ;QACV,CAAC,CAAC,MAAM,CAAC;AACb,CAAC;AAED;;;;;;;;;;;GAWG;AACH,MAAM,8BAA8B,GAEhC;IACF,cAAc,EAAE,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,CAAC;IAC9C,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,sBAAsB,CAAC;QAChD,GAAG,EAAE,eAAe,CAAC,EAAE,EAAE,wBAAwB,CAAC;KACnD,CAAC;IAEF,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,6DAA6D;IAC7D,qEAAqE;IACrE,iCAAiC;IAEjC,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;QACD,OAAO,EAAE;YACP,KAAK,EAAE,UAAU;YACjB,KAAK,EAAE,QAAQ;YACf,YAAY,EAAE,gBAAgB;SAC/B;KACF,CAAC;IAEF,IAAI,mBAAmB,CAAC,EAAE,CAAC,EAAE,CAAC;QAC5B,oEAAoE;QACpE,oEAAoE;QACpE,+CAA+C;QAC/C,uEAAuE;QACvE,gEAAgE;QAChE,wEAAwE;QACxE,OAAO,CAAC,GAAG,GAAG;YACZ,QAAQ,EAAE,WAAW;YACrB,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,IAAI,EAAE,MAAM;IACZ,MAAM,EAAE,QAAQ;IAChB,MAAM,EAAE,QAAQ;IAChB,IAAI,EAAE,MAAM;IACZ,IAAI,EAAE,MAAM;CACb,CAAC;AAEF,SAAS,sBAAsB,CAAC,IAI/B;IACC,MAAM,QAAQ,GAAG,IAAI,CAAC,eAAe,CAAC,OAAO,CAAC,SAAS,EAAE,EAAE,CAAC,CAAC;IAC7D,MAAM,EAAE,GAAG,CACT,EAAyB,EACzB,KAA0C,EAC1C,KAAa,EACb,WAAmB,EACnB,MAA4C,EAChB,EAAE,CAAC,CAAC;QAChC,EAAE;QACF,KAAK;QACL,QAAQ,EACN,EAAE,KAAK,QAAQ;YACb,CAAC,CAAC,IAAI,CAAC,eAAe;YACtB,CAAC,CAAC,GAAG,QAAQ,IAAI,YAAY,CAAC,EAAE,CAAC,OAAO;QAC5C,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;KACP,CAAC,CAAC;IAEH,OAAO;QACL,gBAAgB,EAAE,QAAQ;QAC1B,QAAQ,EAAE;YACR,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC;YAC5C,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,CAAC,EAAE,CAAC,EAAE,WAAW,CAAC;YACxC,EAAE,CAAC,QAAQ,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC;YAC5C,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,IAAI,EAAE,SAAS,CAAC;YAC1C,EAAE,CAAC,MAAM,EAAE,OAAO,EAAE,IAAI,EAAE,GAAG,EAAE,SAAS,CAAC;SAC1C;KACF,CAAC;AACJ,CAAC;AAED,SAAS,YAAY,CAAC,EAAgB;IACpC,MAAM,WAAW,GAAG,eAAe,CAAC,EAAE,CAAC,CAAC;IACxC,QAAQ,EAAE,EAAE,CAAC;QACX,KAAK,cAAc;YACjB,OAAO,GAAG,WAAW,gEAAgE,CAAC;QACxF,KAAK,YAAY;YACf,OAAO,GAAG,WAAW,oEAAoE,CAAC;QAC5F,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,MAAM,SAAS,GAAG,CAAC,EAAgB,EAA6B,EAAE,CAChE,GAAG,EAAE,UAAU,CAAC;AAElB,2EAA2E;AAC3E,qFAAqF;AACrF,kFAAkF;AAClF,gFAAgF;AAChF,MAAM,aAAa,GAWf;IACF,cAAc,EAAE;QACd,OAAO,EAAE,0BAA0B;QACnC,MAAM,EAAE,MAAM;QACd,MAAM,EAAE,IAAI;QACZ,QAAQ,EAAE,CAAC;QACX,MAAM,EAAE,OAAO;KAChB;IACD,YAAY,EAAE;QACZ,OAAO,EAAE,wBAAwB;QACjC,MAAM,EAAE,MAAM;QACd,MAAM,EAAE,IAAI;QACZ,QAAQ,EAAE,CAAC;QACX,MAAM,EAAE,OAAO;KAChB;CACF,CAAC;AAEF,SAAS,gBAAgB,CAAC,EAAgB;IACxC,MAAM,OAAO,GAAG,aAAa,CAAC,EAAE,CAAC,CAAC;IAClC,IAAI,CAAC,OAAO,EAAE,CAAC;QACb,MAAM,IAAI,KAAK,CAAC,mCAAmC,EAAE,EAAE,CAAC,CAAC;IAC3D,CAAC