hugo/DS4Server
1
0
Fork 0
Code Issues 2 Pull Requests Actions Packages Projects Releases Wiki Activity

Prepare exact Qwen3.8 Flash Next Bare Speed artifacts #93

New Issue
Closed
opened 2026-09-01 11:53:12 +00:00 by hugo · 1 comment
hugo commented 2026-09-01 11:53:12 +00:00
Owner
Copy Link

Area

Prepare and publish the exact DS4Server artifact set for Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed. This issue owns artifact provenance, deterministic repackaging, precision preservation, and hashes. It does not implement inference.

The attached research plan and the completed MTPLX/official-source review are inputs to this implementation issue.

Sequence

Step 1 of 7. Next: #94. Do not begin #94 until this issue has frozen the artifact layout and hashes.

Implementation

  • Pin the exact Hugging Face revision of Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed, its Qwen/Qwen3.8-Flash-Next base revision, the Qwen Community license, tokenizer files, chat template, and mtplx_runtime.json serving contract.
  • Produce a deterministic DS4Server artifact set using repository-owned Rust tooling: resident transformer core, separate PLE table, separate MTP head, tokenizer/config files, and one signed or hashed manifest consumed by the existing download verifier.
  • Preserve the Bare Speed precision map exactly: every MoE expert and dense matrix at 4-bit with 64-weight groups; GDN convolution/recurrent parameters, norms, QSA indexer, and MTP head at 16-bit; PLE as the separate streamed table. Repackaging may change containers and tensor layout only when the numerical interpretation remains identical.
  • Preserve every tensor needed for text inference, PLE, and MTP. Vision tensors may be excluded from the DS4Server text artifact, but the exclusion and resulting hashes must be explicit.
  • Record tensor names, shapes, dtypes, quantization parameters, byte sizes, source URLs, source commits, output hashes, and converter version. Conversion must fail on missing, duplicated, renamed, or unexpectedly quantized tensors.
  • Reconcile current llama.cpp b356fa2624643b6d5753162ae43efff8cdd4d8cb: its Qwen converter exports PLE but deliberately drops MTP, so it cannot be used unchanged as the artifact authority.
  • Keep MTPLX Python/MLX code out of the product and converter. Independent implementation from the published model data and format contracts is required unless a separate provenance/license decision explicitly allows a source fragment.

Verification

  • Run conversion twice from a clean cache and prove identical output hashes.
  • Compare the complete source/output tensor inventories and precision assignments.
  • Compare representative dequantized rows from every tensor class, including experts, dense matrices, QSA, GDN, PLE, and MTP.
  • Verify the manifest can download and authenticate every file without loading model weights.
  • Record total download size and the separate resident-core, PLE, and MTP sizes.

Completion requirements

  • The exact Bare Speed source revision and license are pinned.
  • Reproducible DS4Server artifacts and their immutable hashes exist.
  • The precision map and representative tensor values match the MTPLX Bare Speed pack.
  • PLE and MTP are present as independently addressable verified artifacts.
  • No Python/MLX runtime or foreign executor is added.

Non-goals

  • Model registration, tokenizer execution, Metal inference, vision support, or performance tuning.
## Area Prepare and publish the exact DS4Server artifact set for `Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed`. This issue owns artifact provenance, deterministic repackaging, precision preservation, and hashes. It does not implement inference. The attached [research plan](https://git.rfc1437.de/attachments/0ce49e73-3337-4a28-949d-177a8cfb4e82) and the completed MTPLX/official-source review are inputs to this implementation issue. ## Sequence Step 1 of 7. Next: #94. Do not begin #94 until this issue has frozen the artifact layout and hashes. ## Implementation - Pin the exact Hugging Face revision of `Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed`, its `Qwen/Qwen3.8-Flash-Next` base revision, the Qwen Community license, tokenizer files, chat template, and `mtplx_runtime.json` serving contract. - Produce a deterministic DS4Server artifact set using repository-owned Rust tooling: resident transformer core, separate PLE table, separate MTP head, tokenizer/config files, and one signed or hashed manifest consumed by the existing download verifier. - Preserve the Bare Speed precision map exactly: every MoE expert and dense matrix at 4-bit with 64-weight groups; GDN convolution/recurrent parameters, norms, QSA indexer, and MTP head at 16-bit; PLE as the separate streamed table. Repackaging may change containers and tensor layout only when the numerical interpretation remains identical. - Preserve every tensor needed for text inference, PLE, and MTP. Vision tensors may be excluded from the DS4Server text artifact, but the exclusion and resulting hashes must be explicit. - Record tensor names, shapes, dtypes, quantization parameters, byte sizes, source URLs, source commits, output hashes, and converter version. Conversion must fail on missing, duplicated, renamed, or unexpectedly quantized tensors. - Reconcile current llama.cpp `b356fa2624643b6d5753162ae43efff8cdd4d8cb`: its Qwen converter exports PLE but deliberately drops MTP, so it cannot be used unchanged as the artifact authority. - Keep MTPLX Python/MLX code out of the product and converter. Independent implementation from the published model data and format contracts is required unless a separate provenance/license decision explicitly allows a source fragment. ## Verification - Run conversion twice from a clean cache and prove identical output hashes. - Compare the complete source/output tensor inventories and precision assignments. - Compare representative dequantized rows from every tensor class, including experts, dense matrices, QSA, GDN, PLE, and MTP. - Verify the manifest can download and authenticate every file without loading model weights. - Record total download size and the separate resident-core, PLE, and MTP sizes. ## Completion requirements - The exact Bare Speed source revision and license are pinned. - Reproducible DS4Server artifacts and their immutable hashes exist. - The precision map and representative tensor values match the MTPLX Bare Speed pack. - PLE and MTP are present as independently addressable verified artifacts. - No Python/MLX runtime or foreign executor is added. ## Non-goals - Model registration, tokenizer execution, Metal inference, vision support, or performance tuning.
QWEN38FLASH.md
40 KiB
hugo added the idea label 2026-09-01 11:53:12 +00:00
hugo changed title from support qwen 3.8 flash next to Prepare exact Qwen3.8 Flash Next Bare Speed artifacts 2026-09-01 18:36:51 +00:00
hugo added enhancement and removed idea labels 2026-09-01 18:41:42 +00:00
hugo commented 2026-09-03 17:16:24 +00:00
Author
Owner
Copy Link

