cuda: a keyed graph cache and a device arena, off until a family installs them; failed launches name their kernel (DeepSeek-V4.1 stack 2/14)

RUNBOOK.md

@@ -168,6 +168,10 @@ The server logs `heat_wait_s`, a cancel ends the wait, and the reply's tokens do
 On three GB10s serving GLM-5.3-Flash NVFP4 at TP=3, a 1,035,295-token prompt without the bands kept one host's hottest
 zone at 94 C or more for most of the prompt (97.7 C at the top); `TF_HEAT_HIGH=92` and `TF_HEAT_LOW=88` held all three
 at 93.0 C or under, and the first token came 10% later.
+`TF_GRAPH_CACHE=0` (or `off`, any case) runs every step eagerly in any family that installs the keyed CUDA graph cache,
+without a rebuild. Set it alike on every tensor-parallel rank: an off rank never joins the cache's per-miss vote, so a
+rank left on would wait at its first miss, and only the off rank logs the switch. Ranks that share a cache compare
+`Settings.digest` of their effective settings at start, before building it, and refuse to start when they differ.
 Archives whose names contain `host-only` are CPU verification artifacts and cannot serve CUDA inference.
 
 ## Flash Next and paired Metal serving

zig/src/cuda/abi.zig

@@ -39,6 +39,8 @@ pub const DeviceAttribute = enum(c_int) {
     compute_capability_minor = 76,
     max_shared_memory_per_multiprocessor = 81,
     max_shared_memory_per_block_optin = 97,
+    virtual_memory_management_supported = 102,
+    gpu_direct_rdma_with_cuda_vmm_supported = 110,
 };
 
 pub const FunctionAttribute = enum(c_int) {
@@ -242,3 +244,10 @@ comptime {
     std.debug.assert(@sizeOf(KernelNodeParams) == 72 and @offsetOf(KernelNodeParams, "params") == 40);
     std.debug.assert(@sizeOf(ExecUpdateResultInfo) == 24);
 }
+
+test "CUDA device attribute matches the RDMA VMM capability ABI" {
+    // CUDA's CUdevice_attribute assigns 110 to RDMA VMM support; 126 is the memory sync domain count.
+    try @import("std").testing.expectEqual(@as(c_int, 110), @backingInt(
+        DeviceAttribute.gpu_direct_rdma_with_cuda_vmm_supported,
+    ));
+}