Uploaded using `kernel-builder`.
Browse files
README.md
ADDED
|
@@ -0,0 +1,41 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: kernels
|
| 3 |
+
license: other
|
| 4 |
+
---
|
| 5 |
+
|
| 6 |
+
This is the repository card of kernels-community/msa that has been pushed on the Hub. It was built to be used with the [`kernels` library](https://github.com/huggingface/kernels). This card was automatically generated.
|
| 7 |
+
|
| 8 |
+
## How to use
|
| 9 |
+
|
| 10 |
+
```python
|
| 11 |
+
# make sure `kernels` is installed: `pip install -U kernels`
|
| 12 |
+
from kernels import get_kernel
|
| 13 |
+
|
| 14 |
+
kernel_module = get_kernel("kernels-community/msa", version=0)
|
| 15 |
+
sparse_atten_func = kernel_module.sparse_atten_func
|
| 16 |
+
|
| 17 |
+
sparse_atten_func(...)
|
| 18 |
+
```
|
| 19 |
+
|
| 20 |
+
## Available functions
|
| 21 |
+
- `sparse_atten_func`
|
| 22 |
+
- `sparse_atten_nvfp4_kv_func`
|
| 23 |
+
- `sparse_decode_atten_func`
|
| 24 |
+
- `SparseDecodePagedAttentionWrapper`
|
| 25 |
+
- `fp4_indexer_block_scores`
|
| 26 |
+
- `build_k2q_csr`
|
| 27 |
+
- `SparseK2qCsrBuilderSm100`
|
| 28 |
+
- `Nvfp4QuantizedTensor`
|
| 29 |
+
- `quantize_bf16_to_nvfp4_128x4`
|
| 30 |
+
- `quantize_kv_bf16_to_nvfp4_128x4`
|
| 31 |
+
- `dequantize_nvfp4_128x4_to_bf16`
|
| 32 |
+
- `swizzle_nvfp4_scale_to_128x4`
|
| 33 |
+
- `nvfp4_global_scale_from_amax`
|
| 34 |
+
|
| 35 |
+
## Benchmarks
|
| 36 |
+
|
| 37 |
+
No benchmark available yet.
|
| 38 |
+
|
| 39 |
+
## Source code
|
| 40 |
+
|
| 41 |
+
Source code of this kernel originally comes from https://github.com/MiniMax-AI/MSA and it was repurposed for compatibility with `kernels`.
|