gdn chunk / Volta recurrence: name the sm_70-only kernels only in the host pass and the sm_70 device pass (a 60;70 or 70;86 build failed in the other device passes)

src/prefill/gdn_chunk.cu

@@ -532,7 +532,7 @@ void gdn_chunk_recurrence(float* state, const float* h, const float* gate, const
     (void) state; (void) h; (void) gate; (void) beta; (void) oc; (void) T; (void) stream;
     std::fprintf(stderr, "gdn_chunk_recurrence: not built for HIP\n");
     std::exit(1);
-#else
+#elif !defined(__CUDA_ARCH__) || __CUDA_ARCH__ >= 700   // the sm_60 device pass of a multi-arch build has no wmma kernel to name
     if (T <= 0) return;
     const int64_t T64 = T - T % CH;
     if (T64 <= 0) return;

src/prefill/kernels.cu

@@ -2116,7 +2116,7 @@ void gdn_recurrence_variant(int variant, float* state, const float* h, const flo
     (void) variant;   // the quad recurrence kernels are AMD-only
 #endif
     static const bool serial = std::getenv("STRATA_GDN_REC_HEADS") != nullptr;   // the one-block-per-head kernel (A/B)
-#if defined(STRATA_EXPERIMENTAL_SM60)
+#if defined(STRATA_EXPERIMENTAL_SM60) && (!defined(__CUDA_ARCH__) || __CUDA_ARCH__ == 700)   // the kernels exist in the host pass and the sm_70 device pass only
     // v100/gdn-chunk: the chunked recurrence (src/prefill/gdn_chunk.cu) on Volta, where the token chain is the
     // cost.  STRATA_GDN_CHUNK=0 forces the recurrence, =1 forces the chunked path (the parity test's A/B switch);
     // unset: cc == 7.0 only, every other architecture keeps the recurrence bit-for-bit as before.  Read per call