transcription flow cilalama, bugfixes

.gitignore

@@ -14,4 +14,6 @@ venv/
 # Cython
 *.pyd
 
-
+logs/**
+User/**
+data/**

data

@@ -0,0 +1 @@
+/media/raid/agents/tools/mcp-transcriptor

mcp.logs

@@ -0,0 +1,6 @@
+{"jsonrpc":"2.0","id":1,"result":{"protocolVersion":"2025-03-26","capabilities":{"experimental":{},"prompts":{"listChanged":false},"resources":{"subscribe":false,"listChanged":false},"tools":{"listChanged":false}},"serverInfo":{"name":"fast-whisper-mcp-server","version":"1.9.4"}}}
+INFO:mcp.server.lowlevel.server:Processing request of type ListToolsRequest
+{"jsonrpc":"2.0","id":2,"result":{"tools":[{"name":"get_model_info_api","description":"\n    Get available Whisper model information\n    ","inputSchema":{"properties":{},"title":"get_model_info_apiArguments","type":"object"}},{"name":"transcribe","description":"\n    Transcribe audio files using Faster Whisper\n\n    Args:\n        audio_path: Path to the audio file\n        model_name: Model name (tiny, base, small, medium, large-v1, large-v2, large-v3)\n        device: Execution device (cpu, cuda, auto)\n        compute_type: Computation type (float16, int8, auto)\n        language: Language code (such as zh, en, ja, etc., auto-detect by default)\n        output_format: Output format (vtt, srt, json or txt)\n        beam_size: Beam search size, larger values may improve accuracy but reduce speed\n        temperature: Sampling temperature, greedy decoding\n        initial_prompt: Initial prompt text, can help the model better understand context\n        output_directory: Output directory path, defaults to the audio file's directory\n\n    Returns:\n        str: Transcription result, in VTT subtitle or JSON format\n    ","inputSchema":{"properties":{"audio_path":{"title":"Audio Path","type":"string"},"model_name":{"default":"large-v3","title":"Model Name","type":"string"},"device":{"default":"auto","title":"Device","type":"string"},"compute_type":{"default":"auto","title":"Compute Type","type":"string"},"language":{"default":null,"title":"Language","type":"string"},"output_format":{"default":"vtt","title":"Output Format","type":"string"},"beam_size":{"default":5,"title":"Beam Size","type":"integer"},"temperature":{"default":0.0,"title":"Temperature","type":"number"},"initial_prompt":{"default":null,"title":"Initial Prompt","type":"string"},"output_directory":{"default":null,"title":"Output Directory","type":"string"}},"required":["audio_path"],"title":"transcribeArguments","type":"object"}},{"name":"batch_transcribe_audio","description":"\n    Batch transcribe audio files in a folder\n\n    Args:\n        audio_folder: Path to the folder containing audio files\n        output_folder: Output folder path, defaults to a 'transcript' subfolder in audio_folder\n        model_name: Model name (tiny, base, small, medium, large-v1, large-v2, large-v3)\n        device: Execution device (cpu, cuda, auto)\n        compute_type: Computation type (float16, int8, auto)\n        language: Language code (such as zh, en, ja, etc., auto-detect by default)\n        output_format: Output format (vtt, srt, json or txt)\n        beam_size: Beam search size, larger values may improve accuracy but reduce speed\n        temperature: Sampling temperature, 0 means greedy decoding\n        initial_prompt: Initial prompt text, can help the model better understand context\n        parallel_files: Number of files to process in parallel (only effective in CPU mode)\n\n    Returns:\n        str: Batch processing summary, including processing time and success rate\n    ","inputSchema":{"properties":{"audio_folder":{"title":"Audio Folder","type":"string"},"output_folder":{"default":null,"title":"Output Folder","type":"string"},"model_name":{"default":"large-v3","title":"Model Name","type":"string"},"device":{"default":"auto","title":"Device","type":"string"},"compute_type":{"default":"auto","title":"Compute Type","type":"string"},"language":{"default":null,"title":"Language","type":"string"},"output_format":{"default":"vtt","title":"Output Format","type":"string"},"beam_size":{"default":5,"title":"Beam Size","type":"integer"},"temperature":{"default":0.0,"title":"Temperature","type":"number"},"initial_prompt":{"default":null,"title":"Initial Prompt","type":"string"},"parallel_files":{"default":1,"title":"Parallel Files","type":"integer"}},"required":["audio_folder"],"title":"batch_transcribe_audioArguments","type":"object"}}]}}
+INFO:mcp.server.lowlevel.server:Processing request of type CallToolRequest
+INFO:model_manager:GPU test passed: NVIDIA GeForce RTX 3060 (12.5GB)
+INFO:model_manager:Loading Whisper model: large-v3 device: cuda compute type: float16

model_manager.py

@@ -4,7 +4,7 @@ Model Management Module
 Responsible for loading, caching, and managing Whisper models
 """
 
