Update README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# fa4-cute-runtime
|
| 2 |
+
|
| 3 |
+
Forward-only, SM100-family FlashAttention-4 CuTe DSL package used by FlashRT's
|
| 4 |
+
production GROOT N1.7 and PI0.5 Thor runtimes. In addition to the standard
|
| 5 |
+
D48/D72/D128 path, it includes the dedicated D256 2CTA forward kernel needed
|
| 6 |
+
by PI0.5's PaliGemma encoder.
|
| 7 |
+
|
| 8 |
+
The package exposes `flash_attn_func`, `flash_attn_varlen_func`, and the
|
| 9 |
+
preallocated `forward_static` CUDA Graph entry, including `seqused_k` for
|
| 10 |
+
fixed-shape padded K/V. It keeps the vendored
|
| 11 |
+
implementation in the private `flashrt_fa4` namespace and vendors the required
|
| 12 |
+
quack helper subset, so it cannot replace or mutate upstream `flash_attn`
|
| 13 |
+
imports.
|
| 14 |
+
|
| 15 |
+
See [CARD.md](CARD.md) for usage and [VALIDATION.md](VALIDATION.md) for the
|
| 16 |
+
release gate.
|