@atomic-chat/core 0.5.1

package/dist/speculative/dflash-registry.js

@@ -0,0 +1,258 @@
+/**
+ * Ported verbatim from: extensions/llamacpp-upstream-extension/src/dflashRegistry.ts
+ * @file Static registry mapping a DFlash *target* model to the separate DFlash
+ * *draft* GGUF it needs for upstream llama.cpp multi-layer draft speculative
+ * decoding (`--model-draft <draft> --spec-type draft-dflash`, PR
+ * ggml-org/llama.cpp#22105, commit d1b34251b, first tagged release `b9831`).
+ *
+ * DFlash ships the draft as a second, small GGUF that must be downloaded next
+ * to the target model (the same way `mmproj` and Gemma 4 MTP heads are handled
+ * — see gemmaMtpRegistry.ts). The upstream MLX DFlash drafts are published as
+ * safetensors and would need conversion; these entries instead point at
+ * GGUF repos built for mainline llama.cpp so the file can be consumed
+ * directly. Fork-only `dflash-draft` / `qwen35-dflash-draft` architectures
+ * are deliberately excluded.
+ *
+ * Draft files were verified against the live Hugging Face API (sha256 + size
+ * pinned so the downloader validates them). Atomic Chat's verified Q8_0
+ * conversions are the default for every target-model quantization when
+ * available; compatible community conversions provide the remaining draft
+ * choices and cover families without an Atomic Chat conversion.
+ */
+function hasQwenFamily(normalizedId, family) {
+    const compact = normalizedId.replace(/[\s._-]+/g, '');
+    return compact.includes(`qwen${family}`);
+}
+function compactId(normalizedId) {
+    return normalizedId.replace(/[\s._-]+/g, '');
+}
+function isBaseQwen3(normalizedId) {
+    const compact = compactId(normalizedId);
+    return (compact.includes('qwen3') &&
+        !compact.includes('qwen35') &&
+        !compact.includes('qwen36') &&
+        !compact.includes('qwen3coder'));
+}
+function atomicQ8Draft(repo, draftFilename, draftSha256, draftSize) {
+    return {
+        quant: 'Q8_0',
+        repo,
+        draftFilename,
+        draftSha256,
+        draftSize,
+    };
+}
+const REGISTRY = [
+    {
+        matches: (id) => compactId(id).includes('qwen3codernext'),
+        drafts: [
+            atomicQ8Draft('AtomicChat/Qwen3-Coder-Next-DFlash-GGUF', 'Qwen3-Coder-Next-DFlash.Q8_0.gguf', '139081af2f9f634b3b8bce8f75a0449f92bcb75ae7a2e32abaf1bcdf6f4a7952', 509674720),
+        ],
+    },
+    {
+        matches: (id) => {
+            const compact = compactId(id);
+            return compact.includes('qwen3coder') && compact.includes('30b') && compact.includes('a3b');
+        },
+        drafts: [
+            atomicQ8Draft('AtomicChat/Qwen3-Coder-30B-A3B-DFlash-GGUF', 'Qwen3-Coder-30B-A3B-DFlash.Q8_0.gguf', 'abd0e078f1e5683ba83c4ced92836b136d9e2e6b2626715b699088fd5c3a51fc', 509674848),
+        ],
+    },
+    {
+        matches: (id) => compactId(id).includes('qwen3527b'),
+        drafts: [
+            atomicQ8Draft('AtomicChat/Qwen3.5-27B-DFlash-GGUF', 'Qwen3.5-27B-DFlash.Q8_0.gguf', 'e24126896291f03a8dd6da7da8678ea4d7b019d6fc3719ecf6136e392f361645', 2272886944),
+        ],
+    },
+    {
+        matches: (id) => compactId(id).includes('qwen354b'),
+        drafts: [
+            atomicQ8Draft('AtomicChat/Qwen3.5-4B-DFlash-GGUF', 'Qwen3.5-4B-DFlash.Q8_0.gguf', 'f0689eadab0d46468bc2629bb2706f4aba7f794b63e1c9f6eb291269d572a52b', 685133984),
+        ],
+    },
+    {
+        matches: (id) => isBaseQwen3(id) && compactId(id).includes('qwen38b'),
+        drafts: [