-import os
+import os; print(os.environ.get("WHISPER_MODEL_DIR"))
 import time
 import logging
 from typing import Dict, Any
@@ -17,6 +17,30 @@ logger = logging.getLogger(__name__)
 # Global model instance cache
 model_instances = {}
 
+def test_gpu_driver():
+    """Simple GPU driver test"""
+    try:
+        if not torch.cuda.is_available():
+            logger.error("CUDA not available in PyTorch")
+            raise RuntimeError("CUDA not available")
+        
+        gpu_count = torch.cuda.device_count()
+        if gpu_count == 0:
+            logger.error("No CUDA devices found")
+            raise RuntimeError("No CUDA devices")
+        
+        # Quick GPU test
+        test_tensor = torch.randn(10, 10).cuda()
+        _ = test_tensor @ test_tensor.T
+        
+        device_name = torch.cuda.get_device_name(0)
+        memory_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
+        logger.info(f"GPU test passed: {device_name} ({memory_gb:.1f}GB)")
+        
+    except Exception as e:
+        logger.error(f"GPU test failed: {e}")
+        raise RuntimeError(f"GPU initialization failed: {e}")
+
 def get_whisper_model(model_name: str, device: str, compute_type: str) -> Dict[str, Any]:
     """
     Get or create Whisper model instance
@@ -46,9 +70,8 @@ def get_whisper_model(model_name: str, device: str, compute_type: str) -> Dict[s
         raise ValueError(f"Invalid device: {device}. Valid devices: cpu, cuda")
 
     if device == "cuda" and not torch.cuda.is_available():
-        logger.warning("CUDA not available, automatically switching to CPU")
-        device = "cpu"
-        compute_type = "int8"
+        logger.error("CUDA requested but not available")
+        raise RuntimeError("CUDA not available but explicitly requested")
 
     if compute_type not in ["float16", "int8"]:
         raise ValueError(f"Invalid compute type: {compute_type}. Valid compute types: float16, int8")
@@ -65,8 +88,9 @@ def get_whisper_model(model_name: str, device: str, compute_type: str) -> Dict[s
         logger.info(f"Using cached model instance: {model_key}")
         return model_instances[model_key]
 
-    # Clean GPU memory (if using CUDA)
+    # Test GPU driver before loading model and clean
     if device == "cuda":
+        test_gpu_driver()
         torch.cuda.empty_cache()
 
     # Instantiate model
@@ -173,4 +197,4 @@ def get_model_info() -> str:
             "memory_available": f"{torch.cuda.get_device_properties(0).total_memory / 1e9 - torch.cuda.memory_allocated() / 1e9:.2f} GB"
         }
 
-    return json.dumps(info, indent=2)
+    return json.dumps(info, indent=2)

run_server.sh

@@ -10,12 +10,13 @@ USER_ID=$(id -u)
 GROUP_ID=$(id -g)
 
 # Set environment variables
+export CUDA_VISIBLE_DEVICES=1
 export WHISPER_MODEL_DIR="/home/uad/agents/tools/mcp-transcriptor/data/models"
-export TRANSCRIPTION_OUTPUT_DIR="/home/uad/agents/tools/mcp-transcriptor/data/transcripts"
-export TRANSCRIPTION_BATCH_OUTPUT_DIR="/home/uad/agents/tools/mcp-transcriptor/data/transcripts/batch"
-export TRANSCRIPTION_MODEL="base"
-export TRANSCRIPTION_DEVICE="auto"
-export TRANSCRIPTION_COMPUTE_TYPE="auto"
+export TRANSCRIPTION_OUTPUT_DIR="/media/raid/agents/tools/mcp-transcriptor/outputs"
+export TRANSCRIPTION_BATCH_OUTPUT_DIR="/media/raid/agents/tools/mcp-transcriptor/outputs/batch"
+export TRANSCRIPTION_MODEL="large-v3"
+export TRANSCRIPTION_DEVICE="cuda"
+export TRANSCRIPTION_COMPUTE_TYPE="cuda"
 export TRANSCRIPTION_OUTPUT_FORMAT="txt"
 export TRANSCRIPTION_BEAM_SIZE="2"
 export TRANSCRIPTION_TEMPERATURE="0.0"
@@ -24,6 +25,7 @@ export TRANSCRIPTION_FILENAME_PREFIX="test_"
 
