@atomic-chat/core 0.5.1

package/dist/speculative/mlx-dflash-registry.js

@@ -0,0 +1,230 @@
+/**
+ * Ported verbatim from: extensions/mlx-extension/src/dflashRegistry.ts
+ * DFlash draft model registry.
+ *
+ * A pure, network-free static manifest mapping base MLX model identifiers to
+ * matching `z-lab/*-DFlash*` draft repos from the curated collection at
+ * https://huggingface.co/collections/z-lab/dflash.
+ *
+ * The MLX extension is responsible for the local-first resolution and direct
+ * `/resolve/main/<file>` downloads — this module only carries the static
+ * lookup data so the toggle path never touches `huggingface.co/api/...`.
+ */
+/// `dflash.model_mlx.load_draft` (in the dflash package) only reads
+/// `config.json` and any `*.safetensors` from the draft directory; tokenizer
+/// files belong to the *target* model. Most `z-lab/*` drafts ship as a single
+/// `model.safetensors`, so we only require those two files. Sharded drafts
+/// (rare; e.g. >2B params) need a per-repo override.
+const DEFAULT_REQUIRED = ['config.json', 'model.safetensors'];
+/// Currently unused — kept so per-repo overrides can opt in to additional
+/// downloads (e.g. `model.safetensors.index.json` + shards) without
+/// changing the type shape.
+const DEFAULT_OPTIONAL = [];
+/// Build the required-file set for a sharded draft (weights split across N
+/// `model-XXXXX-of-YYYYY.safetensors` files plus a `model.safetensors.index.json`
+/// shard map) instead of a single `model.safetensors`. All shards and the
+/// index must be present for the draft directory to load.
+function shardedRequired(total) {
+    const pad = (n) => String(n).padStart(5, '0');
+    const files = ['config.json', 'model.safetensors.index.json'];
+    for (let i = 1; i <= total; i++) {
+        files.push(`model-${pad(i)}-of-${pad(total)}.safetensors`);
+    }
+    return files;
+}
+/// Snapshot of https://huggingface.co/collections/z-lab/dflash.
+///
+/// Keys are normalized base-model ids (lowercase, separators preserved as
+/// `-`) derived via `normalizeBaseId`. Values describe the matching draft
+/// repo and the file set that needs to be present locally.
+export const STATIC_DRAFT_MAP = {
+    'qwen3.5-122b-a10b': {
+        repo: 'z-lab/Qwen3.5-122B-A10B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.6-35b-a3b': {
+        repo: 'z-lab/Qwen3.6-35B-A3B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.6-27b': {
+        repo: 'z-lab/Qwen3.6-27B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    /// Gemma 4 also ships DFlash drafts (z-lab), giving Gemma 4 targets a third
+    /// speculative path alongside MTP (`*-assistant`) and EAGLE-3. Keys mirror
+    /// `eagle3Registry.ts` — `-it` is stripped by `TRAIL_HINT_RE`.
+    'gemma-4-31b': {
+        repo: 'z-lab/gemma-4-31B-it-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gemma-4-26b-a4b': {
+        repo: 'z-lab/gemma-4-26B-A4B-it-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    /// Kimi drafts ship sharded safetensors, so they need explicit per-shard
+    /// required sets (the default single `model.safetensors` would 404).
+    'kimi-k2.5': {
+        repo: 'z-lab/Kimi-K2.5-DFlash',
+        required: shardedRequired(2),
+        optional: DEFAULT_OPTIONAL,
+    },
+    'kimi-k2.6': {
+        repo: 'z-lab/Kimi-K2.6-DFlash',
+        required: shardedRequired(3),
+        optional: DEFAULT_OPTIONAL,
+    },
+    'minimax-m2.5': {
+        repo: 'z-lab/MiniMax-M2.5-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'minimax-m2.7': {
+        repo: 'z-lab/MiniMax-M2.7-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.5-4b': {
+        repo: 'z-lab/Qwen3.5-4B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.5-9b': {
+        repo: 'z-lab/Qwen3.5-9B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.5-35b-a3b': {
+        repo: 'z-lab/Qwen3.5-35B-A3B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.5-27b': {
+        repo: 'z-lab/Qwen3.5-27B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3-coder-next': {
