diff options
| author | skal <pascal.massimino@gmail.com> | 2026-02-12 12:11:53 +0100 |
|---|---|---|
| committer | skal <pascal.massimino@gmail.com> | 2026-02-12 12:11:53 +0100 |
| commit | eaf0bd855306e70ca03f2d6579b4d6551aff6482 (patch) | |
| tree | 62316af1143db1e59e1ad62e70b9844e324cda55 /workspaces | |
| parent | e8344bc84ec0f571e5c5aafffe7c914abe226bd6 (diff) | |
TODO: 8-bit weight quantization for 2× size reduction
- Add QAT (quantization-aware training) notes
- Requires training with fake quantization
- Target: ~1.6 KB weights (vs 3.2 KB f16)
- Shader unpacking needs adaptation (4× u8 per u32)
Diffstat (limited to 'workspaces')
| -rw-r--r-- | workspaces/main/shaders/cnn_v2_compute.wgsl | 1 |
1 files changed, 1 insertions, 0 deletions
diff --git a/workspaces/main/shaders/cnn_v2_compute.wgsl b/workspaces/main/shaders/cnn_v2_compute.wgsl index f9eb556..b19a692 100644 --- a/workspaces/main/shaders/cnn_v2_compute.wgsl +++ b/workspaces/main/shaders/cnn_v2_compute.wgsl @@ -46,6 +46,7 @@ fn pack_channels(values: array<f32, 8>) -> vec4<u32> { // Get weight from storage buffer (f16 packed as u32 pairs) // Buffer layout: [header: 4 u32][layer_info: N×5 u32][weights: packed f16] +// TODO: Support 8-bit quantized weights (4× per u32) for 2× size reduction fn get_weight(idx: u32) -> f32 { // Skip header (16 bytes = 4 u32) and layer info // Weights start after header + layer_info, but weight_offset already accounts for this |