Separate SparkRing subsystems and add GLM-5.3 DCP routes

.gitattributes

@@ -1,7 +1,10 @@
 *.sh text eol=lf
 *.patch text eol=lf
 runtime/** text eol=lf
+performance/receipts/glm53-flash/dflash7-snapshot-v1-safe/*.json text eol=lf
+performance/receipts/glm53-flash/split-page-shared-base-c8-20260830/*.json text eol=lf
 runtime/exl3/patches/*.patch whitespace=-trailing-space
+runtime/glm53-flash-split-page-sparkcache/patches/*.patch whitespace=-trailing-space
 
 # Runtime receipts hash these source files byte-for-byte before copying them,
 # and the plugin vendors byte-identical copies whose SHA-256 it records in

.github/workflows/ci.yml

@@ -53,7 +53,7 @@ jobs:
             --index-url https://download.pytorch.org/whl/cpu \
             "torch==${TORCH_VERSION}"
       - name: Test maintained Python trees
-        run: python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-adaptive-mtp-python-overlay runtime/glm53-flash-dflash7-python-overlay runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
+        run: python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-jj-r7-gb10 runtime/glm53-flash-split-page-sparkcache runtime/glm53-flash-adaptive-mtp-python-overlay runtime/glm53-flash-dflash7-python-overlay runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
 
   docs-links:
     name: docs links

AGENTS.md

@@ -27,7 +27,7 @@ conversation or development history.
 - Keep operator instructions focused on what to do and what result to expect.
 - Put lane, maturity, hardware, and evidence metadata in one table or callout
   instead of repeating it throughout the prose.
-- Prefer `the new settings are still being tested` over phrases such as
+- Prefer `the selected settings are still being tested` over phrases such as
   `candidate target changes`, `silent promotion`, `historical qualification
   scope`, or `revalidate the composition` when the plain statement is accurate.
 - Use exact formal vocabulary only where a machine-readable contract or release
@@ -40,10 +40,11 @@ SparkRing supports four model families across ten deployment profiles:
 
