chipwright-llm: Qwen2.5-1.5B Q8_0 GGUF (llama.cpp-HTP, v73 silicon-verified)

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+qwen2.5-1.5b-Q8_0.gguf filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,56 @@
+---
+license: apache-2.0
+base_model: Qwen/Qwen2.5-1.5B-Instruct
+tags:
+  - chipwright
+  - qualcomm
+  - hexagon
+  - npu
+  - llama.cpp
+  - gguf
+  - q8_0
+library_name: gguf
+---
+
+# chipwright-llm — Qwen2.5-1.5B-Instruct Q8_0 (llama.cpp-HTP, Hexagon v73)
+
+Artifacts hosted for [**chipwright**](https://github.com/nameissakthi25/chipwright) — a verified package
+manager for Qualcomm NPUs (`resolve · run · build`). This repo holds the **llama.cpp-HTP runtime** variant:
+a `Q8_0` GGUF that dispatches on the Hexagon NPU through the llama.cpp Hexagon backend.
+
+**Why this exists:** classic Genie is w4a16-only. This is the **>4-bit integer route Genie structurally
+lacks** — `Q8_0` MUL_MAT running on the HTP — and chipwright treats *runtime* as a first-class resolver
+axis (spec 004).
+
+## What's here
+| file | sha256 | notes |
+|---|---|---|
+| `qwen2.5-1.5b-Q8_0.gguf` | `cedde4175edae2e9467d8b98d70843e57c8fb101afd54ca773bce022e80f7e4e` | 1.6 GB, llama.cpp build `b3278e921` |
+
+## Silicon verification (chipwright Constitution Principle I — verify on the device)
+Re-verified **2026-10-24** on **QCS9075 / Dragonwing IQ-9075** (soc_id 676, **Hexagon v73**) through the
+chipwright llama.cpp-HTP recipe:
+
+- **Dispatch (SC-RT03):** `Q8_0` MUL_MAT dispatched to **HTP0** on the compute-DSP (`_dom=cdsp`), **not** a
+  CPU fallback — prefill **592.82 t/s** on HTP0 vs **217.55 t/s** CPU-only (`GGML_HEXAGON_NDEV=0`) =
+  **2.72× prefill**. (The earlier QCS8550/v73 sweep showed 266.74 vs 152.93 = 1.74×.)
+- **Quality:** WikiText perplexity (16 chunks) = **11.6631 ± 0.494**, near-lossless vs the model's own f16
+  baseline (11.6388).
+
+## How to run it (the recipe)
+On a **v73+** Hexagon board with fastrpc + `libcdsprpc.so`, using the llama.cpp Hexagon backend
+(`libggml-htp-v73.so`) under an injected glibc loader:
+
+```sh
+ADSP_LIBRARY_PATH=<pkg>/lib \
+  ./glibc238/ld-linux-aarch64.so.1 --library-path ./glibc238:<pkg>/lib:/usr/lib:/lib \
+  <pkg>/bin/llama-bench -m qwen2.5-1.5b-Q8_0.gguf --device HTP0 -p 64 -n 32 -r 2
+# CPU reference: prefix GGML_HEXAGON_NDEV=0 and drop --device
+```
+
+`chipwright resolve qwen2.5-1.5b-instruct --target htpv73 --runtime llamacpp-htp` returns this variant
+(`USE`), and refuses it **by name** on v68 (`FAIL axis=runtime` — needs v73-or-newer Hexagon). CLI
+on-device execution of this runtime is tracked as spec 004 T-RT11.
+
+## License
+Weights: Apache-2.0 (inherited from Qwen/Qwen2.5-1.5B-Instruct). GGUF quantized with llama.cpp.

qwen2.5-1.5b-Q8_0.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cedde4175edae2e9467d8b98d70843e57c8fb101afd54ca773bce022e80f7e4e
+size 1646572640