DL Compilers

Compiler-oriented projects around graph capture, IR lowering, fusion, scheduling, generated kernels, and deployment-oriented compilation.

PyTorch Compiler & LLM Inference Performance Analysis Correctness-gated torch.compile suite on RTX 5050 with FX capture, isolated caches, and fresh-process profiling. Replaced tensor-dependent control flow with torch.cond to reduce three FX regions to one full graph. Measured 2.07–2.12× FP32 pointwise speedup and 13.1–16.2% higher latency in a compiled 4.7M-parameter MLP.

PyTorch → ONNX → QNN Static ONNX opset-16 export for MobileCLIP2-S2, FastViT reparameterization, and normalization/SDPA/EOS rewrites for Qualcomm QNN deployment. Validated 56 image and 211 text device outputs with unchanged R@10 of 0.897470.

Further Explorations

MLIR IR design, dialects, lowering passes, and code generation pipelines

XLA Graph compilation, HLO-style transformations, and accelerator execution

TVM Tensor program optimization, scheduling, autotuning, and lowering