diff --git a/config/template-pins.json b/config/template-pins.json index 26273a27b..62f5ad5b6 100644 --- a/config/template-pins.json +++ b/config/template-pins.json @@ -1,5 +1,25 @@ { "pins": { + "qwen3-0.6b-npu-1bit": { + "sha256": "a136047ba140347b359d17b6ff8fa1bf8cb29d23f0c07bef7e0bf479b8680793", + "source": "FastFlowLM/Qwen3-0.6B-NPU2" + }, + "qwen3-0.6b-npu2": { + "sha256": "a136047ba140347b359d17b6ff8fa1bf8cb29d23f0c07bef7e0bf479b8680793", + "source": "FastFlowLM/Qwen3-0.6B-NPU2" + }, + "qwen3-0.6b-q8_0.gguf": { + "sha256": "57f1fd00f0013a2be96aa79b857391f27e23df5b5f847072b524c897e24d0361", + "source": "1bit-MONSTER/Qwen3-0.6B-GGUF" + }, + "qwen3-4b-instruct-2507-1bit": { + "sha256": "c979e0e71a3e21b8f208e6ab120d5cb29327885f29d2a8b18fda67a723798e18", + "source": "unsloth/Qwen3-4B-Instruct-2507-GGUF" + }, + "qwen3-4b-instruct-2507-q4_k_m.gguf": { + "sha256": "c979e0e71a3e21b8f208e6ab120d5cb29327885f29d2a8b18fda67a723798e18", + "source": "unsloth/Qwen3-4B-Instruct-2507-GGUF" + }, "qwen3-coder-30b-a3b-instruct-q4_k_m.gguf": { "sha256": "87710339d25b4e789c1d723f93c91ee861a86d305bb3d20a845536f251d6ea8a", "source": "Qwen/Qwen3-Coder-30B-A3B-Instruct" diff --git a/docs/serve.md b/docs/serve.md index ff49c003a..9ec19fab9 100644 --- a/docs/serve.md +++ b/docs/serve.md @@ -796,6 +796,12 @@ says nothing, because nothing runs (this keeps the header-only route fixtures on {"pins": {"zaya1-8b-q4_k_m.gguf": {"sha256": "b55ecfbbbcbb238a…", "source": "Zyphra/ZAYA1-8B"}}} ``` +The table carries the rows the appliance itself ships — the stick's `Qwen3-4B-Instruct-2507-1bit` GGUF and +its `Qwen3-0.6B-NPU-1bit` container, whose digest is the `chat_template` string from +`tokenizer_config.json` ([npu-chat-template.md](npu-chat-template.md)) — plus the demo and scoreboard rows and all four `laya-first` routing rows. +A model with no row is logged, not checked, which is why a row is added with its measured digest rather +than guessed (`models/pins.json` records that digest for every file we serve and publish). + A pin is matched against the model id serve serves (`--alias`, i.e. Lemonade's catalog entry) first, then the file's own name, case-insensitively; `"none"` pins a model that must carry no template. A model whose pin does not match prints a second line and **still serves**: diff --git a/tests/template_pin.sh b/tests/template_pin.sh index 3a4563862..77982e46d 100755 --- a/tests/template_pin.sh +++ b/tests/template_pin.sh @@ -53,7 +53,9 @@ base = [st("general.architecture", "qwen3"), u32("qwen3.block_count", 28), u32(" u32("qwen3.embedding_length", 1024), u32("qwen3.attention.head_count", 16), u32("qwen3.attention.head_count_kv", 8)] for name, kv in (("tpl", [st("tokenizer.chat_template", sys.argv[2])]), ("bare", []), - ("zaya1-8b-q4_k_m", [st("tokenizer.chat_template", "{% for m in messages %}{{ m }}{% endfor %}")])): + ("zaya1-8b-q4_k_m", [st("tokenizer.chat_template", "{% for m in messages %}{{ m }}{% endfor %}")]), + ("qwen3-4b-instruct-2507-q4_k_m", + [st("tokenizer.chat_template", "{% for m in messages %}{{ m }}{% endfor %}")])): all_kv = base + kv with open(f"{sys.argv[1]}/{name}.gguf", "wb") as f: f.write(b"GGUF" + struct.pack("