diskformer 0.5.4 → 0.6.6

package/CHANGELOG.md

@@ -1,5 +1,69 @@
 # Changelog
 
+## 0.6.6 (2026-10-24)
+- The chat demo (`ggml/chat.html`) lists Ling 3.0 Flash 124B (2 parts, 77 GB, 6 GB budget) and downloads a model
+  published in parts part by part, each under its own name; stop and resume work across parts, and Remove deletes every
+  part. Checked in the browser: the download starts on part 1 of 2 from Hugging Face, a stop shows 0.34 GB of 77.0 GB
+  stored with Resume, and Remove leaves no file in the page's OPFS.
+
+## 0.6.5 (2026-10-24)
+- The browser build loads a split GGUF (`name-00001-of-0000n.gguf`, as published): the worker opens every part from
+  OPFS and mounts them together, `df_read_file_batch` and `df_upload_batch_js` read experts from the part that holds
+  them, and `df_plan` sums the paged bytes over all parts. `gate.html` and `scripts/chrome-gate.mjs` take every part.
+  Ling 3.0 Flash (124B-A5B, `bailingmoe3`, 2 parts, 77 GB) in Chrome at a 6 GB budget: copied into OPFS in 126 s,
+  loaded in 60 s (22 slots), 4.1-4.9 tok/s; 6/9 replies match llama.cpp on the CPU (4 identical, 2 where the CPU reply
+  is a prefix of Chrome's), the other 3 are coherent and differ where two words score almost the same. Paging exact
+  natively, 9/9 at 8 vs 32 slots; Granite 4.0 H Tiny rechecked 9/9 in Chrome.
+- Native `df-chat <gguf> <slots> --record <refs.json> <out.json>` renders each gate conversation with the model's own
+  template and records 64-token replies, for models whose reference cannot come from llama-server on Metal.
+- `scripts/record-refs.mjs`: CPU references (`--ngl 0`) keep the weights memory-mapped and unrepacked
+  (`-lm mmap --no-repack`), so a model larger than RAM loads; a server killed by a signal now stops the run instead of
+  leaving it waiting for 30 minutes.
+
+## 0.6.4 (2026-10-23)
+- `engines/lfm2_5.js` (LFM2.5 230M/1.2B, `Lfm2Mobile`), `engines/gemma-4-e2b.js` (Gemma 4 E2B/E4B, `Gemma4Mobile`) and
+  `engines/ternary_bonsai_2_27b.js` (`TernaryBonsai2`), from LocalMind main a02f06e. Held back in 0.6.0 because their
+  webml-community Spaces showed no license; Chirag wrote them and released them under MIT (2026-10-23). All three
+  import in Node; their GPU checks are LocalMind's (`scripts/bench-engines.mjs`), not rerun here. With these,
+  diskformer holds every LocalMind engine.
+
+## 0.6.3 (2026-10-23)
+- `LingMoeSsd.load` reads the two parts' headers (20 MB from Hugging Face) only when an ingest follows; a complete
+  store loads without them (found by LocalMind's live check through anyformer). Verified in a real browser by LocalMind
+  (after a 77 GB ingest through anyformer v0.2.5, a reload made no requests to the part URLs; answers correct before and
+  after). `qwen3_moe_ssd.js` exports `STORE_FORMAT` and `STORE_ROOT`.
+- `scripts/sync-localmind.mjs` runs diskformer → LocalMind for every file now, engines included (imports rewritten
