anyformer 0.4.17 → 0.4.19

package/CHANGELOG.md

@@ -1,5 +1,59 @@
 # Changelog
 
+## 0.4.19 — 2026-10-31
+
+### Changed
+- diskformer pin 314362e → 317d2a0 (v0.6.18): with thinking off, diskformer's llama.cpp chat closes a chat template's
+  always-open `<think>` block and bans the `<think>` token for the reply, so LFM2.5-24B-A2B (base and
+  `LFM2_5_24B_PI`) answers directly. This resolves 0.4.18's known limitation. diskformer's gates (Chrome, 4 GB):
+  thinking off 9/9 plain replies with and without the Pi adapter; thinking on unchanged (8/9 against llama.cpp).
+
+### Verified
+- `LFM2_5_24B_PI` through anyformer, 4 turns, thinking off (browser `chat?entry=chat.LFM2_5_24B_PI&local=1&turns=4&tokens=96`,
+  Chrome 156, this Mac, 12:12 IST, base GGUF and adapter from local files, diskformer 317d2a0): PASS. Loaded in 101.1 s;
+  every reply on topic with no `<think>` text (CPU, RAM, which is faster with latency figures, a one-sentence
+  summary); 2.6–3.1 tok/s.
+- LFM2.5 230M, 4 turns, on 0.4.18's pin 314362e (browser `chat?local=1&turns=4`, `localmind`, 12:13 IST): PASS, load
+  8.7 s, 4/4 turns on topic.
+
+## 0.4.18 — 2026-10-31
+
+### Added
+- LoRA adapters: `load(repo, { adapter: { repo, file, revision, size, scale } })` (or an entry carrying `adapter`)
+  downloads the adapter GGUF into OPFS beside the model and applies it at `scale` on `diskformer-llama`, the only back
+  end with `adapters: true`; the others drop out of the plan. Needs diskformer v0.6.17.
+- `MODELS.chat.LFM2_5_24B_PI`: LFM2.5-24B-A2B (QAD-Q4_0) with harness-hub's Pi coding-agent LoRA
+  (`harness-hub/LFM2.5-24B-A2B-pi-GGUF` at a994cf7, 37.4 MB), 4 GB budget, `experimental`, `license: 'lfm1.0'`.
+  diskformer's gate (v0.6.17): 8/9 replies identical to llama-server --lora (7/9 without the adapter, the path's own
+  baseline). The adapter was trained inside the Pi agent; elsewhere expect base-model behaviour or worse. Hub fixture
+  at 0912d33.
+
+### Changed
+- diskformer pin f17009c → 314362e (v0.6.17, which includes v0.6.16). v0.6.16: Chrome 156 rejects mixed f16/f32 subgroup-matrix kernels on Apple GPUs,
+  so Ternary Bonsai 2 27B, Gemma 4 E2B, LFM2.5 and the IQ-quant dense path failed to load ("Mixed precision subgroup
+  matrix multiply is not supported by the device"). diskformer's new `engines/subgroup-matrix-guard.js` hides the
+  feature when a tiny mixed kernel fails to compile, and those engines take their non-matrix prefill (slower prefill).
+- Browser tests: `chat` takes `?turns=N` (up to 4 follow-up turns, each checked on topic); `vision` takes `?backend=`
+  and `?local=1` (MiniCPM-V 4.7 on diskformer, image as packed RGB after `ensureVision()`); `run.mjs` prints each
+  failed network request with its reason. `vision?local=1` serves `cats.jpg` from `--models` as well. `chat` also takes
+  `?entry=<family>.<KEY>` (a MODELS entry, adapter included) and `?tokens=N`; its `?local=1` serves diskformer-llama
+  GGUFs, adapters and (through the engines' fetch) localmind GGUFs by file name.
+
+### Known limitation
+- LFM2.5-24B-A2B's chat template always opens a `<think>` block and ignores `enable_thinking`, so on
+  `diskformer-llama` a reply with thinking off still starts with visible reasoning, then `</think>`, then the answer.
