Spaces:
Paused
Paused
Commit History
-
sync: bring the Space up to the current GitHub tree
The Space had drifted well behind main — missing whole feature directories (beliefs, connectors, agentic loop, self-consistency, eval calibration / citation-eval / judge / gate, contextual retrieval, web-URL ingest, VLM) as well as this week's work: the Ollama/vLLM/AirLLM serving-backend selector, HTTP-based model pull with streamed progress, and the composer popover fix. Preserved as Space-specific and deliberately NOT overwritten: - root Dockerfile (ollama binary, prebuilt llama-cpp wheel, HF cache bake) - README.md (the HF Space frontmatter lives there) - wheels/, .gitattributes - deploy/huggingface/entrypoint.sh — keeps `ollama serve` for the local fallback link and `--workers 2`. That worker count is load-bearing: the /query endpoints are async but the provider call is blocking, so a single event loop freezes under concurrency (the old HTTP 000 outages). The hosted HF router call has exactly the same blocking shape. Pre-flight before pushing: `npm ci && npm run build` pa…656439d
-
fix(api): run 2 uvicorn workers — a blocking LLM call in an async endpoint was freezing the single event loop (HTTP 000 outages)
Verified signature9318680
-
fix(gpu): warm LLM into VRAM at startup + health-check grace period — stops the cold-load restart loop
Verified signaturecc2baaf -
fix(gpu): ollama v0.32.1 .tar.zst (correct URL) + zstd; extract to /usr; GPU CUDA libs baked
Verified signaturede23710 -
feat(gpu): Ollama LLM server (qwen2.5:3b) for GPU inference + on-demand T4 with 15min auto-sleep
Verified signature39e5349
-
fix(docker): keep wheel filename so pip accepts it
Verified signature95a521a -
perf(llm): install prebuilt AVX2 llama-cpp wheel (no HF-build compile) + Qwen2.5-3B for citations
Verified signaturea406137 -
perf(llm): AVX2/FMA source build of llama-cpp + Qwen2.5-3B (reliable [n] citations) — utilizes the paid CPU's vector units, 3B follows citation instructions
MHamdan3d0f1e8 -
fix(slm): thread count from cgroup CPU quota, not host os.cpu_count() — 64 threads on an 8-vCPU container thrashed (slower); now uses the real limit
MHamdane8c6ada -
perf(slm): use all CPU cores (n_threads=cpu_count, n_batch=512) — cuts prefill/TTFT on the paid multi-core tier
MHamdanf20cfdd -
fix(web): raise /api proxy timeout to 300s — CPU LLM generation exceeds Next's 30s default on non-streaming endpoints
MHamdana88ea41 -
feat(llm): local Qwen2.5-1.5B GGUF via llama-cpp (prebuilt CPU wheel, model baked) — real generative answers on the paid cpu-upgrade tier
MHamdanb4c429e -
fix(space): HF_HUB_OFFLINE=1 — baked models must load from cache, never network (prevents relaunch health timeouts)
MHamdan555a43b
-
perf(voice): shorter spoken summary (snappier turns) + startup model warmup (no cold-start 5xx)
MHamdan33e1d6c -
feat(voice): NotebookLM-style voice UX
MHamdan- ASR: base.en + greedy/no-word-timestamps/no-conditioning → ~3-5x faster - Voice-out autoplay fixed: shared <audio> unlocked on the mic gesture so the spoken answer actually plays after transcription - "Hear the answer" button on every assistant message (POST /speak → kokoro) - Male/female voice selector (af_heart / am_michael), threaded to /voice + /speak - Spoken summary trimmed + markdown/citation-cleaned; full answer stays on screen5a0feef -
feat(voice): speak a short, clean summary (fast TTS); full answer stays on screen
Verified signatureacacd40 -
feat(voice): ASR small + Kokoro TTS (spoken replies); prebake models
Verified signature32f283e -
fix(voice): install faster-whisper (free, torch-free) + prebake Whisper base; enable ASR
Verified signatureb67cacd
-
HardwareGate: detect the visitor's real device client-side
Verified signature4de5f82 -
ModelFit: clarify server-vs-device on hosted Space + enable wiki
Verified signature9f1969b
-
Install .[ingest,llm] so the huggingface LLM provider (openai SDK) loads
Verified signaturef78dd4f -
Deploy Auralynq RAG (Llama-3.3-70B via HF Inference Providers)
Verified signature8c1b9fe -
initial commit
Verified signature4fd8f25