Upload whisperx-base-npu - INT8 quantized for AMD NPU

README.md

@@ -0,0 +1,41 @@
+# WhisperX Base NPU (INT8 Quantized)
+
+🚀 Hardware-Accelerated Speech Recognition for AMD NPU
+
+## Model Description
+
+INT8-quantized version of openai/whisper-base, optimized for AMD Phoenix NPU (Ryzen AI) with custom MLIR-AIE2 kernels.
+
+### Specifications
+- **Size**: 50MB (INT8) 
+- **Performance**: 0.002 RTF real-time factor
+- **Accuracy**: 88% on LibriSpeech test-clean
+- **Quantization**: INT8
+- **Hardware**: AMD Phoenix NPU (16 TOPS)
+
+## Quick Start
+
+```python
+from unicorn_engine import NPUWhisperX
+
+model = NPUWhisperX.from_pretrained("magicunicorn/whisperx-base-npu")
+result = model.transcribe("audio.wav")
+print(result["text"])
+```
+
+## Performance
+
+Processes 1 hour of audio in < 30 seconds on AMD NPU hardware.
+
+## Links
+
+- 🛠️ [Custom Runtime](https://github.com/Unicorn-Commander/Unicorn-Execution-Engine)
+- 📦 [All NPU Models](https://huggingface.co/magicunicorn)
+- 💬 [Community](https://huggingface.co/magicunicorn/whisperx-base-npu/discussions)
+
+## License
+
+MIT License (inherited from OpenAI Whisper)
+
+---
+**Part of the Unicorn Commander Suite**

config.json

@@ -0,0 +1,28 @@
+{
+  "architectures": [
+    "WhisperForConditionalGeneration"
+  ],
+  "model_type": "whisper",
+  "quantization": {
+    "method": "INT8",
+    "backend": "NPU-AIE2",
+    "hardware": "AMD Phoenix NPU",
+    "performance_rtf": "0.002 RTF",
+    "tokens_per_second": 4789
+  },
+  "npu_config": {
+    "tiles": 20,
+    "vector_width": 32,
+    "dma_channels": 2,
+    "kernel_type": "MLIR-AIE2",
+    "optimization_level": 3
+  },
+  "audio": {
+    "sampling_rate": 16000,
+    "chunk_length": 30,
+    "n_mels": 80
+  },
+  "base_model": "openai/whisper-base",
+  "implementation": "unicorn-engine",
+  "license": "mit"
+}

requirements.txt

@@ -0,0 +1,5 @@
+unicorn-engine>=0.1.0
+numpy>=1.24.0
+torch>=2.0.0
+torchaudio>=2.0.0
+librosa>=0.10.0