+            atomicQ8Draft('AtomicChat/Qwen3-8B-DFlash-GGUF', 'Qwen3-8B-DFlash.Q8_0.gguf', '5be4f6b1bfd5c2c1aa753d4c03e30700114654fefbbf29f02257ef37adb00bf0', 1120250336),
+        ],
+    },
+    {
+        matches: (id) => isBaseQwen3(id) && compactId(id).includes('qwen34b'),
+        drafts: [
+            atomicQ8Draft('AtomicChat/Qwen3-4B-DFlash-GGUF', 'Qwen3-4B-DFlash.Q8_0.gguf', '2913a8a5619b30233c0745ccd58de8fbe144b831b54770eb04d4a371221ca273', 577047008),
+        ],
+    },
+    {
+        // Qwen3.5-9B (dense). e.g. Qwen3.5-9B-Instruct-Q4_K_M, qwen3.5-9b-...
+        matches: (id) => hasQwenFamily(id, '35') && id.includes('9b'),
+        drafts: [
+            {
+                quant: 'Q4_K_M',
+                repo: 'onion515/Qwen3.5-9B-DFlash-GGUF',
+                draftFilename: 'qwen3.5-9b-dflash-Q4_K_M.gguf',
+                draftSha256: 'd9999163059b9ae0e87b0f503d76451d922b5483c981d9afae250d82804b5f0d',
+                draftSize: 765960384,
+            },
+            {
+                quant: 'Q5_K_M',
+                repo: 'onion515/Qwen3.5-9B-DFlash-GGUF',
+                draftFilename: 'qwen3.5-9b-dflash-Q5_K_M.gguf',
+                draftSha256: 'aa7521c7c210d1fd56e4e7520b6cbb5df6d0c7f813c8ddddff0861c4f8663045',
+                draftSize: 913809600,
+            },
+            {
+                quant: 'Q6_K',
+                repo: 'onion515/Qwen3.5-9B-DFlash-GGUF',
+                draftFilename: 'qwen3.5-9b-dflash-Q6_K.gguf',
+                draftSha256: '4c9a1e123944e008ebc55677ba9d51a63dc81fcdb80e44a9a3e405e57563f8b0',
+                draftSize: 1070899392,
+            },
+            {
+                quant: 'Q8_0',
+                repo: 'AtomicChat/Qwen3.5-9B-DFlash-GGUF',
+                draftFilename: 'Qwen3.5-9B-DFlash.Q8_0.gguf',
+                draftSha256: '27b9d18e605aea9c50ef506e4b63921e0dc624cc91c63c60617a8295f153d436',
+                draftSize: 1383768224,
+            },
+            {
+                quant: 'BF16',
+                repo: 'onion515/Qwen3.5-9B-DFlash-GGUF',
+                draftFilename: 'qwen3.5-9b-dflash-bf16.gguf',
+                draftSha256: '9c1f55bc887fc5efb565748580a6526a5c1ef9bdd1893002ccdca7be4d1213a5',
+                draftSize: 2594873536,
+            },
+        ],
+    },
+    {
+        // Qwen3.6-35B-A3B (MoE). Require "35b" + "a3b" so only the MoE matches;
+        // checked before the 27B entry (both are 3.6 but the size tokens disjoin).
+        matches: (id) => hasQwenFamily(id, '36') && id.includes('35b') && id.includes('a3b'),
+        drafts: [
+            {
+                quant: 'IQ4_XS',
+                repo: 'williamliao/Qwen3.6-35B-A3B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-35B-A3B-DFlash-IQ4_XS.gguf',
+                draftSha256: '13866fb2301f34cc31b5b8cda9ee1059bb7468ba5ef636be7e5e98d6fd371cdc',
+                draftSize: 218030368,
+            },
+            {
+                quant: 'Q4_K_M',
+                repo: 'williamliao/Qwen3.6-35B-A3B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-35B-A3B-DFlash-Q4_K_M.gguf',
+                draftSha256: '9f8a75a850d9e207ca8a8de2cfa691aef61c63895d224defc00abfc34d5779bd',
+                draftSize: 235692320,
+            },
+            {
+                quant: 'Q5_K_M',
+                repo: 'williamliao/Qwen3.6-35B-A3B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-35B-A3B-DFlash-Q5_K_M.gguf',
+                draftSha256: '397df52ae2bb4afd88c1272a122001dc56ad2546153038aee468f6371b0045e9',
+                draftSize: 280256800,
+            },
+            {
