NVIDIA: STRATA_MMQ_KQ=1 is on by default for compute capability 8.0 and up (owner yes 11-03): K-quant experts (UD-Q4/Q5/Q6_K_XL) read the prompt on the MMQ kernels

kNvKqDefaultOn is true; STRATA_MMQ_KQ=0, STRATA_NV_KQ_DEFAULT=0 and STRATA_DEFAULTS=026 turn it off, a STRATA_MMQ_KQ the user sets is kept, Volta / Turing / Pascal keep the FP16 prompt path. A pack without K-quant experts never reads the variable.
Start check (RTX PRO 6000, UD-Q4_K_XL, no STRATA variable set): the start prints the K-quant MMQ line and loads libstrata_kq; 4K prompt GPU timeline 1,047 ms (dequant 7.3%) against 1,500 ms (dequant 25.1%) with STRATA_MMQ_KQ=0, prompt 2,854 vs 2,233 tok/s (one request each).

include/strata/core/nv_defaults.hpp

@@ -36,12 +36,13 @@ std::vector<std::string> apply_nv_defaults(int major, int minor);
 
 namespace strata::core {
 
-/// NVIDIA candidate default that changes the answers (round 0212, NVSINGLE; NOT on until the owner says yes): STRATA_MMQ_KQ=1, the K-quant MMQ prompt
+/// NVIDIA default that changes the answers (round 0212, NVSINGLE; the owner said yes 11-03): STRATA_MMQ_KQ=1, the K-quant MMQ prompt
 /// kernels for the experts of Unsloth's UD-Q4/Q5/Q6_K_XL files (Q4_K / Q5_K / Q5_1 / Q6_K), which otherwise are dequantized to FP16 and multiplied by
 /// cuBLAS.  A pack without K-quant experts never reads the variable.  Cards of compute capability 8.0 and up only (measured on sm_86 and sm_120).
-/// Today the rule is behind STRATA_NV_KQ_DEFAULT=1; flipping kNvKqDefaultOn makes it a default with STRATA_MMQ_KQ=0, STRATA_NV_KQ_DEFAULT=0 and
-/// STRATA_DEFAULTS=026 as the off switches.  min_major / min_minor are the lowest compute capability among the visible cards.
-constexpr bool kNvKqDefaultOn = false;
+/// On by default; STRATA_MMQ_KQ=0, STRATA_NV_KQ_DEFAULT=0 and STRATA_DEFAULTS=026 are the off switches (kNvKqDefaultOn=false would leave it behind
+/// STRATA_NV_KQ_DEFAULT=1).  Measured: RTX PRO 6000 prompt +26..38%, A6000 +2..11%, first-token KL 0.024 / 0.035, 61-task suite 54 vs 53.
+/// min_major / min_minor are the lowest compute capability among the visible cards.
+constexpr bool kNvKqDefaultOn = true;
 std::vector<std::pair<std::string, std::string>> nv_kq_default_env(int min_major, int min_minor, bool defaults026);
 std::vector<std::pair<std::string, std::string>> nv_kq_default_env(int min_major, int min_minor);
 /// Sets what nv_kq_default_env returns where the user has not set the variable; prints one line to stderr when it set anything.

src/core/nv_defaults_test.cpp

@@ -66,9 +66,10 @@ int main() {
         CHECK(!p.empty() && p == q);
         CHECK(!find(a, "STRATA_PF_FUSED") && !find(p, "STRATA_PF_FUSED"));
     }
-    // round 0212 candidate: STRATA_MMQ_KQ=1 for compute capability 8.0+, behind STRATA_NV_KQ_DEFAULT=1 until kNvKqDefaultOn flips
+    // round 0212: STRATA_MMQ_KQ=1 is a default for compute capability 8.0+ (owner yes 11-03), with its own off switches
     unset("STRATA_NV_KQ_DEFAULT");
-    CHECK(kNvKqDefaultOn || nv_kq_default_env(12, 0, false).empty());
+    CHECK(kNvKqDefaultOn);
+    for (int major : {8, 9, 10, 12}) CHECK(nv_kq_default_env(major, 0, false).size() == 1);   // on without any switch
     put("STRATA_NV_KQ_DEFAULT", "1");
     for (int major : {8, 9, 10, 12}) {
         const auto k = nv_kq_default_env(major, 0, false);

src/program/generate.cpp

@@ -3799,7 +3799,7 @@ int main(int argc, char** argv) {
                        cudaDeviceGetAttribute(&n, cudaDevAttrComputeCapabilityMinor, d) == cudaSuccess && m == m0 && n == n0;
             }
             if (same) { strata::core::apply_nv_defaults(m0, n0); strata::core::apply_spec_adapt_default(m0, n0); }
-            {   // round 0212: the K-quant MMQ prompt kernels (flag STRATA_NV_KQ_DEFAULT=1 until the owner says yes); the lowest compute capability decides
+            {   // round 0212: the K-quant MMQ prompt kernels, on by default for compute capability 8.0+ (STRATA_MMQ_KQ=0 or STRATA_NV_KQ_DEFAULT=0 turns them off); the lowest compute capability decides
                 int lo_m = m0, lo_n = n0;
                 for (int d = 1; d < nd; ++d) {
                     int m = 0, n = 0;