Spaces:
Running
Running
Commit History
-
fix(tts): improve performance of send_audio function by avoiding repeated np.concatenate Running
CrazyMonkey0c56327c
-
fix(tts): remove tokenizer_tts from app.state
CrazyMonkey0431435a -
feat(tts): switch TTS model from mms-tts-eng to Kokoro-82M
CrazyMonkey0d6cde26
-
feat(nlp): change nlp model to Qwen/Qwen2.5-1.5B-Instruct
CrazyMonkey04d18a16 -
feat(nlp): change nlp model to microsoft/Phi-3.5-mini-instruct
CrazyMonkey0f854f33
-
feat(asr): add local ASR endpoint using faster-whisper with async lock
CrazyMonkey00ee88ba -
fix(asr): fix loading model
CrazyMonkey00d30253 -
feat(asr): replace Whisper HF with faster-whisper for CPU-friendly transcription
CrazyMonkey0c84acef
-
feat(api): add sending generated audio to external backend
CrazyMonkey020a7446 -
test(tts): checking whether the tts model is working correctly
CrazyMonkey065952f6 -
feat(nlp): switch NLP model to Qwen2.5-0.5B-Instruct
CrazyMonkey0df63d34 -
fix(nlp): remove options n_threads=os.cpu_count() in load_model_nlp
CrazyMonkey00bdf4f1 -
feat(nlp): change version to qwen2.5-1.5b-instruct-q3_k_m.gguf
CrazyMonkey0dc74289 -
test(nlp): comment part of tts in nlp.py
CrazyMonkey0018ff86 -
fix: resolve TypeError in TTS audio generation and optimize model performance
CrazyMonkey0- Fix scipy.io.wavfile.write() incorrect format parameter - Add audio normalization to int16 for WAV format - Optimize llama-cpp parameters (n_ctx=512, n_threads=2, n_batch=256) - Add error handling for TTS failures - Reduce max_tokens from 128 to 100 for faster responsese288dcc -
fix(tts): swapping the model and tokenizer in the return function of load_models_tts
CrazyMonkey04784a54 -
fix: optimize FastAPI + Qwen2.5-1.5B for CPU, reduce max_tokens, increase timeout
CrazyMonkey088284a4 -
feat(nlp): switch Qwen2.5 model to 1.5B GGUF q6_k version
CrazyMonkey05d51d0f -
fix(asr): load audio from in-memory buffer instead of disk
CrazyMonkey0Replaced file-based audio loading with io.BytesIO to handle uploaded audio directly in memory. Librosa/SoundFile reads the buffer, resamples to 16kHz if needed, and feeds Whisper ASR model.9ea2744 -
fix(chat): use llm() directly instead of create_chat_completion
CrazyMonkey03ad9eac -
feat(chat): return NLP response with in-memory TTS audio
CrazyMonkey0- /chat endpoint now generates audio in-memory using TTS and returns it with the text. - Uses multipart/form-data to send both text and audio in a single response. - Removes disk I/O for audio, improving performance.245cf59 -
feat(tts): migrate Kokoro TTS to Hugging Face facebook/mms-tts-eng with in-memory optimization
CrazyMonkey0- Replaced previous Kokoro-based TTS pipeline with Hugging Face AutoModelForTextToWaveform. - Updated to generate WAV in-memory without saving to disk. - Uses tokenizer + model from app.state for FastAPI conventions. - Returns audio as bytes or URL (depending on implementation), improving performance and reducing I/O.2a3f624 -
fix(nlp): Adding chat_handler for handling the Qwen2.5-3B-Instruct-GGUF model in llama-cpp-python
CrazyMonkey05f3ceca
-
fix(nlp):add ',' to fix an error in response generation
CrazyMonkey02d6bfd5 -
perf: implement lazy loading to fix startup timeouts
CrazyMonkey0- Load model on first request instead of startup - Increase token limits and Gunicorn timeout - Add stop token for cleaner responses8f110eb -
fix: resolve model loading and state management issues
CrazyMonkey0- Fix load_model_nlp() to return only model (not tuple) - Update startup_event to assign single model value - Replace direct llm() call with create_chat_completion() - Add proper error handling and logging - Comment out unimplemented model loaders (TTS, ASR, Translation) - Add health check endpoint to verify model loading statusbf1dc5f
-
refactor(chat): migrate from transformers to llama-cpp-python using Qwen 3B
CrazyMonkey06151d5f
-
feat(nlp): optimize NLP model for CPU
CrazyMonkey0d5d8ff1 -
feat(nlp): Optimize CPU usage for Hugging Face Spaces Free Tier
CrazyMonkey075451ba -
feat(nlp): reintroduce Qwen2.5-1.5B-Instruct model and migrate back to Transformers
CrazyMonkey094cf754 -
feat(llama): another attempt to integrate llama-cpp with the Qwen3-8B-Q4_K_M.gguf model
CrazyMonkey09bb78b3
-
test(models): downloading models from transformers
CrazyMonkey0fe8b413 -
fix(nlp): update Llama loading to use from_pretrained()
CrazyMonkey0f7ec4f4 -
Fix(nlp): NLP model download and build fix
CrazyMonkey089865a6 -
feat(nlp): add lama.cpp support for Qwen3-8B-Q5_K_M.gguf and download models
CrazyMonkey0b2565e9 -
Initial APP
CrazyMonkey07eb3110