Add VocalRender VoxCPM2 finetune and conversion script

.gitattributes

@@ -249,3 +249,4 @@ Voxtral-Mini-4B-Realtime-2602-GGUF/voxtral-mini-4b-realtime-2602-q4_k.gguf filte
 Voxtral-Mini-4B-Realtime-2602-GGUF/voxtral-mini-4b-realtime-2602-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
 Kokoro-82M-GGUF/kokoro-82m-multilingual-q8_0.gguf filter=lfs diff=lfs merge=lfs -text
 LatentSync-1.6-GGUF/latentsync-1.6-f32-f16.gguf filter=lfs diff=lfs merge=lfs -text
+VoxCPM2-GGUF/finetune/vocalrender-pro-orig.gguf filter=lfs diff=lfs merge=lfs -text

VoxCPM2-GGUF/finetune/convert_vocalrender.py

@@ -0,0 +1,96 @@
+#!/usr/bin/env python3
+"""Convert a VocalRender checkpoint into a standalone VoxCPM2 GGUF.
+
+Requires torch, safetensors, and an audiocpp_gguf executable built with VoxCPM2.
+Example:
+    python convert_vocalrender.py --input ./VocalRender-Pro \
+        --output ./vocalrender-pro-orig.gguf --converter ./audiocpp_gguf
+"""
+
+import argparse
+import json
+import os
+from pathlib import Path
+import shutil
+import subprocess
+import tempfile
+
+import torch
+from safetensors import safe_open
+from safetensors.torch import save_file
+
+
+def main():
+    parser = argparse.ArgumentParser(description=__doc__)
+    parser.add_argument("--input", type=Path, required=True, help="Source checkpoint directory.")
+    parser.add_argument("--output", type=Path, required=True, help="Output orig-precision GGUF.")
+    parser.add_argument("--converter", default="audiocpp_gguf", help="Converter executable path or PATH command.")
+    parser.add_argument("--model-spec", type=Path, help="Optional voxcpm2 spec; otherwise use the converter catalog.")
+    parser.add_argument("--overwrite", action="store_true")
+    args = parser.parse_args()
+    source = args.input.expanduser().resolve(strict=True)
+    output = args.output.expanduser().resolve()
+    converter = shutil.which(os.path.expanduser(args.converter))
+    if converter is None:
+        parser.error("Converter not found; provide --converter or add audiocpp_gguf to PATH.")
+    if output.exists() and not args.overwrite:
+        parser.error("Output already exists; use --overwrite to replace it.")
+    weights = source / "model.safetensors"
+    with safe_open(weights, framework="pt", device="cpu") as tensors:
+        shape = tensors.get_slice("base_lm.embed_tokens.weight").get_shape()
+    config = json.loads((source / "config.json").read_text(encoding="utf-8"))
+    tokenizer = json.loads((source / "tokenizer.json").read_text(encoding="utf-8"))
+    token_config = json.loads((source / "tokenizer_config.json").read_text(encoding="utf-8"))
+    ids = [*tokenizer["model"]["vocab"].values(), *[t["id"] for t in tokenizer["added_tokens"]]]
+    if len(shape) != 2 or shape[1] != config["lm_config"]["hidden_size"]:
+        raise ValueError("Embedding shape does not match the model configuration.")
+    if min(ids) < 0 or max(ids) >= shape[0]:
+        raise ValueError("Tokenizer IDs exceed the checkpoint embedding capacity.")
+    config["lm_config"]["vocab_size"] = shape[0]
+    tokenizer["model"]["merges"] = [
+        " ".join(merge) if isinstance(merge, list) else merge
+        for merge in tokenizer["model"]["merges"]
+    ]
+    token_config["added_tokens_decoder"] = {
+        str(t["id"]): {key: value for key, value in t.items() if key != "id"}
+        for t in tokenizer["added_tokens"]
+    }
+    special = {key: token_config[key] for key in ("bos_token", "eos_token", "unk_token")}
+    output.parent.mkdir(parents=True, exist_ok=True)
+    with tempfile.TemporaryDirectory(prefix="vocalrender-convert-", dir=output.parent) as temporary:
+        root = Path(temporary) / source.name
+        root.mkdir()
+        for name, data in (("config.json", config), ("tokenizer.json", tokenizer),
+                           ("tokenizer_config.json", token_config), ("special_tokens_map.json", special)):
+            (root / name).write_text(json.dumps(data, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
+        checkpoint = torch.load(source / "audiovae.pth", map_location="cpu", weights_only=True)
+        state = checkpoint.get("state_dict", checkpoint)
+        tensors = {}
+        for name, tensor in state.items():
+            if not isinstance(name, str) or not isinstance(tensor, torch.Tensor):
+                raise ValueError("AudioVAE state_dict must contain named tensors only.")
+            tensors[name] = tensor.detach().cpu().contiguous()
+        if not tensors:
+            raise ValueError("AudioVAE checkpoint contains no tensors.")
+        vae = root / "audiovae.safetensors"
+        save_file(tensors, vae)
+        with safe_open(vae, framework="pt", device="cpu") as saved:
+            if set(saved.keys()) != set(tensors):
+                raise ValueError("AudioVAE tensor names changed during conversion.")
+            for name, expected in tensors.items():
+                actual = saved.get_tensor(name)
+                if actual.dtype != expected.dtype or actual.shape != expected.shape or not torch.equal(actual, expected):
+                    raise ValueError(f"AudioVAE tensor changed during conversion: {name}")
+        command = [str(Path(converter).resolve()), "--family", "voxcpm2", "--type", "orig",
+                   "--input", f"weights={weights}", "--input", f"audiovae_weights={vae}",
+                   "--root", str(root), "--output", str(output), "--log"]
+        if args.model_spec:
+            command.extend(["--model-spec", str(args.model_spec.expanduser().resolve(strict=True))])
+        if args.overwrite:
+            command.append("--overwrite")
+        subprocess.run(command, check=True)
+    print(f"Converted: {output}")
+
+
+if __name__ == "__main__":
+    main()

VoxCPM2-GGUF/finetune/vocalrender-pro-orig.gguf

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1c9148d67d3df67a0b04a8dbf9201d61bb7cac346bbfc9650b42f9d387a3df05
+size 9544158368