README: compare the four D1A builds
Browse files
README.md
CHANGED
|
@@ -23,6 +23,33 @@ Use it where the GPU should stay free or power matters, not for speed.
|
|
| 23 |
- Photo, voice and video (the vision and audio encoders are separate models, still to convert).
|
| 24 |
- Inputs longer than 512 tokens: 2,048-token programs give wrong answers on the M4 Neural Engine.
|
| 25 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
## Files
|
| 27 |
|
| 28 |
| Path | What |
|
|
|
|
| 23 |
- Photo, voice and video (the vision and audio encoders are separate models, still to convert).
|
| 24 |
- Inputs longer than 512 tokens: 2,048-token programs give wrong answers on the M4 Neural Engine.
|
| 25 |
|
| 26 |
+
## D1A builds compared
|
| 27 |
+
|
| 28 |
+
v0.6 is the current D1A. E2B v0.6 is not trained yet (jonpol01/d1a#229), so the E2B columns show the v0.6 plan and
|
| 29 |
+
today's E2B build, v0.2.1.
|
| 30 |
+
|
| 31 |
+
| | E2B · MLX | E2B · Core ML | E4B · MLX | E4B · Core ML |
|
| 32 |
+
|---|---|---|---|---|
|
| 33 |
+
| Status | v0.6 planned; v0.2.1 released | v0.6 planned; a v0.2.1 text pilot passed, not uploaded | **v0.6 released** | **v0.6 uploaded** (private) |
|
| 34 |
+
| Repository | JohnP1/d1a-e2b-mlx-q8 (public) | JohnP1/d1a-e2b-coreml (private) | JohnP1/d1a-e4b-mlx-q8 (public) | JohnP1/d1a-e4b-coreml (private) |
|
| 35 |
+
| Runs on | Mac GPU | Neural Engine: iPhone, iPad, Mac | Mac GPU | Neural Engine (tested on Macs only) |
|
| 36 |
+
| Text | yes | yes | yes | yes |
|
| 37 |
+
| Photo | yes | planned for v0.6 | yes | not yet |
|
| 38 |
+
| Voice | yes | planned for v0.6 | yes | not yet |
|
| 39 |
+
| Video (16 frames, no sound) | yes | planned for v0.6 | yes | not yet |
|
| 40 |
+
| Input length | long documents | ≤ 512 tokens | long documents (trained up to 5,120 tokens) | ≤ 512 tokens |
|
| 41 |
+
| Questions | yes/no, choice, score, with a calibrated probability for each option | same | same | same |
|
| 42 |
+
| Languages | questions in English; documents in English or Japanese | same | same | same |
|
| 43 |
+
| PR labeling (v0.3–v0.6 skills) | arrives with v0.6 | arrives with v0.6 (PRs ≤ 512 tokens) | yes | yes, for PRs that fit in 512 tokens |
|
| 44 |
+
| API | System One API (`d1a.serving.serve`), with `/v1/systemone/media` | Python reference only; no iOS app yet | System One API, served on the Mac mini | Python reference only (one process per chunk) |
|
| 45 |
+
| Size | 3.8 GB + 1.0 GB media | not built | 6.6 GB + 1.0 GB media | 6.1 GB (3.9 GB chunks + 2.1 GB embeddings) |
|
| 46 |
+
| Text speed (median) | ~0.4 s per 6-question request, M1 Max | pilot: ~0.95 s at 512 tokens, M4 | 0.63 s M4, 0.8 s M1 Max ¹ | 1.8 s M4, 2.9 s M1 Max ¹; load 52 s / 106 s |
|
| 47 |
+
| Media speed (median, M1 Max, v0.2.1) | photo 0.57 s, voice 0.19 s, video 1.94 s | not built | not measured | not built |
|
| 48 |
+
| Answer changes vs full precision | 5 of 274 (v0.2, vs PyTorch bf16) | pilot: 0 clear-cut | 0 of 38 clear-cut ¹ | 0 of 38 clear-cut ¹ |
|
| 49 |
+
|
| 50 |
+
¹ The same 28 real requests (42 questions, all ≤ 512 tokens), against the fp32 PyTorch build of v0.6. "Clear-cut"
|
| 51 |
+
means the reference's top answer leads by more than 0.05.
|
| 52 |
+
|
| 53 |
## Files
|
| 54 |
|
| 55 |
| Path | What |
|