README: commit, commands, hashes

README.md

@@ -1,59 +1,35 @@
----
-license: other
-license_name: qwen
-license_link: https://huggingface.co/Qwen/Qwen3.8-Flash-Next
-base_model: Qwen/Qwen3.8-Flash-Next
-tags: [unbiasedstrata, strata, mtp, speculative-decoding, gguf]
----
-
 # UnbiasedStrataPacks
 
-Pre-built, host-independent artifacts for serving Qwen3.8 Flash-Next GSQ-RCO quants with
-[UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata). Each one is a deterministic
-product of upstream's tools; building it on every host means a multi-GB download and a quantize
-step, so it is built once here and copied.
-
-## mtp/qwen3.8-flash-next-q2_0 — the MTP draft head (787 MB)
-
-The runtime directory the engine loads with `--mtp DIR`. The GSQ-RCO GGUFs ship no MTP head; it
-comes from the BF16 checkpoint.
-
-| file | bytes | sha256 |
-|---|---|---|
-| dense.bin | 116,099,072 | c724dc0b0822ada5d2977bf5bde821605feabaa64ea2e0045b67ca656329070a |
-| dense.txt | 1,851 | 991456cf4eb266eb8a888c349e548d11bc640f5dbd9149439aa9786af0f732fd |
-| draft_vocab.bin | 425,196 | b1e1d3a7a9e4bf862dcd5923ce661fb59bbd07907e594df5cf86a62ac235cb91 |
-| experts.bin | 707,788,800 | 09398406be61f1f54c93861f449e48b8df0bfccbc9ec9b2b7636775a6ea9244f |
-| mtp-manifest.json | 11,643 | 1ad281ed87ad9cd42f9c048e9c49172165ea714a600d15f1d80cd0a02db12328 |
-
-Provenance:
-
-- Source: the 31 `mtp.*` tensors of `Qwen/Qwen3.8-Flash-Next` at revision
-  `de4b8e4d43b917e7706784d8bb445c9af86a3540` (the revision pinned in `tools/mtp_fetch.py`), 5.21 GB
-  of BF16 read by HTTP range requests; `mtp-manifest.json` lists each tensor's shard, byte range and
-  sha256 as the fetch tool verified them.
-- Build: UnbiasedStrata main `11b04edbb62dbe15acbd287da0fd5c6077be64f1` (2026-10-20):
-  `tools/mtp_fetch.py fetch` → `tools/mtp_pack.py --experts q2_0` → `tools/mtp_rt.py`;
-  `draft_vocab.bin` is the fork's `data/draft_vocab.bin`.
-- Reproducibility: the same four files, byte for byte, came out of three independent builds
-  (homelab 4090 VM 2026-10-19, RunPod RTX 3090 container 2026-10-20, homelab 4090 VM 2026-10-20).
-
-Use: `--mtp /path/to/qwen3.8-flash-next-q2_0 --spec 4 --spec-min-p 0.5` (the engine checks the
-directory against the model at load). Works with every GSQ-RCO quant of Flash-Next, including the
-REAP'd Coder shard.
-
-## profiles/expert-profile-coder-shard-bba88856.bin — Coder expert profile (98 KB)
-
-The expert ranking (`--expert-profile`) for `ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF`
-at revision `bba88856f8c3c0c9e40e45f82d05e228b814ac43`, whose shard 1 (sha256 `bfa3644d…4016c211`,
-replaced upstream 2026-10-17) keeps a different set of 256-of-512 experts from the earlier shard.
-ISTA ships no profile for it. This one maps the fork's full-model ranking (`data/expert-profile.bin`)
-onto the shard's kept experts, read from its `rco-allocation.txt`, and renumbers them in ascending
-order. The same mapping reproduces ISTA's shipped profile for the previous shard bit for bit
-(`scripts/pod16/map_coder_profile.py --check` in the regina repo). sha256
-`8c1d130d3f3aa64547307679c14c3783668f9f35a5344004d8354e1c3f6c7ac9`.
-
-## Licence
-
-The MTP head is derived from Qwen3.8 Flash-Next and is distributed under that model's licence. The
-profile is a ranking table derived from the same model.
+Pre-built artifacts for [UnbiasedStrata](https://github.com/robertkarl/UnbiasedStrata). Built once, copied to each host.
+
+## mtp/qwen3.8-flash-next-q2_0 (787 MB) — `--mtp DIR`
+
+Built with UnbiasedStrata `11b04edbb62dbe15acbd287da0fd5c6077be64f1`:
+
+```
+STRATA_SOURCE=huggingface python3 tools/mtp_fetch.py fetch --out mtp-src      # Qwen/Qwen3.8-Flash-Next @ de4b8e4d43b917e7706784d8bb445c9af86a3540
+python3 tools/mtp_pack.py --src mtp-src --experts q2_0 --out mtp-src/mtp-q2_0.gguf
+python3 tools/mtp_rt.py --gguf mtp-src/mtp-q2_0.gguf --out mtp
+cp data/draft_vocab.bin mtp/draft_vocab.bin
+```
+
+```
+c724dc0b0822ada5d2977bf5bde821605feabaa64ea2e0045b67ca656329070a  dense.bin
+991456cf4eb266eb8a888c349e548d11bc640f5dbd9149439aa9786af0f732fd  dense.txt
+b1e1d3a7a9e4bf862dcd5923ce661fb59bbd07907e594df5cf86a62ac235cb91  draft_vocab.bin
+09398406be61f1f54c93861f449e48b8df0bfccbc9ec9b2b7636775a6ea9244f  experts.bin
+1ad281ed87ad9cd42f9c048e9c49172165ea714a600d15f1d80cd0a02db12328  mtp-manifest.json   (the fetch tool's per-tensor record)
+```
+
+## profiles/expert-profile-coder-shard-bba88856.bin (98 KB) — `--expert-profile FILE`
+
+For `ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF` @ `bba88856f8c3c0c9e40e45f82d05e228b814ac43` (shard 1 sha256 `bfa3644d…4016c211`).
+Built with regina `scripts/pod16/map_coder_profile.py` against UnbiasedStrata `11b04edbb62dbe15acbd287da0fd5c6077be64f1`:
+
+```
+python3 scripts/pod16/map_coder_profile.py --full data/expert-profile.bin --alloc rco-allocation.txt --out expert-profile-coder-shard-bba88856.bin
+```
+
+```
+8c1d130d3f3aa64547307679c14c3783668f9f35a5344004d8354e1c3f6c7ac9  expert-profile-coder-shard-bba88856.bin
+```