Expose live audio generation API route

Pass Qwen prompt through async job

Fix Qwen multimodal prompt content

Diagnose Qwen generation termination

Enable Qwen reasoning by default

Add torchvision for Qwen processor

Add CPU Qwen3.8 inference job

Normalize TinyLlama chat template inputs

Fix CPU TinyLlama PyTorch initialization

Accept JSON body for TinyLlama inference

Restore Gradio ZeroGPU Space metadata

Add Transformers runtime for TinyLlama

Add CPU TinyLlama inference to Gradio Space

Remove incompatible Docker entrypoint

Switch Space to Docker experiment

Add standalone CPU server entrypoint

Test standalone llama server container

Test RAM loaded GGUF without pipe deadlock

Rebuild llama.cpp with stable CPU backend

Limit CPU thread pools to cgroup CPUs

Prevent llama output pipe deadlock

Probe llama CPU process state

Run loader diagnosis as background job

Add CPU model loader diagnostics

Run CPU benchmarks asynchronously

Diagnose single-thread CPU model loading

Add tiny GGUF CPU control test

Use RAM loaded CPU model benchmark

Separate model loading from CPU warmup

Add CPU model smoke diagnostics

Serialize CPU benchmark requests

Keep local model benchmarks CPU only

Select llama benchmark executable

Fix llama benchmark binary cache

Use llama bench for bounded CPU measurements

Fix CPU benchmark diagnostics

Restore stable Space server launch

Add share=True to bypass broken proxy

Restore working create_app monkey-patch with staticmethod

Background thread inject routes into running Gradio server

Fix: inject routes via background thread after launch

Fix: patch App.__init__ to inject routes into the actual served app

Fix: monkey-patch without staticmethod wrapper, spaces before torch, ssr_mode=False

Fix route injection: try demo.app directly, fallback to monkey-patch

Add model filter param to benchmark endpoint

Fix: build llama.cpp with static linking, find binary after build

Build llama.cpp from source instead of pre-built binaries

Add explicit libllama.so verification logging

Fix: detect stale cache, re-extract if libs missing

Fix: extract all llama.cpp libs, set LD_LIBRARY_PATH