+                quant: 'Q6_K',
+                repo: 'williamliao/Qwen3.6-35B-A3B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-35B-A3B-DFlash-Q6_K.gguf',
+                draftSha256: 'c1fe996a6f198c867b3d422b9bcf5c385f1dcc29b877628ccb2af1f9e278f649',
+                draftSize: 327606560,
+            },
+            {
+                quant: 'Q8_0',
+                repo: 'williamliao/Qwen3.6-35B-A3B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-35B-A3B-DFlash-Q8_0.gguf',
+                draftSha256: 'fc79ee2c84688674e4f134d5797a6baee4884167659a6eec8866c7789ee1318b',
+                draftSize: 421060896,
+            },
+            {
+                quant: 'F16',
+                repo: 'williamliao/Qwen3.6-35B-A3B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-35B-A3B-DFlash-F16.gguf',
+                draftSha256: 'bd5385344d049ae9770972b9b694dfcc18005d7f4b9dfabde04f9b30ce21b631',
+                draftSize: 782819616,
+            },
+        ],
+    },
+    {
+        // Qwen3.6-27B (dense).
+        matches: (id) => hasQwenFamily(id, '36') && id.includes('27b'),
+        drafts: [
+            {
+                quant: 'IQ4_XS',
+                repo: 'williamliao/qwen3.6-27B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-27B-DFlash-IQ4_XS.gguf',
+                draftSha256: '23861020ce5313567cb8e8fe56ce01b5f8e467df7d23ebecfe17fd19136f97d7',
+                draftSize: 934435136,
+            },
+            {
+                quant: 'Q4_K_M',
+                repo: 'williamliao/qwen3.6-27B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-27B-DFlash-Q4_K_M.gguf',
+                draftSha256: '47901c650191c72a73a3d69e1f7cfa5744bce2bb46f82e6ebeaf729f90f5c3a3',
+                draftSize: 1033066816,
+            },
+            {
+                quant: 'Q5_K_M',
+                repo: 'williamliao/qwen3.6-27B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-27B-DFlash-Q5_K_M.gguf',
+                draftSha256: 'ece22ee82e7a9acbd7b9617b4b252b029e3f3cd8a7a6a398362ebdd5b56b520f',
+                draftSize: 1225742656,
+            },
+            {
+                quant: 'Q6_K',
+                repo: 'williamliao/qwen3.6-27B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-27B-DFlash-Q6_K.gguf',
+                draftSha256: 'ff4197923ce8d9c56f2dc745cfa95a1820d8d63a51aab66d4ad88d0827b6d8a4',
+                draftSize: 1430460736,
+            },
+            {
+                quant: 'Q8_0',
+                repo: 'AtomicChat/Qwen3.6-27B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-27B-DFlash.Q8_0.gguf',
+                draftSha256: 'b9be52167d0b4f5cbaffba6658bcb20dd814ad4e5846532353c165ccf97e47a1',
+                draftSize: 1849481472,
+            },
+            {
+                quant: 'F16',
+                repo: 'williamliao/qwen3.6-27B-DFlash-GGUF',
+                draftFilename: 'Qwen3.6-27B-DFlash-F16.gguf',
+                draftSha256: '6bbbca1f2391ee5adf76e01ee9b7614787dbf0cecc5b67b45f5a20003915c081',
+                draftSize: 3471497536,
+            },
+        ],
+    },
+];
+/**
+ * Every draft this registry downloads — the files the app itself puts on disk
+ * next to a target. The web app's model-file filter is tested against them so
+ * a local scan never offers one back as a model (ATO-523).
+ */
+export const DFLASH_DRAFT_FILENAMES = REGISTRY.flatMap((entry) => entry.drafts.map((draft) => draft.draftFilename));
+export const DEFAULT_DFLASH_DRAFT_QUANT = 'Q8_0';
+function normalizeId(modelId) {
+    return modelId.toLowerCase();
