Separate SparkRing subsystems and add GLM-5.3 DCP routes

[FujitsuPolycom]

Status

Implemented and TP4 smoke-verified. This draft publishes operator contracts for two immutable Linux/ARM64 GB10 images. The bounded smoke covers one four-host TP4/DCP1 C4 workload. It does not establish general throughput, soak behavior, or every model and concurrency shape.

Published artifacts

• Base runtime: ghcr.io/fujitsupolycom/sparkring-glm53-runtime@sha256:11922064b342de1fc98f0ef85e6648843c8fa7eb3e4f4353c6ad82d6e457dde0
• SparkCache runtime: ghcr.io/fujitsupolycom/sparkring-glm53-sparkcache@sha256:f012dd915c0fff0be384820c2d72cd015b83b9b33c3f980445dd718a807cd0c5
• Platform: linux/arm64
• Entrypoint: vllm serve; the launcher supplies only serving arguments.

The images contain no model checkpoints. The run contract pins the Local Inference Lab GLM-5.3 Flash NVFP4 target and the incoai BF16 DFlash2 draft at proposal depth seven. Inherited MXFP8 plumbing labels are recorded separately and do not identify the mounted draft.

Repository documentation boundary

Generic SparkRing pages now describe cluster topology, transport, collective admission, fallback behavior, runtime composition, and evidence routing without using a model as the subsystem identity. Model names, tensor geometries, checkpoint IDs, launch settings, and measurements remain in profile registries, recipes, quickstarts, runtime artifacts, and evidence records.

Resulting operator surface

• docs/GLM53_JJ_R7_GB10_TP4_QUICKSTART.md provides one route for the base and SparkCache variants.
• runtime/glm53-flash-jj-r7-gb10/runtime.env.example exposes common topology, capacity, batching, model, cache, CUDA-graph, and backend settings.
• launch-rank.sh verifies the immutable image ID and identity-bearing target and draft files before starting a rank.
• Checkpoints can be downloaded once on rank 0 and copied to the other ranks with rsync.
• One OCI image can be pulled on rank 0, saved once, and streamed to the other ranks over SSH.
• Base and SparkCache launches use distinct cache namespaces. Modified settings receive launch-specific container labels and do not inherit the bounded smoke status.
• Prior GLM-5.3 artifact procedures remain available as explicitly historical records rather than competing quickstarts.

Operator defaults

The launcher defaults to a 524,288-token request limit, 8,192 batched tokens, and a 524,288-token SparkCache publication span. These limits are implemented but unqualified. The bounded C4 smoke used 262,144, 4,096, and a 262,144-token span, so default launches are labeled implemented-unqualified-configuration rather than inheriting the smoke claim.

DCP2 and DCP4 routes

The cache-disabled launcher supports DCP1, DCP2, and DCP4. Automatic mode selects CP interleave 1 with full-CKV gather disabled for DCP1, and CP interleave 4 with B12X full-CKV gather enabled for DCP2/DCP4. At the 524,288-token request limit and 8,192 batched tokens, bounded DCP2 and DCP4 live requests completed with exact semantics.

The immutable published SparkCache image remains restricted to DCP1 because it does not contain SparkCache commit 65895c8. A source-overlay deployment completed exact DCP2 and DCP4 publication, full process restart, and SparkCache CUDA restore. runtime/glm53-flash-jj-r7-gb10/SPARKCACHE_DCP_SOURCE_OVERLAY.md records the boundary. A replacement pullable image remains required before DCP2/DCP4 can be offered through the public-image SparkCache route.

Exact composition

• SparkRing image builder: ca91fa72a4cf7e1edaad9875a1a99ab4f71c49af
• Active vLLM Python composition: 331573d20bd47e78327ed8d8b4d2e6d350bbb1ab, tree 927f52a0085bcecfd2ba679e5abebe1a62623daf
• Local Inference Lab B12X: 6255090a03b12c3f7d552102a02fac0b542fb8c9, tree 0bb58d0dcc10e29e00ff9850c0d719fca1aba5ad
• Switchless NCCL library SHA-256: 5f1c3f10d5ace66d4ba584415bbfe42b6ac1a0a9116a3b81dcbe50516ad924b3
• SparkCache source: dcbe040d339f243621163b0c6ed4ce96462403d8
• SparkCache CUDA placement library SHA-256: d57509052b73853bcc8e3c3f47bb81748d87b9cbd8d908fc20d4c79a09aa400c

The receipt distinguishes active Python source from compiled vLLM extensions retained from lower image layers. It does not claim that the active Python revision rebuilt every native extension.

Bounded TP4 result

The SparkCache image published four exact 16K contexts, restarted all four ranks, and restored all four requests with exact codewords. Every restored request reported a 100% external prefix hit. Client time was 0.561595–1.582937 seconds; worker cache service was approximately 277–394 ms.

This C4 smoke used complete 16K snapshots. It does not independently prove the separate C8 shared-base result recorded by the historical exact artifact.

Validation

python -m pytest ...
2015 passed, 9 skipped

ruff check --select E,F,W --ignore E501 spark_transport runtime scripts performance
All checks passed

JSON contracts, shell syntax, repository-relative links, release-safety checks, and whitespace checks also pass.

Compatibility and limits

• The SparkCache image keeps cache use disabled until explicitly selected and configured.
• Cache identities remain bound to exact model content, topology, physical rank, schema, geometry, and speculative policy.
• MTP with SparkCache, C16 page-delta restore, sustained soak, and live fault injection remain unqualified.
• The images and source contracts are immutable inputs; changing operator settings creates a different deployment claim.

This pull request remains a draft for human review. It updates the existing PR #150 branch and does not merge or close any pull request.

[FujitsuPolycom]

Superseded by draft PR #153, which is based directly on public main and includes the exact R8 ARM64 builder, DCP1/DCP2/DCP4 launcher, topology-aware KV defaults, SparkCache/vLLM-only modes, deep-context evidence, direct-fabric fan-out, and current quickstart. Closing the older stacked draft avoids two competing review surfaces.