Build Your Trace
Build Your Trace: Show HN: Reconstruct distributed LLM training traces
Last verified:
What is Build Your Trace?
Build Your Trace is an interactive educational platform for learning distributed training optimization by building and analyzing GPU execution traces. Users reconstruct training workflows by placing kernels and collectives in dependency order, progressing from single-GPU operations through various parallelism strategies (TP, FSDP, EP, CP, MoE).
Build Your Trace pricing
Pricing model: Freemium
Build Your Trace pros
- Progressive challenge-based learning with 15 trace builds, 8 comparison labs, and 7 recognition rounds
- Comprehensive coverage of distributed training concepts—from single-GPU execution to expert parallelism and MoE
- Interactive visual interface showing timing, dependencies, and overlaps in GPU execution traces
- Hands-on practice with kernel scheduling and collective operations (AllGather, ReduceScatter, AllReduce, AllToAll)
Build Your Trace cons
- Appears to be educational/learning tool with no commercial features or professional output
- Limited to browser-based interaction with no API, export, or integration capabilities mentioned
- No apparent community features, peer review, or professional certification
- Content focused on conceptual understanding rather than production profiling or debugging tools
Frequently asked questions about Build Your Trace
What is a training trace?
A timeline visualization of CPU and GPU operations recorded during profiling. Each row represents an execution context (e.g., CUDA stream), and blocks show operation timing and dependencies.
How many challenges are available?
15 trace building challenges organized by concept (MoE, Data Parallel, Tensor Parallel, FSDP, Expert Parallel, Context Parallel), plus 8 comparison labs and 7 recognition rounds.
What parallelism strategies are covered?
Data parallelism (DDP), tensor parallelism (TP), sequence parallelism, fully sharded data parallel (FSDP), expert parallelism (EP), context parallelism (CP), and mixture-of-experts (MoE).