| |
| """ |
| voice — Lift a model's "voice" (its output-projection tensor) and cast it onto |
| another model, in either direction between safetensors and GGUF. |
| |
| voice get <model> [tensors...] Get a voice: one tensor (or head) to the registry, |
| several/--multi to voicepack files |
| voice graft <pulse> [donor] Graft one voice onto another (DELLA single graft) |
| voice chorus [pulse] [donors...] Blend voices: pulse + support + lead (DELLA chorus) |
| voice cast <target> <voice> [tensors...] Cast a voice (or chosen tensors) onto a file |
| voice delta [voice] Make a reusable style (voice − base) |
| voice add <safetensors> <json> Add a voice from local voice.safetensors + voice.json |
| voice list [base] List stored voices (or cached bases) |
| voice info <voice> Show voice details |
| voice remove <voice> Remove a stored voice |
| voice cleanup Remove incomplete downloads |
| voice report Save a report locally and optionally share |
| voice auth Save your Hugging Face token (optional, faster) |
| voice path Put `voice` on your PATH |
| voice doctor Check your setup |
| |
| Every extract/cast action is appended to ~/.voices/operations.log. |
| """ |
|
|
| import argparse |
| import getpass |
| import json |
| import os |
| import re |
| import shutil |
| import struct |
| import sys |
| import textwrap |
| import time |
| import urllib.error |
| import urllib.request |
| from pathlib import Path |
| from urllib.parse import urlparse |
|
|
| |
| |
| _HERE = os.path.dirname(os.path.realpath(__file__)) |
| if _HERE not in sys.path: |
| sys.path.insert(0, _HERE) |
| from vlib.ctx import VOICES_DIR, HF_TOKEN_PATH, OPERATIONS_LOG, REPORTS_JSONL, REPORTS_OUTBOX, REPORTS_DATASET_ID, REPORTS_REMOTE_PREFIX, VERSION, DTYPE_BYTES, VERBOSE |
| from vlib import ctx |
| from vlib.compute import _load_voice, _orient, _cast_st_prepare, _cast_core, _is_head_name, _parse_quant_override, _quant_from_key, _quantize_to_Q, _qerr, _nearest_encodable, _fuse_delta, _gguf_fuse_one, _gguf_write_repl, _read_gguf_arch, _confirm_arch_mismatch, _cast_gguf, _cast_gguf_pack, _cast_pack, _parse_tensor_list, _parse_exclude_list, _resolve_one_tensor, _tensor_src_meta, _load_tensor, _pack_aliases, _resolve_pack_targets |
| from vlib.net import _HF_TOKEN_CACHE, _get_hf_token, _HF_TRUSTED_HOSTS, _is_hf_url, _hf_auth_headers, _PreserveAuthRedirectHandler, _parse_retry_after, _bounded_retry_after, _timeout, http_range, http_get_json, _safe_id, _acquire_parts_lock, download_range, _fmt_eta, _cleanup_tmp |
| from vlib.tensors import _validate_safetensors_header, read_safetensors, write_safetensors, _write_safetensors_streaming, rewrite_safetensors_tensor, _safetensors_write_repl, _DTYPE_BYTES_MAP, dtype_bytes, bf16_to_f32, decode_to_f32, encode_from_f32, _gguf_tensor_f32, _gguf_raw_to_f32, _gguf_raw_slice_to_f32, _DELTA_CHUNK_ELEMS, _gguf_nbytes, _parse_gguf_header, _is_gguf_file, _QUANT_DTYPES, _is_ggml_quant, _is_quant_fetched, _is_delta_safetensors, _write_single_tensor_gguf |
| from vlib.sources import OUTPUT_ALIASES, EMBEDDING_ALIASES, _family, _nested, detect_output_tensor, largest_2d, _hf_config, _hf_license, _hf_model_index, _normalize_hf_repo, _base_paths, _hf_api_base_model, _hf_mergekit_base, _detect_base, TensorRef, SafetensorsSource, GgufSource, open_source, _source_from_repo, _pick_gguf, _sanitize_voice_name, _NAME_RULES, _public_source, _slugify_voice_name, _triage_gguf_files, _gguf_unit_census, _gguf_dtype_rank, _gguf_source_for_repo, role_of, is_tied, find_head, find_embed, pair_role |
| from vlib.fetch import _source_arch, _resolve_output_name, _stream_safetensors_tensor, _fetch_tensor, _stream_gguf_tensor, _fetch_pack, _write_fetched, _write_voicepack_json |
