Make measured NVFP4 the default; retain latest EXL3 and video evidence - #1
Merged
Merged
Conversation
Provenance: - Copied source tree from SparkGLM local measured revision 6769921bc0f1e9a66a7b5a94dec233267f541bf5. Existing upstream lineage, including https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks at eb0469fbb2b49fd7c025f594a3339a121e58f7a9, remains in the source ledger and notices. Original work: - Public snapshot on reviewed ancestry; no source-byte changes and no new performance claim. Verification: - Git tree identity equals the measured source snapshot. Publication checks run before pushing.
Provenance: - Copied source tree from SparkGLM local measured revision 45942d36d87c531981a6638f6e2c7f31398cea4e. Existing upstream lineage, including https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks at eb0469fbb2b49fd7c025f594a3339a121e58f7a9, remains in the source ledger and notices. Original work: - Public snapshot on reviewed ancestry; no source-byte changes and no new performance claim. Verification: - Git tree identity equals the measured source snapshot. Publication checks run before pushing.
Provenance: - Copied source tree from SparkGLM local measured revision 80b85461737059fd00ba8c7eedafd9749e267628. Existing upstream lineage, including https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks at eb0469fbb2b49fd7c025f594a3339a121e58f7a9, remains in the source ledger and notices. Original work: - Public snapshot on reviewed ancestry; no source-byte changes and no new performance claim. Verification: - Git tree identity equals the measured source snapshot. Publication checks run before pushing.
Provenance: - Copied source tree from SparkGLM local measured revision 45b03d696ad2ac865d576f9d6c08ad86ff126b49. Existing upstream lineage, including https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks at eb0469fbb2b49fd7c025f594a3339a121e58f7a9, remains in the source ledger and notices. Original work: - Public snapshot on reviewed ancestry; no source-byte changes and no new performance claim. Verification: - Git tree identity equals the measured source snapshot. Publication checks run before pushing.
Provenance: - Copied source tree from SparkGLM local measured revision 69794a914cb43dbdd55971bbd157d8759dc5bcc0. Existing upstream lineage, including https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks at eb0469fbb2b49fd7c025f594a3339a121e58f7a9, remains in the source ledger and notices. Original work: - Public snapshot on reviewed ancestry; no source-byte changes and no new performance claim. Verification: - Git tree identity equals the measured source snapshot. Publication checks run before pushing.
A fresh source install now selects the September 8 NVFP4 recipe through LLooM: native CUTLASS, MXFP8 DFlash2 TP2 k7, 512K context, 9 GiB KV per rank, 2K chunks and mixed scheduling. An explicit skip option reproduces the second video; --profile exl3 builds the latest corrected concurrent E3 recipe. Preserve the old standalone launcher and its regression coverage, the measured source trees, and checksum-bound comparison/capacity/video receipts. Provenance: - Adapted managed recipe metadata from https://github.com/Enntity/lloom at 75e08ca0923e7eb58f087b53032c82f88175ea3a, retaining its MIT notice. The public dependency supplies the same entrypoint plus the measured long-prefill transport fix. - Measured engine snapshots retain MiaAI-Lab lineage at https://github.com/MiaAI-Lab/GLM-5.3-Flash-EXL3-2x-DGX-Sparks revision eb0469fbb2b49fd7c025f594a3339a121e58f7a9 and all existing notices. Public/original tree identities are mapped in provenance/2026-09-08-publication-map.json. Original work: - Source-layer orchestration and default/profile selection, immutable two-rank recipe installation, publication mapping, measured-result presentation and X draft. No new numerical optimization, weights, compiled binaries or videos are included. - Maintainer-selected source-preview default; original qualification records keep their states. This does not claim G5 certification or bit-identical rebuilt images. Verification: - scripts/check.sh all, publication privacy/history and commit-provenance checks; immutable-layer and both-rank recipe tests; legacy regression coverage retained. Existing hardware receipts cover the selected source/settings; no new performance measurements are claimed by this packaging change.
Provenance: original implementation Original work: update helper dispatch and its fixture after the default launcher switch; mark earlier NVFP4 bring-up notes as historical. Verification: scripts/check.sh all passed, including the tinyGLM candidate-image override regression and publication audit.
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.
Make measured NVFP4 the public default and retain both comparison lanes
A fresh source install now selects the September 8 NVFP4 recipe through LLooM: native CUTLASS, MXFP8 DFlash2 TP2 k7, 512K context, 9 GiB KV per rank, 2K chunks and mixed scheduling. An explicit skip option reproduces the second video; --profile exl3 builds the latest corrected concurrent E3 recipe. Preserve the old standalone launcher and its regression coverage, the measured source trees, and checksum-bound comparison/capacity/video receipts.
Provenance:
Original work:
Verification: