--- a/src/hipconv/cmake/GpuArch.cmake
+++ b/src/hipconv/cmake/GpuArch.cmake
@@ -85,11 +85,12 @@
     # Which sources compile never varies, so the host code is the same in every
     # build and only the offload bundles differ, as RFC0008 sharding requires. An
     # architecture the build serves no GPU for still needs a non-empty
-    # HIP_ARCHITECTURES; it takes the first GPU in the build and its bundles go
-    # unused. See docs/multi-arch-convergence.md.
+    # HIP_ARCHITECTURES; it takes its own first GPU and its bundles go unused.
+    # (Compiling for a foreign GPU would trip target-feature checks on the
+    # arch-specific builtins, e.g. gfx90a MFMA on an RDNA target.)
     set(offload_gpus "${served_gpus}")
-    if(NOT offload_gpus AND CMAKE_HIP_ARCHITECTURES)
-        list(GET CMAKE_HIP_ARCHITECTURES 0 offload_gpus)
+    if(NOT offload_gpus)
+        list(GET ARG_TARGETS 0 offload_gpus)
     endif()
     if(offload_gpus)
         set_target_properties(hipconv_arch_${name} PROPERTIES
--- a/src/hipconv/src/arch/cdna4/grouped/mfma_dispatch.h
+++ b/src/hipconv/src/arch/cdna4/grouped/mfma_dispatch.h
@@ -19,9 +19,10 @@
 {
     return __builtin_amdgcn_mfma_f32_16x16x16f16(a, b, c, 0, 0, 0);
 }
+// The _1k bf16 builtins take i16 vectors; __bf16x4 is bit-identical.
 __device__ inline fp32x4_t mfma_16x16x16(bf16x4_t a, bf16x4_t b, fp32x4_t c)
 {
-    return __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(a, b, c, 0, 0, 0);
+    return __builtin_amdgcn_mfma_f32_16x16x16bf16_1k(__builtin_bit_cast(int16x4_t, a), __builtin_bit_cast(int16x4_t, b), c, 0, 0, 0);
 }
 
 __device__ inline fp32x4_t mfma_4x4x4_16b(fp16x4_t a, fp16x4_t b, fp32x4_t c)
@@ -30,7 +31,7 @@
 }
 __device__ inline fp32x4_t mfma_4x4x4_16b(bf16x4_t a, bf16x4_t b, fp32x4_t c)
 {
-    return __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(a, b, c, 0, 0, 0);
+    return __builtin_amdgcn_mfma_f32_4x4x4bf16_1k(__builtin_bit_cast(int16x4_t, a), __builtin_bit_cast(int16x4_t, b), c, 0, 0, 0);
 }
 
 // TF32 (simulated): inputs are fp32, computed via 3x BF16 MFMA.
--- a/src/hipconv/src/bunnies.hpp
+++ b/src/hipconv/src/bunnies.hpp
@@ -20,6 +20,9 @@
 using uint32x4  = __attribute__((ext_vector_type(4))) uint32_t;
 using uint32x8  = __attribute__((ext_vector_type(8))) uint32_t;
 using uint32x16 = __attribute__((ext_vector_type(16))) uint32_t;
+using int32x4   = __attribute__((ext_vector_type(4))) int32_t;
+using int32x8   = __attribute__((ext_vector_type(8))) int32_t;
+using int32x16  = __attribute__((ext_vector_type(16))) int32_t;
 using floatx4   = __attribute__((ext_vector_type(4))) float;
 using floatx8   = __attribute__((ext_vector_type(8))) float;
 using fp16x2    = __attribute__((ext_vector_type(2))) _Float16;
--- a/src/hipconv/src/arch/cdna4/bunnies_cdna4.hpp
+++ b/src/hipconv/src/arch/cdna4/bunnies_cdna4.hpp
@@ -238,8 +238,9 @@
                                     matrix<fpfmt::e8m23, 16, 16, use::Acc>& c)
         {
             constexpr int scale = 0;
+            // The builtin takes i32 vectors; uint32x8 storage is bit-identical.
             d.data              = __builtin_amdgcn_mfma_scale_f32_16x16x128_f8f6f4(
-                a.data, b.data, c.data, 0, 0, 0, scale, 0, scale);
+                __builtin_bit_cast(int32x8, a.data), __builtin_bit_cast(int32x8, b.data), c.data, 0, 0, 0, scale, 0, scale);
         }
     };
 
