| # Provenance — ARustyCoder93 |
|
|
| Everything below was recorded on the build host at build time or checked |
| against the files afterwards. Dates are UTC unless marked. |
|
|
| ## Output |
|
|
| | file | bytes | sha256 | |
| |---|---|---| |
| | `arustycoder93.gguf` (built as `deepseek4-keep93.gguf`, 2026-08-16 19:25 UTC) | 62,213,283,424 | `90a4c8a8a32fc664d822f1df093d5e228111c94eebf9398ae5c5737391ecb058` | |
|
|
| GGUF header: arch `deepseek4`, `expert_count = 93`, `expert_used_count = 6`, |
| 43 MoE layers (3 hash-routed + 40 router layers), all other tensors and |
| metadata copied byte-for-byte from the source. Coarse ftype label as shown by |
| llama.cpp is not a quant family — the file is mixed (MXFP4 experts, Q8/F16 |
| elsewhere), exactly as in the source. |
|
|
| ## Source model |
|
|
| | item | value | |
| |---|---| |
| | repo | https://huggingface.co/antirez/deepseek-v4-gguf (MIT), base `deepseek-ai/DeepSeek-V4-Flash` (MIT) | |
| | file | `DeepSeek-V4-Flash-MXFP4Experts-F16HC-F16Compressor-F16Indexer-Q8Attn-Q8Shared-Q8Out-chat-v2-mxfp4-0731.gguf` | |
| | bytes | 155,976,458,848 (matches the HF tree listing) | |
| | sha256 | `0e3a161b670f686128ec5f92a601dfde616a37bf5e7e48999fa2d32471b57ec6` | |
| | arch | deepseek4, 43 layers × 256 routed experts, top-6 + 1 shared; layers 0-2 hash-routed via `ffn_gate_tid2eid` | |
|
|
| ## Build host |
|
|
| | item | value | |
| |---|---| |
| | GPU | NVIDIA RTX PRO 6000 Blackwell Workstation Edition, 96 GB (97887 MiB), power limit 350 W; one GPU used (`CUDA_VISIBLE_DEVICES=0`) | |
| | driver / CUDA | 610.57.04 / CUDA 13.3 (nvcc V13.3.73) | |
| | OS | Ubuntu 24.04.4 LTS, kernel 6.8.0-137-generic | |
| | llama.cpp | `ggml-org/llama.cpp` commit `22b8e310b921d568e013e4533002be5a8fe53f17` (2026-08-15, "server: re-design yield_to_queue thread model (#27133)") + `llama.cpp-patch/wants-census.patch`; built Release, `GGML_CUDA=ON`, `GGML_NATIVE=ON`, gcc/g++ from Ubuntu 24.04 | |
| | python | 3.12.3 (venv), `gguf==0.19.0`, `numpy==2.5.2`, `tqdm==4.70.0`, `PyYAML==6.0.3` | |
|
|
| ## Scripts (as run; identical to the copies in `scripts/`) |
|
|
| | file | sha256 | |
| |---|---| |
| | `prune_experts.py` | `f1ea4f50c49c354da32a64e232b47e2a8e41ec178483824a6ef1765ca0bf5962` | |
| | `compare_sets.py` | `0d6992654b3cd2a08c0bbba643e83c638583ff0c53c888dba2c0be6db40bb76c` | |
| | `apply_wants_patch.py` | `592f3f34a76a0786fbbb512af9120252cbab59d93359aca102b8c3e7916b4406` | |
| | `grow.sh` | `1d49a05c619be943552660312a1e5391cdb1664c9c3ba1e5964f10f88458ff3c` | |
| | `grow2.sh` | `259ed5e0c78a15b1b85f67dc9cf5dd3d79b48eba84fce2df7bc29df1912602f1` | |
| | `grow-coder.sh` | `9dd532ca77f56b5f1eb1c84dab4ec0f28839f6042a3fed8f737d36ed6c4b32fd` | |
| | `build_corpus.py` | `a516e863acf76136334e417dfe3590ef2d4fce3a8bb9ff73c5a17d9745b7b3c4` | |
| | `build_coder_corpus.py` | `36a762ae06ba5ae5a9f26327b5e74d0276b765d9ed0f95203d3895bf8cc8ec6b` | |
| | `merge_wants.py` | ``62427369e07ccdb03afc7d0bce3d58c133f770b4275b91a884e52eead0158a4e`` — written after the fact (2026-08-17) as the re-implementation of the inline merge; verified top-93-identical per layer against `merged-wants.tsv` | |
| | `grow-ppl-paired.sh` | ``c2272bcc1edb5d1d427542cdc23fe1deb5268baf9df1edc2834952a54c97cafc`` | |
| | `grow-ppl-ci.sh` | ``e0350ea7d3490d3b98c626366e37f37f0acb29f6f5f2a3553e9f7e5eaafaf00f`` | |
| | `ppl-paired.py` | ``349be3a0af9f7c53e79d1bf620d919b396a65bdcf84f063eaa27046b5addee57`` (the paired-test tool; results/RESULTS.md) | |
| | `code-smoke.py` | ``cd2fb11c48f43fb02887ef717ece92ac632723ad4c128249c8164154ee093620`` | |
|
|
| `grow-coder.sh` as published differs from the as-run copy in two comment/ |
| provenance lines (an internal repo path and a wiki reference were removed); |
| the sha256 above is the as-run file. Likewise `configs/grow-coder/config.txt` |
| had one appended correction paragraph shortened and the hostname redacted in |
| all three `config.txt`; nothing numeric changed. |
|
|
| ## Census / ranking files (in `wants/`) |
|
|
| | file | sha256 | role | |
| |---|---|---| |
| | `coder-census-64.wants.tsv` (was `grow-coder/census-64.imatrix.wants.tsv`) | `bead82b2976d9c48ae4339c1f2ae2c1d32e72cc1386b530078258e572658cd0a` | final code-grow census, input A | |
| | `general-census-64.wants.tsv` (was `grow/census-64.imatrix.wants.tsv`) | `f20041fac879f2c4553f416fc03383d4bd30ea0435b38144668d97ffb665621b` | final general-grow census, input B | |
| | `merged-wants.tsv` | `8152442949c297b0dc159533f71f883165a709df5b3deaa4debf07b3b02b263b` | the ranking the model was pruned with (`PRUNE_WANTS`) | |
|
|
| Final prune command (from `configs/grow-union/config.txt`): |
|
|
| ``` |
| PRUNE_WANTS=merged-wants.tsv python prune_experts.py SRC.gguf deepseek4-keep93.gguf 93 6 |
| ``` |
|
|
| ## Corpora (not redistributed; rebuild with the builders + these sources) |
|
|
| | file | bytes | sha256 | built from | |
| |---|---|---|---| |
| | `coder-train.txt` | 4,197,328 | `f1698b26cc486134027b9207eaf0109a1b716e2c0643c74049a7681685466122` | `build_coder_corpus.py`: ripgrep `crates/**/*.rs`, llama.cpp `src/ggml/tools` C/C++ + `gguf-py`/`scripts` Python + headers, Rust book `src/*.md`, OWASP CheatSheetSeries, Linux man pages; 4000-char chunks, `random.seed(7)`, every 10th chunk to heldout | |
| | `coder-heldout.txt` | 468,010 | `18314cd45eccf14cda9fee14542a6ee0b7bf9c9022bf27776bef2a6b5e13f7b0` | same | |
| | `domain-train.txt` | — | `a3e3fd1a326e6568f7ea944d186cc8826a198bdec0fcd14a92d1c2e1f1b6e045` | `build_corpus.py`: Rust book, ripgrep, OWASP, man pages, GSM8K (first 1800 rows), plus ~0.7 MB of our own model reasoning traces (`agentic.txt`, sha256 `ae4f8f386dd9fdff1cd72385561dfaae2eb84ee82d2b958c51c6d7a5b4723e05`, not published) | |
| | `domain-heldout.txt` | — | `78e379ccaa5b5dfcb0bbb2f106c78aa10940069e484d238afb499ec683f111aa` | same | |
| | `gsm8k.jsonl` | 4,166,206 | `17f347dc51477c50d4efb83959dbb7c56297aba886e5544ee2aaed3024813465` | GSM8K train split, one JSON object per line (`question`, `answer`) | |
