PyTorch Profiler (Kineto) & HTA GPU Bubble & Overlap Inspector
Stop zooming into millions of tiny boxes in Perfetto. Paste a PyTorch Kineto trace snippet below to measure Pure Compute vs. Exposed NCCL Collectives vs. Idle GPU Bubbles and get the exact FSDP prefetch or CUDA Graph fix.
100% client-side evaluation in your browser. For multi-gigabyte traces, run python3 denpex.py --kineto-trace trace.json on the cluster node.
Step Time & Collective Overlap Verdict
Only 31.3% of your ReduceScatter is hidden behind backward compute; setting backward_prefetch=BACKWARD_PRE will recover ~26.3% step speed.
Recommended Fix: Enable FSDP backward_prefetch=BackwardPrefetch.BACKWARD_PRE (or DDP bucket_cap_mb=50 with gradient_as_bucket_view=True) and remove per-step .item() / torch.cuda.synchronize() host waits.
Correlate GPU Bubbles with Multi-Rank Stragglers & MoE Hot Experts
Stream live step telemetry into the Fleet Training tab or unlock 50 cloud diagnoses/day for 30 days with your work email.
Try Denpex for 30 days. FREE.
No credit card. No automatic subscription. See whether Denpex helps with an incident you already understand.
Start your free 30-day Team evaluation
Up to 50 diagnoses a day for 30 days. Verify your work email to activate an eligible company evaluation, then replay a resolved incident and compare the evidence and next steps with your team's confirmed cause.