Fetching metadata from the HF Docker repository...
Jun 12 23:52 GPU final: model resident on cuda (no per-call transfer) + 240s budget + one-shot — fixes empty-output timeout 23:42 REAL MODEL on reliable CPU: llama.cpp 4B Nemotron, one-shot, compact prompt; cpu-basic hardware 23:05 match the working bench EXACTLY: .to(cuda) inside the GPU call every time (no caching) — fixes the persistent NVML on the live single call 22:58 ONE-SHOT plan generation for ZeroGPU: whole wish in a single @spaces.GPU call (ends the multi-call NVML crash, far faster) + deterministic town fallback so a wish never builds nothing 22:50 FIX multi-call NVML crash: eager .to(cuda) under spaces at load (manages GPU residency across a wish's many calls) — not lazy-per-call 22:44 GPU recovered — back to ZeroGPU (fast+free) with retry resilience; restore transformers stack 22:27 disable flaky LLM judge layer (default-denied every clean wish on llama.cpp); wordlist layers 1-2 remain the gate 22:23 CPU deployment: trim torch/transformers (llama.cpp only) for fast light cpu-basic builds 22:18 back to zerogpu (CPU-only app is killed on ZeroGPU hardware) + retry transient NVML/CUDA failures on a fresh GPU schedule 21:57 switch live backend to llama.cpp (CPU, GBNF-constrained) — reliable, ends the ZeroGPU NVML flakiness; max threads 21:51 FORGIVING TOWNS: map any building kind (never reject), district reroute, restore god's voice (content-only moderation), town few-shot pushes build_district+roads, honest fail on empty 21:41 FIX zerogpu NVML crash: defer cuda placement into the GPU context (no startup CUDA init); reject failed wishes honestly instead of empty-granting 20:54 Town Mode: close town view by default, build_district + place_road, bank/market/house, grow-one-town steering, behold-the-world reveal, tamed needle 19:48 City Update: towers/warehouses/cafes, spawn_life (carts/birds/fireflies), cinematic descent, visible deltas, camera gaze, bigger ticker, counter fix 18:57 security: re-moderate model text, identity-bound rite token, XFF/queue-freeze hardening, OOM guard; README truth fixes 18:22 inline ticker flow, forgiving water args, wish-counter from traces, canonical namespaced tags 18:10 readable tool docs (engine as_dict), water syntax note, translate-not-refuse guidance 18:01 strip echoed READING label; fix status sky suffix accretion 17:58 fix your_turn race: buffer early turns + re-announce in thirds 17:41 browser-handshake unit tests 17:34 planet owns the root (evict gradio SPA index) 17:30 example coordinates stay in the example 17:25 sharpen example/live wish boundary in prompts 17:21 gradio Server mode + browser-invoked granting (the her idiom) 17:00 transformers v5 in-tree stack 16:55 switch to 9B-v2 in-tree, kernel-free 16:49 ship libcudart via pip + kernels 16:45 drop mamba kernels (no libcudart at startup) 16:43 preload libcudart for mamba kernel import 16:40 trust_remote_code for nvidia repos 16:39 prebuilt llama-cpp wheel index + field notes 16:34 pin sdk 6.16 stack (mirrors her) 16:31 GODSEED v0 16:30 initial commit 
