Curated tracks instead of a wall of problems. Each is a hand-ordered curriculum — start at the top, finish at the bottom, end up shipping production ML code. Sign in to track progress across tracks.
The math primitives every ML engineer implements at least once: softmax, cross-entropy, normalization, dropout.
Hand-derive backward passes for every layer. Verified against numerical gradients via gradcheck — no autograd shortcuts.
Attention mechanisms, positional encodings, RoPE, multi-head, KV cache, and FlashAttention — the architecture powering modern LLMs.
KV caching, sampling strategies, speculative decoding, prefix caching, paged attention — what runs in vLLM and TGI.
Sizing models on real hardware: parameter counts, KV cache, activation memory, ZeRO, DDP, FSDP.
Real-world bugs planted in real-world code. No hints, no checklists — find what an experienced reviewer would catch.
Tracks are evolving. New problems land in the catalog all the time; tracks get curated additions when they fit the curriculum. Suggestions welcome — email support@torchedup.dev.