PassNet provides a dataset of 18K graphs and PassBench for LLM-generated compiler passes, with fine-tuned models achieving 2.67x gains on long-tail tasks where TorchInductor underperforms.
STARK: Strategic team of agents for refining kernels
8 Pith papers cite this work. Polarity classification is still indexing.
8
Pith papers citing it
years
2026 8representative citing papers
Step-TP is a dataset providing grounded, atomic step-level IR transitions and CoT supervision to enable reliable multi-step LLM-guided tensor program optimization instead of end-to-end imitation.
KernelBenchX benchmark shows task category explains nearly three times more variance in LLM kernel correctness than method choice, iterative refinement boosts correctness but reduces performance, and quantization remains unsolved.
Hawk raises NPU kernel generation accuracy from 49.4% to 80% and yields up to 2.2× speedups by retrieving and distilling structured hardware-aware knowledge without any model training.
Boundary-only shape sampling is the safe default for tensor-kernel testing (78% recall, 0% false positives); regular 'representative' shapes miss all softmax tail-mask bugs.
SpecGen introduces speculative generation to fork non-reasoning kernel candidates during LLM reasoning traces, enabling early termination and parallel profiling to reduce end-to-end optimization time on H200 GPUs.
Fixed-shape allclose checks in LLM kernel benchmarks can certify seeded buggy Triton kernels as correct; a schema-aware fp64 fuzzer catches all seeded bugs on five GPUs and passes all correct controls.
AscendOptimizer combines kernel rewinding for reusable experience with evolutionary search on hardware feedback to optimize Ascend NPU operators, delivering 1.21x geometric-mean speedup and faster performance on 53.47% of 101 tested operators versus baseline.
citing papers explorer
-
PassNet: Scaling Large Language Models for Graph Compiler Pass Generation
PassNet provides a dataset of 18K graphs and PassBench for LLM-generated compiler passes, with fine-tuned models achieving 2.67x gains on long-tail tasks where TorchInductor underperforms.
-
Step-TP: A Grounded, Step-Level Dataset with Chain-of-Thought Reasoning for LLM-Guided Tensor Program Optimization
Step-TP is a dataset providing grounded, atomic step-level IR transitions and CoT supervision to enable reliable multi-step LLM-guided tensor program optimization instead of end-to-end imitation.
-
KernelBenchX: A Comprehensive Benchmark for Evaluating LLM-Generated GPU Kernels
KernelBenchX benchmark shows task category explains nearly three times more variance in LLM kernel correctness than method choice, iterative refinement boosts correctness but reduces performance, and quantization remains unsolved.
-
Hawk: Harnessing Hardware-Aware Knowledge for High-Performance NPU Kernel Generation
Hawk raises NPU kernel generation accuracy from 49.4% to 80% and yields up to 2.2× speedups by retrieving and distilling structured hardware-aware knowledge without any model training.
-
Test-Input Generation for Tensor Programs: What Actually Finds Kernel Bugs
Boundary-only shape sampling is the safe default for tensor-kernel testing (78% recall, 0% false positives); regular 'representative' shapes miss all softmax tail-mask bugs.
-
SpecGen: Accelerating Agentic Kernel Optimization with Speculative Generation
SpecGen introduces speculative generation to fork non-reasoning kernel candidates during LLM reasoning traces, enabling early termination and parallel profiling to reduce end-to-end optimization time on H200 GPUs.
-
The Correctness Illusion in LLM-Generated GPU Kernels
Fixed-shape allclose checks in LLM kernel benchmarks can certify seeded buggy Triton kernels as correct; a schema-aware fp64 fuzzer catches all seeded bugs on five GPUs and passes all correct controls.
-
AscendOptimizer: Episodic Agent for Ascend NPU Operator Optimization
AscendOptimizer combines kernel rewinding for reusable experience with evolutionary search on hardware feedback to optimize Ascend NPU operators, delivering 1.21x geometric-mean speedup and faster performance on 53.47% of 101 tested operators versus baseline.