ROCm/composable_kernel
1
0
Fork 0
mirror of https://github.com/ROCm/composable_kernel.git synced 2026-05-15 02:27:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge commit '232523d9fa70a9ce85573bc714078181f2ffc11e' into develop

Browse Source
This commit is contained in:
assistant-librarian[bot]
2025-10-16 01:40:26 +00:00
parent d10f7d00a4
commit a17e7e734b
1 changed files with 10 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
example/ck_tile/38_block_scale_gemm/README.md
View File

@@ -4,9 +4,18 @@ This folder contains examples of quant GEMMs using the ck_tile tile-programming
- AQuant kernel with blocks of A matrix sharing scales: custom GEMM pipeline
- BQuant kernel with blocks of B matrix sharing scales: custom GEMM pipeline
- Row and Column-wise scaled: All of the rowwise elements in A Matrix and columwise elements in B Matrix will share the same quantization element and the elementwisde operation will complete in epilogue.
- Row and Column-wise scaled: All of the row-wise elements in A Matrix and column-wise elements in B Matrix will share the same quantization element and the element-wise operation will complete in epilogue.
- Tensor-wise scaled: Share the same scalar scale across the whole tensor of A or B
## Quantization Mode Comparison
| Quant Mode | A Matrix Organization | A Scale Shape | B Matrix Organization | B Scale Shape |
|------------|----------------------|---------------|----------------------|---------------|
| **AQuant** | Blocks along K dimension<br/>Each M×GroupSize block shares one scale | `[M, K/GroupSize]` | Not quantized | N/A |
| **BQuant** | Not quantized | N/A | Blocks along K dimension<br/>Each GroupSize×N block shares one scale | `[K/GroupSize, N]` |
| **RowColQuant** | Per-row quantization<br/>All K elements in each row share one scale | `[M, 1]` | Per-column quantization<br/>All K elements in each column share one scale | `[1, N]` |
| **TensorQuant** | Tensor-wise quantization<br/>All M×K elements share one scale | `[1]` | Tensor-wise quantization<br/>All K×N elements share one scale | `[1]` |
---
## Features