FJK

Optimize: n_batch=2048, n_ubatch=2048, max_inputs=11, scaledown=28s

Add Modal GPU T4 deployment files for faster inference

Revert n_threads to 4 - 8 threads causes hangs

Increase n_threads from 4 to 8 for better CPU performance

Fix content-type header logic for all body methods

Add debug logging for backend response headers and body

fix: Remove else block causing indentation error

debug: Temporarily disable auth to diagnose 400 error

fix: Serve HTML directly for public routes instead of proxying to backend

fix: Allow public access to non-API routes (HF Spaces Log Viewer)

fix: Add missing valid_key variable for auth check

fix: Read body before auth check to avoid UnboundLocalError

trigger rebuild: force deploy to fix UnboundLocalError

fix: Only read body for POST/PUT/PATCH methods to avoid UnboundLocalError

fix: Accept multiple auth formats for HF Spaces compatibility

fix: Add health endpoint without auth, add debug logging

debug: Add logging to trace 400 error

fix: Strip authorization header before forwarding to backend

feat: Add API key authentication for private Space access

fix: Simplified proxy with better path handling and error handling

fix: Use lifespan instead of deprecated on_event

fix: Add FastAPI proxy to strip HF Spaces __sign parameter

🔧 Add libgomp (OpenMP) for CPU multi-threading support

🔧 Add libstdc++ to Alpine (required by pre-compiled C++ wheel)

🔧 Switch to Alpine (python:3.11-alpine) for musl-compatible wheel

🔧 Try pre-compiled wheel v0.3.19 (avoids timeout + OOM, likely has LFM2)

🔧 Compile with MinSizeRel (reduces linker RAM ~2.5GB→~1.2GB, avoids OOM)

🔧 Force pre-compiled wheel v0.3.14, --only-binary :all: (no compilation)

🔧 Use pre-compiled CPU wheel (no compilation, includes CPU backend)

🔧 Fix --flash-attn flag syntax: use 'auto' instead of bare flag

🔧 Bulletproof tar extraction: flatten llama-server + .so libs into llama-cpp/

🔧 Keep dynamic libs (libmtmd.so.0), set LD_LIBRARY_PATH

🔧 Add libcurl4 (fixes exit code 127: libcurl.so.4 not found)

🔧 Use precompiled llama.cpp b7633 binary (no compilation, no OOM)

🔧 Fix lfm2 support: use latest llama-cpp-python (0.3.7 was too old)

🔧 Add huggingface_hub to Dockerfile (needed for model download)

🔧 Remove proxy: use llama-server directly (simpler, same as Colab)

🔄 Revert to llama-cpp-python server (same approach as Colab, CPU mode)

🔧 Use latest llama-cpp-python[server] for best architecture support

🔧 Install llama-cpp-python[server] with all server dependencies

🔧 Add sse-starlette (required by llama-cpp-python server)

🔧 Remove rm -rf /root/.cache (permission denied as user)

🔧 Add git to Dockerfile (required by llama-cpp-python build)

🔄 Switch to llama-cpp-python server (pre-built binary had missing CPU backend)

🔧 Fix llama-server path: strip-components=1 puts binary in root

🔄 Rebuild with HF_TOKEN secret configured

📥 Download model at runtime from private HF repo

🔧 Fix llama-server binary URL: use b8783 ubuntu-x64 tar.gz

🔧 Remove unnecessary llama-server auth (only accessed locally by proxy)