Upload h/mxfp8/README.md with huggingface_hub
Browse files- h/mxfp8/README.md +15 -2
h/mxfp8/README.md
CHANGED
|
@@ -45,11 +45,23 @@ Against the local bf16 MLX bundle at `512x512` on 12 WALDO crop images:
|
|
| 45 |
|
| 46 |
| Metric | Mean | Min |
|
| 47 |
| --- | ---: | ---: |
|
| 48 |
-
| Summary cosine | 0.
|
| 49 |
-
| Spatial cosine | 0.
|
| 50 |
|
| 51 |
This is lower precision than the 8-bit affine bundle. Treat this as experimental.
|
| 52 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
## Usage
|
| 54 |
|
| 55 |
```sh
|
|
@@ -59,6 +71,7 @@ cradio-mlx embed \
|
|
| 59 |
--image image.jpg \
|
| 60 |
--image-size 512 \
|
| 61 |
--dtype bfloat16 \
|
|
|
|
| 62 |
--save-npz embedding.npz
|
| 63 |
```
|
| 64 |
|
|
|
|
| 45 |
|
| 46 |
| Metric | Mean | Min |
|
| 47 |
| --- | ---: | ---: |
|
| 48 |
+
| Summary cosine | 0.990272 | 0.974978 |
|
| 49 |
+
| Spatial cosine | 0.988784 | 0.976665 |
|
| 50 |
|
| 51 |
This is lower precision than the 8-bit affine bundle. Treat this as experimental.
|
| 52 |
|
| 53 |
+
## Measured Speed
|
| 54 |
+
|
| 55 |
+
Fast-kernel compiled-forward MLX measurements at `512x512`, batch 1:
|
| 56 |
+
|
| 57 |
+
| Runtime | p50 latency | Throughput |
|
| 58 |
+
| --- | ---: | ---: |
|
| 59 |
+
| packed | 52.6 ms | 19.0 images/s |
|
| 60 |
+
| dequantize at load | 45.4 ms | 22.0 images/s |
|
| 61 |
+
|
| 62 |
+
`packed` keeps weights low-bit at runtime but is slower for this ViT encoder. Use
|
| 63 |
+
`--quantized-runtime dequantize` when latency matters; it expands weights to bf16 at load.
|
| 64 |
+
|
| 65 |
## Usage
|
| 66 |
|
| 67 |
```sh
|
|
|
|
| 71 |
--image image.jpg \
|
| 72 |
--image-size 512 \
|
| 73 |
--dtype bfloat16 \
|
| 74 |
+
--quantized-runtime dequantize \
|
| 75 |
--save-npz embedding.npz
|
| 76 |
```
|
| 77 |
|