Merge pull request #1130 from jamiepine/prep/pr-1112

Fix infinite HF retry storm when loading a cached model offline

backend/backends/base.py

@@ -15,6 +15,7 @@ from typing import Callable, List, Optional, Tuple
 import numpy as np
 
 from ..utils.audio import normalize_audio, load_audio
+from ..utils.hf_offline_patch import force_offline_if_cached
 from ..utils.progress import get_progress_manager
 from ..utils.hf_progress import HFProgressTracker, create_hf_progress_callback
 from ..utils.tasks import get_task_manager
@@ -319,7 +320,8 @@ def model_load_progress(
         )
 
     try:
-        yield tracker_context
+        with force_offline_if_cached(is_cached, model_name):
+            yield tracker_context
     except Exception as e:
         # Report error to both managers
         progress_manager.mark_error(model_name, str(e))

backend/backends/chatterbox_backend.py

@@ -29,11 +29,17 @@ logger = logging.getLogger(__name__)
 
 CHATTERBOX_HF_REPO = "ResembleAI/chatterbox"
 
-# Files that must be present for the multilingual model
+# The files ChatterboxMultilingualTTS.from_pretrained() downloads, as of
+# chatterbox-tts 0.1.7 (mtl_tts.py allow_patterns). The load runs with HF
+# offline mode forced when this reports cached, so a partial snapshot must not
+# count as cached -- if upstream adds a file to that list, add it here too.
 _MTL_WEIGHT_FILES = [
     "t3_mtl23ls_v2.safetensors",
     "s3gen.pt",
     "ve.pt",
+    "grapheme_mtl_merged_expanded_v1.json",
+    "conds.pt",
+    "Cangjie5_TC.json",
 ]
 
 

backend/backends/chatterbox_turbo_backend.py

@@ -29,11 +29,19 @@ logger = logging.getLogger(__name__)
 
 CHATTERBOX_TURBO_HF_REPO = "ResembleAI/chatterbox-turbo"
 
-# Files that must be present for the turbo model
+# The files ChatterboxTurboTTS.from_local() reads, as of chatterbox-tts 0.1.7:
+# the three weight files, the GPT-2 tokenizer files AutoTokenizer needs, and
+# the built-in voice. The load runs with HF offline mode forced when this
+# reports cached, so a partial snapshot must not count as cached -- if upstream
+# starts reading another file, add it here too.
 _TURBO_WEIGHT_FILES = [
     "t3_turbo_v1.safetensors",
     "s3gen_meanflow.safetensors",
     "ve.safetensors",
+    "tokenizer_config.json",
+    "vocab.json",
+    "merges.txt",
+    "conds.pt",
 ]
 
 

backend/backends/hume_backend.py

@@ -35,6 +35,9 @@ logger = logging.getLogger(__name__)
 
 # HuggingFace repos
 TADA_CODEC_REPO = "HumeAI/tada-codec"
+# TADA hardcodes the gated meta-llama/Llama-3.2-1B tokenizer; we load it from
+# this ungated mirror instead (see load_model).
+TADA_TOKENIZER_REPO = "unsloth/Llama-3.2-1B"
 TADA_1B_REPO = "HumeAI/tada-1b"
 TADA_3B_ML_REPO = "HumeAI/tada-3b-ml"
 
@@ -52,6 +55,12 @@ _TADA_CODEC_WEIGHT_FILES = [
     "encoder/model.safetensors",
 ]
 
+_TADA_TOKENIZER_FILES = [
+    "tokenizer.json",
+    "tokenizer_config.json",
+    "special_tokens_map.json",
+]
+
 
 class HumeTadaBackend:
     """HumeAI TADA TTS backend for high-quality voice cloning."""
@@ -80,7 +89,8 @@ class HumeTadaBackend:
         repo = TADA_MODEL_REPOS.get(model_size, TADA_1B_REPO)
         model_cached = is_model_cached(repo, required_files=_TADA_MODEL_WEIGHT_FILES)
         codec_cached = is_model_cached(TADA_CODEC_REPO, required_files=_TADA_CODEC_WEIGHT_FILES)
-        return model_cached and codec_cached
+        tokenizer_cached = is_model_cached(TADA_TOKENIZER_REPO, required_files=_TADA_TOKENIZER_FILES)
+        return model_cached and codec_cached and tokenizer_cached
 
     async def load_model(self, model_size: str = "1B") -> None:
         """Load the TADA model and encoder."""
@@ -140,7 +150,7 @@ class HumeTadaBackend:
             # local cache path so we can point TADA at it directly.
             logger.info("Downloading Llama tokenizer (ungated mirror)...")
             tokenizer_path = snapshot_download(
-                repo_id="unsloth/Llama-3.2-1B",
+                repo_id=TADA_TOKENIZER_REPO,
                 token=None,
                 allow_patterns=["tokenizer*", "special_tokens*"],
             )