--- a/src/hipconv/src/arch/cdna5/bunnies_mi400.hpp
+++ b/src/hipconv/src/arch/cdna5/bunnies_mi400.hpp
@@ -236,8 +236,9 @@
                                     matrix<fpfmt::e4m3, 128, 16, use::B>& b,
                                     matrix<fpfmt::e8m23, 16, 16, use::Acc>& c)
         {
+            // The builtin takes i32 vectors; uint32x16 storage is bit-identical.
             d.data = __builtin_amdgcn_wmma_f32_16x16x128_fp8_fp8(
-                a.data, b.data, 0, c.data, A_reuse, B_reuse);
+                __builtin_bit_cast(int32x16, a.data), __builtin_bit_cast(int32x16, b.data), 0, c.data, A_reuse, B_reuse);
         }
         __device__ static void wmma_scale(matrix<fpfmt::e8m23, 16, 16, use::Acc>& d,
                                           matrix<fpfmt::e4m3, 16, 128, use::A>& a,
@@ -247,9 +248,9 @@
                                           matrix<fpfmt::ue8m0, 4, 16, use::B>& b_scale)
         {
             d.data = __builtin_amdgcn_wmma_scale_f32_16x16x128_f8f6f4(0,
-                                                                      a.data,
+                                                                      __builtin_bit_cast(int32x16, a.data),
                                                                       0,
-                                                                      b.data,
+                                                                      __builtin_bit_cast(int32x16, b.data),
                                                                       0,
                                                                       c.data,
                                                                       0,
@@ -277,9 +278,9 @@
                     constexpr int opsel_b = nb;
                     constexpr int opsel_a = mb;
                     *dsub                 = __builtin_amdgcn_wmma_scale_f32_16x16x128_f8f6f4(0,
-                                                                             *asub,
+                                                                             __builtin_bit_cast(int32x16, *asub),
                                                                              0,
-                                                                             *bsub,
+                                                                             __builtin_bit_cast(int32x16, *bsub),
                                                                              0,
                                                                              *csub,
                                                                              opsel_a,
@@ -878,6 +879,17 @@
         uint32x8 data = 0;
     };
 
+    // The tensor builtin takes signed int vectors; the descriptor groups
+    // store uint32 vectors. Bit-identical.
+    __device__ inline static void tdm_load_to_lds(uint32x4 d0, uint32x8 d1, uint32x4 d2,
+                                                  uint32x4 d3, uint32x8 d4)
+    {
+        __builtin_amdgcn_tensor_load_to_lds(d0, __builtin_bit_cast(int32x8, d1),
+                                            __builtin_bit_cast(int32x4, d2),
+                                            __builtin_bit_cast(int32x4, d3),
+                                            __builtin_bit_cast(int32x8, d4), 0);
+    }
+
     enum class scope : int
     {
         WGP = 0,
--- a/src/hipconv/src/arch/cdna5/tdm_desc.h
+++ b/src/hipconv/src/arch/cdna5/tdm_desc.h
@@ -49,7 +49,7 @@
         d0.lds_addr = lds_offset_bytes;
         d1.set_tensor_dim1(tensor_dim1);
         d1.tile_dim1 = tile_dim1;
-        __builtin_amdgcn_tensor_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data, 0);
+        bunnies::arch_mi400::tdm_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data);
     }
 };
 
--- a/src/hipconv/src/arch/cdna5/direct/kernel.hpp
+++ b/src/hipconv/src/arch/cdna5/direct/kernel.hpp
@@ -387,7 +387,7 @@
     __syncthreads();
     // Pre-load; no pre-load for wave_id_j=1
     if(wave_id_j == 0)
-        __builtin_amdgcn_tensor_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data, 0);
+        bunnies::arch_mi400::tdm_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data);
 
     // Input LDS layout for bn::lds_load
     auto const in_lds_layout = [&](int buf, int r, int s, int part_c) {
@@ -462,8 +462,7 @@
                     d0.set_global_addr(tdm_base_addr + cnext);
                     d1.set_tensor_dim0(c_per_group - cnext);
                     d0.set_lds_addr(tdm_lds_addr + tdm_lds_stride * (buf_in ^ 1));
-                    __builtin_amdgcn_tensor_load_to_lds(
-                        d0.data, d1.data, d2.data, d3.data, d4.data, 0);
+                    bunnies::arch_mi400::tdm_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data);
                 }
             }
             __builtin_amdgcn_s_barrier();
@@ -483,8 +482,7 @@
                             d2.set_tensor_dim2(0 <= h && h < hi - h0 ? 1 : 0);
                         const int hring = h % tile_size_h_ring;
                         d0.set_lds_addr(tdm_lds_addr + in_lds_view(hring, 0, 0, buf_in, 0));
-                        __builtin_amdgcn_tensor_load_to_lds(
-                            d0.data, d1.data, d2.data, d3.data, d4.data, 0);
+                        bunnies::arch_mi400::tdm_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data);
                     }
                 }
 #pragma nounroll
@@ -514,8 +512,7 @@
                                     d1.set_tensor_dim0(rem_c);
                                 }
                                 d0.set_lds_addr(tdm_lds_addr + tdm_lds_stride * (buf_wei ^ 1));
-                                __builtin_amdgcn_tensor_load_to_lds(
-                                    d0.data, d1.data, d2.data, d3.data, d4.data, 0);
+                                bunnies::arch_mi400::tdm_load_to_lds(d0.data, d1.data, d2.data, d3.data, d4.data);
                             }
                         }
                     }
