data-angel · Oct 13, 2026 · Oct 13, 2026 · Oct 13, 2026
@@ -6,6 +6,8 @@ All notable changes to LLooM will be documented in this file. The format follows
 
 ### Added
 
+- Atlas SparkQwen RC16 two-Spark recipe (`linux-nvidia-dgx-spark-2x-qwen38-flash-next-atlas`, profile `8x262k`) with a fail-closed pin gate and a digest-only image installer, and explicit per-node `${fabricHca}` settings for distributed recipes. Both Atlas recipes pin the published images with exact rail selection; missing settings and ambiguous discovery fail clearly.
+
 - LLooM Hear CPU audio analysis, with structured estimates, optional dashboard images, and opt-in upstream interpretation. File and URL inputs require operator configuration.
 
 - Four standalone NVIDIA Spark audio recipes (Qwen3-TTS 1.7B CustomVoice, Base voice cloning and VoiceDesign, and Whisper large-v3-turbo). Each runs one pinned, integrity-checked checkpoint in its own container from a locally built `spark-audio` image that replaces a host-local prebuilt image.

@@ -186,3 +186,28 @@ The model `nvidia/GLM-5.3-Flash-NVFP4` (MIT per SparkGLM's licensing notes)
 and the DFlash2 drafter `incoai/GLM-5.3-Flash-DFlash2` (**CC BY-NC-ND 4.0,
 non-commercial**) are downloaded from their publishers under their own terms.
 See the [backend README](backends/atlas-sparkglm/README.md#credits-and-licenses).
+
+## Atlas SparkQwen engine image (not included)
+
+The optional Atlas SparkQwen lane in `backends/atlas-sparkqwen` and
+`recipes/linux-nvidia-dgx-spark-2x-qwen38-flash-next-atlas.json` pulls the
+container image `ghcr.io/enntity/atlas-sparkqwen` built from
+[Enntity/sparkqwen](https://github.com/Enntity/sparkqwen) at
+`85325e5104b2a6864e0cf31ec470a8c3327d4de4` (install tree `656234c336af`).
+LLooM starts that image as a separate program; no engine source or binary is
+included in the LLooM npm package, and LLooM's files for the lane remain MIT.
+
+- Engine: [Atlas](https://github.com/Atlas-Inf/atlas) (Atlas-Inf) with
+  SparkGLM's upstream series and SparkQwen's Qwen3.8-Flash-Next work on the
+  [Enntity/atlas](https://github.com/Enntity/atlas) fork, at
+  [`1c43247d93ca050f02cb67b0603e88edfe3d92a5`](https://github.com/Enntity/atlas/tree/1c43247d93ca050f02cb67b0603e88edfe3d92a5)
+  (recorded in the image's `/opt/atlas/source-manifest.json`)
+- License: AGPL-3.0-only. The corresponding source is Enntity/sparkqwen at the
+  revision above plus that Enntity/atlas commit; SparkQwen's `NOTICE` and
+  `docs/LICENSING.md` list what the image build fetches. The image ships
+  CUTLASS's license and the Rust crate notices under `/opt/atlas/notices/`;
+  the engine's license is at `/LICENSE`.
+
+The model `nvidia/Qwen3.8-Flash-Next-NVFP4` (NVIDIA Open Model License per its
+model card) is downloaded from its publisher under its own terms. See the
+[backend README](backends/atlas-sparkqwen/README.md#credits-and-licenses).
@@ -3,7 +3,7 @@
 LLooM-managed two-node Atlas SparkGLM candidate for a directly connected pair
 of NVIDIA DGX Spark systems. This directory is **MIT orchestration only**: no
 Atlas engine source is committed here. The engine image is built from immutable
-revision `3c7d67146929f350ed75fcdfd5e5349db843b483` of `Enntity/sparkglm` by
+revision `7a67d9bce4b5f408ad67f2a958ef1a8787136fdd` of `Enntity/sparkglm` by
 that repository's `install/build.sh`. The engine inside the image is
 AGPL-3.0-only; see [Credits and licenses](#credits-and-licenses). Live hardware
 and full serving qualification remain pending.
@@ -26,7 +26,7 @@ and full serving qualification remain pending.
   install-tree label is a contents check, not an artifact identity.
