D1: Parameterize HEAD_DIM in FmhaKernel (64→512)

- Promote HEAD_DIM from module constant to constructor parameter
- FmhaKernel(head_dim=64, s_k=128, ...) — default 64 for regression
- All references to HEAD_DIM replaced with self.head_dim
- PV MMA tiler, V layout, softmax corr_tiles all parameterized
- TMEM budget warning when num_tmem_alloc_cols > 512
- New test: test_fmha_v3_stage_d1.py tests hd=64 (regression) and hd=512
- Stage C test preserved as-is for reference

tests/unit/test_fmha_v3_stage_d1.py

@@ -0,0 +1,78 @@
+"""
+FMHA v3 Stage D1: Parameterized HEAD_DIM (64 → 512).
+
+Tests the FmhaKernel class from dsv4.kernels.attention.fmha with variable head_dim.
+- HEAD_DIM=64: regression test (must match Stage C results)
+- HEAD_DIM=512: DSV4 production config (TMEM budget is the key risk)
+"""
+import torch, math, sys
+import cutlass.cute as cute
+import cutlass.torch as ct
+import cuda.bindings.driver as cuda
+from dsv4.kernels.attention.fmha import FmhaKernel
+
+def test_head_dim(hd, n_kv):
+    """Test FMHA kernel at given head_dim and KV length."""
+    m = 128  # M tile is always 128
+    torch.manual_seed(42)
+
+    q = torch.randn(m, hd, 1, dtype=torch.bfloat16, device='cuda')
+    k = torch.randn(n_kv, hd, 1, dtype=torch.bfloat16, device='cuda')
+    v = torch.randn(n_kv, hd, dtype=torch.bfloat16, device='cuda')
+    v_kernel = v.unsqueeze(-1)
+    c = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
+
+    # FP32 reference
+    qf = q[:, :, 0].float()
+    kf = k[:, :, 0].float()
+    scale = 1.0 / math.sqrt(hd)
+    attn = qf @ kf.T * scale
+    attn = torch.softmax(attn, dim=-1)
+    ref = attn @ 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_kernel).mark_layout_dynamic(leading_dim=ct.get_leading_dim(v_kernel))
+    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 = FmhaKernel(head_dim=hd, s_k=n_kv)
+    print(f'hd={hd}, n={n_kv}: 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()
+    max_abs = (out - ref).abs().max().item()
+    status = "PASS" if cos >= 0.97 else "FAIL"
+    print(f'hd={hd}, n={n_kv}: cos {cos:.6f}  max_abs {max_abs:.4f}  {status}')
+    if cos < 0.97:
+        print(f'  out[0,:4]={out[0,:4].tolist()}')
+        print(f'  ref[0,:4]={ref[0,:4].tolist()}')
+    return cos
+
+
+def test():
+    print("=== Stage D1: Parameterized HEAD_DIM ===\n")
+
+    # Regression: hd=64 must match Stage C results (cos ~0.973)
+    print("--- Regression: HEAD_DIM=64 ---")
+    cos64_128 = test_head_dim(64, 128)
+    cos64_256 = test_head_dim(64, 256)
+
+    # DSV4 production: hd=512
+    print("\n--- Production: HEAD_DIM=512 ---")
+    cos512_128 = test_head_dim(512, 128)
+
+    # Summary
+    print("\n=== Summary ===")
+    print(f"hd=64, n=128: cos={cos64_128:.6f} {'PASS' if cos64_128 >= 0.97 else 'FAIL'}")
+    print(f"hd=64, n=256: cos={cos64_256:.6f} {'PASS' if cos64_256 >= 0.97 else 'FAIL'}")
+    print(f"hd=512, n=128: cos={cos512_128:.6f} {'PASS' if cos512_128 >= 0.97 else 'FAIL'}")
+
+
+if __name__ == '__main__':
+    test()