Glossary¶
A bilingual mirror of the repository's ubiquitous-language file (
CONTEXT.md). Definitions only — no implementation detail, decisions, or budget. Cross-references use →. Terms are kept in English throughout the site, even in Chinese pages, to match interview and source-reading contexts.
Inference Flow¶
- Prefill — the stage that processes the whole input prompt at once, computes its KV, and emits the first output token; usually → compute-bound. → Decode, → TTFT
- Decode — the autoregressive stage that generates subsequent tokens one at a time, computing a single new token per step and appending its KV; usually → memory-bound. → Prefill, → TPOT
- Autoregressive — each new token is generated conditioned on all previous tokens.
- Sampling parameters — temperature, top-p / top-k and related knobs that control token selection at each → Decode step; greedy (temperature=0) vs sampling, affecting output diversity and also batching / throughput.
Memory & Cache¶
- KV cache — cached Key/Value tensors of already-generated tokens, avoiding step-by-step recomputation of historical attention; the central tension behind inference memory footprint and the throughput ceiling. → PagedAttention, → GQA
- HBM / SRAM — the GPU's high-bandwidth memory (HBM) versus on-chip cache/registers (SRAM); their bandwidths differ by an order of magnitude, the premise of IO-aware optimization.
- Memory-bound / Compute-bound — whether the bottleneck is data movement or computation. → Roofline
Architecture¶
- MHA / MQA / GQA — Multi-Head / Multi-Query / Grouped-Query Attention; the number of KV heads decreases in turn, directly shrinking the → KV cache and raising the throughput ceiling.
- FFN / MLP — the feed-forward layers that carry most of the FLOPs and weight memory in a Transformer.
- RoPE — rotary position embedding; its extrapolation properties underpin → long-context inference.
- MoE — Mixture-of-Experts, a sparse structure that activates only some experts per token. → Expert parallelism
Metrics¶
- TTFT — Time To First Token; dominated by → Prefill.
- TPOT / ITL — Time Per Output Token / Inter-Token Latency; dominated by → Decode.
- Throughput — tokens or requests processed per unit time.
- Goodput — effective throughput under → SLO constraints, not raw throughput.
Single-GPU Performance¶
- Roofline / Arithmetic Intensity — a model that judges whether an operator is compute- or bandwidth-limited from its "compute / memory-traffic" ratio.
- FlashAttention — an IO-aware attention algorithm that cuts HBM reads/writes with tiling + online softmax.
- CUDA graphs — record a sequence of kernel launches into a graph and replay it, amortizing launch overhead in the → Decode stage.
- Kernel fusion — merge multiple operators into one kernel to reduce memory traffic and launch overhead.
GPU Programming¶
- SM / Warp / Occupancy — Streaming Multiprocessor / a 32-thread scheduling unit / occupancy.
- Coalescing / Shared memory / Bank conflict — memory-access coalescing / on-chip shared memory / bank conflicts.
- Triton — a Python-based GPU kernel language.
Quantization¶
- PTQ / QAT — Post-Training Quantization / Quantization-Aware Training.
- Weight-only vs Weight+Activation — quantizing only weights vs quantizing both weights and activations.
- Per-tensor / per-channel / per-group — quantization granularity.
- GPTQ / AWQ / SmoothQuant / FP8 / LLM.int8() — mainstream quantization method families.
- KV-cache quantization — quantizing the → KV cache itself to save memory.
Serving & Throughput¶
- Static / Dynamic / Continuous batching — continuous batching (Orca-style) admits requests as they arrive and evicts them as they finish; the key lever for inference throughput.
- PagedAttention — manages the → KV cache in blocks, allocated like virtual-memory paging, eliminating fragmentation and raising utilization.
- Chunked prefill — split a long prefill into chunks and interleave it with → Decode scheduling, balancing TTFT and throughput.
- PD disaggregation — split → Prefill and → Decode onto different resources to optimize each separately.
- Prefix caching — reuse the KV of a shared prefix, skipping repeated prefill.
- Speculative decoding — a small draft model guesses several tokens, the large model verifies them in one pass, speeding up → Decode.
- Preemption (recompute vs swap) — when the → KV cache pool is exhausted, the scheduler pauses some sequences and reclaims their KV, choosing on resume to recompute or swap (out/in); the mechanism that preserves fairness and throughput under concurrency. → PagedAttention
Advanced Topics¶
- LoRA / Multi-LoRA serving — low-rank adapters; a serving form with one base + multiple adapters swapped dynamically.
- Guided / Structured decoding — constrain output with JSON / regex / grammar.
- Long-context inference — → RoPE extrapolation, attention sink, KV compression, and the memory/scheduling problems of long sequences.
Distributed¶
- Tensor / Pipeline / Data / Expert Parallelism — the four parallelism axes. → MoE
- Collective communication — all-reduce / all-gather / reduce-scatter and other primitives (NCCL).
- TP degree — the number of GPUs a tensor-parallel split spans.
Production¶
- SLO — Service Level Objective; drives → Goodput and tuning.
- Knee — the point where the throughput-vs-concurrency curve starts to degrade; the key spot load tests hunt for.
- KV-cache aware routing — route requests by which instance already caches their prefix, improving prefix-cache hit rate.