Upload folder using huggingface_hub

Dockerfile

@@ -0,0 +1,17 @@
+# tinkersnot bake-off inference Space — official llama.cpp server, no compile.
+# Serves an OpenAI-compatible /v1/chat/completions with native tool-calling (--jinja).
+FROM ghcr.io/ggml-org/llama.cpp:server
+
+# llama-server reads LLAMA_ARG_* env vars for its flags — no CMD/entrypoint wrestling.
+# Swap LLAMA_ARG_HF_REPO to change the candidate model (repo:quant tag).
+#   Real candidate  : unsloth/Qwen3.5-0.8B-GGUF:Q8_0   (hybrid VLM, tool-calling)
+#   Smoke-test/fallb: ggml-org/Qwen3-0.6B-GGUF:Q8_0    (std arch, guaranteed to load)
+# If the 0.8B errors on the hybrid arch or demands a vision projector, flip to the 0.6B line.
+ENV LLAMA_ARG_HOST=0.0.0.0 \
+    LLAMA_ARG_PORT=7860 \
+    LLAMA_ARG_HF_REPO=unsloth/Qwen3.5-0.8B-GGUF:Q8_0 \
+    LLAMA_ARG_CTX_SIZE=8192 \
+    LLAMA_ARG_JINJA=1 \
+    HF_HOME=/tmp/hf
+
+EXPOSE 7860

README.md

@@ -1,10 +1,12 @@
 ---
-title: Bench Server
-emoji: 🦀
-colorFrom: gray
-colorTo: gray
+title: Tinkersnot Bench Server
+emoji: 🦙
+colorFrom: blue
+colorTo: green
 sdk: docker
+app_port: 7860
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+OpenAI-compatible llama.cpp server for base-model bake-off eval.
+Endpoint: `/v1/chat/completions`. Model set via `LLAMA_ARG_HF_REPO` in the Dockerfile.