Attempt fix windows (#45)

* Update README.md

* feat: Add custom prompt support for improved transcription accuracy

This commit implements custom prompt functionality to improve Whisper transcription
accuracy for domain-specific vocabulary, names, and technical terms.

## New Features

### CLI Arguments
- --initial_prompt: Direct prompt specification
- --prompt_file: Load prompt from text file

### Python API
- Added initial_prompt parameter to transcribe() function
- Maintains backward compatibility

### Backend Support
- Works with all Whisper backends (cpu, cuda, insane, mps)
- Updated whisper_mac.py to support other_args parameter
- Prompts are passed as --initial_prompt to underlying Whisper engines

## Examples and Documentation
- Added example prompt files for common domains:
  - examples/prompts/technical_terms.txt (AI, programming terms)
  - examples/prompts/medical_terms.txt (medical terminology)
  - examples/prompts/business_names.txt (company names)
- Updated README with comprehensive usage examples
- Added demo script showing various use cases

## Testing
- Added unit tests for prompt functionality
- Verified CLI argument parsing
- Tested API parameter integration

## Usage Examples

CLI:
transcribe-anything video.mp4 --initial_prompt "AI, machine learning, neural networks"
transcribe-anything video.mp4 --prompt_file examples/prompts/technical_terms.txt

Python API:
transcribe("video.mp4", initial_prompt="Custom vocabulary terms here")

This enhancement addresses the need for better recognition of:
- Technical terminology
- Proper names and company names
- Domain-specific jargon
- Industry-specific vocabulary

Fixes: Improves transcription accuracy for specialized content

* refactor: Remove example files and demo scripts

Keep only the core functionality for custom prompt support:
- Removed example prompt files
- Removed demo scripts
- Updated README to remove references to example files
- Focused on clean, minimal implementation

* feat: integrate aj47/whisper-mps fork with initial_prompt support

- Update whisper-mps dependency to use aj47/whisper-mps fork
- Add native --initial_prompt support for MPS backend
- Remove filtering of --initial_prompt argument for MPS backend
- Add test for MPS backend with initial_prompt functionality
- Enable custom vocabulary support with Apple Silicon acceleration

Fixes issue where --initial_prompt was not supported by whisper-mps backend.
Now users can use custom vocabulary with fast MPS acceleration.

* Refactor Mac implementation from whisper-mps to lightning-whisper-mlx

- Replace whisper-mps backend with lightning-whisper-mlx for 10x performance improvement
- Add support for multiple languages (remove English-only restriction)
- Add support for custom vocabulary via --initial_prompt parameter
- Add support for both transcribe and translate tasks
- Maintain backward compatibility with run_whisper_mac_english function
- Update environment setup to use lightning-whisper-mlx dependencies
- Enhance argument parsing to support new MLX features
- Update tests to cover new functionality and maintain backward compatibility
- Update documentation to reflect new capabilities and performance improvements
- Bump version to 3.1.0

This addresses the user's request for custom vocabulary support while significantly
improving performance and expanding language support on Apple Silicon Macs.

* Fix MLX refactor: update lightning-whisper-mlx version requirement and improve error handling

- Fix version requirement from >=0.1.0 to >=0.0.10 (latest available version)
- Add better error handling with stderr/stdout capture for debugging
- Add warnings for unsupported features (initial_prompt, word_timestamps, temperature)
- All tests passing and transcription working correctly
- README formatting improvements

* Rename MPS to MLX throughout codebase

- Change Device enum from MPS to MLX in api.py
- Update command-line argument parser to accept 'mlx' as device option
- Maintain backward compatibility by accepting 'mps' as alias for 'mlx'
- Update README.md to use --device mlx instead of --device mps
- Rename test file from test_insanely_fast_whisper_mps.py to test_insanely_fast_whisper_mlx.py
- Update all documentation and examples to use MLX terminology
- All tests passing with new MLX naming convention

* Integrate aj47/lightning-whisper-mlx fork with initial_prompt support

- Updated MLX environment to use aj47/lightning-whisper-mlx fork
- Added initial_prompt parameter support to MLX transcription
- Updated .gitignore to exclude model files and test outputs
- Enables custom vocabulary and context prompting for better transcription accuracy

* move mlx model location

* Fix: Cache NVIDIA detection to prevent repeated torch downloads

- Add caching to has_nvidia_smi() function to ensure consistent results
- Cache based on system fingerprint to handle hardware changes properly
- Add debug logging to track pyproject.toml content changes
- Add --clear-nvidia-cache command line option for cache management
- Prevents uv-iso-env from triggering unnecessary 2.2GB+ torch reinstalls
- Fixes Windows development environment performance issues

Files modified:
- src/transcribe_anything/util.py: Enhanced NVIDIA detection with caching
- src/transcribe_anything/whisper.py: Added debug logging
- src/transcribe_anything/insanley_fast_whisper_reqs.py: Added debug logging
- src/transcribe_anything/whisper_mac.py: Added debug logging
- src/transcribe_anything/_cmd.py: Added clear cache CLI option
- tests/test_nvidia_cache.py: Comprehensive test coverage
- NVIDIA_CACHE_FIX.md: Detailed documentation of the fix

