✕ Clear all filters
13 articles

📰 Dev.to · Ingero Team

13 articles · Updated every 3 hours · View all reads

All Articles 66,559Blog Posts 99,451Tech Tutorials 16,028Research Papers 13,806News 12,499 ⚡ AI Lessons
From Kernel Scheduler to Python Source Line: Tracing a GPU Stall End to End
Dev.to · Ingero Team 1d ago
From Kernel Scheduler to Python Source Line: Tracing a GPU Stall End to End
TL;DR A GPU that reports 97% utilization can still be the slowest part of a training step,...
Read article → + Playlist
AllReduce Stalls Are Network Stalls. Most Tools See Neither.
Dev.to · Ingero Team 3d ago
AllReduce Stalls Are Network Stalls. Most Tools See Neither.
A slow AllReduce on rank 5 lines up against TCP retransmits on rank 5’s NIC, four ms before the...
Read article → + Playlist
What GitHub Uses eBPF For (and the Layer They Have Not Ported Yet)
Dev.to · Ingero Team 5d ago
What GitHub Uses eBPF For (and the Layer They Have Not Ported Yet)
Three eBPF patterns hyperscalers run in production today, mapped to the equivalent patterns on the...
Read article → + Playlist
GPU Observability for Workloads That Cannot Phone Home
Dev.to · Ingero Team 1w ago
GPU Observability for Workloads That Cannot Phone Home
For an air-gapped GPU host, the trace is only useful if collection, storage, and query all happen...
Read article → + Playlist
One Kernel, Zero Sidecars: Tracing AI Workloads Without an Agent on Every Host
Dev.to · Ingero Team 1w ago
One Kernel, Zero Sidecars: Tracing AI Workloads Without an Agent on Every Host
Per-host overhead multiplied across N hosts, vs. one kernel-level instrumentation per host. The...
Read article → + Playlist
Same eBPF, Different Vendor: Tracing libhip Calls on AMD ROCm
Dev.to · Ingero Team 2w ago
Same eBPF, Different Vendor: Tracing libhip Calls on AMD ROCm
libhip.so is to ROCm what libcudart.so is to CUDA: the user-mode runtime API the framework calls...
Read article → + Playlist
What Inference-Platform Benchmark Posts Leave Out
Dev.to · Ingero Team 2w ago
What Inference-Platform Benchmark Posts Leave Out
DCGM stops at host-level GPU counters. Kernel-side eBPF adds the per-rank, per-tenant signals...
Read article → + Playlist
MCP Shows What the Agent Did. eBPF Shows Why the GPU Stalled.
Dev.to · Ingero Team ☁️ DevOps & Cloud ⚡ AI Lesson 2w ago
MCP Shows What the Agent Did. eBPF Shows Why the GPU Stalled.
MCP exposes the agent’s tool calls. eBPF exposes the kernel events that explain why those tool...
View lesson → + Playlist
MCP Tools Are New API Surfaces. eBPF Sees What They Actually Touch.
Dev.to · Ingero Team 3w ago
MCP Tools Are New API Surfaces. eBPF Sees What They Actually Touch.
An MCP tool call is a tiny line of agent code that fans out to syscalls, library calls, and kernel...
Read article → + Playlist
A Cluster Stall Looks Healthy on Every Host. The Cause Is in the Pattern Across Hosts.
Dev.to · Ingero Team 3w ago
A Cluster Stall Looks Healthy on Every Host. The Cause Is in the Pattern Across Hosts.
Eight ranks on two hosts. Every per-host metric reads healthy. Rank 5 enters the barrier 290ms...
Read article → + Playlist
26 Seconds to Find a Straggler: Fleet v0.10 End-to-End on A100 and GH200
Dev.to · Ingero Team 1mo ago
26 Seconds to Find a Straggler: Fleet v0.10 End-to-End on A100 and GH200
TL;DR Ingero Fleet v0.10 FOSS is live. We validated the full pipeline end-to-end on two...
Read article → + Playlist
124x Slower: What PyTorch DataLoader Actually Does at the Kernel Level
Dev.to · Ingero Team 1mo ago
124x Slower: What PyTorch DataLoader Actually Does at the Kernel Level
TL;DR: PyTorch's DataLoader can be 50-124x slower than direct tensor indexing for in-memory GPU...
Read article → + Playlist
Tracing a 13x PyTorch Slowdown to a Hidden NumPy Synchronization
Dev.to · Ingero Team 2mo ago
Tracing a 13x PyTorch Slowdown to a Hidden NumPy Synchronization
TL;DR: A .cpu().numpy() call buried inside a forward pass was forcing a full CPU-GPU synchronization...
Read article → + Playlist