Spaces:
Running
Running
NakliTechie commited on
Commit ·
8d9f687
1
Parent(s): 53ebc1b
Sync from GitHub 2026-10-22T21:50:34Z
Browse files- index.html +45 -17
index.html
CHANGED
|
@@ -9307,10 +9307,12 @@ self.addEventListener('message', async (e) => {
|
|
| 9307 |
// on-device storage once (~37 GB of disk). Experimental: rung 2b of the SSD-streaming ladder,
|
| 9308 |
// decode speed is bound by SSD reads (~9 tok/s on an M4 Pro), and prompts are read one token at
|
| 9309 |
// a time. Greedy output matches llama.cpp Metal on layer cuts (plan/history.md, 2026-10-04).
|
|
|
|
|
|
|
| 9310 |
'qwen36-35b-a3b-ssd': {
|
| 9311 |
id: 'unsloth/Qwen3.6-35B-A3B-GGUF',
|
| 9312 |
label: 'Qwen3.6 35B-A3B · SSD-streamed (experimental)',
|
| 9313 |
-
backend: '
|
| 9314 |
size: '~37 GB',
|
| 9315 |
// The pinned GGUF (QWEN36_35B_A3B in qwen35_moe_ssd.js): "Load .gguf from disk…" routes a file of exactly
|
| 9316 |
// this size here and ingests it instead of downloading. test-local-gguf-ssd.mjs keeps the two in step.
|
|
@@ -17777,6 +17779,7 @@ self.addEventListener('message', async (e) => {
|
|
| 17777 |
'gemma4moe-ssd': { engine: 'gemma4_moe_ssd.js', exportName: 'Gemma4MoeSsd', label: 'Gemma 4 26B-A4B (Q4_0)' },
|
| 17778 |
'lfm2moe-ssd': { engine: 'lfm2_moe_ssd.js', exportName: 'Lfm2MoeSsd', label: 'LFM2.5 24B-A2B (Q4_0)' },
|
| 17779 |
'flint-ssd': { engine: 'flint_moe_ssd.js', exportName: 'FlintMoeSsd', label: 'Flint 40B-A2.5B (Q4_1)' },
|
|
|
|
| 17780 |
'anyformer-ling3flash': { engine: 'anyformer_models.js', exportName: 'Ling3Flash', label: 'Ling 3.0 Flash (Q4_K_M, via anyformer)' },
|
| 17781 |
'anyformer-glm47flash': { engine: 'anyformer_models.js', exportName: 'Glm47Flash', label: 'GLM-4.7-Flash (Q4_K_M, via anyformer)' },
|
| 17782 |
'qwen3dense': { engine: 'qwen35_dense.js', exportName: 'Qwen3Dense', label: 'Qwen3 4B (Q8_0) + FlashMap' },
|
|
@@ -19395,7 +19398,7 @@ self.addEventListener('message', async (e) => {
|
|
| 19395 |
const ENGINE_STORES = {
|
| 19396 |
'transformers': ['transformers-cache'], 'wllama': ['wllama'], 'lfm2-webgpu': ['gguf-v1'],
|
| 19397 |
'bonsai2-webgpu': ['gguf-cache-v1'], 'gemma4-webgpu': ['safetensors-cache-v1', 'bonsai-pipeline-v1'],
|
| 19398 |
-
'qwen35moe-ssd': ['on disk (OPFS)'], 'minicpmv-ssd': ['on disk (OPFS)'], 'gemma4moe-ssd': ['on disk (OPFS)'], 'lfm2moe-ssd': ['on disk (OPFS)'], 'flint-ssd': ['on disk (OPFS)'], 'anyformer-glm47flash': ['on disk (OPFS)'], 'anyformer-ling3flash': ['on disk (OPFS)'], 'qwen3dense': ['on disk (OPFS)'], 'qwen35dense': ['on disk (OPFS)'],
|
| 19399 |
};
|
| 19400 |
function parseSizeBytes(size) {
|
| 19401 |
const m = /([\d.]+)\s*(GB|MB)/i.exec(String(size || ''));
|
|
@@ -24018,17 +24021,20 @@ self.addEventListener('message', async (e) => {
|
|
| 24018 |
* anyformer picks the back end per model; LocalMind sees one call shape.
