Views
No views yet
drbaph/s2-pro-fp8 (FP8 weight-only quant of fishaudio/s2-pro, ~12 GB VRAM vs ~24 GB bf16).VOCENCE_S2_MODEL_PATH=fishaudio/s2-pro for the original (gated; needs HF token).sgl-omni fails at runtime.chute_config.yml, tts_engine.py, and s2pro_tts.yaml for deployment.sgl-venv site-packages with a stub numpy and minimal_engine.py (no S2, no Parler, no real model load). That only proves import sandboxing — runtime behavior must be tested on Chutes or a full local GPU stack. Run from vocence_s2pro_miner/:python3 scripts/test_chutes_sandbox_local.pydrbaph/s2-pro-fp8. If warmup fails, the engine falls back to Parler-TTS so /speak still works. GET /health includes s2_warmup_error, s2_log_excerpt (tail of /app/sgl-omni.stderr.log), and s2_model_path when the engine recorded an S2 failure—no SSH required. The Chute keeps allow_external_egress=False to match subnet validator rules (no third-party outbound HTTP from inference). That means weights must be available without runtime egress (e.g. baked into the image build or read from a local path you ship). Do not turn on egress “for testing” to hit Hugging Face from production. Other common causes: readiness timeout (raise S2PRO_SERVER_READY_TIMEOUT_SEC), OOM / CUDA (see log excerpt). Set VOCENCE_REQUIRE_S2=1 to disable fallback and surface the real error at startup (Chute will not become healthy until S2 works).s2_warmup_error shows code=-11: that is SIGSEGV (a native crash inside sgl-omni/CUDA, not a missing HF token or a problem specific to the drbaph/s2-pro-fp8 weight files—the HF model card documents FP8 storage with bf16 compute via dequant on all CUDA GPUs). Upstream sglang-omni hardcodes disable_cuda_graph=False for the Fish S2 SGLang stage; CUDA graph capture is a common source of this class of crash on some hosts. The Chutes image build applies a small patch (see chute_config.yml after extracting the pinned sglang-omni tarball) so disable_cuda_graph=True is installed into /app/sgl-venv. Rebuild/redeploy after changing that patch. If it still faults: use s2_log_excerpt from /health, note GPU model (nvidia-smi), and try S2PRO_SERVE_EXTRA_ARGS per sglang-omni’s merge syntax or open a Chutes ticket with instance id + excerpt.y prompts (build + deploy)chutes Python package uses input() in two places you care about:chutes build (chutes/entrypoint/build.py): after listing build-context files, it may ask “show the rest? (y/n)” and always asks “Confirm submitting build context? (y/n)”.printf 'n\ny\n'.printf 'y\n'. Sending n first when only one prompt exists aborts the build.chutes deploy (chutes/entrypoint/deploy.py): asks “upload … and deploy …, confirm? (y/n)” before the API call. --accept-fee only sets the accept_fee query param for the deployment fee; it does not skip that prompt.1cd vocence_s2pro_miner
2# Build (many files in repo → use two lines). Use --include-cwd so remote build has context.
3printf 'n\ny\n' | chutes build vocence_chute:chute --wait --include-cwd
4# Deploy
5printf 'y\n' | chutes deploy vocence_chute:chute --accept-feescripts/chutes_release.sh (edit/delete steps as needed).vocence_chute.py)VocenceSandboxImporter only allows imports whose code lives under stdlib, normal site-packages, or your tts_engine.py. The S2 Pro Docker image also uses /app/sgl-venv/.../site-packages and /app/pyextra; those paths are whitelisted. For local simulation, set VOCENCE_EXTRA_IMPORT_ROOTS to colon-separated dirs (see scripts/test_chutes_sandbox_local.py).