[Tsury]

Older DeepSeek-V4.1 GGUF conversions can have supported tensors but fail during packing. The pinned mrfakename Q2_K uses engram.key_length and engram.pad_id; current packing expects head_dim and pad_token_id, so write_engram raises TypeError: list indices must be integers or slices, not NoneType. Those files also omit the sparse KV/index source lists and candidate geometry entirely.

The packer now accepts those two older Engram names. An explicit --config-json option fills omitted sparse geometry from the exact source model configuration. It checks existing core/sparse geometry for disagreement instead of replacing it. This does not rewrite GGUF metadata or tensors and does not change engine code or inference defaults.

Example:

python tools/dsv41/pack_gguf.py --model /path/to/Q2_K --out /path/to/pack --config-json /path/to/source-model/config.json

The pinned source config agrees with the official 1017 config: Engram dimension256/pad2; KV source layers2,8,14,20; candidate layer20/top-k2048/block8. This is the source of the missing values, not a guess based on architecture size.

Validation on Linux/Python3.14, base 7dd5cf18e68c639a8773e096c5c79f971ea2a9f5:

• Passed: python3 tests/terminal/test_pack_gguf_engram.py (5 tests). The new legacy-Engram case first reproduced the TypeError on the base implementation. It now writes the expected DSE1 header and identical bytes to current names. Canonical values retain precedence. The source-config case checks restored geometry and rejects a conflict.
• Passed: full pack CLI with all ten real Q2_K tensor directories/metadata retained from bounded HTTP Range reads. Temporary sparse files had the published logical sizes but contained no weight data. Packing produced47,286 tensor entries, correct source/candidate geometry, and two Engram layers with24 hash columns. The resulting Engram hash-table binary was byte-identical to the current-format Orca1017 header's table.
• Passed: target Python3.15.1 packer --help exposes the option; syntax checks and git diff --check.
• Not run: live model loading or inference for this repair. Weights are still downloading. Header/packing checks do not establish numerical correctness or performance. No backend code changed, so CUDA/HIP/SYCL builds are outside this Python metadata fix.

[Tsury]

Follow-up validation: the metadata repair was used successfully with the unmodified signed DeepSeek engine 0.1.49 after converting this artifact’s separately unsupported dense Q2/Q3 tensors to BF16 and Engram Q2_K tables to Q6_K through stock GGML reference functions. Routed expert bytes were unchanged. The engine loaded and produced coherent, correct arithmetic (17 × 19 = 323).

The unconverted Q2_K artifact still cannot load as published: the dense loader falls through to the generic “not an EXL3 matrix” error, and the Engram reader does not accept Q2_K. This PR fixes metadata compatibility; it does not add those tensor formats or claim general model-quality validation.

[Niko1221]

Thank you. Fixed in 0.2.7: Accept older DeepSeek GGUF metadata names. Your name is on the commit.

Release: https://github.com/Niko1221/Strata/releases/tag/v0.2.7