Zhenyu Sha
Notes on ML infrastructure, distributed systems, large-scale training data pipelines, and practical architecture trade-offs.
Posts
-
Why a GPU Profiling Capture Turned into a Distributed Systems Problem
Starting from the ideas behind Meta’s MAIProf, I redesigned capture ownership, process discovery, and trace delivery for Kubernetes.
-
Starting from the GPU Roofline: Tuning the Data Path for Multimodal Video Training
In a previous post, I wrote about design patterns for large-scale multimodal training data pipelines: metadata + blobs, distributed streaming DAGs, and pre-sharded training artifacts.
-
How to Get Tensor Core Metrics from a SageMaker Training Job
W&B will happily tell you your GPU is at “100% utilization.” It will not tell you whether the workload actually hit the tensor-instruction path, whether the SMs were meaningfully occupied, or whether the memory subsystem was the real bottleneck.
-
Is Kubernetes Autoscaling Missing a Capacity Intent Layer?
A friend recently made a simple observation about Kubernetes autoscaling:
-
When We Talk About Workflow, What Are We Actually Talking About?
Airflow, Argo, Kubeflow Pipelines, Flyte, Dify, and Google ADK all call themselves workflow systems. They are not interchangeable. The useful question is not which one is the best workflow engine, but what execution contract each family represents. -
MultiKueue Solves Dispatch, Not Multi-Cloud GPU Placement
MultiKueue is useful Kubernetes-native plumbing for multi-cluster job dispatch, but it should not be over-sold as a complete multi-cloud GPU placement system. -
Design Patterns for Large-Scale Multimodal Training Data Pipelines
I wrote this mostly to clarify my own thinking.
subscribe via RSS