;IACD,OAAO;QACL,EAAE,EAAE,SAAS,CAAC,EAAE,CAAC;QACjB,WAAW,EAAE,GAAG,eAAe,CAAC,EAAE,CAAC,UAAU;QAC7C,MAAM,EAAE,eAAe;QACvB,YAAY,EAAE,kBAAkB,CAAC,EAAE,CAAC;QACpC,QAAQ,EAAE,UAAU,CAAC,EAAE,EAAE,OAAO,CAAC,OAAO,CAAC;QACzC,MAAM,EAAE,OAAO,CAAC,MAAM;QACtB,KAAK,EAAE,kCAAkC;QACzC,MAAM,EAAE,OAAO,CAAC,MAAM;QACtB,QAAQ,EAAE,OAAO,CAAC,QAAQ;QAC1B,QAAQ,EAAE,MAAM;QAChB,MAAM,EAAE,OAAO,CAAC,MAAM;QACtB,eAAe,EAAE,QAAQ;QACzB,iBAAiB,EAAE,IAAI;QACvB,WAAW,EAAE,aAAa;QAC1B,mBAAmB,EAAE,EAAE;QACvB,KAAK,EAAE,8CAA8C;KACtD,CAAC;AACJ,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,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;SAC5B,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,4EAA4E;AAC5E,oEAAoE;AACpE,iFAAiF;AACjF,oEAAoE;AACpE,MAAM,kBAAkB,GAAmB,gBAAgB,CAAC,MAAM,CAChE,CAAC,EAAE,EAAE,EAAE,CAAC,aAAa,CAAC,EAAE,CAAC,CAC1B,CAAC,GAAG,CAAC,CAAC,EAAE,EAAE,EAAE,CAAC,gBAAgB,CAAC,EAAE,CAAC,CAAC,CAAC;AAEpC,MAAM,UAAU,gBAAgB,CAAC,EAAU;IACzC,OAAO,CACL,aAAa,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC;QACtC,kBAAkB,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,EAAE,KAAK,EAAE,CAAC,CAC5C,CAAC;AACJ,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,IAAI,GACR,OAAO,CAAC,GAAG,CAAC,iBAAiB,EAAE,IAAI,EAAE,CAAC,OAAO,CAAC,MAAM,EAAE,EAAE,CAAC;QACzD,wBAAwB,CAAC;IAC3B,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/gpu-profiles.d.ts

@@ -0,0 +1,110 @@
+/**
+ * Per-GPU deployment profiles for single-GPU Eliza-1 servers.
+ *
+ * Scope: ONE GPU per host. We do not target NVLink, tensor-parallel splits,
+ * or multi-tenant datacenter workloads. The product target is "one
+ * conversation at a time on a single GPU box" — `--parallel N` can still go
+ * higher than on a phone, but the model is not split across cards.
+ *
+ * These profiles are *recommendations* — the runtime selects a bundle by id
+ * and then maps the host's GPU to a `GpuProfile` to fill in llama-server
+ * flags (KV cache types, n-gpu-layers, batch sizing, MTP draft range,
+ * mlock). A non-NVIDIA host returns `null` and the runtime falls back to
+ * the catalog defaults.
+ *
+ * The KV cache type names match `LocalRuntimeKernel` / the llama.cpp fork
+ * kernel handles (`qjl1_256`, `q4_polar`, `turbo3_0`, `turbo4_0`, `q8_0`,
+ * `f16`) so that `appendOptimizationFlags` + `applyGpuProfile` can pass
+ * them straight through to `--cache-type-k` / `--cache-type-v`.
+ *
+ * The active mobile/local release exposes Eliza-1 0_8b, 2b, 4b, 9b, and 27b
+ * tiers, with larger cards recommended toward the biggest installed bundle
+ * that leaves memory headroom.
