Labyricorn/voicebox
1
0
Fork 0
mirror of https://github.com/jamiepine/voicebox.git synced 2026-10-03 00:55:14 -07:00
Code Issues Packages Projects Releases Wiki Activity

fix gated Llama tokenizer error by redirecting to ungated mirror

Browse Source 
TADA hardcodes 'meta-llama/Llama-3.2-1B' as its tokenizer source in
both the Aligner and TadaForCausalLM.from_pretrained(). That repo is
gated and requires accepting Meta's license on HuggingFace.

Monkey-patch AutoTokenizer.from_pretrained during model loading to
redirect Llama tokenizer requests to 'unsloth/Llama-3.2-1B', an
ungated mirror with identical tokenizer files. The patch is scoped
to model loading only and restored immediately after.
This commit is contained in:
James Pine
2026-03-17 02:22:26 -07:00
parent b02ce8e2f3
commit 7a90290a76
1 changed files with 46 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files
backend/backends/hume_backend.py
+46 -14
View File
@@ -130,27 +130,59 @@ class HumeTadaBackend:
allow_patterns=["*.safetensors", "*.json", "*.txt", "*.bin", "*.model"], allow_patterns=["*.safetensors", "*.json", "*.txt", "*.bin", "*.model"],
) )
# Pre-download the Llama tokenizer from an ungated mirror.
# TADA hardcodes "meta-llama/Llama-3.2-1B" which is gated;
# we redirect to unsloth's ungated copy at load time.
logger.info("Downloading Llama tokenizer (ungated mirror)...")
snapshot_download(
repo_id="unsloth/Llama-3.2-1B",
token=None,
allow_patterns=["tokenizer*", "special_tokens*"],
)
# Determine dtype — use bf16 on CUDA for ~50% memory savings # Determine dtype — use bf16 on CUDA for ~50% memory savings
if device == "cuda" and torch.cuda.is_bf16_supported(): if device == "cuda" and torch.cuda.is_bf16_supported():
model_dtype = torch.bfloat16 model_dtype = torch.bfloat16
else: else:
model_dtype = torch.float32 model_dtype = torch.float32
# Load encoder (only needed for voice prompt encoding) # TADA hardcodes "meta-llama/Llama-3.2-1B" as the tokenizer
from tada.modules.encoder import Encoder # source in its Aligner and TadaForCausalLM.from_pretrained().
logger.info("Loading TADA encoder...") # That repo is gated (requires Meta license acceptance on HF).
self.encoder = Encoder.from_pretrained( # Monkey-patch AutoTokenizer.from_pretrained to redirect to an
TADA_CODEC_REPO, subfolder="encoder" # ungated mirror that ships the identical tokenizer files.
).to(device) from transformers import AutoTokenizer
self.encoder.eval() _orig_from_pretrained = AutoTokenizer.from_pretrained.__func__
# Load the causal LM (includes decoder for wav generation) @classmethod # type: ignore[misc]
from tada.modules.tada import TadaForCausalLM def _patched_from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs):
logger.info(f"Loading TADA {model_size} model...") if "meta-llama/Llama-3.2" in str(pretrained_model_name_or_path):
self.model = TadaForCausalLM.from_pretrained( pretrained_model_name_or_path = "unsloth/Llama-3.2-1B"
repo, torch_dtype=model_dtype kwargs.setdefault("token", None)
).to(device) logger.info("Redirecting Llama tokenizer to ungated mirror: unsloth/Llama-3.2-1B")
self.model.eval() return _orig_from_pretrained(cls, pretrained_model_name_or_path, *args, **kwargs)
AutoTokenizer.from_pretrained = _patched_from_pretrained
try:
# Load encoder (only needed for voice prompt encoding)
from tada.modules.encoder import Encoder
logger.info("Loading TADA encoder...")
self.encoder = Encoder.from_pretrained(
TADA_CODEC_REPO, subfolder="encoder"
).to(device)
self.encoder.eval()
# Load the causal LM (includes decoder for wav generation)
from tada.modules.tada import TadaForCausalLM
logger.info(f"Loading TADA {model_size} model...")
self.model = TadaForCausalLM.from_pretrained(
repo, torch_dtype=model_dtype
).to(device)
self.model.eval()
finally:
# Restore original to avoid affecting other code
AutoTokenizer.from_pretrained = _orig_from_pretrained
logger.info(f"HumeAI TADA {model_size} loaded successfully on {device}") logger.info(f"HumeAI TADA {model_size} loaded successfully on {device}")