+  The fix belongs in diskformer's chat driver (close the block when thinking is off).
+
+### Verified
+- `LFM2_5_24B_PI` through anyformer (browser `chat?entry=chat.LFM2_5_24B_PI&local=1&tokens=400`, Chrome 156, this
+  Mac, 11:56 IST, base GGUF and adapter from local files, adapter sha256 5e96d059…, diskformer 314362e): loaded with
+  the adapter in 208.9 s; reasoning, `</think>`, then a correct two-sentence answer about the CPU; 162 tokens at
+  2.73 tok/s. The page reports FAIL because its check rejects `</think>` text (the limitation above). A 4-turn run at
+  64 tokens per reply spent each reply on reasoning (1.5–3.0 tok/s).
+- MiniCPM-V 4.7 image round trip through anyformer (browser `vision`, Chrome 156, this Mac, 11:12 IST, GGUF, vision
+  encoder and image from local files, 6.5 GB budget, diskformer f17009c): "There are **two cats** in this picture",
+  then a correct description, 40 tokens at 18.3 tok/s.
+
 ## 0.4.17 — 2026-10-30
 
 ### Added

package/README.md

@@ -76,8 +76,8 @@ and `threw`. Every other failure is an `AnyformerError` too, with a code from `E
 ## Commands
 
 ```js
-plan(repo, { budget, task })      // what would run it, and why the others would not; loads nothing
-load(repo, { budget, task, file, maxCtx, poolBytes, localFile, fetch, ple, speculative, onProgress, signal })  // → { backend, task, plan, run(input), generate?, think?, prime?, reset?, speculative?, dispose() }
+plan(repo, { budget, task })      // what would run it (adapter: a LoRA GGUF; only diskformer-llama applies one), and why the others would not; loads nothing
+load(repo, { budget, task, file, adapter, maxCtx, poolBytes, localFile, fetch, ple, speculative, onProgress, signal })  // → { backend, task, plan, run(input), generate?, think?, prime?, reset?, speculative?, dispose() }
 pipeline(task, repo, { budget })  // transformers.js's call shape: gen(messages) → [{ generated_text }]
 probe()                           // WebGPU, shader-f16, subgroups, limits, OPFS quota, memory64
 describe(repo)                    // Hub repo → formats, architecture, task, tags, files

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "anyformer",
-  "version": "0.4.17",
+  "version": "0.4.19",
   "description": "One call to run a model in a browser tab: anyformer reads the Hub repo, probes the browser, and picks the back end (transformers.js, LocalMind's WebGPU kernels, diskformer, kohra) that fits the model's format, architecture and your GPU budget.",
   "type": "module",
   "main": "./index.js",

package/src/backends/diskformer-llama.js

@@ -112,6 +112,8 @@ function textOnly(messages) {
 export const diskformerLlamaBackend = {
   id: 'diskformer-llama',
   tasks: ['text-generation', 'image-text-to-text'],
+  // Applies a LoRA GGUF over the model (llama.load({ lora, loraScale }), diskformer v0.6.17+), as llama-server --lora.
+  adapters: true,
   source: SOURCES.diskformerLlama,
   // Where a GGUF downloads from; a test points it at a local server.
   url: (model, file) => `${model.endpoint || 'https://huggingface.co'}/${model.repo}/resolve/${model.revision || 'main'}/${file.name.split('/').map(encodeURIComponent).join('/')}`,
@@ -131,7 +133,8 @@ export const diskformerLlamaBackend = {
   },
 
   // maxCtx sizes the KV cache; 1024 is the context diskformer's gates run at.
-  async load(model, { budget, file = null, maxCtx = 1024, onProgress, signal }) {
+  // adapter: { repo, file, revision, size, scale } downloads into OPFS beside the model and applies at scale (default 1).
+  async load(model, { budget, file = null, adapter = null, maxCtx = 1024, onProgress, signal }) {
     const f = pickGguf(model, file);
     const { Llama } = await import(diskformerLlamaBackend.source);
     const llama = new Llama();
@@ -140,23 +143,27 @@ export const diskformerLlamaBackend = {
     signal?.addEventListener('abort', stop, { once: true });
     let info;
     try {
-      // Each part and side file downloads under its own name; progress counts bytes over all of them.