+        repo: 'z-lab/Qwen3-Coder-Next-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gpt-oss-20b': {
+        repo: 'z-lab/gpt-oss-20b-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gpt-oss-120b': {
+        repo: 'z-lab/gpt-oss-120b-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3-4b': {
+        repo: 'z-lab/Qwen3-4B-DFlash-b16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3-8b': {
+        repo: 'z-lab/Qwen3-8B-DFlash-b16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3-coder-30b-a3b': {
+        repo: 'z-lab/Qwen3-Coder-30B-A3B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    /// The published mlx-community repos carry an `-Instruct` infix
+    /// (`Qwen3-Coder-30B-A3B-Instruct-*`) that `normalizeBaseId` does not strip,
+    /// so alias the `-instruct` form to the same draft.
+    'qwen3-coder-30b-a3b-instruct': {
+        repo: 'z-lab/Qwen3-Coder-30B-A3B-DFlash',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'llama3.1-8b-instruct': {
+        repo: 'z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'llama-3.1-8b-instruct': {
+        repo: 'z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'meta-llama-3.1-8b-instruct': {
+        repo: 'z-lab/LLaMA3.1-8B-Instruct-DFlash-UltraChat',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+};
+/// Quantization / format suffixes that should be stripped from a model id
+/// before attempting registry lookup. Order matters: longer first.
+/// `qat` (Quantization-Aware Training) and `unquantized` are included so
+/// QAT-quantized targets (e.g. `gemma-4-E4B-it-qat-4bit`) normalize to the
+/// same key as the standard float builds (`gemma-4-e4b`).
+const QUANT_SUFFIX_RE = /-(?:4bit|8bit|6bit|5bit|3bit|2bit|q4_k_m|q4_k_s|q5_k_m|q5_k_s|q6_k|q8_0|fp16|bf16|fp8|mxfp4|mxfp8|mlx|qat|unquantized)(?:-[a-z0-9]+)?$/i;
+/// Trailing dimensionality / dataset hints that occasionally appear in MLX
+/// repo names (e.g. `-128k`, `-it`, `-chat`). Stripped after quant suffixes.
+const TRAIL_HINT_RE = /-(?:it|chat|128k|256k|1m|hf|base)$/i;
+/**
+ * Normalize an arbitrary MLX model id into the registry-key form.
+ *
+ *   "mlx-community/Qwen3-4B-4bit" -> "qwen3-4b"
+ *   "Qwen3.5-9B-mlx-bf16"         -> "qwen3.5-9b"
+ *   "gpt-oss-20b"                 -> "gpt-oss-20b"
+ */
+export function normalizeBaseId(modelId) {
+    let id = modelId.trim();
+    const slashIdx = id.lastIndexOf('/');
+    if (slashIdx !== -1) {
+        id = id.slice(slashIdx + 1);
+    }
+    id = id.toLowerCase();
+    for (let i = 0; i < 3; i++) {
+        const next = id.replace(QUANT_SUFFIX_RE, '').replace(TRAIL_HINT_RE, '');
+        if (next === id)
+            break;
+        id = next;
+    }
+    return id;
+}
+/**
+ * Resolve a DFlash draft repo for the given base model.
+ *
+ * Pure / synchronous — relies solely on `STATIC_DRAFT_MAP` and never touches
+ * the network. Returns `null` when the base model is not in the curated
+ * z-lab collection.
+ */
+export function resolveDflashDraft(modelId) {
+    /// If the caller already passed a fully-qualified `z-lab/<repo>` id, do a
+    /// reverse lookup against the map values so we still pull the correct
+    /// manifest (and not whatever `normalizeBaseId` would mangle it into).
+    if (/^z-lab\//i.test(modelId)) {
+        const target = modelId.toLowerCase();
+        for (const manifest of Object.values(STATIC_DRAFT_MAP)) {
+            if (manifest.repo.toLowerCase() === target) {
+                return {
+                    repo: manifest.repo,
+                    required: manifest.required,
+                    optional: manifest.optional ?? [],
+                };
+            }
+        }
+        return null;
+    }
+    const baseId = normalizeBaseId(modelId);
+    const manifest = STATIC_DRAFT_MAP[baseId];
+    if (!manifest)
+        return null;
+    return {
+        repo: manifest.repo,
+        required: manifest.required,
+        optional: manifest.optional ?? [],
+    };
+}
+//# sourceMappingURL=mlx-dflash-registry.js.map