| | `man.txt` | 1,602,349 | `a95d16db24d8ae616ef582c657c900f522fe77b26be6313fa0a644b3a740d213` | concatenated `man` output from the build host (starts with `SYSTEMD(1)`) | |
| | `ds4flash-domain-200.imatrix` | — | `26ceeec5a0f1c6f29809c4154c4eb1872f5518eddfd9125732dc3a354f4b701e` | full-model "truth" census over `domain-train.txt` (205 chunks); used only for comparison | |
| | `wiki.test.raw` | — | `173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08` | wikitext-2 test, PPL reference only | |
|
|
| Corpus source checkouts on the build host: |
|
|
| | repo | commit | |
| |---|---| |
| | github.com/BurntSushi/ripgrep | `3fce3b5bb0236da2df6d99672afb8a719642eca7` | |
| | github.com/OWASP/CheatSheetSeries | `07111ee754e832e335377ac64fd0f8f848d9029c` | |
| | github.com/rust-lang/book | `917544888a55e4da7109bdba8c88c893c0da70f4` | |
| | github.com/ggml-org/llama.cpp (code corpus source) | `22b8e310b921d568e013e4533002be5a8fe53f17` (working tree with the wants patch applied) | |
|
|
| ## Grow runs |
|
|
| **General grow** (`grow.sh`, then `grow2.sh` after a pruner parser crash at |
| the keep-16 stage; 2026-08-15 21:57 → 23:28 UTC per `configs/grow-general/grow.log` and file mtimes): random keep-8 |
| (`PRUNE_SEED=1`, `PRUNE_FULL_ROUTER=1`) → census 100 chunks of |
| `domain-train.txt` at `-c 2048 -b 2048 -ub 2048` → keep-16 → keep-32 → |
| keep-64; final `grown-keep64.gguf` (45,531,490,912 bytes, from `ls`; no sha256 recorded) built from |
| `census-64.imatrix.wants.tsv` with a normal pruned router. Per-stage logs in |
| `configs/grow-general/`. This run predates the config-record habit, so it has |
| no `config.txt`; the driver and logs are the record. The keep-8 stage ran with |
| the pre-fix pruner (the bug was in `PRUNE_WANTS` parsing, which the random |
| keep-8 stage does not use); every later stage used the pruner at the sha256 |
| above. |
|
|
| **Code grow** (`grow-coder.sh`, 2026-08-15 23:44 → 2026-08-16 01:16 UTC): |
| same schedule on `coder-train.txt`; full record in |
| `configs/grow-coder/config.txt` (final `deepseek4-coder-keep64.gguf` |
| 45,531,490,912 bytes, sha256 |
| `d9630be3be692823db1460d544035dda737f30d72bb84c8da2c9d0071e5fb186`). |
|
|
| **Union build** (2026-08-16 19:25 UTC): `configs/grow-union/config.txt` + |
| `prune.log`. Bytes in 145.3 GiB → out 57.9 GiB (tensor payload); file |
| 62,213,283,424 bytes. |
|
|
| **merge-128** (`configs/grow-merge/`, 2026-08-16 07:43 UTC): the earlier |
| union + ~40 filler experts at keep-128. Generated cleanly but did badly on an |
| internal 92-question reasoning eval (most cases hit the token cap; of the |
| 17 it finished within budget it got 9 right where the unpruned base got 16) |
| and was deleted; kept here because it is where the +1e7 |
| union-bonus rule was found and verified. |
|
|
| ## Serving (as run for the smoke tests) |
|
|
| ``` |
| llama-server -m arustycoder93.gguf -ngl 99 -c 262144 -fa on |
| ``` |
|
|
| ~67 GB resident on the PRO 6000 at 256K context; ~46 tok/s decode, ~470 |
| tok/s prompt processing (from the server log, single request). |
|
|
