Pinned Loading

1. roofline-llama roofline-llama Public

   A from-scratch implementation of Llama-3.2-1B in PyTorch, decode-latency benchmarks on three GPUs (T4, L4, A100), three weight-only quantization methods (RTN, GPTQ, AWQ) measured against both, and …

   Jupyter Notebook

2. triton-attention-lab triton-attention-lab Public

   FlashAttention-2 forward and Flash-Decoding kernels written from scratch in Triton, with grouped-query attention (GQA) and causal masking, benchmarked against PyTorch SDPA and a naive eager impleme…

   Jupyter Notebook

3. LlamaDistill-Sentiment LlamaDistill-Sentiment Public

   LlamaDistill-Sentiment (LLM-NEO v2) is a complete end-to-end machine learning system for distilling a large LLM (Meta-Llama-3-8B) into a compact, efficient student model (Llama-3.2-1B)

   Jupyter Notebook

4. MishrTok MishrTok Public

   MishrTok (मिश्र = mixed): 32k BPE tokenizer for code-mixed Hinglish (Roman) + Hindi (Devanagari) — ~14% fewer tokens vs GPT-4o.

   Jupyter Notebook