package/dist/speculative/mlx-dflash-registry.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"mlx-dflash-registry.js","sourceRoot":"","sources":["../../src/speculative/mlx-dflash-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;GAWG;AAoBH,oEAAoE;AACpE,6EAA6E;AAC7E,8EAA8E;AAC9E,2EAA2E;AAC3E,qDAAqD;AACrD,MAAM,gBAAgB,GAAG,CAAC,aAAa,EAAE,mBAAmB,CAAC,CAAA;AAE7D,0EAA0E;AAC1E,oEAAoE;AACpE,4BAA4B;AAC5B,MAAM,gBAAgB,GAAa,EAAE,CAAA;AAErC,2EAA2E;AAC3E,kFAAkF;AAClF,0EAA0E;AAC1E,0DAA0D;AAC1D,SAAS,eAAe,CAAC,KAAa;IACpC,MAAM,GAAG,GAAG,CAAC,CAAS,EAAE,EAAE,CAAC,MAAM,CAAC,CAAC,CAAC,CAAC,QAAQ,CAAC,CAAC,EAAE,GAAG,CAAC,CAAA;IACrD,MAAM,KAAK,GAAG,CAAC,aAAa,EAAE,8BAA8B,CAAC,CAAA;IAC7D,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,IAAI,KAAK,EAAE,CAAC,EAAE,EAAE,CAAC;QAChC,KAAK,CAAC,IAAI,CAAC,SAAS,GAAG,CAAC,CAAC,CAAC,OAAO,GAAG,CAAC,KAAK,CAAC,cAAc,CAAC,CAAA;IAC5D,CAAC;IACD,OAAO,KAAK,CAAA;AACd,CAAC;AAED,gEAAgE;AAChE,GAAG;AACH,0EAA0E;AAC1E,0EAA0E;AAC1E,2DAA2D;AAC3D,MAAM,CAAC,MAAM,gBAAgB,GAAsC;IACjE,mBAAmB,EAAE;QACnB,IAAI,EAAE,gCAAgC;QACtC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,iBAAiB,EAAE;QACjB,IAAI,EAAE,8BAA8B;QACpC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,aAAa,EAAE;QACb,IAAI,EAAE,0BAA0B;QAChC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,4EAA4E;IAC5E,2EAA2E;IAC3E,+DAA+D;IAC/D,aAAa,EAAE;QACb,IAAI,EAAE,6BAA6B;QACnC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,iBAAiB,EAAE;QACjB,IAAI,EAAE,iCAAiC;QACvC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,yEAAyE;IACzE,qEAAqE;IACrE,WAAW,EAAE;QACX,IAAI,EAAE,wBAAwB;QAC9B,QAAQ,EAAE,eAAe,CAAC,CAAC,CAAC;QAC5B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,WAAW,EAAE;QACX,IAAI,EAAE,wBAAwB;QAC9B,QAAQ,EAAE,eAAe,CAAC,CAAC,CAAC;QAC5B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,cAAc,EAAE;QACd,IAAI,EAAE,2BAA2B;QACjC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,cAAc,EAAE;QACd,IAAI,EAAE,2BAA2B;QACjC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,YAAY,EAAE;QACZ,IAAI,EAAE,yBAAyB;QAC/B,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,YAAY,EAAE;QACZ,IAAI,EAAE,yBAAyB;QAC/B,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,iBAAiB,EAAE;QACjB,IAAI,EAAE,8BAA8B;QACpC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,aAAa,EAAE;QACb,IAAI,EAAE,0BAA0B;QAChC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,kBAAkB,EAAE;QAClB,IAAI,EAAE,+BAA+B;QACrC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,aAAa,EAAE;QACb,IAAI,EAAE,0BAA0B;QAChC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,cAAc,EAAE;QACd,IAAI,EAAE,2BAA2B;QACjC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,UAAU,EAAE;QACV,IAAI,EAAE,2BAA2B;QACjC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,UAAU,EAAE;QACV,IAAI,EAAE,2BAA2B;QACjC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,qBAAqB,EAAE;QACrB,IAAI,EAAE,kCAAkC;QACxC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,gEAAgE;IAChE,6EAA6E;IAC7E,oDAAoD;IACpD,8BAA8B,EAAE;QAC9B,IAAI,EAAE,kCAAkC;QACxC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,sBAAsB,EAAE;QACtB,IAAI,EAAE,6CAA6C;QACnD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,uBAAuB,EAAE;QACvB,IAAI,EAAE,6CAA6C;QACnD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,4BAA4B,EAAE;QAC5B,IAAI,EAAE,6CAA6C;QACnD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;CACF,CAAA;AAED,0EAA0E;AAC1E,mEAAmE;AACnE,yEAAyE;AACzE,2EAA2E;AAC3E,0DAA0D;AAC1D,MAAM,eAAe,GACnB,yIAAyI,CAAA;AAE3I,2EAA2E;AAC3E,6EAA6E;AAC7E,MAAM,aAAa,GAAG,qCAAqC,CAAA;AAE3D;;;;;;GAMG;AACH,MAAM,UAAU,eAAe,CAAC,OAAe;IAC7C,IAAI,EAAE,GAAG,OAAO,CAAC,IAAI,EAAE,CAAA;IAEvB,MAAM,QAAQ,GAAG,EAAE,CAAC,WAAW,CAAC,GAAG,CAAC,CAAA;IACpC,IAAI,QAAQ,KAAK,CAAC,CAAC,EAAE,CAAC;QACpB,EAAE,GAAG,EAAE,CAAC,KAAK,CAAC,QAAQ,GAAG,CAAC,CAAC,CAAA;IAC7B,CAAC;IAED,EAAE,GAAG,EAAE,CAAC,WAAW,EAAE,CAAA;IAErB,KAAK,IAAI,CAAC,GAAG,CAAC,EAAE,CAAC,GAAG,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC;QAC3B,MAAM,IAAI,GAAG,EAAE,CAAC,OAAO,CAAC,eAAe,EAAE,EAAE,CAAC,CAAC,OAAO,CAAC,aAAa,EAAE,EAAE,CAAC,CAAA;QACvE,IAAI,IAAI,KAAK,EAAE;YAAE,MAAK;QACtB,EAAE,GAAG,IAAI,CAAA;IACX,CAAC;IAED,OAAO,EAAE,CAAA;AACX,CAAC;AAED;;;;;;GAMG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAe;IAChD,0EAA0E;IAC1E,sEAAsE;IACtE,uEAAuE;IACvE,IAAI,WAAW,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QAC9B,MAAM,MAAM,GAAG,OAAO,CAAC,WAAW,EAAE,CAAA;QACpC,KAAK,MAAM,QAAQ,IAAI,MAAM,CAAC,MAAM,CAAC,gBAAgB,CAAC,EAAE,CAAC;YACvD,IAAI,QAAQ,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,MAAM,EAAE,CAAC;gBAC3C,OAAO;oBACL,IAAI,EAAE,QAAQ,CAAC,IAAI;oBACnB,QAAQ,EAAE,QAAQ,CAAC,QAAQ;oBAC3B,QAAQ,EAAE,QAAQ,CAAC,QAAQ,IAAI,EAAE;iBAClC,CAAA;YACH,CAAC;QACH,CAAC;QACD,OAAO,IAAI,CAAA;IACb,CAAC;IAED,MAAM,MAAM,GAAG,eAAe,CAAC,OAAO,CAAC,CAAA;IACvC,MAAM,QAAQ,GAAG,gBAAgB,CAAC,MAAM,CAAC,CAAA;IACzC,IAAI,CAAC,QAAQ;QAAE,OAAO,IAAI,CAAA;IAE1B,OAAO;QACL,IAAI,EAAE,QAAQ,CAAC,IAAI;QACnB,QAAQ,EAAE,QAAQ,CAAC,QAAQ;QAC3B,QAAQ,EAAE,QAAQ,CAAC,QAAQ,IAAI,EAAE;KAClC,CAAA;AACH,CAAC"}

