chanderbalaji commited on
Commit
3506f28
·
verified ·
1 Parent(s): c9c0f7e

Update benchmark notes

Browse files
Files changed (1) hide show
  1. README.md +21 -4
README.md CHANGED
@@ -31,11 +31,28 @@ model remains in the upstream Hugging Face repository.
31
 
32
  | Variant | Folder | Quantization | Size | Best fit |
33
  | --- | --- | --- | ---: | --- |
34
- | MLX 8-bit | `mlx-8bit/` | affine, group size 64 | 34 GB | Highest-quality local MLX run. |
35
- | MLX 6-bit | `mlx-6bit/` | affine, group size 64 | 26 GB | Balanced memory and quality. |
36
- | MLX 4-bit | `mlx-4bit/` | affine, group size 64 | 18 GB | Smallest footprint and easiest local run. |
37
 
38
- Benchmarks will be added later.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
39
 
40
  ## Usage
41
 
 
31
 
32
  | Variant | Folder | Quantization | Size | Best fit |
33
  | --- | --- | --- | ---: | --- |
34
+ | MLX 8-bit | `mlx-8bit/` | affine, group size 64 | 36.85 GB | Highest-quality local MLX run. |
35
+ | MLX 6-bit | `mlx-6bit/` | affine, group size 64 | 28.19 GB | Balanced memory and quality. |
36
+ | MLX 4-bit | `mlx-4bit/` | affine, group size 64 | 19.53 GB | Smallest footprint and easiest local run. |
37
 
38
+ ## Local benchmark notes
39
+
40
+ Initial local testing was performed on a Mac Studio with an M4 Max and 64 GB
41
+ unified memory using oMLX.
42
+
43
+ | Variant | Result | Notes |
44
+ | --- | --- | --- |
45
+ | MLX 8-bit | Not loaded under default memory cap | oMLX projected 54.62 GB total memory use against a 51.84 GB effective ceiling. The model files themselves are 36.85 GB; the higher runtime estimate includes the current oMLX process footprint, MLX runtime/allocator overhead, buffers, and KV/cache planning. |
46
+
47
+ The 8-bit variant should be retested after raising the Apple GPU wired-memory
48
+ cap and restarting the local serving process, for example:
49
+
50
+ ```bash
51
+ sudo sysctl iogpu.wired_limit_mb=59392
52
+ ```
53
+
54
+ Throughput numbers are not published yet. This section will be updated after a
55
+ successful full benchmark run.
56
 
57
  ## Usage
58