+  from `../src/` to `./`); `--ref` is gone, since LocalMind no longer writes engines.
+- Native driver reads a split GGUF (`name-0000k-of-0000n.gguf`, as published) part by part, so it is never merged.
+  Ling 3.0 Flash (124B-A5B, 2 parts, 77 GB) on the llama.cpp path natively: replies identical at 8 and 32 slots on 3
+  prompts, 32 tokens. The browser build refuses a split GGUF with a message for now. Rebuilt browser build rechecked:
+  Granite 4.0 H Tiny 9/9 in Chrome.
+
+## 0.6.2 (2026-10-23)
+- `engines/ling_moe_ssd.js`: the Ling 3.0 Flash engine (inclusionAI, 124B-A5B, `bailingmoe3`: KDA linear attention +
+  MLA, 512 experts), handed over by LocalMind (d210986); the first engine diskformer adopts rather than mirrors.
+  `splitFetch` ingests the two-part GGUF as one file. `engines/iq_quants.js` gains Q6_K and Q5_0. LocalMind's gate:
+  6-layer cut 8/8 vs llama.cpp Metal; full model 5/8 vs llama.cpp CPU over 32 tokens, layer 0 within 5 decimals of a
+  float64 reference (CPU is 1-3% off); decode 5.53 tok/s, GPU 2.51 GB dense + 4.29 GB pool (M4 Pro). Here: imports in
+  Node, LocalMind's no-GPU test (`test-ling.mjs`) passes on this copy; the GPU gate was not rerun here.
+
+## 0.6.1 (2026-10-22)
+- `Llama.chat(messages, { thinking: false })` renders the chat template with reasoning off (llama.cpp's
+  `enable_thinking`), for models that reason first: Qwen3.6 35B-A3B natively answers directly instead of opening with a
+  thinking process. Default unchanged (the template's own choice). Asked for by anyformer, whose other engines pass
+  `enableThinking: false`. Native driver: `df-chat model slots prompt n --no-think`. In Chrome (through anyformer's
+  diskformer-llama): Gemma 4 26B's explain-tcp chat with `thinking: false` is identical to `gemma.json`'s reply.
+
+## 0.6.0 (2026-10-22)
+- `engines/` is the one home of LocalMind's hand-written engines (Chirag 2026-10-22: diskformer holds every custom
+  engine; anyformer routes; LocalMind loads through anyformer). New: `lfm2_moe_ssd.js`, `hy_v3_moe_ssd.js`,
+  `flint_moe_ssd.js`, `iq_quants.js`, `qwen35_dense.js`, `ple-opfs.js`, `flashmap_drafter.js`,
+  `ternary_bonsai_2_dflash.js`, `ternary_bonsai_2_mtp.js`, copied from LocalMind main b5c289d with imports rewritten;
+  each was gated in LocalMind and imports cleanly here (Node). Not copied: `lfm2_5.js`, `gemma-4-e2b.js` and
+  `ternary_bonsai_2_27b.js`, vendored from webml-community Spaces that publish no license.
+- `scripts/sync-localmind.mjs --ref <commit>` reads LocalMind's engines from a commit instead of its working tree.
+- `scripts/chrome-gate.mjs --env K=V,…` passes driver settings to the gate page.
+
 ## 0.5.4 (2026-10-22)
 - `diskformer/llama`: the GPU budget plan picks the physical batch size (llama.cpp's `n_ubatch`) as the largest of
   512, 256, 128 and 64 that leaves room for one token's experts. The compute buffers scale with it (Gemma 4 26B-A4B:

