æ:// GLOCAL CUDA · HERMES AGENT
BRANDED æææ:// TOWER
NVIDIA GLOCAL
æ
GLOCAL CUDA
HERMES AGENT × NVIDIA · distributed-local compute
#opensourceware
#hermiphicationisinevitable
NVIDIA glocal
TensorRT-LLM
Mixture of Devices
CAPABILITY · GLOCAL CUDASPEC
Hermes Agent (conductor, pc://conductor) owns the route tower.
Compute is a Mixture of Devices (MoD) — addressed, not administered.
brain droplet 129.212.180.252:3000 — public identity/data, rtx3050:// front door
│ reverse SSH tunnel (ed25519, key-only)
▼
hands Victus RTX 3050 6GB 127.0.0.1:3050 — sovereign CUDA, compute-capability 8.6
│ +æ://route register vps://cloud <ep>
▼
burst cloud pod (RunPod/Vast/DO-gpu) — same vps_node, per-job cost
All three are MoD nodes. One +æ://route tower. Codemode: a device is a language-op.
The droplet holds zero keys — secrets resolve via +æ://secrets at runtime.
Brain (droplet) is the public surface. No GPU, no keys — it routes.
Hands (RTX 3050) run the silicon: nvcc 13.3, nvidia-smi, TensorRT-LLM ready. Reached only via tunnel.
Burst (cloud) is optional, codemode-ready: provision → register → dispatch → destroy.
BRING UP THE BRIDGEFLOW
bash rtx-bridge.sh up # self-healing GPU endpoint + reverse tunnel
bash rtx-bridge.sh test # E2E: matmul droplet → tunnel → RTX 3050
bash rtx-bridge.sh install # systemd --user: survives Victus reboot
bash rtx-bridge.sh down # tear down
Prove it — a request to the public droplet runs a kernel on the local GPU:
ssh root@129.212.180.252 \
"curl -s 'http://127.0.0.1:3050/xrpc/ae.vps.rtx?op=matmul&n=512'"
# → {"ok":true,"probe":{"name":"NVIDIA GeForce RTX 3050 6GB Laptop GPU",
# "memory_total_mib":6144,"utilization_gpu":0},"run":{"host_ms":~496,"n":512}}
ROUTE TOWER CONTRACT+æ://route
+æ://route # list MoD: Victus + neuromitosis + cloud slot
+æ://route register vps://cloud <ep> --host <h> # add burst pod as addressable route
vps://neuromitosis # probe droplet liveness + records + compute
vps://cloud # probe burst pod (after register)
dispatch (Hermes Agent → GPU):
vps_node.dispatch_rtx(op, n) → gpu_mcp GPUAgent: probe → compile_kernel → run_kernel
op ∈ {matmul, ...} n = matrix dim (extend toward TensorRT-LLM inference)
PROVISION A CLOUD HANDCODEMODE
Pick a pod host — RunPod/Vast.ai (cheapest burst) or DO-gpu (native, needs GPU-access request + heavier commit). NVIDIA glocal: burst hands appear on demand.
Deploy the identical vps_node.py + secret_source.py — same chassis, different physical silicon. Keys resolve via +æ://secrets at runtime; never bake a token into the image.
+æ://route register vps://cloud http://<pod-ip>:3000 --host runpod-rtx4090
curl -s 'http://<pod-ip>:3000/xrpc/ae.vps.rtx?op=matmul&n=1024' # verify burst hands
PROVEN PROFILERTX 3050
hands : RTX 3050 6GB Laptop GPU · compute-capability 8.6 · nvcc 13.3 · driver 592.27
surface: python vps_node.py --host 127.0.0.1 --port 3050 (ThreadingHTTPServer, JSON)
tools : probe_gpu · compile_kernel (matmul.cu) · run_kernel · rust_build_wasm
proven : matmul 512×512 → host_ms ~496 on real silicon, zero cloud rent
stack : Hermes Agent conductor → +æ://route → droplet tunnel → gpu-mcp → CUDA
BRAND · æææ:// TOWERNAMESPACE
our built backbone is branded under the æææ:/// (triple-ashe) namespace:
æææ://neuromitosis → GLOCAL backbone (droplet brain: identity/data + rtx front door)
æææ://cuda → NemoClaw GLOCAL CUDA tower (Hermes Agent conducts)
both delegate to the same vps:// + NVIDIA:// engines already verified.
DNS: æææ.com (punycode xn--6caa.com) is registered + live, but presently
points at 185.134.245.113 — not yet our droplet (129.212.180.252).
Repoint A/AAAA → droplet to make æææ.com actually serve this tower. Until then the
scheme is real + routed; the domain is the one remaining step.
PITFALLSAVOID
Zombie endpoint — a stale vps_node holding :3050 serves old code and masks fixes. Kill by PID (taskkill) before re-probing. The while true wrapper restarts on death; a leftover orphan does not.
GPU context contention — two processes can't both own the CUDA context. One endpoint per GPU. A foreground test + background server collide → list index out of range.
Cloud cost creep — a forgotten pod is the only recurring bill. Destroy on idle; the route slot stays codemode-ready with no spend.
Keys on the pod — resolve secrets at runtime via +æ://secrets broker, same as Victus. Never image-bake.
Droplet is brain, not hands — it has no GPU; rtx3050://ready is an advertisement. Real silicon is Victus/cloud behind the tunnel.
C:\ae vs C:\æ — unicode path is real. Latin silently fails (same trap as secrets).
#hermiphicationisinevitable