Write-ups on my research experiments and engineering projects.
vLLM's reload copy-back only protects registered parameters and buffers, so any other tensor a kernel holds rebinds under captured CUDA graphs on every RL weight sync. The obvious end-to-end test passes on broken code, so everything here rests on pointer identity instead. 88/88 tensors moved in Marlin (PR #48438, nine sites) and again in Machete (PR #48539), and the allocator freed 72/72 capture-time storages in FlashInfer CUTLASS MoE. Five confirmed kernels, a livelock that pinned a GPU at 100% for ten minutes, and the case for detecting the rest with an oracle independent of any repair.
I built Modal's snapshot-restore model from scratch on gVisor, along with the content-addressed image format, blob store, lazy FUSE, scheduler, and Python SDK underneath it. On a real workload, resnet50 over a 661 KB JPEG, cold p50 lands at 774 ms and warm at 125 ms. runsc restore took about 280 ms at every stage of the project, including the stages that measured 5.7 seconds, so every tax I found sat somewhere around restore rather than inside it. The parts worth reading are the ones that didn't work the first time.
A multi-agent pipeline for generating production-shaped and cross-language code where frontier-model priors are weak. Matched-budget GRPO on synthetic tasks shows monotonic +8.9pp learning where the same setup on public OSS oscillates and ends at −1.3pp.
A paired RLVR experiment on Qwen3-4B with algorithm and training budget held constant; only the task domain varied. Designed to test whether RLVR cross-domain transfer is task-agnostic or task-dependent. The two checkpoints produce opposite-sign GPQA Diamond effects, evidence that task choice is a real design lever in a post-training mix.
Boolean circuit minimization (Espresso as verifier) improves GPQA Diamond by +3.5pp, in the same ballpark as NP-Engine's classical-NP results. Monotonic across checkpoints (+1.5 at step 50, +3.5 at step 250), arguing against pure sampling noise.
RNA inverse folding (ViennaRNA as verifier) on a domain with zero pretraining overlap: strong in-domain learning (30% → 52% perfect solve), but cross-domain transfer is −4.1pp on GPQA Diamond, opposite sign at matched algorithm, model, and budget.