a/attnc
Playground Pipeline GitHub ↗
Python DSL → optimized IR → fused CUDA

Attention variants,
compiled—not hand-written.

Compose causal masks, sliding windows, GQA, softcaps, and ALiBi. attnc lowers the composition into one kernel and proves its semantics against a NumPy oracle.

Open compiler playground ↓ Read the source ↗
160differential cases
2Python frontends
3tile states
1fused forward kernel
01 / LIVE EXPLORER

Forge a kernel.

This browser demo mirrors attnc's lowering rules. Flip a feature and inspect the compiler artifacts change together.


          

          

          
        


        

          
STATIC TILE PLAN

          
query tiles ↓KV tiles →

          

          
fast pathpredicatedskipped

          
Causal + window bounds become a closed-form key interval. Fully masked tiles issue no memory traffic.

        

      

    


    

      
02 / COMPILER PIPELINE

      
Small compiler. Real compiler.
Both frontends converge on one IR, then pass through conservative optimizations before runtime dispatch.

      

        
01Capture
Immutable combinators or proxy-object tracing record score and mask semantics.
score - slopes[h] * Δ

        
02Optimize
Constant folding, boolean simplification, bound inference, and dead-branch removal.
causal ∧ window(4096)

        
03Classify
Every key tile becomes fully masked, fully unmasked, or partial.
skip · fast · predicate

        
04Emit + JIT
Splice expressions into an online-softmax skeleton, compile with NVRTC, and cache.
hash(IR, shape, arch)

      

    


    

      

        
✓
Correctness is a first-class backend
The same IR runs through an independent NumPy interpreter, enabling differential fuzzing across variants, shapes, and seeds.

        
160 / 160CPU differential cases passing