package/dist/speculative/mlx-eagle3-registry.d.ts

@@ -0,0 +1,24 @@
+/**
+ * Ported verbatim from: extensions/mlx-extension/src/eagle3Registry.ts
+ * EAGLE-3 draft model registry.
+ *
+ * A pure, network-free static manifest mapping base MLX model identifiers to
+ * the matching `RedHatAI/*-speculator.eagle3` draft checkpoints (Speculators /
+ * SGLang EAGLE-3 format), which `mlx_vlm`'s `eagle3` drafter loads directly.
+ *
+ * Mirrors the shape of `dflashRegistry.ts` / `mtpRegistry.ts` so the MLX
+ * extension routes all three speculative-decoding kinds through the same
+ * local-first resolution and direct `/resolve/main/<file>` download path. This
+ * module only carries the static lookup data; no `huggingface.co/api/...`
+ * calls happen here.
+ */
+import { type DraftRepoManifest, type DraftResolution } from './mlx-dflash-registry.js';
+export declare const STATIC_EAGLE3_MAP: Record<string, DraftRepoManifest>;
+/**
+ * Resolve an EAGLE-3 draft repo for the given base model.
+ *
+ * Pure / synchronous — relies solely on `STATIC_EAGLE3_MAP` and never touches
+ * the network. Returns `null` when the base model has no canonical EAGLE-3
+ * pairing.
+ */
+export declare function resolveEagle3Draft(modelId: string): DraftResolution | null;

package/dist/speculative/mlx-eagle3-registry.js

@@ -0,0 +1,81 @@
+/**
+ * Ported verbatim from: extensions/mlx-extension/src/eagle3Registry.ts
+ * EAGLE-3 draft model registry.
+ *
+ * A pure, network-free static manifest mapping base MLX model identifiers to
+ * the matching `RedHatAI/*-speculator.eagle3` draft checkpoints (Speculators /
+ * SGLang EAGLE-3 format), which `mlx_vlm`'s `eagle3` drafter loads directly.
+ *
+ * Mirrors the shape of `dflashRegistry.ts` / `mtpRegistry.ts` so the MLX
+ * extension routes all three speculative-decoding kinds through the same
+ * local-first resolution and direct `/resolve/main/<file>` download path. This
+ * module only carries the static lookup data; no `huggingface.co/api/...`
+ * calls happen here.
+ */
+import { normalizeBaseId } from './mlx-dflash-registry.js';
+/// Speculators EAGLE-3 checkpoints ship a single-file `model.safetensors`
+/// plus a `config.json` (a sibling `config.py` carries the upstream loader
+/// shim and is NOT needed by mlx-vlm's own `Eagle3Config`). Tokenizer files
+/// belong to the target model.
+const DEFAULT_REQUIRED = ['config.json', 'model.safetensors'];
+/// `model.safetensors.index.json` is grabbed best-effort so a future sharded
+/// EAGLE-3 head keeps working through the local-first path; absence is not
+/// fatal.
+const DEFAULT_OPTIONAL = ['model.safetensors.index.json'];
+/// Snapshot of the canonical Gemma 4 EAGLE-3 speculators published by
+/// `RedHatAI`. Only the two large Gemma 4 targets have an EAGLE-3 head; the
+/// E2B / E4B targets use the Gemma 4 MTP "assistant" drafter instead (see
+/// `mtpRegistry.ts`).
+///
+/// Keys are normalized base-model ids (lowercase, separators preserved as
+/// `-`) derived via `normalizeBaseId`. The trailing `-it` is stripped by
+/// `TRAIL_HINT_RE` in `dflashRegistry.ts`, so the keys here are
+/// `gemma-4-{31b,26b-a4b}` without the `-it` suffix.
+export const STATIC_EAGLE3_MAP = {
+    'gemma-4-31b': {
+        repo: 'RedHatAI/gemma-4-31B-it-speculator.eagle3',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gemma-4-26b-a4b': {
+        repo: 'RedHatAI/gemma-4-26B-A4B-it-speculator.eagle3',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+};
+/**
+ * Resolve an EAGLE-3 draft repo for the given base model.
+ *
+ * Pure / synchronous — relies solely on `STATIC_EAGLE3_MAP` and never touches
+ * the network. Returns `null` when the base model has no canonical EAGLE-3
+ * pairing.
+ */
+export function resolveEagle3Draft(modelId) {
+    /// Reverse lookup: caller already passed a fully-qualified
+    /// `*-speculator.eagle3` id — match against manifest values rather than
+    /// running it through `normalizeBaseId` (which would mangle the
+    /// `-speculator.eagle3` suffix).
+    if (/\.eagle3(?:-|$)/i.test(modelId) || /^redhatai\//i.test(modelId)) {
+        const target = modelId.toLowerCase();
+        for (const manifest of Object.values(STATIC_EAGLE3_MAP)) {
+            if (manifest.repo.toLowerCase() === target) {
+                return {
+                    repo: manifest.repo,
+                    required: manifest.required,
+                    optional: manifest.optional ?? [],
+                };
+            }
+        }
+        return null;
+    }
+    const baseId = normalizeBaseId(modelId);
+    const manifest = STATIC_EAGLE3_MAP[baseId];
+    if (!manifest)
+        return null;
+    return {
+        repo: manifest.repo,
+        required: manifest.required,
+        optional: manifest.optional ?? [],
+    };
+}
+//# sourceMappingURL=mlx-eagle3-registry.js.map

