Merge pull request #1 from susmitsingh01/main

Author: puneethgv

SUMMARY.md

@@ -0,0 +1,35 @@
+# Project Summary
+
+**Agentic Kernel Generation** is an LLM-agent system that writes, verifies,
+and evolves Triton GPU kernels for LLM-inference operators — with a benchmark
+harness designed so the reported speedups can't be faked.
+
+## The idea
+
+An LLM generates a Triton kernel. A sandboxed harness checks it against a
+PyTorch reference on unseen random inputs and profiles it. A reflection agent
+turns the profile into the next optimization, and an evolutionary controller
+breeds the best kernels across generations.
+
+## What makes it different
+
+- **Reward-hack-proof measurement.** Correctness is a hard gate, the kernel
+  never sees the reference output, and every timed iteration uses fresh inputs.
+- **Learns from its own failures.** A knowledge base distilled from 1,000+ real
+  Triton errors feeds verified fixes back into the refine loop.
+
+## Results
+
+Runs on a single 8 GB laptop GPU (RTX 3070, Ampere). Highlights vs PyTorch eager:
+sliding-window attention 174.7×, W4A16 matmul 24.5×, flash attention 11.5× —
+and ahead of PyTorch's own fused kernels (FlexAttention 1.48×, SDPA 1.17×) on
+selected workloads.
+
+## Structure
+
+Two nested loops: an inner loop that refines one candidate serially
+(evaluate → reflect → refine), and an outer evolutionary loop that selects
+parents from a MAP-Elites archive and breeds offspring by mutation or crossover.
+Fitness is speedup, hard-gated to zero unless the kernel is correct.
+
+See the [README](README.md) for full results, methodology, and quickstart.