Publish the reproducible GLM-5.3 R8 ARM64 runtime

[FujitsuPolycom]

Result

SparkRing can build, verify, distribute, and launch one GLM-5.3 Flash R8
Linux/ARM64 image with SparkCache across four GB10 systems. The launcher uses
one environment file for TP4 with DCP1, DCP2, or DCP4 and exposes the 1M
context limit, 8K batched-token budget, topology-aware KV allocation, model
paths, cache limits, speculation, ports, and network settings. One switch
selects persistent SparkCache or vLLM's GPU prefix cache alone.

The immutable Linux/ARM64 image is published at
ghcr.io/fujitsupolycom/sparkring-glm53-sparkcache@sha256:380283a506aeb8f9d486a3c64cd738e44268c3cc21590913ea9e4685869f256a.
Its local Docker image ID is
sha256:b3a13d8003e7de30d7737fd33c8307404e506ba570240819ec7eb4f5c611400f.

Technical reason

The R8 scheduler, DCP-aware SparkCache geometry, CUDA placement library,
switchless NCCL, and B12X runtime must remain one source- and image-bound
composition. A single launcher prevents DCP1, DCP2, and DCP4 from silently
selecting incompatible KV interleaving or full-CKV gather settings.

The direct-fabric image utility downloads one checksum-bound archive, forwards
it through adjacent links, verifies every copy, imports it on all four ranks,
and checks the local image ID.

Compatibility

• The runtime is Linux/ARM64 and targets four NVIDIA GB10 systems.
• DCP1 uses one-token KV interleaving without full-CKV gather.
• DCP2 and DCP4 use four-token KV interleaving with full-CKV gather.
• DCP1 defaults to 26 GiB of FP8 KV per rank; DCP2 and DCP4 default to 30 GiB.
• SparkCache uses the manager-pages-v2 namespace and defaults to complete
  snapshot-v1 publication.
• SPARKCACHE_ENABLED=0 omits persistent publication and restore without
  disabling vLLM prefix caching.
• Model checkpoints and site configuration remain outside the image.

Local Inference Lab provides the GLM-5.3 NVFP4 target quantization, Jovian
Judgement vLLM work, and B12X GB10 kernel integration. The external BF16 draft
comes from incoai/GLM-5.3-Flash-DFlash2. Exact revisions are in
runtime/glm53-flash-jj-r8-gb10/pins.json.

Validation

• python -m pytest runtime/glm53-flash-jj-r8-gb10 scripts/test_fanout_image_archive.py -q
  — 30 passed.
• python -m ruff check runtime/glm53-flash-jj-r8-gb10 scripts/fanout_image_archive.py scripts/test_fanout_image_archive.py
  — passed.
• python -m pytest spark_transport runtime scripts performance/harnesses -q -rs
  — 1,958 passed, 9 skipped.
• python -m ruff check spark_transport runtime scripts performance/harnesses
  — passed.
• The public image was distributed from one archive through three direct
  links and loaded with the same image ID on all four ranks.
• DCP1, DCP2, and DCP4 each completed fresh publication, four-process
  replacement, verified SparkCache CUDA restore, and a semantic marker check.
• The vLLM-only profile retained built-in prefix caching while omitting the
  SparkCache connector.
• The final public image completed a 942,898-token request in 473.4 seconds
  with 26 GiB of FP8 KV per rank and published a 942,592-token SparkCache
  snapshot.

The large snapshot was not replayed after process replacement. The retained
requests do not establish concurrent large-context restore, long-duration
behavior, or general throughput.