SparkGLM 0.1.0 — tool boundaries and product versions

SparkGLM 0.1.0 introduces semantic product versions and fixes GLM tool boundaries.

Fixes short-prose tool suppression and literal tool-XML handling in GLM. Quoted tags and fenced examples remain text; real tool calls retain constrained arguments. Streaming preserves adjacent text and whitespace.

Includes SparkGLM #57, #58, #59 and #60, with engine fixes in Atlas #7 and #8. Issue #55 is closed.

The final engine source passed 32/32 XML controls across OpenAI/Anthropic and blocking/streaming, 15/15 short ToolEvalBench scenarios, and 2,890 native server tests. The earlier loop-only revision passed the 60-case context-pressure sweep. These are correctness checks, not performance claims. Broad Atlas CI retains documented pre-existing failures.

The attached image pin manifest identifies the immutable registry digest, product revision, install tree, engine and checkpoint revisions. Deployment and published-image verification are recorded separately in the release receipt. The managed rollout covers the existing TP2 GLM lane on Sparks 01/02.

Run ./start.sh --version for the checkout version and exact Atlas pin. ./start.sh status also shows the product version and image ID on both running ranks. Images expose the same version with docker run --rm IMAGE --version, in startup logs and in the OCI version label. Install-tree hash tags remain reproducible. See versioning policy.

Build limitation: the canonical image checks passed all test stages, then failed in the existing workspace Clippy gate (a spark-comm test import and two spark-model lints). The release uses the supported runtime-only build path after those checks and the native source qualification above; the full workspace lint gate is not green.

Published-image rollout verification: both managed TP2 ranks run version 0.1.0 from the attached immutable digest. All 32/32 gateway XML/tool checks passed; health, model advertisement and restored routes were verified. The previous image and prefix cache are retained for rollback.

SparkGLM TP2 + TP4: Strict and Spark

SparkGLM now ships one image for two or four DGX Sparks, with Strict and Spark modes, prefix caching, optional native MTP in Strict, and explicit switchless chain or ring on TP4.

Strict preserves checkpoint weight formats. Spark uses the measured corrected-rounding FP4 paths and DFlash2 while retaining native prefill, MLA, dense projections and the vocabulary head. This release also fixes shared-expert arithmetic rounding, native-MTP sampling and near-limit admission/headroom, ring completion ordering, and shared-prefix rewriting during prefill.

Spark decode	TP2	TP4 chain	TP4 ring
Prose	37.489	57.145	58.328
Code	67.166	104.147	106.368
Structured	89.752	137.936	141.363

Tokens/second; two-run medians, uncapped GPU clocks, low reasoning effort, natural completion below 4,096 output tokens. TP2 ran the published image. TP4 qualification used the identical executable, native libraries, launcher and profiles before source metadata was added. Every chain/ring output and reasoning hash matched. Two runs do not establish statistical significance.

The updated arcade cards also show the retained TP2 cold-64K prefill median, 3,008 tok/s. It is explicitly labeled as earlier TP2 qualification, not a new TP4 result.

Install and identity

git clone https://github.com/Enntity/sparkglm.git
cd sparkglm
git checkout unified-20261021
cp .env.example .env  # set WORKER, and MODE=spark if wanted
./start.sh

• Image: ghcr.io/enntity/atlas-sparkglm:e3792e6e8c51
• Immutable image: ghcr.io/enntity/atlas-sparkglm@sha256:0d616370f36f36c8f152a4fcfc68839ab4d12f8a532917ad7b4df71f41f0d010
• Atlas source: sparkglm/atlas-20261021-unified (f4d95a100516338c938a4524e23a65a2c4c3206e)
• TP4 chain/ring instructions

Validation and receipts

Release report · focused TP2/TP4 qualification · retained full TP2 campaign · PR #56

Product CI, clean public-clone launch, anonymous image pulls and live LLooM gateway canary passed. Final Spark serving is verified on 01/02 with normal clock caps; Qwen03/04 is preserved. Raw receipts include failed controls, every repetition and subsequent surgical fixes. The source package contains both exact public source snapshots, all receipt archives, licenses and a checksum-verified cold-build recipe.

Native MTP is opt-in. Strict retains FP8 KV; it is not a cross-engine bitwise promise. Near-1M C4 checks use branches of one shared prefix. Broad TP2 results are reused with their original identities; the release does not claim a fresh full benchmark campaign, an independent cold build, or complete upstream engine CI. The image is a metadata-only repack of the qualified runtime. Check the licensing documentation before using separately downloaded checkpoints or DFlash2 weights.

SparkGLM fabric and refusal-content maintenance

SparkGLM now ships the Atlas #172 fix: blocking Chat Completions retain full assistant content alongside refusal metadata. Per-node RDMA overrides and clear failures for ambiguous discovery are included from PR #48.

Image: ghcr.io/enntity/atlas-sparkglm@sha256:5ec57d5522a95463b600a17471ce60712b315566e04c0a3bb6af1d84a1de0a4a

Atlas: 6f27e8fa0afe65180a169e54361a77565787ae27. Install tree: ed7548dfe442949283b54c7160605429046cd46f.

The native arm64 CUDA build passed all configured image checks, including 25 refusal and 43 streaming tests. Installer/fabric, converter, shell and receipt checks passed. DFlash, display carveout, model pins and profile defaults are unchanged. See results/2026-10-13-maintenance for receipts and limits.

Live GLM activation awaits a coordinated window on the pair, which is currently reserved for MTP experiments. This release does not claim fresh performance benchmarks or upstream Atlas-Inf merge status.

Staging receipt: source and immutable image are present on four Sparks; the Mac product checkout is synchronized. Live fabric inventories on a port-1 TP2 pair passed explicit per-node overrides and rejected ambiguous automatic discovery. Inventories were collected over trusted operator SSH connections because direct node-to-node SSH lacked host-key trust. No model containers were restarted by staging.

Per-node RDMA fabric ports

Launch on either physical port with independent leader and worker overrides. Set LEADER_FABRIC_HCA and WORKER_FABRIC_HCA to exact local RDMA device lists; interface overrides are also supported. ./start.sh fabric performs a read-only preflight. Ambiguous discovery, invalid rail pairs, and missing device selections fail before model or container changes.

Explicit rails are preserved through the image profile without implicit sibling expansion. The Atlas engine pin is unchanged.

Published arm64 image: ghcr.io/enntity/atlas-sparkglm:f95a9af14236

Immutable image: ghcr.io/enntity/atlas-sparkglm@sha256:78420594991e6e7e3914f84a6300ac1cf2f4e18b68e331f8ddc74fd7db905d3c

Validation: public CI, launcher tests, static checks, native image build checks, and a live port-1 fabric preflight and bounded RDMA transfer.