I hate this in rust

Found a good memory diagram

nccl is awesome

Just a small note, The default weights of nn.Linear is uniform distribution, sampled from -sqrt(k) to sqrt(k) where k = 1/n_input_features

Normalisation >> torch.nn.init.kaiming_uniform & torch.nn.init.kaiming_normal

Some Fundamentals: Diff b/w Parallelism & Concurrency: Parallelism: Doing 2 or more tasks at the exact same time. Concurrency: Looks like doing many tasks at same time, but actually doing one task a little, then switching fast to another, then another, and so on.

I think this explain the whole concept of async and await, atleast its do for me.

Simple as that : Serverless are types of servers without the headache of server management.

I think solving euler problems might help better as compare to plain DSA for logical building.

The PJSUA2 SIP library is great, but it's also very difficult to work with. I'm trying to integrate a large language model (LLM) that can have conversations with people over VoIP, but managing SIP is quite challenging.

I think the voice quality in VoIP should be improved. Currently, AI TTS systems that send audio chunks through VoIP calls use 8000 Hz sampling rate and mono channel, which degrades the quality of the AI-generated voice.

I think if someone wants to build a pipeline that converts audio to text and then back from text to audio, while maintaining both quality and speed,they should use the following : Fast Whisper for speech-to-text MeloTTS + OpenVoice V2 for advanced TTS with voice cloning

Why Voip only support mono channel and 8K hrtz voice transmission, on call, i think it should be increased to 44K with stereo channel

Can anybody with Linux knowledge tell me where this train is going?

Just a quick practice in ml: use torch.load(PATH, mmap=True) to speed up PyTorch model loading. setting mmap=True enables memory-mapped files, mapping file regions directly to virtual memory. This avoids slow I/O operations by accessing data via memory addresses