Repository navigation
FP4-Local release: NVIDIA's FP4-Local checkpoint is the default - #8
Merged
Merged
Conversation
Pin nvidia/Qwen3.8-Flash-Next-FP4-Local @ b9ce18a6 (95.1 GiB) and the engine sparkqwen/atlas-20261016-fp4local @ a8976e86 (tree 6a37ba6b). start-node.sh takes --checkpoint (start.sh passes the pinned one, or the checkpoint .env names) and mounts a persistent --ple-cache at /atlas-ple, where every profile points ATLAS_PLE_PACKED_DIR. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The README leads with FP4-Local against the published RC16 numbers on the NVFP4 checkpoint, adds a Quality section (NVIDIA's evaluation, our probes, exactness, the W8A8-prefill NLL check, what is not measured) and how to choose a checkpoint; the NVFP4-era vLLM table moves out of the README. results/2026-10-16-fp4local holds the RigMark cards and receipts, sq_bench and the gates. bench/nll_probe.py is the NLL driver those results cite. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The build-time tests in a release image compile spark-model's lib tests, and a8976e86 embedded a bench/ fixture the build context leaves out. 9f64bf6b reads it at run time; the serving code is unchanged. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
./start.sh in a fresh clone built ghcr.io/enntity/atlas-sparkqwen:4618128c570b from source, served and answered its smoke question. On that image: the dev binary's prompt-logprob hash bit for bit, tool calling 21/21, and RigMark 9/9 (card raw/recipe/rigmark-FP4L.card.txt). The README links the full RigMark cards for the release image, the final dev run and RC16. 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.
Makes
nvidia/Qwen3.8-Flash-Next-FP4-Local@b9ce18a6(95 GiB) the default checkpoint, served by Enntity/atlassparkqwen/atlas-20261016-fp4local@9f64bf6b. The NVFP4 checkpoint stays available throughCHECKPOINT_REPO/CHECKPOINT_REVISIONin.env, on the same image.Engine: the FP4-Local series on the
sparkqwenlayer:~/.cache/atlas-ple);Install:
checkpoint.jsonpins FP4-Local.start-node.shtakes--checkpointand--ple-cache.ATLAS_PLE_PACKED_DIR..envcan name another checkpoint (both values or neither).4618128c570b.Verified through the release image:
./start.shin a fresh clone on our pair:RigMark on the release image vs RC16 on NVFP4 (published):
The full cards are in
results/2026-10-16-fp4local/raw/recipe/andraw/rigmark/, linked from the README.Quality:
Not covered:
4x262kand the 0.70 sidecar share;./start.sh;The GHCR image is not published yet, so
./start.shbuilds from source until it is.🤖 Generated with Claude Code