Skip to content

SparkRing 2026.09.2

Choose a tag to compare

View all tags
@FujitsuPolycom FujitsuPolycom released this 19 Sep 01:01
shared-2026.09.2
194d23d
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

SparkRing 2026.09.2

Shared ARM64 serving image for NVIDIA GB10 clusters: vLLM, SparkCache and an
isolated SGLang runtime. CUDA 13.3 and PyTorch 2.13.0. Model weights are separate.

docker pull ghcr.io/fujitsupolycom/sparkring@sha256:b58746401f0d51874696eb7fe37f0cffa5bbd1a1aed1dce26ef7be322e8fe123

Tag: ghcr.io/fujitsupolycom/sparkring:shared-2026.09.2.

Changes from 2026.09.0-rc.1

The runtime change from image 2026.09.1 is the CUDA cleanup correction. The list
below also includes improvements since the preceding shared-image GitHub prerelease.

  • Clears unused Qwen sparse-attention request IDs before CUDA-graph execution,
    addressing repeated output and invalid scores caused by stale padding metadata.
  • Prevents a reproduced kernel-timing deadlock: automatic Python cyclic cleanup
    waits until the host releases queued GPU work before unloading CUDA libraries.
    Normal cleanup resumes afterward; libraries are not permanently retained.
  • Includes the B12X #394 sparse-selection fix
    and corrects Qwen multimodal forwarding into the four-rank hyperconnection
    prefill path. Two-rank profiles keep row sharding disabled.
  • Adds a startup configuration audit before API readiness. It reports incompatible
    or suboptimal settings without silently changing them.
  • Gives both Qwen deployment sizes one image for cache-on and cache-off operation,
    distinct -TP2/-TP4 API model names, and dedicated SparkCache disk namespaces.

Managed B12X loading, MTP3, CUDA graphs, checkpoint coalescing and the applicable
communication enhancements remain selected by the Qwen profiles.

Profile status

Quickstart SparkCache Status
Qwen QAD TP2 Off alternative Qualified; bounded checks
Qwen QAD TP2 + SparkCache On Qualified; bounded checks
Qwen QAD TP4 Off Qualified; bounded checks
Qwen QAD TP4 + SparkCache On Qualified; bounded checks

Source checks, installed-file verification and the CUDA cleanup reproducer passed.
All four profiles passed short and 16K text, synthetic images/video, request-order,
concurrent-answer and retained-restart checks. Both SparkCache profiles restored
independent fixtures on every rank. TP2 completed three retained restarts; TP4 two.
The correctness record
identifies the exact image and evidence. Configured 262,144-token context and
16 sequences are not full-context or C16-pressure qualification. No benchmark
results are published.

GLM, DeepSeek and other profiles retain their image pins and are unqualified on
this version. SGLang retained-container restart is unqualified. Included
components are not a compatibility guarantee. APIs require
trusted-network access or an authenticated gateway. Request cache_salt does not
isolate SparkCache disk entries; use separate deployments for tenant isolation.

Sources and licenses

Source records
identify complete vLLM/B12X patches and archive hashes.
Component notices
retain upstream terms, including Mia's AGPL sources in the
inherited source bundle.
The exact inherited NCCL ARM-port patch is unavailable; these assets do not provide
a complete offline rebuild of every foundation library.

Assets 15
Loading