54 lines
1.7 KiB
Markdown
54 lines
1.7 KiB
Markdown
|
# MESSAGE FOR THE-CRAW INFRASTRUCTURE ENGINEER
|
|||
|
|
|
|||
|
|
## CONTEXT:
|
|||
|
|
- Beast has working 1024×1024 with NVMe hybridization
|
|||
|
|
- Original mothballed, NVMe version tested and working
|
|||
|
|
- Need to test on the-craw (GTX 1050)
|
|||
|
|
|
|||
|
|
## FILES AVAILABLE ON BEAST:
|
|||
|
|
1. `D:\openclaw-local\workspace-main\harmonic_scan_sequential\1024x1024\fractal_habit_1024x1024_nvme_proper.cu` - NVMe source
|
|||
|
|
2. `D:\openclaw-local\workspace-main\harmonic_scan_sequential\1024x1024\fractal_habit_nvme_proper.exe` - Compiled binary
|
|||
|
|
3. `D:\openclaw-local\workspace-main\harmonic_scan_sequential\1024x1024\MOTHBALLED_ORIGINAL\` - Original (read-only)
|
|||
|
|
|
|||
|
|
## TEST INSTRUCTIONS FOR THE-CRAW:
|
|||
|
|
|
|||
|
|
### Step 1: Check CUDA/GPU
|
|||
|
|
```bash
|
|||
|
|
/usr/local/cuda/bin/nvcc --version
|
|||
|
|
nvidia-smi
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### Step 2: Copy files from Beast (if possible via node connectivity)
|
|||
|
|
Or compile from source:
|
|||
|
|
```bash
|
|||
|
|
# Get source from Beast path above
|
|||
|
|
# Compile for GTX 1050 (sm_61)
|
|||
|
|
/usr/local/cuda/bin/nvcc -arch=sm_61 -O3 -D_USE_MATH_DEFINES \
|
|||
|
|
fractal_habit_1024x1024_nvme_proper.cu \
|
|||
|
|
-o fractal_habit_nvme_craw -lnvml -lcufft
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### Step 3: Test NVMe checkpointing
|
|||
|
|
```bash
|
|||
|
|
mkdir -p /mnt/nvme/fractal_test # or use /tmp
|
|||
|
|
./fractal_habit_nvme_craw
|
|||
|
|
# Should create checkpoint_00100000.bin at end
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
### Step 4: Report back
|
|||
|
|
1. Compilation success/failure
|
|||
|
|
2. Runtime behavior
|
|||
|
|
3. Checkpoint file creation
|
|||
|
|
4. Any errors
|
|||
|
|
|
|||
|
|
## BEAST STATUS:
|
|||
|
|
- ✅ Original 1024×1024 working (mothballed)
|
|||
|
|
- ✅ NVMe version working (checkpoint saved: 48MB)
|
|||
|
|
- ✅ Three-tiered memory hierarchy implemented
|
|||
|
|
- ✅ Ready for the-craw testing
|
|||
|
|
|
|||
|
|
## URGENCY:
|
|||
|
|
Test NVMe hybridization on actual GTX 1050 hardware to verify:
|
|||
|
|
1. Compilation works on different CUDA version (11.4 vs 12.6)
|
|||
|
|
2. Checkpointing works on Linux/NVMe
|
|||
|
|
3. Performance on lower-power GPU
|