| # Build and verification record |
|
|
| ## Objective |
|
|
| Create a reproducible DS4 package that maximizes target-model quality within a |
| 128 GB unified-memory envelope. Preserve exact native MXFP4 routed experts on |
| the ten sensitivity-selected layers and retain the proven low-bit/Q8 policy |
| everywhere else. |
|
|
| ## Inputs and toolchain |
|
|
| - All weights were regenerated from |
| `apetersson/DeepSeek-V4-Flash-0731-Abliterated-FP8`. |
| - The sparse reference GGUF supplied metadata, tokenizer, tensor order and |
| shapes only. |
| - All 48 source-shard hashes, source metadata and the abliteration manifest |
| were verified before conversion. |
| - The routed-expert imatrix supplied exact coverage for the 129 tensors whose |
| output types consume calibration data. |
| - DS4 source revision: `d516d4eeb82c454aeb2831af1b1961801d6b571b`. |
| - Upstream `ds4f-mxfp4` base: `4893e0c40fba03dbc85555faeb035799aa04e0b6`. |
| - Quantizer SHA-256: `f0a381f4ada808ea2afa740d964354fa327fc1235ba7cebf50874eb89fb97ac5`. |
| - Runtime SHA-256: `2aaf20469b9918d6d6ab8787a02811c11228547cd979787879a03dba8a9e7824`. |
|
|
| ## Quantization policy |
|
|
| - Native MXFP4 gate/up/down routed experts on layers |
| `10, 14, 30, 34, 37, 38, 39, 40, 41, 42`. |
| - IQ2_XXS gate/up and Q2_K down routed experts on the other 33 MoE layers. |
| - Q8 attention, shared-expert and output tensors. |
| - Protected F16 indexer and auxiliary tensors. |
| - DSpark support with IQ2_XXS gate/up and native MXFP4 down projections for |
| target layers 40, 41 and 42. |
| |
| ## Conversion gates |
| |
| - Strict imatrix mode passed for main and DSpark conversions. |
| - Main GGUF contains exactly 1,328 tensors and the intended 30 MXFP4 tensors. |
| - DSpark support contains exactly 81 tensors and three MXFP4 down aggregates. |
| - Every native tensor comparison required an explicit `byte_compare: OK`. |
| - All 30 main and three DSpark MXFP4 tensors independently reproduced the |
| source FP4 codes and scale bytes. |
| - `ds4 --cpu --inspect --dspark-strict` reported zero missing tensors, invalid |
| bindings or metadata errors. |
| - The final GGUF hashes are recorded in `SHA256SUMS`. |
|
|
| ## Result |
|
|
| | Component | Bytes | GiB | |
| | --- | ---: | ---: | |
| | Main GGUF | `102,826,238,912` | `95.7644` | |
| | DSpark support | `7,297,737,120` | `6.7965` | |
| | Combined | `110,123,976,032` | `102.5609` | |
|
|
| Native MXFP4 improves fidelity over a second Q4_K requantization and uses 4.25 |
| bits per weight rather than 4.50. Native MXFP4 Metal and the mixed DSpark path |
| are newer than mature Q4_K kernels, so throughput should be measured on the |
| target system. |
|
|
| ## One-million-token context |
|
|
| Recommended target-only mode: |
|
|
| ```sh |
| ds4 --metal \ |
| -m DeepSeek-V4-Flash-0731-Abliterated-DS4-Quality128.gguf \ |
| --ctx 1048576 --prefill-chunk 2048 |
| ``` |
|
|
| Estimated residency is at most `110.30 GiB`. DSpark can be tested with a 1,024 |
| token prefill chunk; estimated residency is `114.02 GiB`. A 4,096-token chunk |
| with DSpark is estimated at `123.24 GiB` and is not a reliably resident mode on |
| a machine whose Metal recommended working set is approximately `121.60 GiB`. |
|
|
