fix: avoid cuda_bf16.h, use inline PTX for BF16 conversion

dsv4/kernels/attention/fmha_sm100.cuh

@@ -30,9 +30,25 @@
 #pragma once
 
 #include <cuda_runtime.h>
-#include <cuda_bf16.h>
 #include <cstdint>
 
+// NOTE: cuda_bf16.h is NOT included due to CUDA 13.2 C++17 compatibility bug
+// (in_place_from error). __nv_bfloat16 is a built-in CUDA type and works
+// without the header. We define helper functions manually.
+
+// BF16 conversion helpers
+__device__ __forceinline__ __nv_bfloat16 float_to_bfloat16(float f) {
+    __nv_bfloat16 h;
+    asm("cvt.rn.bf16.f32 %0, %1;" : "=h"(h) : "f"(f));
+    return h;
+}
+
+__device__ __forceinline__ float bfloat16_to_float(__nv_bfloat16 h) {
+    float f;
+    asm("cvt.f32.bf16 %0, %1;" : "=f"(f) : "h"(h));
+    return f;
+}
+
 // CUTLASS C++ includes (CUDA device code only)
 #if defined(__CUDA_ARCH__)
 #include <cutlass/cutlass.h>
@@ -559,7 +575,7 @@ fmha_decode_kernel(
                     // Simplified: directly compute and write
                     // (Proper TMEM load + normalize below)
                     float o_val = 0.0f; // placeholder
-                    o_head[j] = __float2bfloat16(o_val / row_sum);
+                    o_head[j] = float_to_bfloat16(o_val / row_sum);
                 }
             }
         }