Repository navigation
SparkRing shared ARM64 runtime 2026.09.0-rc.1
Pre-releaseSparkRing shared ARM64 runtime shared-2026.09.0-rc.1
Candidate for NVIDIA GB10/SM121. Includes vLLM, isolated SGLang, SparkCache and
profile-selectable collective/kernel integrations. Model weights are not included.
Existing profile image pins remain unchanged; this prerelease is not a universal
profile recommendation.
docker pull ghcr.io/fujitsupolycom/sparkring@sha256:9439d39ad4b104c06cad3416a927f0f90852b456f0b06aa6a1dcf9475bf91a83| Configuration | Qualification |
|---|---|
| Qwen3.8-Flash-Next NVFP4-QAD, TP2/DCP1 + SparkCache | Qualified for bounded C1 text and process-restart restore; 5,696 cached tokens per fixture on both ranks |
| Qwen3.8-Flash-Next NVFP4-QAD, TP4/DCP1 + SparkCache | Qualified for bounded C1 text and process-restart restore; 7,200 cached tokens per fixture on all four ranks |
| GLM-5.3-Flash TP2/TP4 + SparkCache | Bounded restore results; research-only with the limitations below |
| DeepSeek-V4-Flash-0731, vLLM TP2 without SparkCache | Qualified only for startup and one C1 text request: 18 prompt tokens, 2 completion tokens |
| DeepSeek-V4.1 isolated SGLang TP4 | Fresh startup and one authenticated C1 text request passed; no SparkCache |
| DeepSeek-V4.1 vLLM and other configurations | Not qualified |
Qwen conditions: checkpoint 629bc3218833a38b475b719f34aa571666f4a03e,
262,144 context limit, MTP3 with one MTP module, C1 requests and aligned recurrent
checkpoints. This does not qualify performance, full context, concurrency pressure,
media correctness, corruption/eviction handling or soak workloads. The dedicated
transport probe retains its separate a75bd02f image identity; it was not rerun on
this candidate. Consult the attached evidence for exact settings and exclusions.
The DeepSeek-0731 check used 131,072 configured context, C8 configured capacity,
4,096 batched tokens, 8GiB KV per rank, B12X loading and DSpark5. It did not test
SparkCache, restart recovery, full context, concurrency pressure, media, throughput
or the public 1M/C32/16GiB profile defaults.
GLM used local checkpoint df116c4fb16b1d37ae43d2cfd624de26ffbc832e.
TP4 restored 7,680/8,192 tokens on all ranks after a repeated worker restart;
an earlier exact-answer mismatch remains unresolved because its raw response was
not retained. TP2 passed restart/restore only in a resource trial: 64K context,
C4, batch4,096, 2GiB KV/node, uniform512 pages and coalescing disabled. Larger
memory envelopes failed. One cold response had the correct key wrapped in
Markdown bold; both replay responses were exact bare keys. These results do not
promote either public GLM profile or establish throughput/full-context safety.
The isolated SGLang check used TP4/EP4, 655,360 configured context, C8 capacity,
4,096 prefill chunks, one prefill admission slot and DSpark5. It reused packed
NVMe Engram tables. Fresh startup does not resolve retained-container restart
failures on the separate baseline image.
Per-request cache-salt isolation is unsupported with SparkCache. OpenAI
cache_salt affects native vLLM GPU block hashes, but does not partition
SparkCache disk entries, whose keys use token/media inputs and static deployment
identity. A salted request is not a cold external-cache control. Use independent
deployments and cache namespaces, or disable external caching, when isolation
is required. No multi-tenant isolation is claimed. Runtime source-contract
integrity checks do not establish request-level isolation; this image does not
change SparkCache's CacheIdentity wire values, digest salts or chunk geometry.
The runtime-build configuration is sha256:768620d56aaed88528941bf97b1d1d8a7f4aec0918a4c2d681689372f3838b32; its labels-only candidate
is sha256:22da81cae0572ae2985a5c34a125db4f3cc58e871fa7e6256dffe59828f1ae5d. Registry manifest digests are separate identities. Source/tooling
adoption commit 3d85d78b16e917dafca915c223f8067594cfcc81 does not relabel the earlier build checkout.
The source archive shared-2026.09.0-rc.1-sources.tar.gz has SHA256 1066426891c461d43cea3562148900dc5734a9aec6205bd8ee5dd87cd8edb9ae and includes complete
modified Mia source, build/install recipes and retained component notices.
Licenses are component-specific: SparkRing integration is Apache-2.0; Mia retains
AGPL-3.0-or-later, upstream MIT and SGLang Apache notices. The image is not licensed
solely under Apache-2.0. The bundle does not claim an offline source-only rebuild
of every foundation library; exact inherited NCCL ARM-port patch bytes remain
unlocated. Existing SGLang restart and unqualified-profile limitations remain.
Source: pinned release inputs.
Publication establishes availability, not profile promotion. No floating latest
tag, profile defaults, or main-branch merge is part of this release operation.