package/dist/speculative/mlx-eagle3-registry.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"mlx-eagle3-registry.js","sourceRoot":"","sources":["../../src/speculative/mlx-eagle3-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;GAaG;AACH,OAAO,EAAgD,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAExG,0EAA0E;AAC1E,2EAA2E;AAC3E,4EAA4E;AAC5E,+BAA+B;AAC/B,MAAM,gBAAgB,GAAG,CAAC,aAAa,EAAE,mBAAmB,CAAC,CAAA;AAE7D,6EAA6E;AAC7E,2EAA2E;AAC3E,UAAU;AACV,MAAM,gBAAgB,GAAa,CAAC,8BAA8B,CAAC,CAAA;AAEnE,sEAAsE;AACtE,4EAA4E;AAC5E,0EAA0E;AAC1E,sBAAsB;AACtB,GAAG;AACH,0EAA0E;AAC1E,yEAAyE;AACzE,gEAAgE;AAChE,qDAAqD;AACrD,MAAM,CAAC,MAAM,iBAAiB,GAAsC;IAClE,aAAa,EAAE;QACb,IAAI,EAAE,2CAA2C;QACjD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,iBAAiB,EAAE;QACjB,IAAI,EAAE,+CAA+C;QACrD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;CACF,CAAA;AAED;;;;;;GAMG;AACH,MAAM,UAAU,kBAAkB,CAAC,OAAe;IAChD,2DAA2D;IAC3D,wEAAwE;IACxE,gEAAgE;IAChE,iCAAiC;IACjC,IAAI,kBAAkB,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,cAAc,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QACrE,MAAM,MAAM,GAAG,OAAO,CAAC,WAAW,EAAE,CAAA;QACpC,KAAK,MAAM,QAAQ,IAAI,MAAM,CAAC,MAAM,CAAC,iBAAiB,CAAC,EAAE,CAAC;YACxD,IAAI,QAAQ,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,MAAM,EAAE,CAAC;gBAC3C,OAAO;oBACL,IAAI,EAAE,QAAQ,CAAC,IAAI;oBACnB,QAAQ,EAAE,QAAQ,CAAC,QAAQ;oBAC3B,QAAQ,EAAE,QAAQ,CAAC,QAAQ,IAAI,EAAE;iBAClC,CAAA;YACH,CAAC;QACH,CAAC;QACD,OAAO,IAAI,CAAA;IACb,CAAC;IAED,MAAM,MAAM,GAAG,eAAe,CAAC,OAAO,CAAC,CAAA;IACvC,MAAM,QAAQ,GAAG,iBAAiB,CAAC,MAAM,CAAC,CAAA;IAC1C,IAAI,CAAC,QAAQ;QAAE,OAAO,IAAI,CAAA;IAE1B,OAAO;QACL,IAAI,EAAE,QAAQ,CAAC,IAAI;QACnB,QAAQ,EAAE,QAAQ,CAAC,QAAQ;QAC3B,QAAQ,EAAE,QAAQ,CAAC,QAAQ,IAAI,EAAE;KAClC,CAAA;AACH,CAAC"}

package/dist/speculative/mlx-mtp-registry.d.ts

@@ -0,0 +1,31 @@
+/**
+ * Ported verbatim from: extensions/mlx-extension/src/mtpRegistry.ts
+ * MTP (Multi-Token Prediction) draft model registry.
+ *
+ * A pure, network-free static manifest mapping base MLX model identifiers to
+ * matching MTP drafters. Three drafter families share the single `mtp`
+ * `--draft-kind` (mlx-vlm auto-detects the exact head from the drafter's
+ * `config.json :: model_type`):
+ *   - Gemma 4 "assistant" heads (`mlx-community/gemma-4-*-it-assistant-bf16`,
+ *     `model_type: gemma4_assistant`) — https://ai.google.dev/gemma/docs/mtp/mtp.
+ *   - Qwen 3.5 / 3.6 MTP heads (`mlx-community/Qwen3.*-MTP-bf16`,
+ *     `model_type: qwen3_5_mtp`) — split out of the UDT target weights and
+ *     republished as standalone drafter repos by `mlx-community`.
+ *   - DeepSeek V4 Flash MTP head (`mlx-community/DeepSeek-V4-Flash-MTP-bf16`,
+ *     `model_type: deepseek_v4_mtp`).
+ *
+ * Mirrors the shape of `dflashRegistry.ts` / `eagle3Registry.ts` so the MLX
+ * extension can route every speculative-decoding kind through the same
+ * local-first resolution and direct-download path. This module only carries
+ * the static lookup data; no `huggingface.co/api/...` calls happen here.
+ */
+import { type DraftRepoManifest, type DraftResolution } from './mlx-dflash-registry.js';
+export declare const STATIC_MTP_MAP: Record<string, DraftRepoManifest>;
+/**
+ * Resolve a Gemma 4 MTP assistant repo for the given base model.
+ *
+ * Pure / synchronous — relies solely on `STATIC_MTP_MAP` and never touches
+ * the network. Returns `null` when the base model has no canonical MTP
+ * pairing.
+ */
+export declare function resolveMtpDraft(modelId: string): DraftResolution | null;