 - GLM-5.2 EXL3 3.5-bpw at four-Spark TP4/DCP4, as a base profile and a
   SparkCache composition, using the R7 runtime and site/candidate contracts.
-- GLM-5.3 Flash with the public BF16 DFlash2 drafter at four-Spark TP4/DCP1,
-  as a cache-disabled base profile and a SparkCache composition, using the
-  immutable source contract in `runtime/glm53-flash/` and the sanitized site
-  and runtime templates in `scripts/config/`.
+- GLM-5.3 Flash with the public BF16 DFlash2 drafter at four-Spark TP4. The
+  cache-disabled profile supports DCP1, DCP2, and DCP4; the SparkCache
+  composition remains DCP1, using the
+  immutable published JJ r7-compatible artifacts and environment contract in
+  `runtime/glm53-flash-jj-r7-gb10/`.
 - DeepSeek-V4-Flash-0731 at two-Spark TP2/DCP1 and four-Spark TP4/DCP1, as
   base profiles and SparkCache compositions, using the published serving image
   and per-rank environment contracts.
@@ -71,8 +72,9 @@ choosing one.
 | Subject | Canonical source |
 |---|---|
 | GLM-5.2 EXL3 runtime build | `runtime/exl3-r7/README.md` |
-| GLM-5.3 Flash runtime, model, DFlash, NCCL, and SparkCache pins | `runtime/glm53-flash/pins.json` |
-| GLM-5.3 Flash four-rank site and runtime profiles | `scripts/config/glm53-flash-tp4-site.example.yaml`, then the selected `scripts/config/glm53-flash-dflash2-bf16-tp4-dcp1*.example.json` |
+| Published JJ r7-compatible GLM-5.3 images, source/model identities, and smoke evidence | `runtime/glm53-flash-jj-r7-gb10/artifacts.json` |
+| Published JJ r7-compatible GLM-5.3 four-rank operator settings | `runtime/glm53-flash-jj-r7-gb10/runtime.env.example` |
+| Historical GLM-5.3 split-page local artifact | `runtime/glm53-flash-split-page-sparkcache/qualified-artifact.json` |
 | Qwen3.8-27B runtime build | `runtime/qwen38/README.md` |
 | GLM/DeepSeek runtime base and model pins | `runtime/faststart-lock.json` |
 | Qwen runtime/source pins and model identities | `runtime/qwen38/pins.json`, then `recipes/qwen38-27b-exl3-k5k6{,-pair}.json` |
@@ -104,7 +106,7 @@ test imports torch:
 python -m pip install -r requirements-dev.txt
 python -m pip install --index-url https://download.pytorch.org/whl/cpu "torch==2.11.0"
 ruff check --select E,F,W --ignore E501 spark_transport runtime scripts performance
-python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-adaptive-mtp-python-overlay runtime/glm53-flash-dflash7-python-overlay runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
+python -m pytest spark_transport runtime/exl3-r7 runtime/glm53-flash runtime/glm53-flash-jj-r7-gb10 runtime/glm53-flash-split-page-sparkcache runtime/glm53-flash-adaptive-mtp-python-overlay runtime/glm53-flash-dflash7-python-overlay runtime/deepseek0731-gb10 runtime/qwen38 runtime/test_public_overlay.py performance/harnesses scripts -q -rs
 ```
 
 The test suite is CPU-only contract coverage. It does not validate CUDA,
@@ -113,8 +115,10 @@ RDMA, live pair/cycle serving, or a performance result.
 ## Runtime and configuration work
 
 `runtime/exl3-r7/` builds the GLM-5.2 EXL3 R7 image.
-`runtime/glm53-flash/` records the GLM-5.3 Flash runtime, model, public BF16
-DFlash2, patched NCCL, SparkCache, and vLLM lease-contract identities.
+`runtime/glm53-flash-jj-r7-gb10/` records the published JJ r7-compatible GLM-5.3 base and
+SparkCache images, source/model identities, operator environment, and bounded
+TP4 smoke evidence. `runtime/glm53-flash/` retains the historical published
+artifact contract.
 `runtime/deepseek0731-gb10/` builds the hardened DeepSeek-V4-Flash-0731 image,
 and `runtime/qwen38/` builds the Qwen3.8-27B ARM64 image.
 `runtime/faststart-lock.json` pins the generic GLM/rollback image, the hardened
@@ -129,12 +133,10 @@ version control. Use `scripts/config/deepseek-v4-flash-0731-pair.env.example`
 for a two-rank pair and `scripts/config/deepseek-v4-flash-0731.env.example` for
 a four-rank cycle.
 
-Use `scripts/config/glm53-flash-tp4-site.example.yaml` with exactly one of the
-two GLM-5.3 runtime-profile templates. Keep resolved site addresses, host
-paths, image IDs, and credentials outside version control. Both GLM-5.3