 - `install.sh` — prepares the image. The normal path reuses an already
   verified local image, otherwise pulls the immutable manifest digest
-  `ghcr.io/enntity/atlas-sparkglm@sha256:e4d313b5e32c47545eab0059eea65a38e65d03fc8e0d5b8c2cfb497c39f0372a`
+  `ghcr.io/enntity/atlas-sparkglm@sha256:78420594991e6e7e3914f84a6300ac1cf2f4e18b68e331f8ddc74fd7db905d3c`
   from GHCR. It never pulls the mutable tag, and a failed pinned digest pull is
   a hard failure with no silent source build. The explicit `--source-build`
   option instead clones `Enntity/sparkglm` at the pinned revision, checks that
@@ -235,7 +235,7 @@ is not part of LLooM:
   `ghcr.io/enntity/atlas-sparkglm`, alongside LLooM (an aggregate); no engine
   source or binary is included in LLooM.
 - **Corresponding source.** [Enntity/sparkglm](https://github.com/Enntity/sparkglm)
-  at the pinned revision (`3c7d6714`) plus the Enntity/atlas commit recorded
+  at the pinned revision (`7a67d9bc`) plus the Enntity/atlas commit recorded
   in the image's `/opt/atlas/source-manifest.json`
   ([`dfb9c26a`](https://github.com/Enntity/atlas/tree/dfb9c26a95c96faa9f94c44e23343b0f2b4f9db6)).
   SparkGLM's `NOTICE` and `docs/LICENSING.md` list everything else it fetches.

@@ -4,14 +4,14 @@
   "note": "Single pin manifest for the Atlas SparkGLM candidate. Portable source, image, model and converter identities are final for this candidate; live hardware and serving qualification remains pending. The normal install path addresses image.reference, the immutable GHCR manifest digest. The tag is retained as source-build metadata (it derives from the git tree of install/ at the pinned revision) and as the local tag produced by an explicit --source-build; the install-tree label is a contents check. A tag or label alone is never the artifact identity.",
   "source": {
     "repo": "Enntity/sparkglm",
-    "revision": "3c7d67146929f350ed75fcdfd5e5349db843b483",
-    "installTree": "246a3a948ab98d19e085bfd39797e554f0ee9261",
+    "revision": "7a67d9bce4b5f408ad67f2a958ef1a8787136fdd",
+    "installTree": "f95a9af14236a781bb13f57dcc9ea0342f339905",
     "cloneUrl": "https://github.com/Enntity/sparkglm.git",
     "buildScript": "install/build.sh"
   },
   "image": {
-    "reference": "ghcr.io/enntity/atlas-sparkglm@sha256:e4d313b5e32c47545eab0059eea65a38e65d03fc8e0d5b8c2cfb497c39f0372a",
-    "tag": "ghcr.io/enntity/atlas-sparkglm:246a3a948ab9",
+    "reference": "ghcr.io/enntity/atlas-sparkglm@sha256:78420594991e6e7e3914f84a6300ac1cf2f4e18b68e331f8ddc74fd7db905d3c",
+    "tag": "ghcr.io/enntity/atlas-sparkglm:f95a9af14236",
     "label": "io.enntity.sparkglm.install-tree",
     "entrypoint": "/opt/atlas/serve.py",
     "profilePath": "/opt/atlas/profiles/4x1m.json",

@@ -0,0 +1,132 @@
+# Atlas SparkQwen under LLooM
+
+LLooM-managed two-node Atlas SparkQwen RC16 lane for a directly connected pair
+of NVIDIA DGX Spark systems, serving `nvidia/Qwen3.8-Flash-Next-NVFP4` as
+published (no overlay conversion, no separate drafter: speculation uses the
+checkpoint's own MTP head). This directory is **MIT orchestration only**: no
+Atlas engine source is committed here. The engine image is built from
+`Enntity/sparkqwen` revision `8b28467324893583838871b6a6b6a28d3059e64f`
+(install tree `4ccd961ecf09`) by that repository's `install/build.sh` and
+published as `ghcr.io/enntity/atlas-sparkqwen:4ccd961ecf09`. The engine inside
+the image is AGPL-3.0-only; see [Credits and licenses](#credits-and-licenses).
+SparkQwen qualified RC16 on its own launcher; no live run through LLooM is
+recorded yet.
