vocence-mark-v1 — Vocence PromptTTS base bundle

HF snapshot layout for subnet netuid 78: checkpoint shards, config.json, Vocence miner.py entrypoint.

Layout (snapshot_download root)

Artifact	Purpose
miner.py	Vocence Miner engine (loads weights from this revision).
chute_config.yml	Builds image + clones your chosen Prompt-TTS warm runtime (VOCENCE_PROMPT_RUNTIME).
vocence_config.yaml	Device / limits / importer strip / backend selection.
config.json	Weight filenames + inference defaults (keep in sync if you rename tensors).
vocence-dit-v1.safetensors	DiT velocity shard (files.transformer).
vocence-audio-components.safetensors	Audio bundle shard (files.audio_components).
assets/silence_latent_frame.pt	Silence latent (files.silence_latent).
WEIGHTS_MANIFEST.json	Optional checksum manifest.

Chutes prerequisite

Set build secret VOCENCE_PROMPT_RUNTIME (HTTPS git clone URL) pointing to whichever LICENSE-compliant codebase your team vets that provides:

• src/inference_server.py (TTSServer),
• sibling directories src/ + ltx2/ + project requirements.txt.

vocence-mark-v1 weights follow Lightricks LTX-2 lineage (LICENSE). Your runtime checkout must legally match those terms.

Never commit personal tokens inside YAML.

Hugging Face secrets

Writable HF_HOME, HF_AUTH_TOKEN typically required for Gemma / text encoder fetches resolved by your runtime stack.

Extra env knobs

Variable	When
VOCENCE_BACKEND	mock forces silence (overrides vocence_config.yaml engine.backend) — smoke only. Remove in production miners.
GEMMA_DIR	Text encoder offline cache anchor if downloader supports it.
LTX_FULL_CHECKPOINT	If conditioners need a multimodal consolidated checkpoint besides **vocence-audio-components**—set explicitly.
VOCENCE_VOICE_REF	Optional WAV path for vocal reference conditioning (runtime-specific feature).

Publish (HF → Chutes → chain)

Orchestration lives one level up under vocence/tools/ — run verify_bundle → upload_hf_model → render_canonical_wrapper, then Chutes CLI and vocence miner commit. See DEPLOY.md.

Vocence importer + runtime path injection

Warm runtime code must be importable. vocence_config.yaml → inference_runtime.strip_vocence_sandbox_importer: true makes miner.py drop VocenceSandboxImporter from sys.meta_path before import inference_server (confirm subnet posture before shipping).

Install root default /opt/vocence-mark-v1-inference (wired in chute_config.yml + miner.py).

Miner contract

Matches MINER_GUIDE:

• Miner.__init__(path_hf_repo)
• warmup()
• generate_wav(instruction, text) → (mono float32, sr) with 16000 ≤ sr ≤ 48000 and ≤ 30 s.

Integrity

From vocence/ checkout root:

python3 tools/verify_bundle.py
python3 tools/verify_bundle.py --model-root ./models/vocence-mark-v1

Licenses

Checkpoint licensing: Lightricks LTX-2 Community License (LICENSE in this folder). Third-party watermark deps (e.g. Perth) observe their respective licenses inside your image build.