resonance-engine/archive/docs/DEPLOYMENT_REPORT.md

DEPLOYMENT REPORT: WSL2 CUDA LBM Daemons

Date: 2026-03-16 (updated)
Host: Beast (RTX 4090, Windows 11)
Status: LIVE — both 512 and 1024 daemons deployed, bridges verified

---

What Was Deployed

Real-time LBM (Lattice Boltzmann Method) CUDA daemon running in WSL2, publishing grid state over ZeroMQ to a Python bridge on the Windows side. The pipeline is:

512 Daemon:  LBM CUDA (WSL2, 100Hz) → ZeroMQ tcp://localhost:5555 → Python Bridge (Windows) → Ollama
1024 Daemon: LBM CUDA (WSL2, 100Hz) → ZeroMQ tcp://localhost:5556 → Python Bridge (Windows) → Ollama

Environment

Component	Detail
WSL Distro	Ubuntu 24.04.2 LTS
GPU	RTX 4090 (sm_89, Ada Lovelace)
NVIDIA Driver	581.57 (CUDA 13.0 capable)
CUDA Toolkit	12.6 (V12.6.85) — minimal install: nvcc, cudart-dev, nvml-dev, cccl
ZeroMQ	libzmq3-dev 4.3.5 (WSL) + pyzmq 27.1.0 (Windows Python)
json-c	libjson-c-dev 0.17
Sudo	Passwordless for user johndev

What Was Fixed Before Deployment

1. Distribution function initialization (critical)

The lbm_cuda_daemon.cu had a stub // ... initialization code ... where d_f (the LBM distribution functions, 512×512×9 floats) was never initialized after cudaMalloc. Running without this would have pumped NaN/garbage metrics through the entire pipeline.

Fix: Added load_etch() function (matching the pattern used in all other .cu files in the project) that:

• Loads from etch_khragixx.bin if present (rho, ux, uy → compute equilibrium distributions)
• Falls back to uniform equilibrium (rho=1, u=0) if file is missing

2. NVML init/shutdown in hot loop

nvmlInit() and nvmlShutdown() were being called every iteration at 100Hz inside the main loop. Moved to one-time setup/teardown outside the loop.

Files Modified

• D:\fractal-brain\beast-build\lbm_cuda_daemon.cu — added load_etch(), fixed NVML lifecycle

Files Created

• D:\fractal-brain\beast-build\setup_wsl_cuda.sh — WSL2 CUDA setup script
• D:\fractal-brain\beast-build\compile_daemon.sh — build script (handles PATH)
• D:\fractal-brain\beast-build\verify_install.sh — dependency verification
• D:\fractal-brain\beast-build\test_bridge.py — ZeroMQ bridge connection test

Binary

• D:\fractal-brain\beast-build\lbm_cuda_daemon (WSL2 ELF, 1MB)
• Compiled: nvcc -O3 -arch=sm_89 -lzmq -ljson-c -lnvidia-ml

Live Metrics (observed at cycle 15000+)

Metric	Range	Notes
Coherence	14.0–15.6	Structural integrity, stable oscillation
H64 (logic)	6.3–8.0	64-cell skeleton dominant
H32 (creative)	0.04–0.05	Creative channel quiet
Asymmetry	5.1–6.0	Left-right grid difference
Vorticity	Published	Curl of velocity field
Power	40–75W	GPU thermal cycling

Grid loaded from etch_khragixx.bin (Khra'gixx etch with live wave injection at 64-cell and 8-cell harmonics).

How To Operate

Start daemon (WSL2)

wsl -d Ubuntu -e bash -c "export PATH=/usr/local/cuda-12.6/bin:\$PATH; cd /mnt/d/fractal-brain/beast-build && ./lbm_cuda_daemon"

Verify bridge (Windows)

cd D:\fractal-brain\beast-build
python test_bridge.py

Run interactive bridge with Ollama (Windows)

cd D:\fractal-brain\beast-build
python lbm_llm_bridge.py

Run contextual bridge (maintains conversation history)

cd D:\fractal-brain\beast-build
python lbm_ollama_bridge_context.py

Recompile after code changes

wsl -d Ubuntu -e bash /mnt/d/fractal-brain/beast-build/compile_daemon.sh

Bridge Architecture

lbm_llm_bridge.py maps LBM physics to inference parameters:

• Temperature = inverse of coherence (high coherence → low temp → structured output)
• Top-p = proportional to asymmetry (high asymmetry → more randomness)
• H64/H32 ratio = logic vs creative balance in token selection

System prompt is dynamically rebuilt from real-time grid state on every query.