-profiles use the same SparkCache-capable image and preserve asynchronous
-scheduling, native prefix caching, and chunked prefill; only the
-SparkCache-named profile enables the external connector.
+Use `runtime/glm53-flash-jj-r7-gb10/runtime.env.example` for the published JJ r7-compatible GLM-5.3
+serving. Keep resolved site addresses, host paths, and credentials outside
+version control. `IMAGE_VARIANT` selects the base or SparkCache digest; only
+the SparkCache variant enables the external connector.
 
 Use the topology-specific Qwen environment in `scripts/config/` with the image
 built by `runtime/qwen38/build-image.sh`, as described in

CONTRIBUTING.md

@@ -5,8 +5,9 @@ code are welcome. You do not need DGX Spark hardware or maintainer approval to
 participate. Partial reports are useful; share what you know and maintainers
 can help identify what is missing.
 
-The supported deployments and their maturity are listed in
-[`README.md`](README.md). This guide intentionally does not repeat that matrix.
+Deployment profiles and their evidence scope are listed in the
+[profile registry](docs/profiles/README.md). This guide intentionally does not
+repeat that matrix.
 
 ## Issues and discussions
 

README.md

@@ -3,123 +3,104 @@
 > Repository branches are mutable. Pin an immutable commit when reproducing a
 > deployment.
 
-  
-  SparkRing is a low-latency collective transport and vLLM-based
-inference-serving stack for switchless clusters of NVIDIA DGX 'Spark' systems
-powered by the GB10 Grace Blackwell Superchip.
+SparkRing is a collective-communication and inference-serving stack for
+switchless clusters of NVIDIA DGX Spark systems. It connects each system
+directly to its neighbours and runs distributed workloads without an external
+Ethernet or InfiniBand fabric switch.
 
-SparkRing supports GB10 pairs and four-node rings. Six-node ring work is
-research-only.
+The repository contains:
 
-Models run as tensor-parallel deployments over the direct fabric without an
-external Ethernet or InfiniBand switch. [SIRCL](https://github.com/FujitsuPolycom/sparkring/blob/main/docs/SIRCL.md) provides custom RDMA collectives
-where applicable, CUDA-graph command rings support repeated decode
-work, and [patched NCCL](https://github.com/FujitsuPolycom/sparkring/blob/main/spark_transport/nccl/README.md) handles communication outside SIRCL's supported paths.
+- cluster discovery, configuration, and diagnostics;
+- native RDMA collectives and vLLM integration;
+- patched NCCL support for communication outside the native transport;
+- reproducible runtime builders and deployment profiles; and
+- benchmark methods, records, and sanitized receipts.
 
-The repository provides launch tooling, model profiles, test evidence, and [performance data](https://github.com/FujitsuPolycom/sparkring/tree/main/performance).
+## Get started
 
-## Setup
+Start on the system that will serve as rank 0:
 
-Start with ssh to node0 and enough disk space for the intended model weights. 
-Use the [bootstrap guide](docs/BOOTSTRAP.md). The model-independent `sparkring cluster
-init` workflow enrolls nodes, discovers management and ConnectX-7 hardware,
-generates the cluster inventory, and launches Ring Doctor before any model
-profile is selected.
+```bash
+curl -fLO \
+  https://raw.githubusercontent.com/FujitsuPolycom/sparkring/main/bootstrap.sh
+less bootstrap.sh
+bash bootstrap.sh
+sparkring cluster init --size 4
+sparkring doctor --verify
+```
 
-## Resources
+The bootstrap workflow enrolls nodes over the management network, discovers
+ConnectX-7 hardware, writes a local cluster inventory, and checks the direct
+fabric before any serving profile is selected.
 
-- [Supported models and profiles](#profiles)
-- [Benchmark results](#benchmark-results)
-- [Deployment prerequisites](docs/PREREQUISITES.md) — then choose a profile quickstart below
+Read the [bootstrap guide](docs/BOOTSTRAP.md) before applying network changes.
+Then choose a deployment from the [profile registry](docs/profiles/README.md).
 
-## Profiles
+## Topologies
 
-### GLM-5.3 Flash profiles
+| Topology | Status | Fabric |
+|---|---|---|
+| Two-system pair | **implemented** | One direct 200 Gb/s link |