package/dist/speculative/mlx-mtp-registry.js

@@ -0,0 +1,139 @@
+/**
+ * Ported verbatim from: extensions/mlx-extension/src/mtpRegistry.ts
+ * MTP (Multi-Token Prediction) draft model registry.
+ *
+ * A pure, network-free static manifest mapping base MLX model identifiers to
+ * matching MTP drafters. Three drafter families share the single `mtp`
+ * `--draft-kind` (mlx-vlm auto-detects the exact head from the drafter's
+ * `config.json :: model_type`):
+ *   - Gemma 4 "assistant" heads (`mlx-community/gemma-4-*-it-assistant-bf16`,
+ *     `model_type: gemma4_assistant`) — https://ai.google.dev/gemma/docs/mtp/mtp.
+ *   - Qwen 3.5 / 3.6 MTP heads (`mlx-community/Qwen3.*-MTP-bf16`,
+ *     `model_type: qwen3_5_mtp`) — split out of the UDT target weights and
+ *     republished as standalone drafter repos by `mlx-community`.
+ *   - DeepSeek V4 Flash MTP head (`mlx-community/DeepSeek-V4-Flash-MTP-bf16`,
+ *     `model_type: deepseek_v4_mtp`).
+ *
+ * Mirrors the shape of `dflashRegistry.ts` / `eagle3Registry.ts` so the MLX
+ * extension can route every speculative-decoding kind through the same
+ * local-first resolution and direct-download path. This module only carries
+ * the static lookup data; no `huggingface.co/api/...` calls happen here.
+ */
+import { normalizeBaseId } from './mlx-dflash-registry.js';
+/// Gemma 4 assistant drafters published by `mlx-community` ship as a
+/// single-file `model.safetensors` plus `config.json`. Sharded layouts
+/// (rare for the 4-layer assistant) need a per-repo override.
+const DEFAULT_REQUIRED = ['config.json', 'model.safetensors'];
+/// `model.safetensors.index.json` is grabbed best-effort so sharded
+/// drafters keep working through the local-first path; absence is not
+/// fatal.
+const DEFAULT_OPTIONAL = ['model.safetensors.index.json'];
+/// Snapshot of the four canonical Gemma 4 target ↔ assistant pairings
+/// from `mlx_vlm/speculative/drafters/gemma4_assistant/README.md`.
+///
+/// Keys are normalized base-model ids (lowercase, separators preserved as
+/// `-`) derived via `normalizeBaseId`. The trailing `-it` is stripped by
+/// `TRAIL_HINT_RE` in `dflashRegistry.ts`, so the keys here are
+/// `gemma-4-{e2b,e4b,26b-a4b,31b}` without the `-it` suffix.
+export const STATIC_MTP_MAP = {
+    'gemma-4-e2b': {
+        repo: 'mlx-community/gemma-4-E2B-it-assistant-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gemma-4-e4b': {
+        repo: 'mlx-community/gemma-4-E4B-it-assistant-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gemma-4-26b-a4b': {
+        repo: 'mlx-community/gemma-4-26B-A4B-it-assistant-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'gemma-4-31b': {
+        repo: 'mlx-community/gemma-4-31B-it-assistant-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    /// Qwen 3.5 / 3.6 MTP heads (`model_type: qwen3_5_mtp`). The drafter is the
+    /// MTP module split out of the corresponding UDT target and republished as a
+    /// standalone single-file `model.safetensors` + `config.json` repo. Verified
+    /// against the `mlx-community` MTP collection. The tokenizer comes from the
+    /// target model, so only the two weight/config files are required.
+    'qwen3.5-4b': {
+        repo: 'mlx-community/Qwen3.5-4B-MTP-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.5-9b': {
+        repo: 'mlx-community/Qwen3.5-9B-MTP-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.6-27b': {
+        repo: 'mlx-community/Qwen3.6-27B-MTP-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    'qwen3.6-35b-a3b': {
+        repo: 'mlx-community/Qwen3.6-35B-A3B-MTP-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+    /// DeepSeek V4 Flash MTP head (`model_type: deepseek_v4_mtp`). Same
+    /// split-and-republish story as the Qwen heads.
+    'deepseek-v4-flash': {
+        repo: 'mlx-community/DeepSeek-V4-Flash-MTP-bf16',
+        required: DEFAULT_REQUIRED,
+        optional: DEFAULT_OPTIONAL,
+    },
+};
+/**
+ * Resolve a Gemma 4 MTP assistant repo for the given base model.
+ *
+ * Pure / synchronous — relies solely on `STATIC_MTP_MAP` and never touches
+ * the network. Returns `null` when the base model has no canonical MTP
+ * pairing.
+ */
+export function resolveMtpDraft(modelId) {
+    /// Reverse lookup: caller already passed a fully-qualified drafter id
+    /// (Gemma `*-it-assistant-*` or a `*-MTP-*` head) — match against manifest
+    /// values rather than running it through `normalizeBaseId` (which would
+    /// strip the `-bf16` suffix and try to look it up as a target).
+    if (/-assistant(?:-|$)/i.test(modelId) || /-mtp(?:-|$)/i.test(modelId)) {
+        const target = modelId.toLowerCase();
+        for (const manifest of Object.values(STATIC_MTP_MAP)) {
+            if (manifest.repo.toLowerCase() === target) {
+                return {
+                    repo: manifest.repo,
+                    required: manifest.required,
+                    optional: manifest.optional ?? [],
+                };
+            }
+        }
+        return null;
+    }
+    /// Quantized targets (e.g. `gemma-4-E4B-it-4bit`) are now allowed.
+    /// The original block was a precaution against biased strict-equality
+    /// verification when the bf16 drafter was paired with a quantized
+    /// target — but the mlx-vlm server forces ``temp=0`` in
+    /// ``_run_speculative`` (greedy verify), so a quantization mismatch
+    /// only reduces acceptance rate (drafter argmax may diverge from
+    /// target argmax), it does not corrupt output. The long-prompt SWA
+    /// mask bug that previously turned this into "Just/Just/Just" cascades
+    /// is fixed upstream by Blaizzy/mlx-vlm#1139 (commit ``2bc33a6``).
+    /// ``normalizeBaseId`` strips the quant + ``-it`` suffixes so e.g.
+    /// ``gemma-4-E4B-it-4bit`` -> ``gemma-4-e4b`` and resolves to the
+    /// same bf16 drafter as the bf16 target.
+    const baseId = normalizeBaseId(modelId);
+    const manifest = STATIC_MTP_MAP[baseId];
+    if (!manifest)
+        return null;
+    return {
+        repo: manifest.repo,
+        required: manifest.required,
+        optional: manifest.optional ?? [],
+    };
+}
+//# sourceMappingURL=mlx-mtp-registry.js.map