| from vlib.registry import log_operation, _atomic_jsonl_append, _write_outbox, _append_report, _iter_voices, _registry_resolve, _read_meta, _registry_name_exists, _registry_folder, _assert_not_symlink, _write_voice, _resolve_name, _out_path, _stem, _check_overwrite |
| from vlib.cache import _base_rev, _cache_validate, _cache_topup_write, _ensure_base_tensors, _ensure_base_tensors_locked, _cache_base, _load_safetensors_or_gguf_f32, _resolve_head_want |
| from vlib.delta import _choose_voice, _resolve_base, _delta_gate, _check_delta_gate, _compute_delta |
| from vlib.reports import _hash_file, _sanitize_note, _sanitize_reporter, _prompt_share, _push_reports_batch, _push_report, _last_operation, cmd_report |
| from vlib.cmds_voices import cmd_get, cmd_cast, cmd_delta, cmd_add |
| from vlib.cmds_chorus import cmd_chorus, cmd_graft |
| from vlib.cmds_misc import cmd_list, _list_bases, _voice_tensors, _verify_voice_file, cmd_info, cmd_remove, cmd_cleanup, cmd_auth, cmd_path, cmd_doctor, _print_version, _print_completion, _entry_script, _ensure_venv |
|
|
| def main(): |
| global VERBOSE |
| parser = argparse.ArgumentParser(prog="voice", description="Lift a model's voice and cast it onto another model.") |
| parser.add_argument("-v", "--verbose", action="store_true", help="Verbose: tensor shapes, decisions, and paths") |
| parser.add_argument("--non-interactive", action="store_true", help="Fail instead of prompting (same as piped stdin; needs --yes to overwrite/remove)") |
| parser.add_argument("--version", action="store_true", help="Show version and environment") |
| parser.add_argument("--completion", choices=["bash", "zsh", "fish"], help="Print shell completion script") |
| sub = parser.add_subparsers(dest="command") |
|
|
| p_get = sub.add_parser("get", help="Get a voice from a model (safetensors repo, .gguf, or URL)") |
| p_get.add_argument("source", help="HF model ID, .gguf/.safetensors path, or URL") |
| p_get.add_argument("tensors", nargs="*", help="Tensor name(s) to fetch (default: auto-detect output head; 2+ or --multi saves a voicepack)") |
| p_get.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_get.add_argument("--name", help="Voice name to save (registry voices only; not with --out)") |
| p_get.add_argument("--multi", action="store_true", help="Fetch the named tensors (1+) as voicepack.safetensors + voicepack.json (not with --name)") |
| p_get.add_argument("--out", help="Write a standalone file instead of a registry voice (default pack name: voicepack.safetensors)") |
| p_get.add_argument("--yes", action="store_true", help="Assume yes to prompts (overwrite without asking)") |
|
|
| p_cast = sub.add_parser("cast", help="Cast a voice onto a .gguf or .safetensors file") |
| p_cast.add_argument("target", help="Path to the target .gguf or .safetensors") |
| p_cast.add_argument("voice", help="Voice name, or a .safetensors/.gguf path") |
| p_cast.add_argument("tensors", nargs="*", help="Only cast these tensors (default: every tensor in the voice)") |
| p_cast.add_argument("--out", help="Output file or folder (default: ~/Voiced/<voice>-<model>/<voice>-<model>.<ext matching target>; a folder invents the filename)") |
| p_cast.add_argument("--yes", action="store_true", help="Assume yes to prompts (overwrite without asking)") |
| p_cast.add_argument("--quant", default="keep", help="Output quant override: keep (default, preserve quality) | target | any quant (q8_0,q4_0,q4_k,q6_k,iq4_nl,f16,f32,bf16,...; safetensors targets: keep|target|f32|f16|bf16)") |
| p_cast.add_argument("--exclude", action="append", dest="excludes", default=[], help="Skip tensors on cast (repeatable: --exclude foo --exclude bar, or --exclude foo,bar)") |
