What changed
AMD has released detailed information on profiling tools designed to analyze performance bottlenecks in the DeepSeek-V4-Flash training process using Primus and madengine. The tools focus on identifying delays in training steps that may arise from framework operators, GPU kernels, communication, or gaps between these components.
What this means in practice
The profiling tools are part of a controlled series of madengine runs aimed at distinguishing among the sources of slow training steps. The workload used is an eight-layer BF16 training proxy derived from the DeepSeek-V4-Flash Primus configuration. It preserves representative V4 attention and mixture-of-experts operations for profiling but is not the full model or an end-to-end pretraining workload. The setup involves a single node with eight AMD Instinct™ MI350X GPUs and the image rocm/primus:v26.7. The parallel configuration remains fixed unless specified otherwise, with tensor-parallel size 1, pipeline-parallel size 1, and expert-parallel size 8.
Limitations and availability
The profiling tools are available as part of madengine v2.3.0 or later. Users must run the tools on one node with eight AMD Instinct MI350X GPUs and the public image rocm/primus:v26.7. The setup requires sufficient free disk space for copied traces, as a rocprofv3 run writes approximately 150 MB of JSON on each rank. The results apply only to the specified single-node, eight-layer training proxy and should not be interpreted as full-model DeepSeek-V4-Flash performance. These results are not official AMD benchmark results.
