deep-learning-compilers

Silicon workbench for chip blueprints, roofline analysis, tile-memory flow, and AI accelerator co-design.

Reproducible study of how tensor layout, reduction length, and pointer alignment drive vendor GEMM dispatch and latency cliffs.

Custom GPT-2 GPU inference engine with fused Triton kernels, KV-cache, and INT8 quantization. Hand-crafted & verified to FP32 correctness.

Advanced static analysis and visual linting tool for Triton GPU IR (TTGIR). Automatically diagnoses shared-memory bank conflicts, global memory coalescing efficiency, register layouts, and SM occupancy limits. Includes an early-pruning autotune hook and an interactive HTML report engine.

Diagnosing non-monotonic activation memory in torch.compile's memory-budget partitioner, with a use-site rematerialization fix.