fix: unblock Thor validation (vocab-ce API, sageattention2 gate, fp8-cross-attention docs)

flashrt-vocab-ce-train/torch-ext/flashrt_vocab_ce_train/__init__.py

@@ -162,10 +162,21 @@ def vocab_ce(
     )
 
 
+def vocab_ce_loss(
+    hidden: torch.Tensor,
+    weight: torch.Tensor,
+    labels: torch.Tensor,
+    z_loss_weight: float = 0.0,
+    ignore_index: int = -100,
+) -> torch.Tensor:
+    """Documented public alias of :func:`vocab_ce` (README/CARD API)."""
+    return vocab_ce(hidden, weight, labels, z_loss_weight, ignore_index)
+
+
 def backend_marker(x: torch.Tensor) -> torch.Tensor:
     if ops is None:
         return x
     return ops._flashrt_training_package_marker(x)
 
 
-__all__ = ["vocab_ce", "reference_vocab_ce", "backend_marker"]
+__all__ = ["vocab_ce", "vocab_ce_loss", "reference_vocab_ce", "backend_marker"]

fp8-cross-attention-blackwell/VALIDATION.md

@@ -20,3 +20,28 @@ both sides of every 128-token boundary are release blockers.
 
 Installed-artifact correctness and `torch.compile(fullgraph=True)` remain
 mandatory after the HF Jobs build. Source validation does not replace them.
+
+## Installed artifact on NVIDIA Thor (SM110)
+
+The installed-artifact gate (missing when this card was written) is now closed
+on Thor `sm_110a`. Build notes:
+
+- Requires CUTLASS **4.4.0** (the package declares `cutlass_4_4`); 4.5.x moved
+  `SM100_MMA_F8F6F4_SS` to a class template and breaks `csrc/fmha77`.
+- `cutlass/util/packed_stride.hpp` is vendored by `fp4-gemm` and must be on the
+  include path (same local-build workaround as `fp8-gemm`).
+- Compiled with `-gencode arch=compute_110a` and the FMHA sm100 kernel runs
+  natively on Thor (no fallback required, unlike `fused-mlp-megakernels`).
+
+Command:
+
+```bash
+python fp8-cross-attention-blackwell/tests/test_fp8_cross_attention_blackwell.py \
+    --backend installed --artifact <installed-dir> --mode full
+```
+
+Result: 9/9 numeric rows, invalid-head rejection, CUDA Graph replay, and
+`torch.compile(fullgraph=True)` all passed on SM110. Worst row across the
+matrix (`B1,Sq786,Sk7984,Hq28,Hkv4,D128`): `max=0.00025749`, `p99=0.00012207`,
+`mean=0.00003710`, `cosine=0.99978602` — comfortably inside the
+`max<=0.004 / cosine>=0.9995` gate. Other rows held `cosine>=0.999778`.

sageattention2-blackwell/tests/test_sageattention2_blackwell.py

@@ -221,8 +221,8 @@ def main() -> None:
 
     if not torch.cuda.is_available():
         raise SystemExit("CUDA is required")
-    major, _minor = torch.cuda.get_device_capability(0)
-    if major < 12:
+    major, minor = torch.cuda.get_device_capability(0)
+    if major not in (10, 11, 12) or (major == 10 and minor not in (0, 3)):
         raise SystemExit("sageattention2-blackwell requires Blackwell-class CUDA capability")
 
     torch.manual_seed(2026)