Skip to content

C
coderai
Commit c8f70fe4 authored by Your Name's avatar Your Name
Browse files
Options

Fix: Skip faster-whisper for GGUF files 

faster-whisper doesn't support GGUF format (it's llama.cpp format).
Now detects GGUF files by extension and goes directly to whispercpp.
parent 11a0fd46
Hide whitespace changes
Inline Side-by-side
Showing with 122 additions and 101 deletions
coderai
View file @ c8f70fe4
......@@ -2383,109 +2383,117 @@ async def create_transcription(
tmp_path = tmp.name
try:
# Try faster-whisper first (requires PyTorch), fall back to whispercpp if it fails
faster_whisper_failed = False
try:
from faster_whisper import WhisperModel
# Determine compute type based on GPU availability
import torch
if torch.cuda.is_available():
compute_type = "float16"
else:
compute_type = "int8"
# Try to load the model (lazy loading)
model_key = f"audio:{model_to_use}"
whisper_model = multi_model_manager.get_model(model_key)
if whisper_model is None:
print(f"Loading faster-whisper model: {model_to_use}")
# Check if model is a GGUF file - faster-whisper doesn't support GGUF format
is_gguf_model = model_to_use.endswith('.gguf') or 'gguf' in model_to_use.lower()
if is_gguf_model:
# Skip faster-whisper for GGUF files - go directly to whispercpp
print("Detected GGUF model - using whispercpp backend")
faster_whisper_failed = True
else:
# Try faster-whisper first
faster_whisper_failed = False
try:
from faster_whisper import WhisperModel
# Check if model_to_use is a URL - download it (with caching)
model_path = None
if model_to_use.startswith('http://') or model_to_use.startswith('https://'):
# Check cache first
cached_path = get_cached_model_path(model_to_use)
if cached_path:
model_to_use = cached_path
print(f"Using cached model: {model_to_use}")
else:
print(f"Downloading model from URL: {model_to_use}")
try:
import requests
import hashlib
# Get cache directory
cache_dir = get_model_cache_dir()
# Extract filename from URL
url_path = model_to_use.split('?')[0]
filename = os.path.basename(url_path)
if not filename.endswith('.bin') and not filename.endswith('.ggml'):
filename = "whisper-model.bin"
# Create safe filename in cache
url_hash = hashlib.sha256(model_to_use.encode()).hexdigest()
cached_filename = f"{url_hash}_{filename}"
model_path = os.path.join(cache_dir, cached_filename)
# Download to cache
response = requests.get(model_to_use, stream=True)
response.raise_for_status()
total_size = int(response.headers.get('content-length', 0))
downloaded = 0
with open(model_path, 'wb') as f:
for chunk in response.iter_content(chunk_size=8192*1024):
if chunk:
f.write(chunk)
downloaded += len(chunk)
if total_size > 0:
percent = (downloaded / total_size) * 100
print(f"Downloaded: {percent:.1f}%", end='\r')
print(f"\nDownloaded and cached to: {model_path}")
model_to_use = model_path
except Exception as e:
print(f"Error downloading model: {e}")
raise
# Determine compute type based on GPU availability
import torch
if torch.cuda.is_available():
compute_type = "float16"
else:
compute_type = "int8"
whisper_model = WhisperModel(
model_to_use,
device="cuda" if torch.cuda.is_available() else "cpu",
compute_type=compute_type
# Try to load the model (lazy loading)
model_key = f"audio:{model_to_use}"
whisper_model = multi_model_manager.get_model(model_key)
if whisper_model is None:
print(f"Loading faster-whisper model: {model_to_use}")
# Check if model_to_use is a URL - download it (with caching)
model_path = None
if model_to_use.startswith('http://') or model_to_use.startswith('https://'):
# Check cache first
cached_path = get_cached_model_path(model_to_use)
if cached_path:
model_to_use = cached_path
print(f"Using cached model: {model_to_use}")
else:
print(f"Downloading model from URL: {model_to_use}")
try:
import requests
import hashlib
