Upload releases/llamafile-v0.10.3+opencoti.c3/RELEASE_NOTES.md with huggingface_hub

releases/llamafile-v0.10.3+opencoti.c3/RELEASE_NOTES.md

@@ -0,0 +1,87 @@
+# opencoti llamafile 0.10.3-c3
+
+Third packaged cut (`c` series) of the opencoti single-file inference
+engine (#613 lane). Built from
+[Mozilla-Ocho/llamafile v0.10.3](https://github.com/Mozilla-Ocho/llamafile/releases/tag/0.10.3)
++ the opencoti patch series (79 patches, `0001`–`0133`) at
+`vendors/patches/llamafile/`.
+
+Version string: `opencoti-0.10.3-c3` — printed by `--version` and exposed
+at runtime.
+
+## Artifact
+
+| File | Size | sha256 |
+|---|---:|---|
+| `opencoti-llamafile-0.10.3-c3-x86_64.llamafile` | 4,812,296,839 B | `8a35b1f87b7bb7f2915d8756450f09008f4cd8331cf4a763e40edbcb8c4cf7e9` |
+
+Embedded `ggml-cuda.so`: 4,756,838,104 B, sha256
+`5e0342ba9b2ded4847280067dfd6ebc4257e1b83a984881988c528c163b23e4a`
+(glibc floor GLIBC_2.29 — runs on Debian 11 / glibc 2.31 and newer).
+
+Single-file Cosmopolitan APE. **Fat binary: runs natively on x86_64 and
+aarch64** (Linux/macOS/Windows/BSD) — on ARM64 hosts (DGX Spark, GH200,
+Apple Silicon) inference is CPU-only in this cut; the embedded CUDA DSO
+is x86_64-linux. GPU-enabled aarch64 (sbsa) is a planned separate
+artifact.
+
+Run as a server:
+
+```bash
+sh ./opencoti-llamafile-0.10.3-c3-x86_64.llamafile --server --port 8080 \
+    -m <model>.gguf -ngl 99 --flash-attn on
+```
+
+On first GPU run the embedded `ggml-cuda.so` (CUDA 13.2, arches
+sm_75/80/86/89/90/120) self-extracts to `~/.llamafile/v/0.10.3/`.
+Without `--server` the binary starts the interactive chat CLI.
+
+## What's new vs c2 (2026-07-05)
+
+**Host binary (patches 0122–0133):**
+- RYS `--repeat-layers` weight-shared layer duplication + `--rys-probe`
+  v2 task battery (#656–#673)
+- bug-858 CLOSED: MTP speculative decode at/above upstream-b9859 parity
+  on both native pairs (spec-clone-cheap + verify GEMM lift, 0128)
+- P2 mixed-KV auto-tail + parallel≥2 fit/checkpoint fixes (0121, 0130)
+- Runtime introspection (#676/#677, 0132): `/props` gains an `"opencoti"`
+  section (kv, residency, dca, speculative, sparse_attn, kv_reuse,
+  repeat_layers, rest_kv); `/slots` gains per-slot lifetime draft
+  counters + session/pool binding
+- Full `--version` string `opencoti-<ver>-<tag>` (#613, 0133)
+
+**CUDA DSO (rebuilt, 4.4 GB):**
+- #674: DCA-off scalar-MMA tensor-core decode generalized to D256/D512
+  (Gemma-4 head dims) — q8q4 KV decode ≈1.15× f16, up to 2.3× native VEC
+- 0128 mmvf f16/bf16 TinyBLAS verify lift (Qwen NextN n3 parity)
+- 0126/0127 P0 DCA-off scalar-MMA D128 foundation
+
+## Verification
+
+`MANIFEST.json` ships alongside the artifact: versionString, gitCommit,
+full patch list, and per-DSO sha256 of the embedded backend. Verify the
+embedded DSO without running:
+
+```bash
+unzip -p opencoti-llamafile-0.10.3-c3-x86_64.llamafile ggml-cuda.so | sha256sum
+```
+
+Zero-dependency smoke performed on this cut: isolated-`HOME` `--version`,
+DSO self-extraction sha-match, `--server` GPU boot on RTX 3090 with
+`/props` opencoti introspection + greedy completion.
+
+## Build provenance
+
+- Repo: `mann1x/opencoti` @ commit `4c3f2ec2e` (dev branch)
+- Base: llamafile 0.10.3 (outer 7fca8b2, nested llama.cpp dbe9c0c) +
+  Mozilla patch overlay + opencoti chain 0001–0133 (replay-proven
+  byte-identical to the shipped tree)
+- Host binary built on solidPC (cosmocc); CUDA DSO built on solidPC
+  (Debian 11, CUDA 13.2) to pin the release glibc floor at 2.29
+  (bug-2190: a bs2-built DSO required glibc ≥2.32 and failed on
+  older hosts)
+
+## License
+
+Apache-2.0 (Mozilla-Ocho/llamafile) + MIT (llama.cpp and bundled
+projects). opencoti patches and packaging are MIT.