---------

Co-authored-by: Zachary Vorhies <[email protected]>

Author: aj47

.gitignore

@@ -138,4 +138,16 @@ transcribe_anything/WHISPER_OPTIONS.json
 !.aider.conf.yml
 !.aiderignore
 activate
-uv.lock
+uv.lock
+
+# MLX model files (downloaded at runtime)
+mlx_models/
+src/mlx_models/
+
+# Test output files
+tests/localfile/output.json
+tests/localfile/text_video_*
+
+# Video files
+*.mp4
+*.wav

NVIDIA_CACHE_FIX.md

@@ -0,0 +1,111 @@
+# NVIDIA Detection Cache Fix
+
+## Problem
+
+On Windows development environments, tests were taking very long due to repeated downloads of the 2.2GB+ torch package. This was caused by:
+
+1. **Inconsistent NVIDIA Detection**: The `has_nvidia_smi()` function was returning different values between runs
+2. **Dynamic PyProject Generation**: Multiple modules generate different `pyproject.toml` content based on NVIDIA GPU detection
+3. **uv-iso-env Behavior**: The `uv-iso-env` package performs a "nuke and pave" reinstall whenever the `pyproject.toml` fingerprint changes
+4. **Repeated Downloads**: Each fingerprint change triggered a complete reinstall including the large torch download
+
+## Root Cause
+
+The issue was that `has_nvidia_smi()` was being called multiple times during test runs, and on Windows systems, the detection could be inconsistent due to:
+- System state changes
+- Process timing issues
+- Environment variable changes
+- Path resolution inconsistencies
+
+This caused different `pyproject.toml` content to be generated between runs, changing the fingerprint and triggering reinstalls.
+
+## Solution
+
+### 1. NVIDIA Detection Caching
+
+Enhanced `has_nvidia_smi()` in `src/transcribe_anything/util.py` to:
+- Cache detection results based on system fingerprint
+- Store cache in `~/.transcribe_anything_nvidia_cache.json`
+- Use system information (platform, machine, version) + nvidia-smi existence as fingerprint
+- Provide consistent results across runs for the same system configuration
+
+### 2. Debug Logging
+
+Added debug logging to environment generation functions:
+- `src/transcribe_anything/whisper.py`
+- `src/transcribe_anything/insanley_fast_whisper_reqs.py`
+- `src/transcribe_anything/whisper_mac.py`
+
+Each now logs the MD5 hash of generated `pyproject.toml` content to help track changes.
+
+### 3. Cache Management
+
+Added command-line option to clear cache when needed:
+```bash
+transcribe-anything --clear-nvidia-cache
+```
+
+### 4. Testing
+
+Created comprehensive tests in `tests/test_nvidia_cache.py` to verify:
+- Caching behavior works correctly
+- Cache clearing functionality
+- Different system fingerprints are handled properly
+
+## Files Modified
+
+- `src/transcribe_anything/util.py` - Enhanced NVIDIA detection with caching
+- `src/transcribe_anything/whisper.py` - Added debug logging
+- `src/transcribe_anything/insanley_fast_whisper_reqs.py` - Added debug logging  
+- `src/transcribe_anything/whisper_mac.py` - Added debug logging
+- `src/transcribe_anything/_cmd.py` - Added clear cache command-line option
+- `tests/test_nvidia_cache.py` - New test file for cache functionality
+
+## Usage
+
+### Normal Operation
+The caching is automatic and transparent. The first run will detect NVIDIA availability and cache the result. Subsequent runs will use the cached result, ensuring consistent `pyproject.toml` generation.
+
+### Debugging
+If you suspect caching issues, you can:
+
+1. **View debug output**: The system will print debug messages showing:
+   - Cached vs fresh NVIDIA detection results
+   - PyProject.toml content hashes for each module
+
+2. **Clear cache**: If hardware changes or you need to force re-detection:
+   ```bash
+   transcribe-anything --clear-nvidia-cache
+   ```
+
+### Expected Behavior
+- **First run**: Detects NVIDIA, caches result, generates environment
+- **Subsequent runs**: Uses cached result, generates identical environment
+- **No more repeated downloads**: Same fingerprint = no reinstall needed
+
+## Benefits
+
+1. **Faster Testing**: Eliminates repeated 2.2GB+ torch downloads
+2. **Consistent Behavior**: Same system configuration always produces same results
+3. **Debuggable**: Clear logging shows what's happening
+4. **Manageable**: Easy cache clearing when needed
+5. **Backward Compatible**: No changes to existing API or behavior
+
+## Technical Details
+
+The cache file (`~/.transcribe_anything_nvidia_cache.json`) stores mappings from system fingerprints to detection results:
+
+```json
+{
+  "Windows-AMD64-10.0.19041-nvidia_smi:true": true,
+  "Linux-x86_64-5.4.0-nvidia_smi:false": false
+}
+```
+
+The system fingerprint includes:
+- Platform system (Windows, Linux, Darwin)
+- Machine architecture (AMD64, x86_64, arm64)
+- Platform version
+- Whether nvidia-smi executable exists
+
+This ensures that hardware or driver changes are properly detected while maintaining consistency for the same configuration.