diskformer 0.6.10 → 0.6.12
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/CHANGELOG.md +18 -0
- package/package.json +1 -1
package/CHANGELOG.md
CHANGED
|
@@ -1,5 +1,23 @@
|
|
|
1
1
|
# Changelog
|
|
2
2
|
|
|
3
|
+
## 0.6.12 (2026-10-30)
|
|
4
|
+
- LFM3-Omni-2B speaks: `generate(messages, { output_modalities: ['text', 'audio'], onAudio })` runs Liquid's
|
|
5
|
+
interleaved mode (6 text tokens, 12 speech frames in turn), the depth decoder's MaskGIT (12 codes a frame) and the
|
|
6
|
+
codec decoder, then an inverse STFT to 24 kHz Float32 PCM, streamed about once a second. Options `audio_temperature`
|
|
7
|
+
(0.4), `audio_top_k`, `audio_repetition_penalty` (1.2), `audio_repetition_window` (64), `audio_chunk_frames`, `seed`;
|
|
8
|
+
`load({ speech: true })` fetches the two speech graphs (depth_decoder 336 MB, codec_decoder 179 MB, fp32) up front,
|
|
9
|
+
else the first speech call does. Text-only output is unchanged and stays the default.
|
|
10
|
+
- Gate `test/browser/lfm3-omni-speech.html`: text identical to Liquid's runner slot for slot; speech codes exact for
|
|
11
|
+
the first frames, then fp16 near-ties on WebGPU flip a code and later frames differ (M4 Pro: frames 0-6 exact), so
|
|
12
|
+
the gate asks for 4 exact frames, the same length and RMS within 10%; a sampled 13.5 s answer took 9.6 s, first
|
|
13
|
+
audio after 2.3 s (speech prompt included). The inverse STFT matches numpy to 5e-10 of peak. `--wav` writes the
|
|
14
|
+
replies for listening.
|
|
15
|
+
|
|
16
|
+
## 0.6.11 (2026-10-30)
|
|
17
|
+
- Republishes 0.6.10's tagged files. npm 0.6.10 was packed from a working tree that held uncommitted, unchecked
|
|
18
|
+
speech-output work in `engines/lfm3_omni.js` (854 lines against the tag's 628); 0.6.11 carries the tag's engine.
|
|
19
|
+
No code change from v0.6.10.
|
|
20
|
+
|
|
3
21
|
## 0.6.10 (2026-10-30)
|
|
4
22
|
- LFM3-Omni-2B (`engines/lfm3_omni.js`, `Lfm3Omni`): LiquidAI's ONNX export (decoder_q4f16, vision_encoder_q8,
|
|
5
23
|
audio_encoder; 2.14 GB) on onnxruntime-web 1.30.0's WebGPU EP, because transformers.js 4.3.2 has no `lfm3_omni`
|
package/package.json
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "diskformer",
|
|
3
|
-
"version": "0.6.
|
|
3
|
+
"version": "0.6.12",
|
|
4
4
|
"description": "Run a model larger than your GPU memory in a browser tab: weights stay in OPFS, WebGPU holds only what each step needs. llama.cpp with paged MoE experts (diskformer/llama), a disk tier, and hand-written MoE engines.",
|
|
5
5
|
"type": "module",
|
|
6
6
|
"license": "MIT",
|