package/README.md

@@ -5,7 +5,7 @@ WebGPU in a worker; the GGUF stays on disk in the browser's private storage (OPF
 needs are paged into GPU memory. 17 architectures are checked against llama.cpp, from Qwen3.6 35B to gpt-oss 20B.
 
 Plain ES modules, no dependencies, no build step. Chrome or Edge with WebGPU, on any origin (CDN or self-hosted).
-Weights stay in the user's own browser storage. The disk tier underneath (`src/`) also runs two hand-written engines.
+Weights stay in the user's own browser storage. The disk tier underneath (`src/`) also runs hand-written engines.
 
 [![npm](https://img.shields.io/npm/v/diskformer?style=flat-square&color=555)](https://www.npmjs.com/package/diskformer)
 [![license](https://img.shields.io/badge/license-MIT-555?style=flat-square)](LICENSE)
@@ -18,7 +18,7 @@ Weights stay in the user's own browser storage. The disk tier underneath (`src/`
 | How | Command |
 |---|---|
 | npm | `npm install diskformer` |
-| CDN | `import { Llama } from 'https://cdn.jsdelivr.net/npm/diskformer@0.5.2/ggml/llama.js'` |
+| CDN | `import { Llama } from 'https://cdn.jsdelivr.net/npm/diskformer@0.6.4/ggml/llama.js'` |
 | Copy | `src/`, `engines/` and `index.js`; nothing else is needed |
 
 ## Quick start
@@ -27,7 +27,7 @@ Paste into any page served over http(s), in Chrome or Edge. The first run downlo
 OPFS (resumable); later loads take seconds. Any MoE GGUF on a [checked architecture](docs/coverage.md) loads the same way.
 ```html
 <script type="module">
-import { Llama } from 'https://cdn.jsdelivr.net/npm/diskformer@0.5.2/ggml/llama.js';
+import { Llama } from 'https://cdn.jsdelivr.net/npm/diskformer@0.6.4/ggml/llama.js';
 const llama = new Llama();
 const name = 'granite-4.0-h-tiny-Q4_K_M.gguf';
 await llama.download({ name, size: 4230976352,
@@ -64,8 +64,9 @@ weights in OPFS, read through sync access handles in workers, and moves only wha
 
 ## The engines
 
-The disk tier also stands alone: your engine keeps its own kernels and decides what a step needs. `engines/` holds three
-from-scratch WebGPU engines on it: Gemma 4 26B-A4B, Qwen3.6 35B-A3B and Qwen3-30B-A3B (at
+The disk tier also stands alone: your engine keeps its own kernels and decides what a step needs. `engines/` is the home of
+LocalMind's hand-written WebGPU engines (since 0.6.0): LFM2 MoE, Hunyuan v3 MoE, Flint, Ling 3.0 Flash, a dense Qwen3.5 layer streamer,
+LFM2.5, Gemma 4 E2B, Ternary Bonsai 2 27B and speculative drafters, each gated in LocalMind; and Gemma 4 26B-A4B, Qwen3.6 35B-A3B and Qwen3-30B-A3B (at
 [/classic/](https://diskformer.naklitechie.com/classic/); `Gemma4MoeSsd.load(null, { gpuBudgetBytes })`). A budget that
 cannot hold the dense part and two tokens' experts is refused with the numbers. Kernels: gemma4-webgpu (Apache-2.0,
 `engines/NOTICE`). Measured on a MacBook M4 Pro, 24 GB, Chrome 154, 2026-10-05, decode the median of

package/engines/NOTICE

@@ -9,9 +9,10 @@ fork of gemma4-webgpu:
   https://github.com/tylerstraub/gemma4-webgpu
 
 Changes from the original: rewritten for mixture-of-experts models whose experts stream from OPFS (Q8_0 and Q4_0
-split planes, slot-indexed expert GEMVs, GPU router top-k, chunked prefill, Q6_K head planes). Everything else in
+split planes, slot-indexed expert GEMVs, GPU router top-k, chunked prefill, Q6_K head planes). `ternary_bonsai_2_dflash.js` is built from NakliTechie/dflash-mlx-bonsai2 (Apache-2.0). `lfm2_5.js`,
+`gemma-4-e2b.js` and `ternary_bonsai_2_27b.js` are built from webml-community Hugging Face Spaces by the same author, MIT. Everything else in
 this folder is MIT (../LICENSE), Copyright (c) 2026 Chirag Patnaik.
 
 The engine files are copies of LocalMind's (https://github.com/NakliTechie/LocalMind), kept in step by
 scripts/sync-localmind.mjs. Model weights are not distributed: each model is downloaded by the user under its own
-terms (Gemma: https://ai.google.dev/gemma/terms; Qwen3.6: Apache-2.0).
+terms (Gemma: https://ai.google.dev/gemma/terms; Qwen3.6: Apache-2.0; each other model under its own license).