Pin SparkQwen shared-host memory support

backends/atlas-sparkqwen/README.md

@@ -5,17 +5,17 @@ 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
+`Enntity/sparkqwen` revision `0cf2436de5211b41edf4d92071180c75cec31dd8`
+(install tree `e7268337a5c8`) by that repository's `install/build.sh` and
+published as `ghcr.io/enntity/atlas-sparkqwen:e7268337a5c8`. 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
+`pins.json` addresses `ghcr.io/enntity/atlas-sparkqwen@sha256:63fcc9f8f87a3963aba91f0a5dbf2c1cf5efa1d809c4ab1b52bd5853979e73ed`, the GHCR
+manifest digest of tag `e7268337a5c8` (recorded 2026-10-13 with
 `docker buildx imagetools inspect`). Both recipe members use the same
 reference as `bootstrap.image`.
 

backends/atlas-sparkqwen/pins.json

@@ -1,11 +1,11 @@
 {
   "manifestVersion": 1,
   "status": "final",
-  "note": "Single pin manifest for the Atlas SparkQwen RC16 lane. Source, install tree, tag and model identities are final. image.reference is the GHCR manifest digest of ghcr.io/enntity/atlas-sparkqwen:4ccd961ecf09, also used as the recipe's bootstrap image. The normal install path addresses image.reference; the tag is source-build metadata (the first 12 characters of the git tree of install/ at the pinned revision) and the install-tree label is a contents check. A tag or label alone is never the artifact identity.",
+  "note": "Single pin manifest for the Atlas SparkQwen RC16 lane. Source, install tree, tag and model identities are final. image.reference is the GHCR manifest digest of ghcr.io/enntity/atlas-sparkqwen:e7268337a5c8, also used as the recipe's bootstrap image. The normal install path addresses image.reference; the tag is source-build metadata (the first 12 characters of the git tree of install/ at the pinned revision) and the install-tree label is a contents check. A tag or label alone is never the artifact identity.",
   "source": {
     "repo": "Enntity/sparkqwen",
-    "revision": "8b28467324893583838871b6a6b6a28d3059e64f",
-    "installTree": "4ccd961ecf09d86c9f13087e42592789b5f56182",
+    "revision": "0cf2436de5211b41edf4d92071180c75cec31dd8",
+    "installTree": "e7268337a5c8e77c874c7cb7be689f80a79946a5",
     "cloneUrl": "https://github.com/Enntity/sparkqwen.git",
     "buildScript": "install/build.sh"
   },
@@ -15,8 +15,8 @@
     "commit": "1c43247d93ca050f02cb67b0603e88edfe3d92a5"
   },
   "image": {
-    "reference": "ghcr.io/enntity/atlas-sparkqwen@sha256:9408884dc4bf2c84b7dd097d62effc73f58fffc25a1d67db73d0b4e69516a7c0",
-    "tag": "ghcr.io/enntity/atlas-sparkqwen:4ccd961ecf09",
+    "reference": "ghcr.io/enntity/atlas-sparkqwen@sha256:63fcc9f8f87a3963aba91f0a5dbf2c1cf5efa1d809c4ab1b52bd5853979e73ed",
+    "tag": "ghcr.io/enntity/atlas-sparkqwen:e7268337a5c8",
     "label": "io.enntity.sparkqwen.install-tree",
     "entrypoint": "/opt/atlas/serve.py",
     "profilePath": "/opt/atlas/profiles/8x262k.json",