-      const files = [...f.parts, ...f.extra];
+      // Each part, side file and adapter downloads under its own name; progress counts bytes over all of them.
+      const lora = adapter && { name: adapter.file.split('/').pop(), size: adapter.size, repo: adapter.repo,
+        url: diskformerLlamaBackend.url({ repo: adapter.repo, revision: adapter.revision, endpoint: model.endpoint }, { name: adapter.file }) };
+      const files = [...f.parts.map((p) => ({ name: opfsName(p), size: p.size, repo: model.repo, url: diskformerLlamaBackend.url(model, p) })),
+        ...f.extra.map((p) => ({ name: opfsName(p), size: p.size, repo: model.repo, url: diskformerLlamaBackend.url(model, p) })),
+        ...(lora ? [lora] : [])];
+      const total = files.reduce((n, x) => n + x.size, 0);
       let before = 0;
-      for (const part of files) {
-        const name = opfsName(part);
+      for (const { name, size, url } of files) {
         const have = (await llama.list()).find((m) => m.name === name && !m.partial);
-        if (!have || have.size !== part.size) {
-          say({ status: 'download', loaded: before, total: f.size });
-          await llama.download({ url: diskformerLlamaBackend.url(model, part), name, size: part.size },
-            (p) => say({ status: 'download', loaded: before + p.copied, total: f.size }));
+        if (!have || have.size !== size) {
+          say({ status: 'download', loaded: before, total });
+          await llama.download({ url, name, size }, (p) => say({ status: 'download', loaded: before + p.copied, total }));
           signal?.throwIfAborted();
         }
-        before += part.size;
+        before += size;
       }
-      await writeIndex({ ...(await readIndex()), ...Object.fromEntries(files.map((p) => [opfsName(p), model.repo])) });
+      await writeIndex({ ...(await readIndex()), ...Object.fromEntries(files.map((p) => [p.name, p.repo])) });
       say({ status: 'load' });
-      info = await llama.load({ name: opfsName(f), budget, ctx: maxCtx, extra: f.extra.map(opfsName) }, (p) => say({ status: 'plan', plan: p }));
+      info = await llama.load({ name: opfsName(f), budget, ctx: maxCtx, extra: f.extra.map(opfsName),
+        ...(lora && { lora: lora.name, loraScale: adapter.scale ?? 1 }) }, (p) => say({ status: 'plan', plan: p }));
       signal?.throwIfAborted();
     } catch (e) {
       llama.terminate();

package/src/load.js

@@ -29,16 +29,19 @@ function noBackend(plan) {
 // What load() would do, without loading: the descriptor, the caps, and every back end's answer.
 // task: a Hugging Face pipeline task ('text-generation', 'feature-extraction', …); null runs the repo's own.
 // file: a weights file in the repo to load (a GGUF name); null lets each back end pick.
-// repo: a Hub id, or an entry from MODELS (its task and budget apply unless opts set them).
+// repo: a Hub id, or an entry from MODELS (its task, budget and adapter apply unless opts set them).
+// adapter: a LoRA GGUF applied over the model, { repo, file, revision, size, scale }; only a back end with
+// `adapters: true` (diskformer-llama) can apply one, so the others drop out of the plan.