+}
+/**
+ * Resolve the DFlash draft for a target model id, or `null` if the model is
+ * not a DFlash-capable target in the registry.
+ */
+export function listDflashDrafts(modelId) {
+    const normalized = normalizeId(modelId);
+    const entry = REGISTRY.find((e) => e.matches(normalized));
+    return entry?.drafts.map((draft) => ({ ...draft })) ?? [];
+}
+export function resolveDflashDraft(modelId, quant = DEFAULT_DFLASH_DRAFT_QUANT) {
+    const drafts = listDflashDrafts(modelId);
+    const exact = drafts.find((draft) => draft.quant === quant);
+    if (exact)
+        return exact;
+    return quant === DEFAULT_DFLASH_DRAFT_QUANT ? (drafts[0] ?? null) : null;
+}
+/** Whether the given model id is a DFlash-capable target. */
+export function checkDflashSupport(modelId) {
+    return listDflashDrafts(modelId).length > 0;
+}
+/** Build the Hugging Face resolve URL for a DFlash draft. */
+export function dflashDraftUrl(draft) {
+    return `https://huggingface.co/${draft.repo}/resolve/main/${draft.draftFilename}`;
+}
+//# sourceMappingURL=dflash-registry.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"dflash-registry.js","sourceRoot":"","sources":["../../src/speculative/dflash-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;GAoBG;AAyBH,SAAS,aAAa,CAAC,YAAoB,EAAE,MAAmB;IAC9D,MAAM,OAAO,GAAG,YAAY,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC,CAAA;IACrD,OAAO,OAAO,CAAC,QAAQ,CAAC,OAAO,MAAM,EAAE,CAAC,CAAA;AAC1C,CAAC;AAED,SAAS,SAAS,CAAC,YAAoB;IACrC,OAAO,YAAY,CAAC,OAAO,CAAC,WAAW,EAAE,EAAE,CAAC,CAAA;AAC9C,CAAC;AAED,SAAS,WAAW,CAAC,YAAoB;IACvC,MAAM,OAAO,GAAG,SAAS,CAAC,YAAY,CAAC,CAAA;IACvC,OAAO,CACL,OAAO,CAAC,QAAQ,CAAC,OAAO,CAAC;QACzB,CAAC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC;QAC3B,CAAC,OAAO,CAAC,QAAQ,CAAC,QAAQ,CAAC;QAC3B,CAAC,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAC,CAChC,CAAA;AACH,CAAC;AAED,SAAS,aAAa,CACpB,IAAY,EACZ,aAAqB,EACrB,WAAmB,EACnB,SAAiB;IAEjB,OAAO;QACL,KAAK,EAAE,MAAM;QACb,IAAI;QACJ,aAAa;QACb,WAAW;QACX,SAAS;KACV,CAAA;AACH,CAAC;AAED,MAAM,QAAQ,GAAmC;IAC/C;QACE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,gBAAgB,CAAC;QACzD,MAAM,EAAE;YACN,aAAa,CACX,yCAAyC,EACzC,mCAAmC,EACnC,kEAAkE,EAClE,SAAS,CACV;SACF;KACF;IACD;QACE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE;YACd,MAAM,OAAO,GAAG,SAAS,CAAC,EAAE,CAAC,CAAA;YAC7B,OAAO,OAAO,CAAC,QAAQ,CAAC,YAAY,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,OAAO,CAAC,QAAQ,CAAC,KAAK,CAAC,CAAA;QAC7F,CAAC;QACD,MAAM,EAAE;YACN,aAAa,CACX,4CAA4C,EAC5C,sCAAsC,EACtC,kEAAkE,EAClE,SAAS,CACV;SACF;KACF;IACD;QACE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,WAAW,CAAC;QACpD,MAAM,EAAE;YACN,aAAa,CACX,oCAAoC,EACpC,8BAA8B,EAC9B,kEAAkE,EAClE,UAAU,CACX;SACF;KACF;IACD;QACE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,SAAS,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,UAAU,CAAC;QACnD,MAAM,EAAE;YACN,aAAa,CACX,mCAAmC,EACnC,6BAA6B,EAC7B,kEAAkE,EAClE,SAAS,CACV;SACF;KACF;IACD;QACE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,WAAW,CAAC,EAAE,CAAC,IAAI,SAAS,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC;QACrE,MAAM,EAAE;YACN,aAAa,CACX,iCAAiC,EACjC,2BAA2B,EAC3B,kEAAkE,EAClE,UAAU,CACX;SACF;KACF;IACD;QACE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,WAAW,CAAC,EAAE,CAAC,IAAI,SAAS,CAAC,EAAE,CAAC,CAAC,QAAQ,CAAC,SAAS,CAAC;QACrE,MAAM,EAAE;YACN,aAAa,CACX,iCAAiC,EACjC,2BAA2B,EAC3B,kEAAkE,EAClE,SAAS,CACV;SACF;KACF;IACD;QACE,sEAAsE;QACtE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,aAAa,CAAC,EAAE,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,IAAI,CAAC;QAC7D,MAAM,