Pinned Loading

1. Agentic-Kernel-Generation Agentic-Kernel-Generation Public

   LLM agents that generate, verify, and evolve Triton GPU kernels. Includes a reward-hack-resistant benchmarking harness with strict correctness verification and fresh-input evaluation. Achieves up t…

   Python 2 1

2. Parallel-Eagle Parallel-Eagle Public

   A lossless speculative decoder: a parallel multi-token drafter + dynamic tree verification.

   Python

3. DiSpec DiSpec Public

   From-scratch LLM inference engine for Qwen2.5/Qwen3: paged attention with prefix caching, continuous batching, a Triton decode kernel, CUDA-graph decode, FlashAttention prefill, speculative decodin…

   Python

4. kvxfer kvxfer Public

   Reuse a small model's prefill in a larger same-family model by mapping its KV cache instead of re-prefilling. Closed-form per-layer maps plus an optional trained residual — 4.6–9.1× faster than vLL…

   Python

5. Schema-Linking-with-SFT Schema-Linking-with-SFT Public

   LoRA fine-tuned Qwen2.5-Coder-1.5B for schema linking in NL-to-SQL: given a question and a database schema, predicts the referenced tables and columns as JSON. Uses self-consistency decoding (k=10)…

   Python

6. rag-context-engineering rag-context-engineering Public

   Retrieval-augmented generation for documentation QA under a 2,000-token context budget — multi-resolution chunking, hybrid BM25+FAISS retrieval, cross-encoder reranking, and grounded generation, w…

   Python