+| Four-system cycle | **implemented** | Four direct links in a closed `0-1-2-3-0` cycle |
+| Six-system cycle | **research-only** for serving | Six direct links in a closed cycle |
 
-| Profile | Deployment | Context | Seqs | Batch | KV / cache | Start here |
-|---|---|---:|---:|---:|---|---|
-| GLM-5.3 Flash + BF16 DFlash2 | 4 Sparks · TP4/DCP1 | 512K | 32 | 8,192 | 12 GiB/rank FP8 KV | [Quickstart](docs/GLM53_FLASH_DFLASH2_BF16_TP4_QUICKSTART.md) |
-| GLM-5.3 Flash + BF16 DFlash2 + SparkCache | 4 Sparks · TP4/DCP1 | 512K | 32 | 8,192 | 12 GiB/rank FP8 KV + 48 GiB nvme/rank SparkCache | [Quickstart](docs/GLM53_FLASH_DFLASH2_BF16_SPARKCACHE_TP4_QUICKSTART.md) |
+Each rank also needs a management-network connection for SSH, rendezvous, and
+the rank-0 API. The management network is not an inference-fabric edge.
 
-### Other model profiles
+## Communication paths
 
-| Profile | Deployment | Context | Seqs | Batch | KV / cache | Start here |
-|---|---|---:|---:|---:|---|---|
-| GLM-5.2 EXL3 3.5-bpw | 4 Sparks · TP4/DCP4 | 1M | 16 | 4,096 | NVFP4 DS-MLA · 9.25 GB/rank | [Quickstart](docs/GLM52_35BPW_QUICKSTART.md) |
-| DeepSeek-V4-Flash-0731 | 2 Sparks · TP2/DCP1 | 1M | 32 | 4,096 | FP8 DS-MLA · 16 GiB/rank | [Quickstart](docs/DEEPSEEK_V4_FLASH_QUICKSTART.md) |
-| DeepSeek-V4-Flash-0731 | 4 Sparks · TP4/DCP1 | 1M | 32 | 4,096 | FP8 DS-MLA · 16 GiB/rank | [Quickstart](docs/DEEPSEEK_V4_FLASH_QUICKSTART.md) |
-| Qwen3.8-27B EXL3 K5/K6 | 2 Sparks · TP2/DCP1 | 1M | 32 | 8,192 | FP8 | [Quickstart](docs/QWEN38_27B_EXL3_K5K6_PAIR_QUICKSTART.md) |
-| Qwen3.8-27B EXL3 K5/K6 | 4 Sparks · TP4/DCP1 | 1M | 64 | 8,192 | FP8 | [Quickstart](docs/QWEN38_27B_EXL3_K5K6_QUICKSTART.md) |
-| GLM-5.2 EXL3 3.5-bpw + SparkCache | 4 Sparks · TP4/DCP4 | 1M | 16 | 4,096 | NVFP4 DS-MLA + SparkCache | [SparkCache compositions](recipes/sparkcache/README.md) |
-| DeepSeek-V4-Flash-0731 + SparkCache | 2 Sparks · TP2/DCP1 | 1M | 32 | 4,096 | FP8 DS-MLA + SparkCache | [SparkCache compositions](recipes/sparkcache/README.md) |
-| DeepSeek-V4-Flash-0731 + SparkCache | 4 Sparks · TP4/DCP1 | 1M | 32 | 4,096 | FP8 DS-MLA + SparkCache | [SparkCache compositions](recipes/sparkcache/README.md) |
+[SIRCL](docs/SIRCL.md) provides persistent RDMA sessions and graph-replayable
+collectives for supported four-rank shapes. [Patched
+NCCL](spark_transport/nccl/README.md) handles other collective shapes and
+phases. A deployment profile states which path it uses.
 
-The public BF16 DFlash2 checkpoint is licensed CC BY-NC-ND 4.0 for research
-and evaluation; review its model card before use. The qualified GLM-5.3 community 
-images are published by immutable digest in the two quickstarts. Both guides also 
-link the complete source-build recipes, SBOM workflow, source commits, applied patches, and license record.
-See the [profile registry](docs/profiles/README.md) for recipe identities and evidence scope.
+See [architecture](docs/ARCHITECTURE.md), [deployment
+prerequisites](docs/PREREQUISITES.md), and [cable
+qualification](spark_transport/CABLE_QUALIFICATION.md) for the system
+contracts.
 
-## Benchmark results
+## Profiles and evidence
 
-### GLM-5.3 Flash research observation
+Model, runtime, topology, memory, and scheduler settings belong to deployment
+profiles rather than the transport definition:
 
-**Research-only — 16K context, single observation.** The SparkCache-enabled
-profile recorded 2,371 tok/s prefill and 36.06 tok/s sustained C1 decode on
-random tokens. No A/B baseline has been completed. C4 and C8 were capacity-limited
-and are omitted rather than reported as throughput results.
+- [profile registry](docs/profiles/README.md);
+- [runtime builders](runtime/README.md);
+- [serving configuration templates](scripts/config/README.md);
+- [benchmark results](docs/RESULTS.md); and
+- [methods, records, and receipts](performance/README.md).
 
-| Profile | Prefill | C1 decode | C8 decode | Highest valid decode | Coding peak |
-|---|---:|---:|---:|---:|---:|
-| [GLM-5.3 Flash + BF16 DFlash2 + SparkCache · 4 Sparks](performance/records/glm53-flash/sparkcache-dflash2-bf16-tp4-16k-run1-20260829.md) | 2,371 | 36.06 | — | C1: 36.06 | — |