| p_cast.add_argument("--speak", action="store_true", help="Voice speaks through output.weight (invented if absent when head/embed geometry survives; GGUF targets only)") |
| p_cast.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_delta = sub.add_parser("delta", help="Make a reusable style (voice − base)") |
| p_delta.add_argument("voicename", nargs="?", help="Voice name (default: pick from registry)") |
| p_delta.add_argument("--base", help="Base HF repo (org/model) or local path to base safetensors") |
| p_delta.add_argument("--yes", action="store_true", help="Assume yes to prompts (overwrite without asking)") |
| p_delta.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_graft = sub.add_parser("graft", help="Graft one voice onto another (DELLA single graft)") |
| p_graft.add_argument("pulse", nargs="?", help="Pulse voice name (default: pick from registry)") |
| p_graft.add_argument("donor", nargs="?", help="Donor voice name (default: pick from registry)") |
| p_graft.add_argument("--density", type=float, default=0.8, help="Keep-fraction for magnitude pruning (default: 0.8)") |
| p_graft.add_argument("--epsilon", type=float, default=0.1, help="Rank-ramp half-width (default: 0.1)") |
| p_graft.add_argument("--lambda_", dest="lambda_", type=float, default=1.0, help="Global rescale after combine (default: 1.0)") |
| p_graft.add_argument("--rescale", default="l1", choices=["l1", "inv_p"], help="Prune rescale: l1 (stable) or inv_p (paper-faithful)") |
| p_graft.add_argument("--seed", type=int, default=42, help="Bernoulli seed for reproducible grafts (default: 42)") |
| p_graft.add_argument("--out", help="Output voice name (default: chorus-<pulse>-<donor>)") |
| p_graft.add_argument("--yes", action="store_true", help="Assume yes to prompts (overwrite without asking)") |
| p_graft.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_chorus = sub.add_parser("chorus", help="Blend voices: pulse + support + lead (DELLA chorus)") |
| p_chorus.add_argument("pulse", nargs="?", help="Pulse voice name (default: pick interactively)") |
| p_chorus.add_argument("donors", nargs="*", help="Support + lead voice names (0-2; 1 = duet at lead density)") |
| p_chorus.add_argument("--config", action="store_true", help="Stage knob screens per graft (Enter accepts each default; screens prefill from flags)") |
| p_chorus.add_argument("--density", type=float, default=0.8, help="First-graft density (default: 0.8)") |
| p_chorus.add_argument("--lead-density", type=float, default=0.9, help="Lead-graft density (default: 0.9)") |
| p_chorus.add_argument("--epsilon", type=float, default=0.1, help="Rank-ramp half-width (default: 0.1)") |
| p_chorus.add_argument("--lead-epsilon", type=float, default=0.05, help="Lead-graft epsilon (default: 0.05; keeps 0.9±eps valid)") |
| p_chorus.add_argument("--lambda_", dest="lambda_", type=float, default=1.0, help="Global rescale after combine (default: 1.0)") |
| p_chorus.add_argument("--rescale", default="l1", choices=["l1", "inv_p"], help="Prune rescale: l1 (stable) or inv_p (paper-faithful)") |
| p_chorus.add_argument("--seed", type=int, default=42, help="Bernoulli seed for reproducible choruses (default: 42)") |
| p_chorus.add_argument("--out", help="Final output voice name (default: chorus-<pulse>-<lead>)") |
| p_chorus.add_argument("--yes", action="store_true", help="Assume yes to prompts (overwrite without asking)") |
| p_chorus.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_add = sub.add_parser("add", help="Add a voice from local voice.safetensors + voice.json") |
| p_add.add_argument("safetensors", help="Path to voice.safetensors (required)") |