EAAE;YACN;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,iCAAiC;gBACvC,aAAa,EAAE,+BAA+B;gBAC9C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,iCAAiC;gBACvC,aAAa,EAAE,+BAA+B;gBAC9C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,iCAAiC;gBACvC,aAAa,EAAE,6BAA6B;gBAC5C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,mCAAmC;gBACzC,aAAa,EAAE,6BAA6B;gBAC5C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,iCAAiC;gBACvC,aAAa,EAAE,6BAA6B;gBAC5C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;SACF;KACF;IACD;QACE,wEAAwE;QACxE,2EAA2E;QAC3E,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,aAAa,CAAC,EAAE,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC;QACpF,MAAM,EAAE;YACN;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,yCAAyC;gBAC/C,aAAa,EAAE,oCAAoC;gBACnD,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,yCAAyC;gBAC/C,aAAa,EAAE,oCAAoC;gBACnD,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,yCAAyC;gBAC/C,aAAa,EAAE,oCAAoC;gBACnD,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,yCAAyC;gBAC/C,aAAa,EAAE,kCAAkC;gBACjD,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,yCAAyC;gBAC/C,aAAa,EAAE,kCAAkC;gBACjD,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,KAAK;gBACZ,IAAI,EAAE,yCAAyC;gBAC/C,aAAa,EAAE,iCAAiC;gBAChD,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;SACF;KACF;IACD;QACE,uBAAuB;QACvB,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,aAAa,CAAC,EAAE,EAAE,IAAI,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC;QAC9D,MAAM,EAAE;YACN;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,qCAAqC;gBAC3C,aAAa,EAAE,gCAAgC;gBAC/C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,SAAS;aACrB;YACD;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,qCAAqC;gBAC3C,aAAa,EAAE,gCAAgC;gBAC/C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;YACD;gBACE,KAAK,EAAE,QAAQ;gBACf,IAAI,EAAE,qCAAqC;gBAC3C,aAAa,EAAE,gCAAgC;gBAC/C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,qCAAqC;gBAC3C,aAAa,EAAE,8BAA8B;gBAC7C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;YACD;gBACE,KAAK,EAAE,MAAM;gBACb,IAAI,EAAE,oCAAoC;gBAC1C,aAAa,EAAE,8BAA8B;gBAC7C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;YACD;gBACE,KAAK,EAAE,KAAK;gBACZ,IAAI,EAAE,qCAAqC;gBAC3C,aAAa,EAAE,6BAA6B;gBAC5C,WAAW,EAAE,kEAAkE;gBAC/E,SAAS,EAAE,UAAU;aACtB;SACF;KACF;CACF,CAAA;AAED;;;;GAIG;AACH,MAAM,CAAC,MAAM,sBAAsB,GAAsB,QAAQ,CAAC,OAAO,CAAC,CAAC,KAAK,EAAE,EAAE,CAClF,KAAK,CAAC,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,aAAa,CAAC,CACjD,CAAA;AAED,MAAM,CAAC,MAAM,0BAA0B,GAAG,MAAM,CAAA;AAEhD,SAAS,WAAW,CAAC,OAAe;IAClC,OAAO,OAAO,CAAC,WAAW,EAAE,CAAA;AAC9B,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,gBAAgB,CAAC,OAAe;IAC9C,MAAM,UAAU,GAAG,WAAW,CAAC,OAAO,CAAC,CAAA;IACvC,MAAM,KAAK,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAA;IACzD,OAAO,KAAK,EAAE,MAAM,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,CAAC,EAAE,GAAG,KAAK,EAAE,CAAC,CAAC,IAAI,EAAE,CAAA;AAC3D,CAAC;AAED,MAAM,UAAU,kBAAkB,CAAC,OAAe,EAAE,KAAK,GAAG,0BAA0B;IACpF,MAAM,MAAM,GAAG,gBAAgB,CAAC,OAAO,CAAC,CAAA;IACxC,MAAM,KAAK,GAAG,MAAM,CAAC,IAAI,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,KAAK,KAAK,KAAK,CAAC,CAAA;IAC3D,IAAI,KAAK;QAAE,OAAO,KAAK,CAAA;IACvB,OAAO,KAAK,KAAK,0BAA0B,CAAC,CAAC,CAAC,CAAC,MAAM,CAAC,CAAC,CAAC,IAAI,IAAI,CAAC,CAAC,CAAC,CAAC,IAAI,CAAA;AAC1E,CAAC;AAED,6DAA6D;AAC7D,MAAM,UAAU,kBAAkB,CAAC,OAAe;IAChD,OAAO,gBAAgB,CAAC,OAAO,CAAC,CAAC,MAAM,GAAG,CAAC,CAAA;AAC7C,CAAC;AAED,6DAA6D;AAC7D,MAAM,UAAU,cAAc,CAAC,KAAkB;IAC/C,OAAO,0BAA0B,KAAK,CAAC,IAAI,iBAAiB,KAAK,CAAC,aAAa,EAAE,CAAA;AACnF,CAAC"}

