# GLM-5.2 NF3-Hybrid build

Image is built from the forked vllm-gb10 repo:

```
git clone https://git.kilork.org/kilork/timothystewart6--vllm-gb10
cd timothystewart6--vllm-gb10
git checkout nf3-hybrid
```

Then build with build-args from versions.env:

```bash
source <(grep -v '#' versions.env | sed 's/^/export /')
args=$(for v in $(compgen -v | grep -E '^[A-Z_]+$'); do
  [ -n "${!v}" ] && echo "--build-arg ${v}=${!v}"
done)
DOCKER_BUILDKIT=1 docker build $args -t vllm-nf3-hybrid:probe -f Dockerfile .
```

Or use the included build-nf3.sh:

```bash
./build-nf3.sh
```

## Pins

| Component | Source | Ref |
|-----------|--------|-----|
| vLLM | local-inference-lab/vllm | codex/dcp-globaltopk-sharddraft-defaults-20260622 @ e232d26 |
| b12x | lukealonso/b12x | @ 9cd63a7 |
| FlashInfer | flashinfer-ai/flashinfer | v0.6.12 |
| cutlass-dsl | nvidia | 4.5.3 (firmware fix) |
| CUDA | nvidia/cuda | 13.2.0-devel-ubuntu24.04 |
| torch | PyTorch cu130 index | 2.11.0 |