package/dist/speculative/mlx-mtp-registry.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"mlx-mtp-registry.js","sourceRoot":"","sources":["../../src/speculative/mlx-mtp-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,OAAO,EAAgD,eAAe,EAAE,MAAM,0BAA0B,CAAA;AAExG,qEAAqE;AACrE,uEAAuE;AACvE,8DAA8D;AAC9D,MAAM,gBAAgB,GAAG,CAAC,aAAa,EAAE,mBAAmB,CAAC,CAAA;AAE7D,oEAAoE;AACpE,sEAAsE;AACtE,UAAU;AACV,MAAM,gBAAgB,GAAa,CAAC,8BAA8B,CAAC,CAAA;AAEnE,sEAAsE;AACtE,mEAAmE;AACnE,GAAG;AACH,0EAA0E;AAC1E,yEAAyE;AACzE,gEAAgE;AAChE,6DAA6D;AAC7D,MAAM,CAAC,MAAM,cAAc,GAAsC;IAC/D,aAAa,EAAE;QACb,IAAI,EAAE,6CAA6C;QACnD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,aAAa,EAAE;QACb,IAAI,EAAE,6CAA6C;QACnD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,iBAAiB,EAAE;QACjB,IAAI,EAAE,iDAAiD;QACvD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,aAAa,EAAE;QACb,IAAI,EAAE,6CAA6C;QACnD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IAED,4EAA4E;IAC5E,6EAA6E;IAC7E,6EAA6E;IAC7E,4EAA4E;IAC5E,mEAAmE;IACnE,YAAY,EAAE;QACZ,IAAI,EAAE,mCAAmC;QACzC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,YAAY,EAAE;QACZ,IAAI,EAAE,mCAAmC;QACzC,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,aAAa,EAAE;QACb,IAAI,EAAE,oCAAoC;QAC1C,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IACD,iBAAiB,EAAE;QACjB,IAAI,EAAE,wCAAwC;QAC9C,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;IAED,oEAAoE;IACpE,gDAAgD;IAChD,mBAAmB,EAAE;QACnB,IAAI,EAAE,0CAA0C;QAChD,QAAQ,EAAE,gBAAgB;QAC1B,QAAQ,EAAE,gBAAgB;KAC3B;CACF,CAAA;AAED;;;;;;GAMG;AACH,MAAM,UAAU,eAAe,CAAC,OAAe;IAC7C,sEAAsE;IACtE,2EAA2E;IAC3E,wEAAwE;IACxE,gEAAgE;IAChE,IAAI,oBAAoB,CAAC,IAAI,CAAC,OAAO,CAAC,IAAI,cAAc,CAAC,IAAI,CAAC,OAAO,CAAC,EAAE,CAAC;QACvE,MAAM,MAAM,GAAG,OAAO,CAAC,WAAW,EAAE,CAAA;QACpC,KAAK,MAAM,QAAQ,IAAI,MAAM,CAAC,MAAM,CAAC,cAAc,CAAC,EAAE,CAAC;YACrD,IAAI,QAAQ,CAAC,IAAI,CAAC,WAAW,EAAE,KAAK,MAAM,EAAE,CAAC;gBAC3C,OAAO;oBACL,IAAI,EAAE,QAAQ,CAAC,IAAI;oBACnB,QAAQ,EAAE,QAAQ,CAAC,QAAQ;oBAC3B,QAAQ,EAAE,QAAQ,CAAC,QAAQ,IAAI,EAAE;iBAClC,CAAA;YACH,CAAC;QACH,CAAC;QACD,OAAO,IAAI,CAAA;IACb,CAAC;IAED,mEAAmE;IACnE,sEAAsE;IACtE,kEAAkE;IAClE,wDAAwD;IACxD,oEAAoE;IACpE,iEAAiE;IACjE,mEAAmE;IACnE,uEAAuE;IACvE,mEAAmE;IACnE,mEAAmE;IACnE,kEAAkE;IAClE,yCAAyC;IACzC,MAAM,MAAM,GAAG,eAAe,CAAC,OAAO,CAAC,CAAA;IACvC,MAAM,QAAQ,GAAG,cAAc,CAAC,MAAM,CAAC,CAAA;IACvC,IAAI,CAAC,QAAQ;QAAE,OAAO,IAAI,CAAA;IAE1B,OAAO;QACL,IAAI,EAAE,QAAQ,CAAC,IAAI;QACnB,QAAQ,EAAE,QAAQ,CAAC,QAAQ;QAC3B,QAAQ,EAAE,QAAQ,CAAC,QAAQ,IAAI,EAAE;KAClC,CAAA;AACH,CAAC"}