docs/recipes.md

@@ -576,9 +576,17 @@ lloom setup --recipe linux-nvidia-dgx-spark-2x-qwen38-flash-next-atlas --additiv
 lloom runtime-start qwen38-flash-next-atlas-cluster
 ```
 
+For shared hosts, the image accepts `SPARKQWEN_GPU_MEMORY_UTILIZATION=0.70`
+through `0.95` in each physical runtime's Docker environment. Lower settings
+leave room for host services and reduce the shared KV pool; the published
+context/concurrency measurements apply to the default 0.88 allocation. Keep
+the logical runtime, physical runtimes, and each placement member's
+`resources.memoryGb` reservation consistent with the chosen deployment.
+Do not raise the host memory safety limit to make an oversized cache fit.
+
 The recipe is additive (no default model, no aliases, `keepWarm` false). Pins
 live in `backends/atlas-sparkqwen/pins.json` (source `Enntity/sparkqwen`
-`8b284673`, install tree `4ccd961ecf09`). The GHCR manifest digest is pinned in the same manifest.
+`0cf2436d`, install tree `e7268337a5c8`). The GHCR manifest digest is pinned in the same manifest.
 `node backends/atlas-sparkqwen/verify-pins.mjs`, setup step
 `verify-atlas-pins` and `install.sh` reject missing or placeholder identities. The
 installer only pulls that digest and verifies the image's arm64 architecture,

recipes/linux-nvidia-dgx-spark-2x-qwen38-flash-next-atlas.json

@@ -14,12 +14,12 @@
   "provenance": {
     "createdBy": "lloom",
     "generatedBy": "manual-curated",
-    "source": "SparkQwen RC16: Enntity/sparkqwen 8b28467324893583838871b6a6b6a28d3059e64f (install tree 4ccd961ecf09d86c9f13087e42592789b5f56182), engine Enntity/atlas sparkqwen/atlas-20261010-rc16 @ 1c43247d (AGPL-3.0-only, recorded in the image's /opt/atlas/source-manifest.json). The image is ghcr.io/enntity/atlas-sparkqwen:4ccd961ecf09, addressed by its GHCR manifest digest in pins.json. The engine argument vector ships inside the image as /opt/atlas/profiles/8x262k.json, selected with SPARKQWEN_PROFILE=8x262k: eight sequences of up to 262144 tokens, BF16 KV (about 3.19M pool tokens at the profile's 0.88 GPU memory utilization), prefix caching, MTP speculation (up to 7 drafts), content-loop steering. Model nvidia/Qwen3.8-Flash-Next-NVFP4 revision fc694b54fb0174e0913e6adf86691ef85a4ead47."
+    "source": "SparkQwen RC16: Enntity/sparkqwen 0cf2436de5211b41edf4d92071180c75cec31dd8 (install tree e7268337a5c8e77c874c7cb7be689f80a79946a5), engine Enntity/atlas sparkqwen/atlas-20261010-rc16 @ 1c43247d (AGPL-3.0-only, recorded in the image's /opt/atlas/source-manifest.json). The image is ghcr.io/enntity/atlas-sparkqwen:e7268337a5c8, addressed by its GHCR manifest digest in pins.json. The engine argument vector ships inside the image as /opt/atlas/profiles/8x262k.json, selected with SPARKQWEN_PROFILE=8x262k: eight sequences of up to 262144 tokens, BF16 KV (about 3.19M pool tokens at the profile's 0.88 GPU memory utilization), prefix caching, MTP speculation (up to 7 drafts), content-loop steering. Model nvidia/Qwen3.8-Flash-Next-NVFP4 revision fc694b54fb0174e0913e6adf86691ef85a4ead47."
   },
   "links": [
     {
       "rel": "upstream",
-      "href": "https://github.com/Enntity/sparkqwen/tree/8b28467324893583838871b6a6b6a28d3059e64f"
+      "href": "https://github.com/Enntity/sparkqwen/tree/0cf2436de5211b41edf4d92071180c75cec31dd8"
     },
     {
       "rel": "engine",
@@ -219,7 +219,7 @@
                 "warmup": false,
                 "bootstrap": {
                   "adapter": "docker",
-                  "image": "ghcr.io/enntity/atlas-sparkqwen@sha256:9408884dc4bf2c84b7dd097d62effc73f58fffc25a1d67db73d0b4e69516a7c0",
+                  "image": "ghcr.io/enntity/atlas-sparkqwen@sha256:63fcc9f8f87a3963aba91f0a5dbf2c1cf5efa1d809c4ab1b52bd5853979e73ed",
                   "pull": false,
                   "pullTimeoutMs": 3600000,
                   "createArgs": [
@@ -320,7 +320,7 @@
                 },
                 "bootstrap": {
                   "adapter": "docker",
-                  "image": "ghcr.io/enntity/atlas-sparkqwen@sha256:9408884dc4bf2c84b7dd097d62effc73f58fffc25a1d67db73d0b4e69516a7c0",
+                  "image": "ghcr.io/enntity/atlas-sparkqwen@sha256:63fcc9f8f87a3963aba91f0a5dbf2c1cf5efa1d809c4ab1b52bd5853979e73ed",
                   "pull": false,
                   "pullTimeoutMs": 3600000,
                   "createArgs": [

test/atlas-sparkqwen-recipe.test.mjs

@@ -323,10 +323,10 @@ assert.throws(() => materialize(missingFabric), /cluster\.nodes\[ennspark04\]\.f
 
 // ---- pins: final identities, pending digest, fail-closed gate ---------------
 assert.equal(pins.source.repo, 'Enntity/sparkqwen');
-assert.equal(pins.source.revision, '8b28467324893583838871b6a6b6a28d3059e64f');
-assert.equal(pins.source.installTree, '4ccd961ecf09d86c9f13087e42592789b5f56182');
+assert.equal(pins.source.revision, '0cf2436de5211b41edf4d92071180c75cec31dd8');
+assert.equal(pins.source.installTree, 'e7268337a5c8e77c874c7cb7be689f80a79946a5');
 assert.equal(pins.engine.commit, '1c43247d93ca050f02cb67b0603e88edfe3d92a5');
-assert.equal(pins.image.tag, 'ghcr.io/enntity/atlas-sparkqwen:4ccd961ecf09');
+assert.equal(pins.image.tag, 'ghcr.io/enntity/atlas-sparkqwen:e7268337a5c8');
 assert.equal(pins.image.tag, imageTagFor(pins));
 assert.equal(pins.image.label, 'io.enntity.sparkqwen.install-tree');
 assert.equal(pins.image.profilePath, '/opt/atlas/profiles/8x262k.json');