|
| 24019 |
*
|
| 24020 |
* AnyformerModel gives an anyformer session the engine interface LocalMind's SSD worker drives (moeSsdWorkerSrc):
|
| 24021 |
-
* static load(null, { maxCtx, onProgress, signal }), generate(messages, { maxNewTokens, signal
|
| 24022 |
-
* { text }, reset(), dispose(), and the think markers. One
|
|
|
|
| 24023 |
*
|
| 24024 |
-
*
|
| 24025 |
-
*
|
| 24026 |
-
*
|
| 24027 |
-
*
|
|
|
|
|
|
|
| 24028 |
*/
|
| 24029 |
const THINK_CLOSE = -2;
|
| 24030 |
|
| 24031 |
-
export const ANYFORMER_URL = 'https://cdn.jsdelivr.net/gh/NakliTechie/anyformer.js@v0.
|
| 24032 |
|
| 24033 |
// anyformer's progress events → the engine events the SSD worker posts. The llama.cpp path reports 'download' (bytes
|
| 24034 |
// into OPFS) and 'load' / 'plan'; a hand engine (diskformer back end) reports its own engine events, 'weights' with
|
|
@@ -24045,27 +24051,40 @@ function toEngineProgress(e) {
|
|
| 24045 |
export class AnyformerModel {
|
| 24046 |
static repo = null;
|
| 24047 |
static file = undefined; // a GGUF in the repo; undefined lets anyformer pick (the file diskformer checked)
|
| 24048 |
-
static budget = 4e9; // GPU bytes
|
|
|
|
| 24049 |
|
| 24050 |
static async load(_id, opts = {}) {
|
| 24051 |
const { load } = await import(ANYFORMER_URL);
|
| 24052 |
const session = await load(this.repo, {
|
| 24053 |
budget: opts.budget || this.budget, maxCtx: opts.maxCtx, signal: opts.signal, ...(this.file ? { file: this.file } : {}),
|
|
|
|
| 24054 |
onProgress: (e) => { const ev = toEngineProgress(e); if (ev && opts.onProgress) opts.onProgress(ev); },
|
| 24055 |
});
|
| 24056 |
return new this(session);
|
| 24057 |
}
|
| 24058 |
|
| 24059 |
-
constructor(session) {
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24060 |
|
| 24061 |
-
get thinkOpenTokenId() { return null; }
|
| 24062 |
-
get thinkCloseTokenId() { return THINK_CLOSE; }
|
| 24063 |
-
get thinkInPrompt() { return true; }
|
| 24064 |
-
get
|
|
|
|
| 24065 |
|
| 24066 |
-
//
|
| 24067 |
-
//
|
|
|
|
| 24068 |
async *generate(messages, { maxNewTokens = 1024, signal, enableThinking = false } = {}) {
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24069 |
let closed = false;
|
| 24070 |
for await (const c of this.session.generate(messages, { maxNewTokens, signal, thinking: enableThinking })) {
|
| 24071 |
const raw = String(c.text || '').replace(/^\s*<think>\n?/, '');
|
|
@@ -24096,6 +24115,15 @@ export class Ling3Flash extends AnyformerModel {
|
|
| 24096 |
static repo = 'inclusionAI/Ling-3.0-flash-GGUF';
|
| 24097 |
static budget = 7e9;
|
| 24098 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24099 |
</script>
|
| 24100 |
<script type="text/plain" data-module="errors.js">/* errors.js — coded errors for the store, ingest and the engines. A host branches on `code`, never on the
|
| 24101 |
* message text and never on `instanceof` (two copies of this module, e.g. a CDN copy and a vendored one, make
|
|
|
|
| 9307 |
// on-device storage once (~37 GB of disk). Experimental: rung 2b of the SSD-streaming ladder,
|
| 9308 |
// decode speed is bound by SSD reads (~9 tok/s on an M4 Pro), and prompts are read one token at
|
| 9309 |
// a time. Greedy output matches llama.cpp Metal on layer cuts (plan/history.md, 2026-10-04).
|
| 9310 |
+
// Since 2026-10-23 loaded through anyformer (backend 'anyformer-qwen36': Qwen36 in anyformer_models.js), which runs
|
| 9311 |
+
// diskformer's copy of the same engine; the store under localmind-ssd/ is reused.
|
| 9312 |
'qwen36-35b-a3b-ssd': {
|
| 9313 |
id: 'unsloth/Qwen3.6-35B-A3B-GGUF',
|
| 9314 |
label: 'Qwen3.6 35B-A3B · SSD-streamed (experimental)',
|
| 9315 |
+
backend: 'anyformer-qwen36',
|
| 9316 |
size: '~37 GB',
|
| 9317 |
// The pinned GGUF (QWEN36_35B_A3B in qwen35_moe_ssd.js): "Load .gguf from disk…" routes a file of exactly
|
| 9318 |
// this size here and ingests it instead of downloading. test-local-gguf-ssd.mjs keeps the two in step.