package/dist/speculative/transcription-registry.d.ts

@@ -0,0 +1,72 @@
+/**
+ * Ported verbatim from: extensions/llamacpp-upstream-extension/src/transcriptionRegistry.ts
+ * The voice model.
+ *
+ * Voice input transcribes on the same bundled llama.cpp engine that runs chat:
+ * its `libmtmd` already carries the `voxtral` audio projector, and `llama-server`
+ * already serves an OpenAI-compatible `/v1/audio/transcriptions`. So the model
+ * is the only thing that has to be fetched, and it goes through the ordinary
+ * download pipeline like any other GGUF.
+ *
+ * One model, deliberately. A tier picker would be a second thing to explain in
+ * the setup wizard, and the sizes here are already dominated by the weights.
+ */
+/** Model id, and the llama-server alias (`-a`) it is loaded under. */
+export declare const TRANSCRIPTION_MODEL_ID = "ggml-org/Voxtral-Mini-3B-2507-Q4_K_M";
+export declare const TRANSCRIPTION_MODEL_HF_REPO = "ggml-org/Voxtral-Mini-3B-2507-GGUF";
+export declare const TRANSCRIPTION_MODEL_URL = "https://huggingface.co/ggml-org/Voxtral-Mini-3B-2507-GGUF/resolve/main/Voxtral-Mini-3B-2507-Q4_K_M.gguf";
+export declare const TRANSCRIPTION_MMPROJ_URL = "https://huggingface.co/ggml-org/Voxtral-Mini-3B-2507-GGUF/resolve/main/mmproj-Voxtral-Mini-3B-2507-Q8_0.gguf";
+/** Exact sizes from the Hugging Face API, so the progress bar has a real total. */
+export declare const TRANSCRIPTION_MODEL_BYTES = 2473001920;
+export declare const TRANSCRIPTION_MMPROJ_BYTES = 715714080;
+export declare const TRANSCRIPTION_TOTAL_BYTES: number;
+/**
+ * Chat template the voice server is loaded with.
+ *
+ * Two things make this mandatory rather than a nicety.
+ *
+ * 1. The template embedded in `ggml-org/Voxtral-Mini-3B-2507-Q4_K_M.gguf` is an
+ *    Unsloth *Devstral* template. It defaults every conversation to a ~1200
+ *    token "You are Devstral, a helpful agentic model … using the OpenHands
+ *    scaffold" system prompt, which llama.cpp then prepends to every
+ *    `/v1/audio/transcriptions` request (that endpoint is a chat completion
+ *    underneath). The model duly behaves like a chat agent: it *hears* the
+ *    audio and answers it — "I'm unable to transcribe audio directly. However,
+ *    I can guide you on how to use Descript, Rev, Amazon Transcribe…" — instead
+ *    of transcribing. Every phrase then trips the plugin's plausibility guard
+ *    and is dropped, so dictation inserts nothing at all.
+ *
+ * 2. Voxtral has a dedicated transcription mode, entered by ending the user
+ *    turn with the `[TRANSCRIBE]` token instead of closing it with `[/INST]`.
+ *    Closing the turn normally — with or without a "transcribe this" prompt —
+ *    leaves the model in conversational mode, where it answers, refuses, or
+ *    returns nothing depending on the clip.
+ *
+ * So: no system prompt, and `[INST]` … `[TRANSCRIBE]` around whatever the
+ * endpoint puts in the user turn (the audio marker, then the caller's `prompt`,
+ * which is where the `lang:xx` directive lands). This server only ever serves
+ * transcription, so treating every message as a transcription turn is honest.
+ *
+ * `String.raw` for consistency with `chatTemplateOverrides.ts`: minja, not JS,
+ * interprets any escapes inside the Jinja source.
+ */
+export declare const VOXTRAL_TRANSCRIPTION_CHAT_TEMPLATE: string;
+/**
+ * Load overrides for the voice server.
+ *
+ * Deliberately small and deterministic: one 30 s `libmtmd` audio window is a few
+ * hundred tokens, a transcript is never long, and this process runs *alongside*
+ * the user's chat model — so it must not size itself to the free VRAM the way a
+ * chat load does.
+ */
+export declare const TRANSCRIPTION_LOAD_OVERRIDES: {
+    readonly fit: false;
+    readonly ctx_size: 4096;
+    readonly n_predict: 512;
+    readonly parallel: 1;
+    readonly cont_batching: false;
+    readonly chat_template: string;
+    readonly extra_args: "--no-warmup";
+};
+/** Unload the voice model after this long without a phrase. */
+export declare const TRANSCRIPTION_IDLE_UNLOAD_MS: number;