-
-### Other model profiles
-
-**Qualified — 16K context.** All values are tokens per second. Prefill uses a
-cold prompt with caching disabled. Decode uses unique, cold prompt contexts at
-temperature 1.0; decode values are aggregate throughput across active streams.
-
-| Profile | Prefill | C1 decode | C8 decode | Highest tested decode | Coding peak |
-|---|---:|---:|---:|---:|---:|
-| [GLM-5.3flash NVFP4 · 4 Sparks](IN PROGRESS) | 2300 | 40 | 130 | C8: 130 | 70 |
-| [GLM-5.2 EXL3 3.5-bpw · 4 Sparks](performance/records/glm-3.5bpw/normalized-base-20260822.md) | 671 | 20.15 | 64.13 | C8: 64.13 | 25.39 |
-| [DeepSeek-V4-Flash DSpark · 2 Sparks](performance/records/deepseek-v4-flash/normalized-tp2-base-temp1-n5-20260823.md) | 1,926 | 58.36 | 162.69 | C32: 307.13 | 59.31 |
-| [DeepSeek-V4-Flash-0731 · 4 Sparks](performance/records/deepseek-v4-flash/normalized-tp4-base-temp1-n5-20260823.md) | 2,488 | 68.84 | 265.16 | C32: 508.11 | 95.77 |
-| [Qwen3.8-27B EXL3 K5/K6 · 2 Sparks](performance/records/qwen38-27b/normalized-tp2-1m-probmtp-temp1-20260823.md) | 1,367 | 29.50 | 142.20 | C8: 142.20 | 39.95 |
-| [Qwen3.8-27B EXL3 K5/K6 · 4 Sparks](performance/records/qwen38-27b/normalized-tp4-1m-probmtp-temp1-20260823.md) | 1,964 | 35.07 | 191.02 | C8: 191.02 | 48.46 |
-
-See [benchmark results and throughput tables](docs/RESULTS.md) for full
-matrices, sample counts, exact settings, and limitations.
-
-## Architecture
-
-Two-Spark profiles use one direct 200 Gb/s cable and patched NCCL. Four-Spark
-profiles use a switchless `0-1-2-3-0` cable cycle. SIRCL serves its tested
-collective paths; patched NCCL handles the remaining paths.
-
-See [architecture](docs/ARCHITECTURE.md), [SIRCL](docs/SIRCL.md), and the
-[deployment prerequisites](docs/PREREQUISITES.md).
+A successful offline test or process start does not establish serving
+correctness, output quality, capacity, or performance. Each profile documents
+its own evidence and limitations.
 
 ## Repository map
 
 | Path | Purpose |
 |---|---|
-| `spark_transport/` | Native transport and vLLM adapters |
-| `runtime/` | Pinned runtime inputs and builders |
-| `scripts/` | Site validation, preflight, launch, and evidence tooling |
-| `recipes/` | Machine-readable serving recipes |
+| `spark_transport/` | Native transport, patched NCCL, and framework adapters |
+| `runtime/` | Pinned runtime inputs and image builders |
+| `scripts/` | Cluster validation, launch, and evidence tooling |
+| `recipes/` | Machine-readable serving compositions |
 | `performance/` | Benchmark methods, records, and sanitized receipts |
-| `docs/` | Profile procedures, architecture, prerequisites, and evidence |
+| `docs/` | Architecture, prerequisites, profiles, and operator procedures |
+
+## Development
+
+See [CONTRIBUTING.md](CONTRIBUTING.md) for contribution and validation
+guidance. Security reports belong in private GitHub security advisories as
+described in [SECURITY.md](SECURITY.md).
 
 ## Acknowledgements
 
-SparkRing builds on work from vLLM, NVIDIA NCCL, B12X, SparkInfer, LMCache,
-ExLlamaV3, and the
-[local inference community](https://github.com/local-inference-lab/).
+SparkRing builds on vLLM, NVIDIA NCCL, B12X, SparkInfer, LMCache, ExLlamaV3,
+and work published by the [Local Inference
+Lab](https://github.com/local-inference-lab/). Model-specific profiles identify
+their exact upstream source, checkpoint, revision, and license.
 
 Detailed third-party attribution is in
-[`THIRD_PARTY_NOTICES.md`](THIRD_PARTY_NOTICES.md).
+[THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).
 
 ## License
 
-Apache-2.0. See [`LICENSE`](LICENSE).
+Apache-2.0. See [LICENSE](LICENSE).

SECURITY.md

@@ -1,6 +1,8 @@
 # Security Policy
 
-SparkRing is pre-release research software. There is no supported-versions table: the only supported version is the latest commit on `main`.
+SparkRing is pre-release research software and does not publish a
+supported-versions table. Security fixes are applied to `main`; deployment
+profiles separately identify the immutable revisions covered by their evidence.
 
 ## Reporting a vulnerability