P5: minimal Python multi-tile test

dsv4/kernels/attention/fmha_multitile_op.py

@@ -130,5 +130,9 @@ def fmha_multitile_decode_raw(
         ctypes.c_float(scale),
     )
     if ret != 0:
+        # Check CUDA error state
+        err = torch.cuda.current_device()
         raise RuntimeError(f"Multi-tile kernel failed: {ret}")
+    # Synchronize to catch async errors
+    torch.cuda.synchronize()
     return o, lse

tests/unit/test_p5_python_minimal.py

@@ -0,0 +1,25 @@
+"""Minimal multi-tile test via Python."""
+import torch, sys, math
+sys.path.insert(0, '/root/dsv4-nvfp4-workspace/kernel')
+
+from dsv4.kernels.attention.fmha_multitile_op import fmha_multitile_decode_raw
+
+torch.manual_seed(42)
+hd = 64
+N = 256
+scale = 1.0 / math.sqrt(hd)
+
+q = torch.randn(1, 1, 1, hd, dtype=torch.bfloat16, device='cuda').contiguous()
+k = torch.randn(1, 1, N, hd, dtype=torch.bfloat16, device='cuda').contiguous()
+v = torch.randn(1, 1, hd, N, dtype=torch.bfloat16, device='cuda').contiguous()
+
+print(f'q align: {q.data_ptr() % 128}, k align: {k.data_ptr() % 128}, v align: {v.data_ptr() % 128}')
+print(f'q shape: {q.shape}, k shape: {k.shape}, v shape: {v.shape}')
+
+try:
+    o, lse = fmha_multitile_decode_raw(q, k, v, scale)
+    print(f'Output[0,0,0,:5]: {o[0,0,0,:5].float()}')
+    print(f'LSE: {lse[0,0,0].item():.4f}')
+    print('SUCCESS')
+except Exception as e:
+    print(f'FAILED: {e}')