Quick test n=128,256

tests/unit/test_quick.py

@@ -0,0 +1,30 @@
+"""Quick test: n=128, 256 only"""
+import torch, math, cutlass, cutlass.cute as cute, cutlass.torch as ct, cuda.bindings.driver as cuda
+from cutlass import Float32, BFloat16, Int32
+from test_fmha_v3_stage_c import FmhaV3StageCMulti
+
+HEAD_DIM = 64
+for n in [128, 256]:
+    torch.manual_seed(42)
+    q = torch.randn(128, HEAD_DIM, 1, dtype=torch.bfloat16, device='cuda')
+    k = torch.randn(n, HEAD_DIM, 1, dtype=torch.bfloat16, device='cuda')
+    v = torch.randn(n, HEAD_DIM, dtype=torch.bfloat16, device='cuda')
+    c = torch.zeros(128, HEAD_DIM, 1, dtype=torch.bfloat16, device='cuda')
+    qf = q[:,:,0].float(); kf = k[:,:,0].float()
+    ref = torch.softmax(qf @ kf.T * (1.0/math.sqrt(HEAD_DIM)), dim=-1) @ v.float()
+    mQ = ct.from_dlpack(q).mark_layout_dynamic(leading_dim=ct.get_leading_dim(q))
+    mK = ct.from_dlpack(k).mark_layout_dynamic(leading_dim=ct.get_leading_dim(k))
+    mV = ct.from_dlpack(v.unsqueeze(-1)).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v.unsqueeze(-1)))
+    mC = ct.from_dlpack(c).mark_layout_dynamic(leading_dim=ct.get_leading_dim(c))
+    stream = cuda.CUstream(torch.cuda.current_stream().cuda_stream)
+    kernel = FmhaV3StageCMulti(s_k=n)
+    print(f'n={n}: Compiling...', flush=True)
+    compiled = cute.compile(kernel, mQ, mK, mV, mC, stream)
+    compiled(mQ, mK, mV, mC, stream)
+    torch.cuda.synchronize()
+    out = c[:,:,0].float()
+    cos = torch.nn.functional.cosine_similarity(out.flatten().unsqueeze(0), ref.flatten().unsqueeze(0)).item()
+    print(f'n={n} ({n//128} tiles): cos {cos:.6f}  {"PASS" if cos >= 0.99 else "FAIL"}')
+    if cos < 0.99 and cos > 0.5:
+        ratio = (out[0,:4] / ref[0,:4]).mean().item()
+        print(f'  out/ref ratio: {ratio:.4f}')