Skip to content
← Quantization & Numerics / 19
expertNewOpenAIAnthropicxAI

Moving rollouts and trainer from BF16 to FP16 shrinks their logprob gap about 8x; deterministic mode does nothing. Why?

One change of number format beats a whole deterministic-inference mode at making the engine agree with the trainer. The answer separates three properties that are easy to conflate, shows with a runnable model where the 8x comes from and when it does not appear, and prices what FP16 gives up.

Updated Nov 2026 · Worked answers on inference mechanisms, with calculations, implementation details, and follow-up questions.

One change of number format beats a whole deterministic-inference mode at making the engine agree with the trainer. The answer separates three properties that are easy to conflate, shows with a runnable model where the 8x comes from and when it does not appear, and prices what FP16 gives up.

10 answers per topic instead of 5, plus saved progress and bookmarks · costs nothingor unlock all 213 remaining answers · $25

Every worked answer and derivation in all 12 topics, every concept page, the full practice-test bank, and every code drill’s edge-case tests and reference solution.

6 months · One payment · No auto-renewal

UP NEXT ON YOUR JOURNEY
Popular right nowWhy is batch-1 decode on a 70B model bandwidth-bound, and what is its latency floor on two H100s?Popular right nowDecode attention intensity is fixed at the query-to-KV head ratio. Derive that, then tell me what MLA does to it.Popular right nowAccount for every gigabyte on a GPU running vLLM: weights, KV pool, activations, CUDA graphs and NCCL buffers.
DISCUSSION · 0

Nobody has commented on this answer. Add a correction or another way to work it.