 const entry = (repo, opts) => (typeof repo === 'object' && repo?.repo
-  ? [repo.repo, { ...(repo.task && { task: repo.task }), ...(repo.budget && { budget: repo.budget }), ...opts }]
+  ? [repo.repo, { ...(repo.task && { task: repo.task }), ...(repo.budget && { budget: repo.budget }), ...(repo.adapter && { adapter: repo.adapter }), ...opts }]
   : [repo, opts]);
 
 export async function plan(repoOrEntry, opts = {}) {
-  let [repo, { budget = DEFAULT_BUDGET, task = null, file = null, backends = BACKENDS, caps, model, fetch }] = entry(repoOrEntry, opts);
+  let [repo, { budget = DEFAULT_BUDGET, task = null, file = null, backends = BACKENDS, caps, model, fetch, adapter = null }] = entry(repoOrEntry, opts);
   caps ??= await probe();
   model ??= await describe(repo, fetch ? { fetch } : {});
-  return { model, caps, budget, task, file, ...(await resolve(model, caps, backends, { budget, task, file })) };
+  if (adapter) backends = backends.filter((b) => b.adapters);
+  return { model, caps, budget, task, file, ...(adapter && { adapter }), ...(await resolve(model, caps, backends, { budget, task, file })) };
 }
 
 export async function load(repoOrEntry, options = {}) {
@@ -50,7 +53,7 @@ export async function load(repoOrEntry, options = {}) {
   const backend = backends.find((b) => b.id === p.chosen);
   let session;
   try {
-    session = await backend.load(p.model, { budget: p.budget, task: p.task, file: p.file, caps: p.caps, maxCtx, dtype, poolBytes, localFile, fetch: engineFetch, ple, speculative, speech, onProgress, signal });
+    session = await backend.load(p.model, { budget: p.budget, task: p.task, file: p.file, adapter: p.adapter, caps: p.caps, maxCtx, dtype, poolBytes, localFile, fetch: engineFetch, ple, speculative, speech, onProgress, signal });
   } catch (e) { throw wrap(e, 'LOAD_FAILED', signal, `${backend.id}: `); }
   // Added onto serialize()'s object, not spread: a spread would read its getters once and freeze them.
   return Object.assign(serialize(session), { backend: backend.id, task: session.task || p.task || p.model.task || null, ...(session.device && { device: session.device }), plan: p });

package/src/models.js

@@ -3,7 +3,7 @@
 // passed. `test` names the page under test/browser/ that proves the entry; test/models.test.js checks it exists.
 // `diffusion: true` marks a text-diffusion model (the reply denoises in blocks, so generate() chunks arrive a block at a time); `experimental: true` marks an entry its
 // app shows as opt-in until that app's own checks pass; `audioOut: true` marks a model that can also answer in speech
-// (generate({ speech: true })).
+// (generate({ speech: true })); `adapter` names a LoRA GGUF applied over the model (diskformer-llama only).
 
 const M = (repo, backend, task, test, extra = {}) => ({ repo, backend, task, test, ...extra });
 
@@ -32,6 +32,17 @@ const NEMOTRON_3_5_BOLT_E4B = M('danyalmalik/NVIDIA-Nemotron-3.5-Bolt-E4B-ONNX-W
 // diskformer checked one hand-written prompt through complete(); generate() renders the GGUF's chat template, which no
 // browser run has checked.
 const DEEPSEEK_V4_1_FLASH = M('ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF', 'diskformer-llama', 'text-generation', 'llama-gate', { budget: 8e9, downloadBytes: 231316411072, experimental: true });
+// LFM2.5-24B-A2B with harness-hub's Pi LoRA (rank 32, alpha 64; asynchronous GRPO inside the Pi coding agent), on
+// diskformer's llama.cpp path, the only back end that applies an adapter. The adapter touches the conv, attention and
+// two dense FFN layers only, so expert paging is unchanged. diskformer v0.6.17's gate: 8/9 replies identical to
+// llama-server --lora (the path's own baseline without the adapter is 7/9). Authors' numbers on 100 data-analysis tasks:
+// 39% fewer tool calls on tasks both solve, pass@1 −4.2 points (noise about ±5); expect base-model behaviour or worse in
+// other agents. Licence lfm1.0 (model and adapter): royalty-free commercial use only below US$10M annual revenue.
