Repository navigation
SparkQwen on the shared Spark engine, with working image prompts (sparkqwen/atlas-20261015-vision) - #7
Merged
Conversation
SparkQwen and SparkGLM now build on one Enntity/atlas engine layer, upstream/atlas-shared-20261015 (fb80cbe0) on Atlas-Inf main 7ffd183f: the GLM-5.3-Flash series, the Qwen3.8-Flash-Next series and the model-independent work both products run. The SparkQwen-only RC15/RC16 commits sit on top as sparkqwen/atlas-20261015-shared (b914c5e1, tree a966cba6); its tree is the prefill-cache release fcf9735f plus the shared layer's newer work, with RC16's merge resolutions restored and every commit compiling. Same-pair A/B against the prefill-cache image: identical prompt-logprob hash and matching decode, prefill, concurrency and warm-cache numbers (results/2026-10-15-shared-engine). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Production and the shared-engine build ran back to back on ennspark01/02, followed by a smoke run of the release image built from the published pin. Raw receipts and SHA256SUMS are included. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-vision) The vision fixes join the one shared layer (upstream/atlas-shared-20261015-vision = fb80cbe0 + 5 commits): every vision tower encodes per chunk into per-sequence state, a malformed vision layout fails only its request, MRoPE positions come from the whole prompt, vision markup in message text encodes as text, and every tower's image prompts are keyed for the prefix cache. Both products pin layers on it. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The release image on the shared engine with the vision fixes passed the text, vision and fill checks on ennspark01/02. RigMark ran with the published command lines. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
data-angel
changed the title
The README's RigMark figures predate the 01/02 clock cap. The headline run lifts the cap for its duration (CLOCKS=uncapped); production stays capped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Alternating production and release starts with fixed prompts: step time and decode are equal to production. The RigMark differences were run-to-run sampling spread. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… per layer Enntity/atlas now has spark-shared, sparkglm and sparkqwen branches. Releases are tags, and the former dated branches are tags of the same name. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Moves SparkQwen onto the one Atlas engine layer it shares with SparkGLM, and makes image prompts work.
Engine:
sparkqwen/atlas-20261015-visionat1954e759. It is Atlas-Infmain7ffd183f, plus the shared layerupstream/atlas-shared-20261015-vision(a9464810), plus the SparkQwen-only commits, each of which compiles. Image tag4e29c8c22671.RC16 content restored: fixes that lived only in RC16's merge resolutions were lost in the first linear replay and are back.
Images:
<|image_pad|>string encodes as text, and a malformed layout fails only its request instead of crashing both ranks.Text: prompt-logprob hash
8713a659d5d3, same as production. sq_bench in line with production.RigMark, RC16's command line at
8x32k/0.88, with the GPU clock cap lifted for the run (production serving stays capped):All three runs pass 9/9 output gates. Each concurrency level is a single sampled run, and the capped pair of runs went the other way at C8. The capped runs are in the result too.
Details and receipts are in
results/2026-10-15-shared-engine. Not deployed and not published to GHCR; both are left for review.🤖 Generated with Claude Code