+ */
+import type { Eliza1TierId } from "./catalog.js";
+export type GpuProfileId = "rtx-3090" | "rtx-4090" | "rtx-5090" | "h200";
+/**
+ * KV cache type names accepted by `llama-server --cache-type-k/-v`. Mirrors
+ * the strings the buun-llama-cpp fork advertises in `CAPABILITIES.json`.
+ * Kept as a string literal union so a profile's choice is checked at
+ * compile time and propagates into the bundle manifest's `requiresKernel`
+ * set.
+ */
+export type KvCacheType = "f16" | "q8_0" | "q4_0" | "qjl1_256" | "q4_polar" | "turbo3_0" | "turbo4_0";
+export interface GpuProfile {
+    id: GpuProfileId;
+    /** Human-readable vendor card name shown to the operator. */
+    displayName: string;
+    /** Total VRAM in GiB. Used for headroom math + recommendations. */
+    vramGb: number;
+    /** CUDA compute capability, e.g. `"sm_86"`. */
+    computeCapability: string;
+    /** Peak HBM/GDDR bandwidth in GB/s — feeds the perf estimator. */
+    memoryBandwidthGBs: number;
+    /** Whether the GPU has hardware FP8 (Ada/Hopper/Blackwell). */
+    fp8: boolean;
+    /**
+     * Recommended catalog bundle ids for single-GPU deployment on this card.
+     * Ordered by quality: first entry is the "best fit" for the card; later
+     * entries are smaller fallbacks that still leave headroom. The recommender
+     * picks the first id that the user has installed.
+     */
+    recommendedBundles: ReadonlyArray<Eliza1TierId>;
+    /** Inference flags. */
+    flashAttn: boolean;
+    /** `--cache-type-k` value. */
+    kvCacheTypeK: KvCacheType;
+    /** `--cache-type-v` value. */
+    kvCacheTypeV: KvCacheType;
+    /**
+     * `--n-gpu-layers` — `-1` for "all layers on GPU". Single-GPU only; we
+     * never split layers across two cards.
+     */
+    nGpuLayers: number;
+    /**
+     * `--ctx-size`. Always sized to the bundle's `contextLength`, but the
+     * profile records the *recommended* max for the card.
+     */
+    contextSize: number;
+    /** `--parallel N` continuous-batching slots. */
+    parallel: number;
+    /** `--batch-size N` logical batch. */
+    batchSize: number;
+    /** `--ubatch-size N` physical micro-batch. */
+    ubatchSize: number;
+    /** MTP speculative-decoding draft range. */
+    mtpDraftMin: number;
+    mtpDraftMax: number;
+    /** `--mlock` — pin model pages in RAM. */
+    mlock: boolean;
+    /** `--no-mmap` — disable mmap loading. */
+    noMmap: boolean;
+    /**
+     * Force KV spill to CPU (`--no-kv-offload`). Only useful on cards that
+     * cannot hold the full KV cache. The 141 GiB H200 sets this to `false`;
+     * the 24 GiB cards can opt in for very long contexts.
+     */
+    kvSpillToCpu: boolean;
+}
+export declare const GPU_PROFILES: Readonly<Record<GpuProfileId, GpuProfile>>;
+export declare const GPU_PROFILE_IDS: ReadonlyArray<GpuProfileId>;
+/**
+ * Match an `nvidia-smi --query-gpu=name` output line to a profile. Returns
+ * `null` when the card is not in the supported set; callers should fall
+ * back to the catalog defaults in that case rather than guess.
+ *
+ * The patterns are intentionally strict to avoid mis-classifying a 3080
+ * as a 3090, or an A100 as a 4090.
+ */
+export declare function matchGpuProfile(gpuName: string): GpuProfileId | null;
+/**
+ * Returns the headroom (in GiB) a profile reserves for the OS / driver /
+ * activations / drafter / N parallel-slot KV. Used for sizing checks
+ * before promoting a bundle to "fits" on a card.
+ *
+ * The figures are deliberate per-tier reserves, not a formula — different
+ * cards have different driver overheads (Windows display drivers steal
+ * ~1 GiB before workloads even start; H200 SXM has none).
+ */
+export declare function reservedHeadroomGb(profile: GpuProfile): number;
+//# sourceMappingURL=gpu-profiles.d.ts.map