Fill in Blank

Fill in the IVF nlist rule of thumb.

Complete the RMSNorm formula

Complete the pre-norm transformer block recipe

Each decode step loads ___ from HBM, regardless of how many tokens have already been generated

End-to-end latency: fill in the queue, prefill and decode components.

FLOPs vs FLOP/s: fill in the counts versus rate distinction.

Qwen 3.5 SFT data: fill the special tokens that wrap an assistant turn.

Fill in the SFT JSONL: the three required fields of a typical chat example.

Compute the effective batch from per-device batch, accum steps, and GPU count

Fill in the DeepSpeed config keys that select a ZeRO stage

Adam optimizer state per parameter: fill in the moment buffers and bytes

Compute the LoRA trainable param count for a 4096x4096 projection at rank r=8

Fill in the memory delta when a 7B weight tensor moves from fp32 to bf16

Each row of the post-softmax attention weight matrix corresponds to which side of the QK product?

Name the two properties softmax guarantees for every row of the attention weight matrix

Fill the blank: for a single head, QK^T has shape (T_query, ___).

Given Q of shape (B, n_heads, T, d_head), the per-head attention output before concatenation has shape ___.

Modern 7B-class LLMs, fill in typical head counts and the resulting per-head dimension

Fill in the blanks: query expansion / query rewriting in RAG and what it costs.

Fill in typical peak LRs for full FT, LoRA, and DPO

Fill in the full-FT memory contributions for a 7B model

Fill in the blanks: the special tokens that control sequence boundaries and conversation structure in modern LLMs.

Complete the claim: the two-axis decomposition of RAG eval that diagnoses which layer failed.

Complete the definition: prompt engineering is the practice of designing the LLM's ___ to shape its output, without modifying model ___.

Complete the key fields in a tools/list tool object and a tools/call request