# Get cache directory
cache_dir = get_model_cache_dir()
# Extract filename from URL
url_path = model_to_use.split('?')[0]
filename = os.path.basename(url_path)
if not filename.endswith('.bin') and not filename.endswith('.ggml'):
filename = "whisper-model.bin"
# Create safe filename in cache
url_hash = hashlib.sha256(model_to_use.encode()).hexdigest()
cached_filename = f"{url_hash}_{filename}"
model_path = os.path.join(cache_dir, cached_filename)
# Download to cache
response = requests.get(model_to_use, stream=True)
response.raise_for_status()
total_size = int(response.headers.get('content-length', 0))
downloaded = 0
with open(model_path, 'wb') as f:
for chunk in response.iter_content(chunk_size=8192*1024):
if chunk:
f.write(chunk)
downloaded += len(chunk)
if total_size > 0:
percent = (downloaded / total_size) * 100
print(f"Downloaded: {percent:.1f}%", end='\r')
print(f"\nDownloaded and cached to: {model_path}")
model_to_use = model_path
except Exception as e:
print(f"Error downloading model: {e}")
raise
whisper_model = WhisperModel(
model_to_use,
device="cuda" if torch.cuda.is_available() else "cpu",
compute_type=compute_type
)
# Store in multi_model_manager
multi_model_manager.add_model(model_key, whisper_model)
# Run transcription
segments, info = whisper_model.transcribe(
tmp_path,
language=language,
initial_prompt=prompt,
temperature=temperature or 0.0,
)
# Store in multi_model_manager
multi_model_manager.add_model(model_key, whisper_model)
# Run transcription
segments, info = whisper_model.transcribe(
tmp_path,
language=language,
initial_prompt=prompt,
temperature=temperature or 0.0,
)
# Collect all segments
text_parts = []
for segment in segments:
text_parts.append(segment.text.strip())
full_text = " ".join(text_parts)
# Collect all segments
text_parts = []
for segment in segments:
text_parts.append(segment.text.strip())
full_text = " ".join(text_parts)
return {"text": full_text}
return {"text": full_text}
except ImportError:
# faster-whisper not available, will try whispercpp below
faster_whisper_failed = True
except Exception as e:
# faster-whisper failed for some other reason (e.g., GGUF file not supported)
print(f"Warning: faster-whisper failed to load model: {e}")
faster_whisper_failed = True
except ImportError:
# faster-whisper not available, will try whispercpp below
faster_whisper_failed = True
except Exception as e:
# faster-whisper failed for some other reason
print(f"Warning: faster-whisper failed to load model: {e}")
faster_whisper_failed = True
# If faster-whisper failed (not installed or couldn't load), try whispercpp
if faster_whisper_failed:
......@@ -3755,8 +3763,19 @@ def main():
should_preload = load_mode in ("loadall", "loadswap") or (model_name is None and args.audio_model)
if should_preload:
print(f"Pre-loading audio model...")
# First try faster-whisper, then fall back to whispercpp if it fails
faster_whisper_failed = False
# Check if model is a GGUF file - faster-whisper doesn't support GGUF format
model_to_use = args.audio_model
is_gguf_model = model_to_use.endswith('.gguf') or 'gguf' in model_to_use.lower()
if is_gguf_model:
# Skip faster-whisper for GGUF files - it doesn't support them
# Go directly to whispercpp
print("Detected GGUF model - using whispercpp backend")
faster_whisper_failed = True
else:
# Try faster-whisper first
faster_whisper_failed = False
try:
# Try faster-whisper first (requires torch)
from faster_whisper import WhisperModel
......@@ -3802,6 +3821,8 @@ def main():
# If faster-whisper failed (not installed or couldn't load), try whispercpp
if faster_whisper_failed:
# Check if model is a GGUF file - whispercpp can handle those
model_is_gguf = model_to_use.endswith('.gguf') or (model_path and model_path.endswith('.gguf'))
try:
import whispercpp
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment