Update CARD.md
Browse files
CARD.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: kernels
|
| 3 |
+
license: bsd-3-clause
|
| 4 |
+
tags: [cuda, blackwell, flash-attention, cute-dsl, inference]
|
| 5 |
+
---
|
| 6 |
+
|
| 7 |
+
# flashrt/fa4-cute-runtime
|
| 8 |
+
|
| 9 |
+
Forward-only FlashAttention-4 CuTe DSL runtime used by FlashRT's GROOT N1.7
|
| 10 |
+
and PI0.5 Thor pipelines. The source is vendored under the private `flashrt_fa4`
|
| 11 |
+
namespace and does not shadow an installed `flash_attn` package.
|
| 12 |
+
|
| 13 |
+
This package adds the dedicated Blackwell D256 2CTA forward path required by
|
| 14 |
+
PI0.5's 8-Q/1-KV-head PaliGemma encoder. The community FlashAttention-4 package
|
| 15 |
+
already covers the D48/D72/D128 family; it does not currently expose D256 on
|
| 16 |
+
SM100/SM110.
|
| 17 |
+
|
| 18 |
+
## Functions
|
| 19 |
+
|
| 20 |
+
- `flash_attn_func`
|
| 21 |
+
- `flash_attn_varlen_func`
|
| 22 |
+
- `forward_static`
|
| 23 |
+
|
| 24 |
+
```python
|
| 25 |
+
from kernels import get_kernel
|
| 26 |
+
|
| 27 |
+
fa4 = get_kernel("flashrt/fa4-cute-runtime", version=1)
|
| 28 |
+
out = fa4.flash_attn_func(q, k, v, causal=False)
|
| 29 |
+
```
|
| 30 |
+
|
| 31 |
+
The vendored forward wrapper returns `(out, lse)`; use `result[0]` when only
|
| 32 |
+
the attention output is needed.
|
| 33 |
+
|
| 34 |
+
For a CUDA Graph hot path, preallocate the output and use the allocation-free
|
| 35 |
+
entry point:
|
| 36 |
+
|
| 37 |
+
```python
|
| 38 |
+
out = torch.empty_like(q)
|
| 39 |
+
fa4.forward_static(q, k, v, out, causal=False)
|
| 40 |
+
```
|
| 41 |
+
|
| 42 |
+
For a padded fixed-shape graph, pass the valid K/V length as a CUDA int32
|
| 43 |
+
tensor. PI0.5 uses this form for its encoder cache:
|
| 44 |
+
|
| 45 |
+
```python
|
| 46 |
+
seqused_k = torch.tensor([valid_k], device="cuda", dtype=torch.int32)
|
| 47 |
+
fa4.forward_static(
|
| 48 |
+
q, k_padded, v_padded, out,
|
| 49 |
+
causal=False,
|
| 50 |
+
pack_gqa=True,
|
| 51 |
+
seqused_k=seqused_k,
|
| 52 |
+
)
|
| 53 |
+
```
|
| 54 |
+
|
| 55 |
+
Inputs follow FlashAttention's `(batch, sequence, heads, head_dim)` contract.
|
| 56 |
+
Qualified model profiles include D72 MHA and D256 GQA (8 Q heads / 1 KV head),
|
| 57 |
+
with both dense and `seqused_k` execution. This package targets SM100-family
|
| 58 |
+
Blackwell forward inference and requires
|
| 59 |
+
CUDA 13 plus `nvidia-cutlass-dsl` 4.4.x, 4.5.x, or 4.6.x. The wrapper selects Thor's
|
| 60 |
+
accepted architecture alias according to the installed DSL version.
|
| 61 |
+
|
| 62 |
+
This is an execution backend rather than a universal SDPA replacement. Select
|
| 63 |
+
it with model-shape profiling; the GROOT causal GQA profile benefits while some
|
| 64 |
+
short vision profiles remain faster on PyTorch SDPA.
|