+
+## Image digest
+
+`pins.json` addresses `ghcr.io/enntity/atlas-sparkqwen@sha256:9408884dc4bf2c84b7dd097d62effc73f58fffc25a1d67db73d0b4e69516a7c0`, the GHCR
+manifest digest of tag `4ccd961ecf09` (recorded 2026-10-13 with
+`docker buildx imagetools inspect`). Both recipe members use the same
+reference as `bootstrap.image`.
+
+## Layout
+
+- `pins.json` — the single pin manifest: SparkQwen source revision and its
+  `install/` git tree, the engine commit, the immutable image reference with
+  the tag kept as metadata, the install-tree label, and the checkpoint revision.
+- `verify-pins.mjs` — fail-closed pin gate. Exits non-zero while the manifest
+  is not `final`, any identity is missing or a DRAFT/TODO placeholder, the
+  revision, install tree or engine commit is not a 40-character git object id,
+  the entrypoint or profile is not a plain path under `/opt/atlas`, the tag is not
+  `ghcr.io/enntity/atlas-sparkqwen:` plus the first 12 characters of the install
+  tree, `image.reference` is not `ghcr.io/enntity/atlas-sparkqwen@sha256:` plus
+  64 lowercase hex characters, an image ID is pinned, or the model leaves
+  `nvidia/Qwen3.8-Flash-Next-NVFP4@fc694b54fb0174e0913e6adf86691ef85a4ead47`.
+- `install.sh` — prepares the image: reuses an already verified local image,
+  otherwise pulls the immutable digest. It never pulls the mutable tag and
+  never builds; a failed pull or check is a hard failure. The image must be
+  arm64 with `io.enntity.sparkqwen.install-tree` equal to the pinned tree and
+  must ship `/opt/atlas/serve.py`, `/opt/atlas/profiles/8x262k.json` and the
+  pinned engine commit in `/opt/atlas/source-manifest.json`. `--check-only`
+  verifies without pulling. For a local build, run `install/build.sh` from
+  Enntity/sparkqwen at the pinned revision; that produces the tag, not the
+  digest this recipe addresses.
+
+## Installation
+
+```sh
+lloom setup --recipe linux-nvidia-dgx-spark-2x-qwen38-flash-next-atlas --additive --apply --yes
+lloom runtime-start qwen38-flash-next-atlas-cluster
+```
+
+The recipe is additive: it sets no default model and adds no aliases. Setup
+steps, in order: `check-docker`, `verify-atlas-pins`, `download-atlas-model`
+(the pinned checkpoint, about 124 GiB, into
+`${modelRoot}/nvidia--Qwen3.8-Flash-Next-NVFP4`), `pull-atlas-image`
+(`install.sh`), and `prepare-atlas-cuda-cache` (creates
+`${installRoot}/atlas-sparkqwen-cuda-cache`, mounted at `/atlas-cuda-cache` so
+compiled CUDA kernels survive container recreation). Run setup on both nodes.
+
+## Container contract
+
+The image ships its own entrypoint and profiles; LLooM supplies environment and
+mounts only, mirroring SparkQwen's reviewed `install/start-node.sh`: host
+network and IPC, 32 GiB shm, `--memory 114g`, all GPUs, `/dev/infiniband`,
+`IPC_LOCK` and `SYS_NICE`, unlimited memlock, `no-new-privileges`, a 60 s stop
+timeout, the checkpoint mounted read-only at the same absolute path, and the
+CUDA JIT cache.