Deliverables Status

Deliverable	Status
512 daemon running in WSL2, publishing to port 5555	DONE
1024 daemon running in WSL2, publishing to port 5556	DONE (NaN fix applied, verified)
Python bridge receiving data (both ports)	DONE
Ollama integration with physics-constrained inference	READY (bridge code exists, requires ollama run llama3.2 to be available)
somatic_dataset.jsonl logging	Requires running interactive/contextual bridge sessions

Known State

• Both daemons deployed, only run one at a time (they share the GPU)
• H32 (creative channel) is very quiet (~0.04–0.05) — the grid is in a structured/logical mode
• Khra'gixx wave injection is active (64-cell + 8-cell harmonics for 512; 128-cell + 16-cell for 1024)
• No streaming kernel — collision-only BGK model (consistent with all other .cu files in project)

---

1024×1024 Daemon (added 2026-03-16)

What it is

16x data density increase over the 512 grid. Testing Granite state under massive scale. Publishes on port 5556 (not 5555).

NaN Bug — Root Cause & Fix

The original lbm_1024x1024.cu was created by another agent and produced NaN for all metrics (coherence, H64, H32) while power and cycle count were normal.

Root cause: Lines 166–170 initialized d_f by reading d_w[j] — a __constant__ GPU memory array — from host code. __constant__ memory lives on the GPU; reading it from the CPU returns garbage/zero. This set rho=0 everywhere → division by zero in collision kernel → NaN propagation through all metrics. Power was normal because NVML reads don't touch the simulation data.

Fix applied:

1. Added load_etch_1024() function with host-side w[] array (not d_w[])
2. Loads from etch_khragixx.bin (512×512 format) and upscales 2x via nearest-neighbor to 1024×1024
3. Falls back to fresh equilibrium (rho=1, u=0) if etch file missing
4. Fixed NVML init/shutdown lifecycle (was in hot loop, same bug the 512 daemon had)

Files Modified

• D:\fractal-brain\beast-build\lbm_1024x1024.cu — added load_etch_1024(), fixed __constant__ read, fixed NVML lifecycle

Binary

• D:\fractal-brain\beast-build\lbm_1024x1024 (WSL2 ELF)
• Compiled: nvcc -O3 -arch=sm_89 lbm_1024x1024.cu -lzmq -ljson-c -lnvidia-ml
• Compile requires explicit PATH: export PATH=/usr/local/cuda-12.6/bin:/usr/bin:/bin:$PATH

Live Metrics (verified at cycle 9400+)

Metric	Range	Notes
Coherence	14.0–15.6	Matches 512 daemon range
H64 (logic)	6.6–8.2	128-cell skeleton (scaled from 64)
H32 (creative)	0.03–0.05	64-cell wave (scaled from 32)
Asymmetry	5.1–5.9	Left-right grid difference
Power	45–80W	GPU thermal cycling

How To Operate

Start 1024 daemon (WSL2)

wsl -d Ubuntu -e bash -c "export PATH=/usr/local/cuda-12.6/bin:/usr/bin:/bin:\$PATH; export LD_LIBRARY_PATH=/usr/local/cuda-12.6/lib64:/usr/lib/x86_64-linux-gnu:\$LD_LIBRARY_PATH; cd /mnt/d/fractal-brain/beast-build && ./lbm_1024x1024"

Verify 1024 bridge (Windows)

cd D:\fractal-brain\beast-build
python test_1024_bridge.py

Recompile 1024 after code changes

wsl -d Ubuntu -e bash -c "export PATH=/usr/local/cuda-12.6/bin:/usr/bin:/bin:\$PATH; cd /mnt/d/fractal-brain/beast-build && nvcc -o lbm_1024x1024 lbm_1024x1024.cu -lzmq -ljson-c -lnvidia-ml -O3 -arch=sm_89"

Kill stale 1024 processes

wsl -d Ubuntu -e bash -c "pkill -9 -f lbm_1024x1024"

IMPORTANT: If the bridge still shows NaN after restart, check for zombie processes holding port 5556 (ps aux | grep lbm_1024). Kill them all before restarting — the bridge connects to whichever process bound the port first.