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

@@ -0,0 +1,48 @@
+/**
+ * Ported verbatim from: extensions/llamacpp-upstream-extension/src/gemmaMtpRegistry.ts
+ * @file Static registry mapping a Gemma 4 MTP *target* model to the separate
+ * MTP *draft* head GGUF it needs for upstream llama.cpp speculative decoding
+ * (`--model-draft <head> --spec-type draft-mtp`, PR ggml-org/llama.cpp#23398,
+ * first tagged release `b9553`).
+ *
+ * Unlike Qwen 3.6 (where the MTP head lives *inside* the same GGUF and no
+ * `--model-draft` is passed), Gemma 4 ships the head as a second, small GGUF
+ * that must be downloaded next to the target model (the same way `mmproj`
+ * files are handled).
+ *
+ * Upstream MTP support covers ONLY Gemma 4 **31B** (dense) and **26B-A4B**
+ * (MoE). E2B/E4B drafters were deferred upstream to a follow-up and are not
+ * supported here; 12B has no MTP head.
+ *
+ * Head files were verified against the live Hugging Face API (sha256 + size
+ * pinned so the downloader validates them):
+ *   - 31B     → am17an/Gemma4-31B-it-GGUF (the PR author's reference head).
+ *   - 26B-A4B → AtomicChat/gemma-4-26B-A4B-it-assistant-GGUF (first-party;
+ *               am17an published no 26B repo, this head is model-identical to
+ *               the upstream reference and guaranteed-stable as our own org).
+ */
+export interface GemmaMtpDraft {
+    /** Hugging Face repo id that hosts the MTP draft head GGUF. */
+    repo: string;
+    /** Exact filename of the draft head inside the repo. */
+    draftFilename: string;
+    /** Pinned sha256 of the head file (lowercase hex) for download validation. */
+    draftSha256: string;
+    /** Pinned byte size of the head file for download validation. */
+    draftSize: number;
+}
+/**
+ * Every head this registry downloads — the files the app itself puts on disk
+ * next to a target. The web app's model-file filter is tested against them so
+ * a local scan never offers one back as a model (ATO-523).
+ */
+export declare const GEMMA_MTP_DRAFT_FILENAMES: readonly string[];
+/**
+ * Resolve the MTP draft head for a Gemma 4 target model id, or `null` if the
+ * model is not a Gemma 4 MTP-capable target (31B / 26B-A4B).
+ */
+export declare function resolveGemmaMtpDraft(modelId: string): GemmaMtpDraft | null;
+/** Whether the given model id is a Gemma 4 MTP-capable target. */
+export declare function checkGemmaMtpSupport(modelId: string): boolean;
+/** Build the Hugging Face resolve URL for a draft head. */
+export declare function gemmaMtpDraftUrl(draft: GemmaMtpDraft): string;

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

@@ -0,0 +1,79 @@
+/**
+ * Ported verbatim from: extensions/llamacpp-upstream-extension/src/gemmaMtpRegistry.ts
+ * @file Static registry mapping a Gemma 4 MTP *target* model to the separate
+ * MTP *draft* head GGUF it needs for upstream llama.cpp speculative decoding
+ * (`--model-draft <head> --spec-type draft-mtp`, PR ggml-org/llama.cpp#23398,
+ * first tagged release `b9553`).
+ *
+ * Unlike Qwen 3.6 (where the MTP head lives *inside* the same GGUF and no
+ * `--model-draft` is passed), Gemma 4 ships the head as a second, small GGUF
+ * that must be downloaded next to the target model (the same way `mmproj`
+ * files are handled).
+ *
+ * Upstream MTP support covers ONLY Gemma 4 **31B** (dense) and **26B-A4B**
+ * (MoE). E2B/E4B drafters were deferred upstream to a follow-up and are not
+ * supported here; 12B has no MTP head.
+ *
+ * Head files were verified against the live Hugging Face API (sha256 + size
+ * pinned so the downloader validates them):
+ *   - 31B     → am17an/Gemma4-31B-it-GGUF (the PR author's reference head).
+ *   - 26B-A4B → AtomicChat/gemma-4-26B-A4B-it-assistant-GGUF (first-party;
+ *               am17an published no 26B repo, this head is model-identical to
+ *               the upstream reference and guaranteed-stable as our own org).
+ */
+const REGISTRY = [
+    {
+        // Gemma 4 31B (dense). e.g. unsloth/gemma-4-31B-it-GGUF, gemma-4-31B-it-Q4_K_M
+        matches: (id) => id.includes('gemma') && id.includes('4') && id.includes('31b'),
+        repo: 'am17an/Gemma4-31B-it-GGUF',
+        draftFilename: 'mtp-gemma-4-31B-it.gguf',
+        draftSha256: '9514a3a9a5f36971580c83212f59ee681a5b9a09d597a6d14e40ecd12f76e8b9',
+        draftSize: 514687200,
+    },
+    {
+        // Gemma 4 26B-A4B (MoE). Require both "26b" and "a4b" so only the MoE
+        // matches. MTP gains on MoE may be marginal — that is expected.
+        matches: (id) => id.includes('gemma') && id.includes('4') && id.includes('26b') && id.includes('a4b'),
+        repo: 'AtomicChat/gemma-4-26B-A4B-it-assistant-GGUF',
+        draftFilename: 'gemma-4-26B-A4B-it-assistant.Q8_0.gguf',
+        // 2026-07-23: assistant repos were regenerated in the post-b10018
+        // gemma4-assistant format (the old heads no longer load on engine
+        // b10018-1.1.1+), so the pin tracks the re-uploaded file.
+        draftSha256: '918d48b5c219035561e0ff791f2926c939cd367d71ee21b002e99222bdecab8b',
+        draftSize: 461766880,
+    },
+];
+/**
+ * Every head this registry downloads — the files the app itself puts on disk
+ * next to a target. The web app's model-file filter is tested against them so
+ * a local scan never offers one back as a model (ATO-523).
+ */
+export const GEMMA_MTP_DRAFT_FILENAMES = REGISTRY.map((entry) => entry.draftFilename);
+function normalizeId(modelId) {
+    return modelId.toLowerCase();
+}
+/**
+ * Resolve the MTP draft head for a Gemma 4 target model id, or `null` if the
+ * model is not a Gemma 4 MTP-capable target (31B / 26B-A4B).
+ */
+export function resolveGemmaMtpDraft(modelId) {
+    const normalized = normalizeId(modelId);
+    const entry = REGISTRY.find((e) => e.matches(normalized));
+    if (!entry)
+        return null;
+    return {
+        repo: entry.repo,
+        draftFilename: entry.draftFilename,
+        draftSha256: entry.draftSha256,
+        draftSize: entry.draftSize,
+    };
+}
+/** Whether the given model id is a Gemma 4 MTP-capable target. */
+export function checkGemmaMtpSupport(modelId) {
+    return resolveGemmaMtpDraft(modelId) !== null;
+}
+/** Build the Hugging Face resolve URL for a draft head. */
+export function gemmaMtpDraftUrl(draft) {
+    return `https://huggingface.co/${draft.repo}/resolve/main/${draft.draftFilename}`;
+}
+//# sourceMappingURL=gemma-mtp-registry.js.map

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

@@ -0,0 +1 @@
+{"version":3,"file":"gemma-mtp-registry.js","sourceRoot":"","sources":["../../src/speculative/gemma-mtp-registry.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;;;;;;;GAsBG;AAsBH,MAAM,QAAQ,GAAqC;IACjD;QACE,+EAA+E;QAC/E,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC;QAC/E,IAAI,EAAE,2BAA2B;QACjC,aAAa,EAAE,yBAAyB;QACxC,WAAW,EAAE,kEAAkE;QAC/E,SAAS,EAAE,SAAS;KACrB;IACD;QACE,sEAAsE;QACtE,gEAAgE;QAChE,OAAO,EAAE,CAAC,EAAE,EAAE,EAAE,CAAC,EAAE,CAAC,QAAQ,CAAC,OAAO,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,GAAG,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC,IAAI,EAAE,CAAC,QAAQ,CAAC,KAAK,CAAC;QACrG,IAAI,EAAE,8CAA8C;QACpD,aAAa,EAAE,wCAAwC;QACvD,kEAAkE;QAClE,kEAAkE;QAClE,0DAA0D;QAC1D,WAAW,EAAE,kEAAkE;QAC/E,SAAS,EAAE,SAAS;KACrB;CACF,CAAA;AAED;;;;GAIG;AACH,MAAM,CAAC,MAAM,yBAAyB,GAAsB,QAAQ,CAAC,GAAG,CAAC,CAAC,KAAK,EAAE,EAAE,CAAC,KAAK,CAAC,aAAa,CAAC,CAAA;AAExG,SAAS,WAAW,CAAC,OAAe;IAClC,OAAO,OAAO,CAAC,WAAW,EAAE,CAAA;AAC9B,CAAC;AAED;;;GAGG;AACH,MAAM,UAAU,oBAAoB,CAAC,OAAe;IAClD,MAAM,UAAU,GAAG,WAAW,CAAC,OAAO,CAAC,CAAA;IACvC,MAAM,KAAK,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,UAAU,CAAC,CAAC,CAAA;IACzD,IAAI,CAAC,KAAK;QAAE,OAAO,IAAI,CAAA;IACvB,OAAO;QACL,IAAI,EAAE,KAAK,CAAC,IAAI;QAChB,aAAa,EAAE,KAAK,CAAC,aAAa;QAClC,WAAW,EAAE,KAAK,CAAC,WAAW;QAC9B,SAAS,EAAE,KAAK,CAAC,SAAS;KAC3B,CAAA;AACH,CAAC;AAED,kEAAkE;AAClE,MAAM,UAAU,oBAAoB,CAAC,OAAe;IAClD,OAAO,oBAAoB,CAAC,OAAO,CAAC,KAAK,IAAI,CAAA;AAC/C,CAAC;AAED,2DAA2D;AAC3D,MAAM,UAAU,gBAAgB,CAAC,KAAoB;IACnD,OAAO,0BAA0B,KAAK,CAAC,IAAI,iBAAiB,KAAK,CAAC,aAAa,EAAE,CAAA;AACnF,CAAC"}

package/dist/speculative/index.d.ts

@@ -0,0 +1,18 @@
+/**
+ * MTP / DFlash / EAGLE-3 draft registries, transcription model constants, chat-template overrides.
+ *
+ * Ported verbatim from: extensions/llamacpp-upstream-extension/src/{dflashRegistry,gemmaMtpRegistry,
+ * transcriptionRegistry,chatTemplateOverrides}.ts and extensions/mlx-extension/src/{dflashRegistry,
+ * eagle3Registry,mtpRegistry}.ts. Every file is pure (static tables + predicates); nothing here does I/O.
+ *
+ * Both DFlash registries export a `resolveDflashDraft`; the llama.cpp one keeps its name and the MLX
+ * one is re-exported as `resolveMlxDflashDraft` (the file itself keeps the app's name).
+ * See PLAN.md §3.2. Public API of this module is exported from this file only.
+ */
+export * from './dflash-registry.js';
+export * from './gemma-mtp-registry.js';
+export * from './transcription-registry.js';
+export * from './chat-template-overrides.js';
+export { STATIC_DRAFT_MAP, normalizeBaseId, resolveDflashDraft as resolveMlxDflashDraft, type DraftRepoManifest, type DraftResolution, } from './mlx-dflash-registry.js';
+export * from './mlx-eagle3-registry.js';
+export * from './mlx-mtp-registry.js';

package/dist/speculative/index.js

@@ -0,0 +1,19 @@
+/**
+ * MTP / DFlash / EAGLE-3 draft registries, transcription model constants, chat-template overrides.
+ *
+ * Ported verbatim from: extensions/llamacpp-upstream-extension/src/{dflashRegistry,gemmaMtpRegistry,
+ * transcriptionRegistry,chatTemplateOverrides}.ts and extensions/mlx-extension/src/{dflashRegistry,
+ * eagle3Registry,mtpRegistry}.ts. Every file is pure (static tables + predicates); nothing here does I/O.
+ *
+ * Both DFlash registries export a `resolveDflashDraft`; the llama.cpp one keeps its name and the MLX
+ * one is re-exported as `resolveMlxDflashDraft` (the file itself keeps the app's name).
+ * See PLAN.md §3.2. Public API of this module is exported from this file only.
+ */
+export * from './dflash-registry.js';
+export * from './gemma-mtp-registry.js';
+export * from './transcription-registry.js';
+export * from './chat-template-overrides.js';
+export { STATIC_DRAFT_MAP, normalizeBaseId, resolveDflashDraft as resolveMlxDflashDraft, } from './mlx-dflash-registry.js';
+export * from './mlx-eagle3-registry.js';
+export * from './mlx-mtp-registry.js';
+//# sourceMappingURL=index.js.map

package/dist/speculative/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../../src/speculative/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AACH,cAAc,sBAAsB,CAAA;AACpC,cAAc,yBAAyB,CAAA;AACvC,cAAc,6BAA6B,CAAA;AAC3C,cAAc,8BAA8B,CAAA;AAC5C,OAAO,EACL,gBAAgB,EAChB,eAAe,EACf,kBAAkB,IAAI,qBAAqB,GAG5C,MAAM,0BAA0B,CAAA;AACjC,cAAc,0BAA0B,CAAA;AACxC,cAAc,uBAAuB,CAAA"}

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

@@ -0,0 +1,39 @@
+/**
+ * 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/...`.
+ */
+export interface DraftRepoManifest {
+    repo: string;
+    required: string[];
+    optional?: string[];
+}
+export interface DraftResolution {
+    repo: string;
+    required: string[];
+    optional: string[];
+}
+export declare const STATIC_DRAFT_MAP: Record<string, DraftRepoManifest>;
+/**
+ * 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 declare function normalizeBaseId(modelId: string): string;
+/**
+ * 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 declare function resolveDflashDraft(modelId: string): DraftResolution | null;