Implemented and pushed as commit 4c83ac0.

The repository now contains a Rust-only qwen38-artifacts tool plus an immutable manifest and complete 2,527-tensor inventory for Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed at 74559cdf34fbfc0b593de72d17e93f37fd4f9ea7, based on Qwen/Qwen3.8-Flash-Next at de4b8e4d43b917e7706784d8bb445c9af86a3540 under the Qwen Community License 1.0.

The tool fetches/resumes immutable revision URLs, authenticates every file, validates safetensors structure and the core index, enforces every effective precision assignment, computes representative dequantized-row receipts, and materializes byte-preserving core/PLE/MTP artifact directories. Vision and processor artifacts are explicitly excluded with hashes. The pinned source truth is preserved where the issue prose was broader: routers and QSA index_qk_proj are 8-bit/group-64, PLE is 4-bit/group-32, and the MTP sidecar contains published 2/4/8-bit affine matrices plus BF16 state.

Verification completed on the full 105,438,842,430-byte selected set. Two clean 29-artifact materializations independently passed verification; normalized SHA-256 listings for all 31 output files were identical with list digest 3e0233c2356b26859de43b25b2a3df08169dd8263071432bdf2fc497da74007c. All six expert/dense/QSA/GDN/PLE/MTP value receipts are pinned in the manifest. The local DS4 tree has no Qwen3.8 or MTPLX implementation, so this is a new DS4Server artifact feature rather than a DS4 behavior port.

Commit gates passed: cargo fmt --all -- --check; cargo clippy --all-targets --all-features -- -D warnings; make bundle; cargo test --all-features (228 passed, 18 environment/hardware tests ignored, plus 4/4 artifact-tool tests passed).

Implemented and pushed as commit 4c83ac0. The repository now contains a Rust-only qwen38-artifacts tool plus an immutable manifest and complete 2,527-tensor inventory for Youssofal/Qwen3.8-Flash-Next-MTPLX-Bare-Speed at 74559cdf34fbfc0b593de72d17e93f37fd4f9ea7, based on Qwen/Qwen3.8-Flash-Next at de4b8e4d43b917e7706784d8bb445c9af86a3540 under the Qwen Community License 1.0. The tool fetches/resumes immutable revision URLs, authenticates every file, validates safetensors structure and the core index, enforces every effective precision assignment, computes representative dequantized-row receipts, and materializes byte-preserving core/PLE/MTP artifact directories. Vision and processor artifacts are explicitly excluded with hashes. The pinned source truth is preserved where the issue prose was broader: routers and QSA index_qk_proj are 8-bit/group-64, PLE is 4-bit/group-32, and the MTP sidecar contains published 2/4/8-bit affine matrices plus BF16 state. Verification completed on the full 105,438,842,430-byte selected set. Two clean 29-artifact materializations independently passed verification; normalized SHA-256 listings for all 31 output files were identical with list digest 3e0233c2356b26859de43b25b2a3df08169dd8263071432bdf2fc497da74007c. All six expert/dense/QSA/GDN/PLE/MTP value receipts are pinned in the manifest. The local DS4 tree has no Qwen3.8 or MTPLX implementation, so this is a new DS4Server artifact feature rather than a DS4 behavior port. Commit gates passed: cargo fmt --all -- --check; cargo clippy --all-targets --all-features -- -D warnings; make bundle; cargo test --all-features (228 passed, 18 environment/hardware tests ignored, plus 4/4 artifact-tool tests passed).
hugo closed this issue 2026-09-03 17:16:28 +00:00
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
Clear labels
cannot reproduce
bug
Something is not working
 critical
issues that are more important than others of the same group
 duplicate
This issue or pull request already exists
 enhancement
New feature
 help wanted
Need some help
 idea
just an idea
 invalid
Something is wrong
 question
More information is needed
 wontfix
This won't be fixed
No Label enhancement
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
hugo
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: hugo/DS4Server#93
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?