|
|
|
|
| 17779 |
'gemma4moe-ssd': { engine: 'gemma4_moe_ssd.js', exportName: 'Gemma4MoeSsd', label: 'Gemma 4 26B-A4B (Q4_0)' },
|
| 17780 |
'lfm2moe-ssd': { engine: 'lfm2_moe_ssd.js', exportName: 'Lfm2MoeSsd', label: 'LFM2.5 24B-A2B (Q4_0)' },
|
| 17781 |
'flint-ssd': { engine: 'flint_moe_ssd.js', exportName: 'FlintMoeSsd', label: 'Flint 40B-A2.5B (Q4_1)' },
|
| 17782 |
+
'anyformer-qwen36': { engine: 'anyformer_models.js', exportName: 'Qwen36', label: 'Qwen3.6 35B-A3B (Q8_0, via anyformer)' },
|
| 17783 |
'anyformer-ling3flash': { engine: 'anyformer_models.js', exportName: 'Ling3Flash', label: 'Ling 3.0 Flash (Q4_K_M, via anyformer)' },
|
| 17784 |
'anyformer-glm47flash': { engine: 'anyformer_models.js', exportName: 'Glm47Flash', label: 'GLM-4.7-Flash (Q4_K_M, via anyformer)' },
|
| 17785 |
'qwen3dense': { engine: 'qwen35_dense.js', exportName: 'Qwen3Dense', label: 'Qwen3 4B (Q8_0) + FlashMap' },
|
|
|
|
| 19398 |
const ENGINE_STORES = {
|
| 19399 |
'transformers': ['transformers-cache'], 'wllama': ['wllama'], 'lfm2-webgpu': ['gguf-v1'],
|
| 19400 |
'bonsai2-webgpu': ['gguf-cache-v1'], 'gemma4-webgpu': ['safetensors-cache-v1', 'bonsai-pipeline-v1'],
|
| 19401 |
+
'qwen35moe-ssd': ['on disk (OPFS)'], 'minicpmv-ssd': ['on disk (OPFS)'], 'gemma4moe-ssd': ['on disk (OPFS)'], 'lfm2moe-ssd': ['on disk (OPFS)'], 'flint-ssd': ['on disk (OPFS)'], 'anyformer-glm47flash': ['on disk (OPFS)'], 'anyformer-ling3flash': ['on disk (OPFS)'], 'anyformer-qwen36': ['on disk (OPFS)'], 'qwen3dense': ['on disk (OPFS)'], 'qwen35dense': ['on disk (OPFS)'],
|
| 19402 |
};
|
| 19403 |
function parseSizeBytes(size) {
|
| 19404 |
const m = /([\d.]+)\s*(GB|MB)/i.exec(String(size || ''));
|
|
|
|
| 24021 |
* anyformer picks the back end per model; LocalMind sees one call shape.
|
| 24022 |
*
|
| 24023 |
* AnyformerModel gives an anyformer session the engine interface LocalMind's SSD worker drives (moeSsdWorkerSrc):
|
| 24024 |
+
* static load(null, { maxCtx, localFile, onProgress, signal }), generate(messages, { maxNewTokens, signal,
|
| 24025 |
+
* enableThinking }) streaming { text, token }, prime(), ensureVision(), reset(), dispose(), and the think markers. One
|
| 24026 |
+
* subclass per model names its repo. anyformer loads from jsDelivr by GitHub tag until it is on npm.
|
| 24027 |
*
|
| 24028 |
+
* Two shapes of session (anyformer v0.3.0):
|
| 24029 |
+
* - a diskformer hand engine: chunks carry the engine's token ids and session.think its markers, so they pass
|
| 24030 |
+
* through unchanged and the worker handles reasoning as it does for LocalMind's own engines;
|
| 24031 |
+
* - diskformer's llama.cpp path: text only, session.think null. With reasoning on, the reply opens inside the
|
| 24032 |
+
* thought; the class drops the literal <think> and </think> from the text and marks the boundary with a chunk
|
| 24033 |
+
* whose token is THINK_CLOSE (the worker's close event).
|
| 24034 |
*/
|
| 24035 |
const THINK_CLOSE = -2;
|
| 24036 |
|
| 24037 |
+
export const ANYFORMER_URL = 'https://cdn.jsdelivr.net/gh/NakliTechie/anyformer.js@v0.3.0/index.js';
|
| 24038 |
|
| 24039 |
// anyformer's progress events → the engine events the SSD worker posts. The llama.cpp path reports 'download' (bytes
|
| 24040 |
// into OPFS) and 'load' / 'plan'; a hand engine (diskformer back end) reports its own engine events, 'weights' with
|
|
|
|
| 24051 |
export class AnyformerModel {
|
| 24052 |
static repo = null;
|
| 24053 |
static file = undefined; // a GGUF in the repo; undefined lets anyformer pick (the file diskformer checked)
|
| 24054 |
+
static budget = 4e9; // GPU bytes: the llama.cpp path plans its expert slots from it; for a hand engine, the fit check
|
| 24055 |
+
static poolBytes = undefined; // a hand engine's expert cache (LocalMind's own engines ran with 4 GiB)
|
| 24056 |
|
| 24057 |
static async load(_id, opts = {}) {
|
| 24058 |
const { load } = await import(ANYFORMER_URL);
|
| 24059 |
const session = await load(this.repo, {
|
| 24060 |
budget: opts.budget || this.budget, maxCtx: opts.maxCtx, signal: opts.signal, ...(this.file ? { file: this.file } : {}),
|
| 24061 |
+
...(this.poolBytes ? { poolBytes: this.poolBytes } : {}), ...(opts.localFile ? { localFile: opts.localFile } : {}),
|
| 24062 |
onProgress: (e) => { const ev = toEngineProgress(e); if (ev && opts.onProgress) opts.onProgress(ev); },
|
| 24063 |
});
|
| 24064 |
return new this(session);
|
| 24065 |
}
|
| 24066 |
|
| 24067 |
+
constructor(session) {
|
| 24068 |
+
this.session = session; this.backend = session.backend; this.think = session.think || null;
|
| 24069 |
+
// Present only where the engine has them (Qwen3.6: both), so the worker's own checks still work.
|
| 24070 |
+
if (session.prime) this.prime = (system) => session.prime(system);
|
| 24071 |
+
if (session.ensureVision) this.ensureVision = (o) => session.ensureVision(o);
|
| 24072 |
+
}
|
| 24073 |
|
| 24074 |
+
get thinkOpenTokenId() { return this.think ? this.think.openId : null; }
|
| 24075 |
+
get thinkCloseTokenId() { return this.think ? this.think.closeId : THINK_CLOSE; }
|
| 24076 |
+
get thinkInPrompt() { return this.think ? this.think.inPrompt ?? undefined : true; }
|
| 24077 |
+
get thinkPreamble() { return this.think ? this.think.preamble ?? '' : ''; }
|
| 24078 |
+
get alwaysThinks() { return this.think ? this.think.always === true : false; }
|
| 24079 |
|
| 24080 |
+
// A hand engine's chunks pass through. On the llama.cpp path: the reply text without its <think> / </think> tags; when
|
| 24081 |
+
// </think> first appears, the text up to it goes out, then the same text again with token THINK_CLOSE (an empty
|
| 24082 |
+
// delta), so the worker closes the thought with no text lost.
|
| 24083 |
async *generate(messages, { maxNewTokens = 1024, signal, enableThinking = false } = {}) {
|
| 24084 |
+
if (this.think) {
|
| 24085 |
+
for await (const c of this.session.generate(messages, { maxNewTokens, signal, thinking: enableThinking })) yield c;
|
| 24086 |
+
return;
|
| 24087 |
+
}
|
| 24088 |
let closed = false;
|
| 24089 |
for await (const c of this.session.generate(messages, { maxNewTokens, signal, thinking: enableThinking })) {
|
| 24090 |
const raw = String(c.text || '').replace(/^\s*<think>\n?/, '');
|
|
|
|
| 24115 |
static repo = 'inclusionAI/Ling-3.0-flash-GGUF';
|
| 24116 |
static budget = 7e9;
|
| 24117 |
}
|
| 24118 |
+
|
| 24119 |
+
// Qwen3.6 35B-A3B: unsloth's Q8_0 GGUF (36.9 GB) on diskformer's Qwen35MoeSsd hand engine, which LocalMind ran itself
|
| 24120 |
+
// until 2026-10-23 (same engine code; the store under localmind-ssd/ is the same one, so a stored copy is reused).
|
| 24121 |
+
// A 4 GiB expert cache as before; the budget covers it, the dense weights and the KV cache.
|
| 24122 |
+
export class Qwen36 extends AnyformerModel {
|
| 24123 |
+
static repo = 'unsloth/Qwen3.6-35B-A3B-GGUF';
|
| 24124 |
+
static budget = 9e9;
|
| 24125 |
+
static poolBytes = 4 * 2 ** 30;
|
| 24126 |
+
}
|
| 24127 |
</script>
|
| 24128 |
<script type="text/plain" data-module="errors.js">/* errors.js — coded errors for the store, ingest and the engines. A host branches on `code`, never on the
|
| 24129 |
* message text and never on `instanceof` (two copies of this module, e.g. a CDN copy and a vendored one, make
|