 # Log start of the script
 echo "$(datetime_prefix) Starting whisper server script..."
+echo "test: $WHISPER_MODEL_DIR"
 
 # Optional: Verify required directories exist
 if [ ! -d "$WHISPER_MODEL_DIR" ]; then
@@ -32,5 +34,4 @@ if [ ! -d "$WHISPER_MODEL_DIR" ]; then
 fi
 
 # Run the Python script with the defined environment variables
-sudo /home/uad/agents/tools/mcp-transcriptor/venv/bin/python \
-  /home/uad/agents/tools/mcp-transcriptor/whisper_server.py
+/home/uad/agents/tools/mcp-transcriptor/venv/bin/python /home/uad/agents/tools/mcp-transcriptor/whisper_server.py 2>&1 | tee /home/uad/agents/tools/mcp-transcriptor/mcp.logs

transcriber.py

@@ -19,9 +19,9 @@ logger = logging.getLogger(__name__)
 # Environment variable defaults
 DEFAULT_OUTPUT_DIR = os.getenv('TRANSCRIPTION_OUTPUT_DIR', None)
 DEFAULT_BATCH_OUTPUT_DIR = os.getenv('TRANSCRIPTION_BATCH_OUTPUT_DIR', None)
-DEFAULT_MODEL = os.getenv('TRANSCRIPTION_MODEL', 'large-v3')
-DEFAULT_DEVICE = os.getenv('TRANSCRIPTION_DEVICE', 'auto')
-DEFAULT_COMPUTE_TYPE = os.getenv('TRANSCRIPTION_COMPUTE_TYPE', 'auto')
+DEFAULT_MODEL = os.getenv('TRANSCRIPTION_MODEL', 'base')
+DEFAULT_DEVICE = os.getenv('TRANSCRIPTION_DEVICE', 'cuda')
+DEFAULT_COMPUTE_TYPE = os.getenv('TRANSCRIPTION_COMPUTE_TYPE', 'base')
 DEFAULT_LANGUAGE = os.getenv('TRANSCRIPTION_LANGUAGE', None)
 DEFAULT_OUTPUT_FORMAT = os.getenv('TRANSCRIPTION_OUTPUT_FORMAT', 'txt')
 DEFAULT_BEAM_SIZE = int(os.getenv('TRANSCRIPTION_BEAM_SIZE', '5'))
@@ -31,7 +31,7 @@ DEFAULT_TEMPERATURE = float(os.getenv('TRANSCRIPTION_TEMPERATURE', '0.0'))
 WHISPER_MODEL_DIR = os.getenv('WHISPER_MODEL_DIR', None)
 
 # File naming configuration
-USE_TIMESTAMP = os.getenv('TRANSCRIPTION_USE_TIMESTAMP', 'true').lower() == 'true'
+USE_TIMESTAMP = os.getenv('TRANSCRIPTION_USE_TIMESTAMP', 'false').lower() == 'true'
 FILENAME_PREFIX = os.getenv('TRANSCRIPTION_FILENAME_PREFIX', '')
 FILENAME_SUFFIX = os.getenv('TRANSCRIPTION_FILENAME_SUFFIX', '')
 
@@ -97,6 +97,7 @@ def transcribe_audio(
 
         # Set transcription parameters
         options = {
+            "verbose": True,
             "language": language,
             "vad_filter": True,
             "vad_parameters": {"min_silence_duration_ms": 500},
@@ -106,7 +107,7 @@ def transcribe_audio(
             "word_timestamps": True,
             "suppress_tokens": [-1],
             "condition_on_previous_text": True,
-            "compression_ratio_threshold": 2.4
+            "compression_ratio_threshold": 2.4,
         }
 
         start_time = time.time()