+const LFM2_5_24B_PI = M('LiquidAI/LFM2.5-24B-A2B-GGUF', 'diskformer-llama', 'text-generation', 'chat', {
+  budget: 4e9, downloadBytes: 13578190880 + 37437888, experimental: true, license: 'lfm1.0',
+  adapter: { repo: 'harness-hub/LFM2.5-24B-A2B-pi-GGUF', file: 'LFM2.5-24B-A2B-harbor-pi-lora-F16.gguf',
+    revision: 'a994cf7f03ab0e572832d5829f42cbe2ef28d472', size: 37437888, scale: 1, trainedFor: 'pi' },
+});
 const MINILM_L6 = M('Xenova/all-MiniLM-L6-v2', 'transformers', 'feature-extraction', 'embed', { downloadBytes: 90e6 });
 const DETR_RESNET_50 = M('Xenova/detr-resnet-50', 'transformers', 'object-detection', 'detect', { downloadBytes: 167e6 });
 
@@ -69,6 +80,8 @@ export const MODELS = {
     LFM3_OMNI_2B,
     /** DeepSeek V4.1 Flash (763B MoE), a tech demo on diskformer's llama.cpp path: 231 GB on disk, very slow. */
     DEEPSEEK_V4_1_FLASH,
+    /** LFM2.5-24B-A2B with harness-hub's Pi coding-agent LoRA (diskformer's llama.cpp path); tuned for the Pi agent; licence lfm1.0. */
+    LFM2_5_24B_PI,
   },
   /** Chat with images in the messages: content [{ type: 'image', image }, { type: 'text', text }]. Gemma 4 takes audio too. */
   vision: { DEFAULT: QWEN3_5_4B, QWEN3_5_4B, GEMMA4_E2B_ONNX, GEMMA4_E4B_ONNX },

package/src/sources.js

@@ -9,6 +9,13 @@ export const PINS = {
   // v0.3.1: seeded block-KV graphs (Nemotron-Labs-Diffusion-3B), chatFormat, split external data; download progress
   // (progress_callback) and the graph files cached in Cache Storage 'kohra-models'.
   kohra: 'bf95d44c3e3cfc001201baa03fa121c2bf9c7d73',
+  // v0.6.18 (317d2a0): with thinking off, the llama.cpp chat closes a template's always-open <think> (LFM2.5-24B) and
+  // bans the <think> token for the reply.
+  // v0.6.17 (314362e): LoRA adapters on the llama.cpp path (llama.load({ lora, loraScale })), and lfm2_moe_ssd.js runs
+  // a LoRA-merged LFM2.5-24B-A2B with Q8_0 dense matrices.
+  // v0.6.16 (467052f): engines/subgroup-matrix-guard.js. Chrome 156 rejects mixed f16/f32 subgroup-matrix kernels
+  // on Apple GPUs; the guard hides the feature when one tiny mixed kernel fails to compile, so Ternary Bonsai 2 27B,
+  // Gemma 4 E2B, LFM2.5 and qwen3_moe_ssd.js take their non-matrix prefill (slower prefill, same output path).
   // v0.6.15 (f17009c): engines/clover_moe_ssd.js, Clover-35B-A3B-Preview (block diffusion on Qwen3.6-35B-A3B; kohra's
   // denoise loop vendored in engines/vendor/). qwen35_moe_ssd.js and qwen3_moe_ssd.js only learn to accept clover_moe.
   // v0.6.14 (93193d0): LFM3-Omni speech out's inverse STFT uses an FFT (same samples), 0.54x → 1.5–1.8x real time on
@@ -23,7 +30,7 @@ export const PINS = {
   // Gemma 4 E2B/E4B and Ternary Bonsai 2) under engines/, and the llama.cpp path under ggml/, which loads split GGUFs
   // and drafters, and reads past 2 GiB (Nemotron 3.5 Bolt E4B needs v0.6.8 or later). The hand engines' ingest
   // streams the GGUF header with progress and resumes a dropped range.
-  diskformer: 'f17009cc31e15361af7ae5536e8f55cb43c5066f',
+  diskformer: '317d2a002d95278e504ed84fab2d0f71dabaa884',
 };
 
 const gh = (repo, sha, path) => `https://cdn.jsdelivr.net/gh/NakliTechie/${repo}@${sha}/${path}`;