| p_add.add_argument("json", help="Path to voice.json (required)") |
| p_add.add_argument("--name", help="Voice name to save (default from voice.json)") |
| p_add.add_argument("--yes", action="store_true", help="Auto-confirm overwrite") |
| p_add.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| p_list = sub.add_parser("list", help="List stored voices") |
| p_list.add_argument("target", nargs="?", choices=["base"], metavar="base", help="Show cached base heads (same as --base)") |
| p_list.add_argument("--base", dest="list_base", action="store_true", help="Show cached base heads instead of voices") |
| p_list.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| p_info = sub.add_parser("info", help="Show voice details") |
| p_info.add_argument("voicename", help="Voice name") |
| p_info.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| p_rm = sub.add_parser("remove", help="Remove a stored voice") |
| p_rm.add_argument("voice", help="Voice name") |
| p_rm.add_argument("--yes", action="store_true", help="Assume yes to prompts (remove without asking)") |
| p_rm.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| p_cleanup = sub.add_parser("cleanup", help="Remove incomplete downloads") |
| p_cleanup.add_argument("--force", action="store_true", help="Also clear in-progress downloads (locked)") |
| p_cleanup.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| p_auth = sub.add_parser("auth", help="Save your HF token (optional)") |
| p_auth.add_argument("--clear", action="store_true", help="Remove the stored token") |
| p_auth.add_argument("--show", action="store_true", help="Show the masked token") |
| p_auth.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| p_report = sub.add_parser("report", help="Save a report locally and optionally share to community") |
| p_report.add_argument("--message", "-m", help="Freeform note (240ch, sanitized)") |
| p_report.add_argument("--voice", help="Voice name context") |
| p_report.add_argument("--target", help="Target file basename context") |
| p_report.add_argument("--reporter", help="Reporter id (optional, 64ch)") |
| p_report.add_argument("--yes", action="store_true", help="Non-interactive: skip preview and auto-share to community") |
| p_report.add_argument("--no-push", action="store_true", help="Local only, skip share prompt") |
| p_report.add_argument("--sync", action="store_true", help="Sync pending outbox to community") |
| p_report.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_report.add_argument("--operation", help=argparse.SUPPRESS) |
|
|
| p_path = sub.add_parser("path", help="Put voice on your PATH") |
| p_path.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
| p_doctor = sub.add_parser("doctor", help="Check your setup") |
| p_doctor.add_argument("-q", "--quiet", action="store_true", help="Quieter: progress and success notes off (warnings and errors still print)") |
|
|
| args = parser.parse_args() |
| VERBOSE = args.verbose |
| ctx.VERBOSE = args.verbose |
| ctx.QUIET = bool(getattr(args, "quiet", False)) |
|
|
| if getattr(args, "completion", None): |
| _print_completion(args.completion) |
| sys.exit(0) |
| if getattr(args, "version", False): |
| _print_version() |
| sys.exit(0) |
| if args.command is None: |
| parser.print_help() |
| sys.exit(2) |
|
|
| _ensure_venv(args.command) |
|
|
| dispatch = { |
| "get": cmd_get, "cast": cmd_cast, "delta": cmd_delta, "add": cmd_add, "list": cmd_list, |
| "info": cmd_info, "remove": cmd_remove, "cleanup": cmd_cleanup, "auth": cmd_auth, |
| "report": cmd_report, "path": cmd_path, "doctor": cmd_doctor, |
| "graft": cmd_graft, "chorus": cmd_chorus, |
| } |
| dispatch[args.command](args) |
|
|
| if __name__ == "__main__": |
| main() |
|
|
