# deepseek-v4-flash-dspark-src — build DeepSeek-V4-Flash DSpark from the b12x fork

EXPERIMENTAL. Builds a DeepSeek-V4-Flash DSpark serving image **from source**,
so we are not locked to the prebuilt `aidendle94/sparkrun-vllm-ds4-gb10` image.

Stack (pinned in `versions.env`):

| Layer | Pin |
|---|---|
| vLLM fork | `local-inference-lab/vllm` @ `feat/gg-ds4-0731-launcher` = `e63a190ee` |
| torch | `2.12.0+cu130` |
| FlashInfer | `ce40d25a` (0.6.15), built from source |
| b12x | `voipmonitor/b12x` @ `284a2ea` |
| NCCL | `5067397c` |
| CUDA base | `nvidia/cuda:13.2.0-devel-ubuntu24.04` @ pinned digest |
| arch | aarch64 / `sm_121a` |

The python locks were regenerated from the **fork's own `requirements/`** at the
pinned commit (torch 2.12.0, flashinfer 0.6.15, cutlass-dsl 4.6.0,
tokenspeed/humming-kernels), not from the GLM-5.2 harness locks (which pin
torch 2.11 / flashinfer 0.6.12 and are therefore invalid for this stack).

## Workflow

1. Reset/refresh the locks (only needed when the fork pins change):
   ```bash
   bash gen-locks.sh          # writes locks/python-build.txt & python-runtime.txt
   ```
2. Build:
   ```bash
   TAG=vllm-ds4-dspark-src:probe bash build.sh
   ```
   via the launcher on a build node:
   ```bash
   sm121 build --recipe deepseek-v4-flash-dspark-src --build-node dgx3.lan
   ```
   This uploads this directory, runs `build.sh` (torch wheels + from-source
   FlashInfer/vLLM + b12x editable) in a tmux build session, and stamps the
   `sm121.checksum` label.

## Notes / caveats
- **Multi-hour build.** Expect 1–3h+ on a DGX Spark; runs in tmux on the node.
- The runtime lock pins flashinfer 0.6.14 (from the fork's `requirements/cuda.txt`)
  but we then install the **0.6.15 from-source wheel** over it — the built wheel
  wins. If pip objects to install-order, pin the runtime lock differently.
- FlashInfer 0.6.15 is built from source and installed with `--no-deps`, so its
  own `install_requires` are never auto-resolved into the lock. `gen-locks.sh`
  therefore re-adds its hard transitive deps explicitly (**cuda-tile==1.5.0**)
  and **nccl4py==0.3.1**) so `pip check` passes. `deep-gemm` (sdist-only) and
  `vllm-flash-attn` (x86-only) have no aarch64 wheel and are not required by the
  fork's vLLM, and `ray` is not needed on the `mp` distributed backend — so all
  three stay out of the aarch64 runtime lock. Note the import names differ from
  the PyPI names: cuda-tile → `cuda.tile`, nccl4py → `nccl`.
- `overlay/` and `patches/` are intentionally empty (GLM-specific pieces dropped);
  add DS4 patches here if the fork needs a local fix.
- Config/flags for serving mirror the fork's `serve-ds4-flash-spark.sh`.