+
+| Variable            | Source                                          | Meaning                                                                                                                                                                                                                      |
+| ------------------- | ----------------------------------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `NODE_RANK`         | `${nodeRank}`                                   | `0` on the leader, `1` on the worker                                                                                                                                                                                         |
+| `MASTER_ADDR`       | `${leaderAddress}`                              | the leader's direct-fabric address                                                                                                                                                                                           |
+| `MASTER_PORT`       | `29510`                                         | distributed rendezvous port                                                                                                                                                                                                  |
+| `FABRIC_INTERFACE`  | `${fabricInterface}`                            | the node's direct-fabric NIC, also `NCCL_SOCKET_IFNAME`/`GLOO_SOCKET_IFNAME`                                                                                                                                                 |
+| `FABRIC_HCA`        | `${fabricHca}`                                  | the node's RoCE device (below); `serve.py` adds the device behind the port's other PCIe half and sets `NCCL_IB_HCA` and `ATLAS_RDMA_RAILS` from both                                                                         |
+| `MODEL_PATH`        | `${modelRoot}/nvidia--Qwen3.8-Flash-Next-NVFP4` | the checkpoint                                                                                                                                                                                                               |
+| `SERVED_MODEL_NAME` | `qwen3.8-flash-next-atlas`                      | gateway model ID                                                                                                                                                                                                             |
+| `SPARKQWEN_PROFILE` | `8x262k`                                        | `/opt/atlas/profiles/8x262k.json`: up to eight requests of up to 262,144 tokens from one shared BF16 KV pool (about 3.19M tokens at the profile's GPU memory utilization of 0.88); the image also ships `8x32k` and `4x262k` |
+| `SPARKQWEN_FP8_GDN` | `0`                                             | the lossy FP8 Gated DeltaNet opt-in stays off                                                                                                                                                                                |
+
+The profile sets every `ATLAS_*` and `NCCL_*` tuning value; `serve.py` drops
+ambient `ATLAS_*` variables so the profile wins. To give the KV pool more of a
+dedicated pair's memory, add `-e SPARKQWEN_GPU_MEMORY_UTILIZATION=0.80-0.95` to
+both members in a local recipe override.
+
+`${fabricHca}` and `${fabricInterface}` are materialized per node by LLooM from
+the node's own config. Because recipes that consume `${fabricHca}` require exact,
+node-local device names, LLooM expects `cluster.nodes.<id>.fabricHca` (a
+comma-separated list of exact local RDMA device names, without spaces or `:port`
+suffixes) and `cluster.nodes.<id>.fabricInterface` on every participating node;
+it never guesses an HCA from the NIC name, since PCI naming and cabling differ by
+host. A pair cabled on the second port therefore needs
+`fabricHca: "rocep1s0f1"` and `fabricInterface: "enp1s0f1np1"` on both nodes.
+Missing or malformed overrides stop recipe materialization with the node and
+field named in the error.
+
+Ports are fixed by the image's `serve.py` and loopback-bound: rank 0 serves
+`127.0.0.1:8893` on the leader and rank 1 binds `127.0.0.1:8894` on the worker.
+The leader's port is what LLooM health-checks and routes to; the worker is
+readiness-checked through its container state. SparkGLM's lane uses the same
+ports, so the two lanes cannot run on the same pair at once (neither fits
+beside the other anyway).
+
+## Distributed startup
+
+The worker starts first (`order` 10, rank 1, `healthStrategy: "container"`, no
+warmup); the leader starts second (`order` 20, rank 0), health-checks
+`/health`, runs a POST warmup and owns routing. The first start compiles CUDA
+kernels into the cache; allow several minutes.
+
+## Credits and licenses
+
+LLooM's files in this directory and the recipe are MIT. The engine they start
+is not part of LLooM:
+
+- **Engine.** [Atlas](https://github.com/Atlas-Inf/atlas) (Atlas-Inf) with
+  SparkGLM's upstream series and SparkQwen's Qwen3.8-Flash-Next work on the
+  [Enntity/atlas](https://github.com/Enntity/atlas) fork, branch
+  `sparkqwen/atlas-20261010-rc16`, licensed **AGPL-3.0-only**. It runs from a
+  separate container image, `ghcr.io/enntity/atlas-sparkqwen`, alongside LLooM
+  (an aggregate); no engine source or binary is included in LLooM.
+- **Corresponding source.** [Enntity/sparkqwen](https://github.com/Enntity/sparkqwen)
+  at the pinned revision plus the Enntity/atlas commit recorded in the image's
+  `/opt/atlas/source-manifest.json`
+  ([`1c43247d`](https://github.com/Enntity/atlas/tree/1c43247d93ca050f02cb67b0603e88edfe3d92a5)).
+  SparkQwen's `NOTICE` and `docs/LICENSING.md` list everything else the build
+  fetches; the image ships CUTLASS's license and the Rust crate notices under
+  `/opt/atlas/notices/` and the engine's license at `/LICENSE`.
+- **Model.** Weights are downloaded from their publisher, not distributed by
+  LLooM. `nvidia/Qwen3.8-Flash-Next-NVFP4` is under the **NVIDIA Open Model
+  License** per its model card (SparkQwen `docs/LICENSING.md`); read the model
+  card before use and before redistributing anything derived from it.