NVIDIA Blackwell and NVFP4 fundamentally change the economics of LLM inference by dramatically increasing effective HBM capacity and enabling much higher concurrency per GPU. Frameworks like PyTorch and inference engines such as vLLM can now drive unprecedented utilization through continuous batching, long-context decoding, and dense multi-tenancy.

However, as memory bandwidth pressure recedes, a new bottleneck emerges: cross-workload interference.
This talk argues that in the Blackwell era, resource governance becomes as critical as kernel optimization. We explore why traditional best-effort batching is no longer sufficient, and why OpenReg-style concepts—explicit admission control, KV-cache isolation, priority-aware scheduling, and policy-driven backpressure—must become first-class concerns in PyTorch inference stacks.
The session concludes with a concrete roadmap for integrating governance into PyTorch-based inference without sacrificing performance—turning raw GPU efficiency into predictable, SLO-compliant systems