A practical model for upload-side and download-side streaming: transport moves bytes, while the application layer defines boundaries, progress, recovery, idempotency, backpressure, and business meaning.

A source-reading walkthrough of the vLLM V1 request path: OpenAI-compatible HTTP entrypoint, serving render, AsyncLLM, EngineCore client, Tensor IPC, scheduler, and one GPUModelRunner forward pass.

Why LLM inference splits into a compute-bound prefill phase and a memory-bandwidth-bound decode phase, and how that explains TTFT, TPOT, batching, KV cache pressure, and serving-engine design.

A series index for LLM attention kernels and GPU primitives: fused softmax, online softmax, FlashAttention, PagedAttention kernels, Triton/CUDA, and memory-access optimization.

A series index for LLM quantization and low-precision serving: INT8/INT4, GPTQ, AWQ, SmoothQuant, NF4, AQLM, KV cache quantization, FP8 serving, and quality/speed/memory tradeoffs.

An LLM inference experiment series index: vLLM/SGLang benchmarks, TTFT/TPOT, prefix cache, chunked prefill, PagedAttention, quantization, and a profiler dashboard.

A vLLM / SGLang source-reading series index: request lifecycle, scheduler, KV cache allocation, block manager, radix cache, and benchmarks.

A series index for core LLM serving mechanisms: prefill/decode, KV cache, PagedAttention, continuous batching, prefix caching, and disaggregated prefill.

A practical survey of LLM quantization, covering linear quantization, codebook quantization, LLM.int8(), SmoothQuant, GPTQ, AWQ, NF4, AQLM, KV cache quantization, and FP8.

A step-by-step explanation of positional encoding in Transformers, from absolute embeddings to sinusoidal encodings, Euler's formula, and rotary position embeddings.

A back-of-the-envelope framework for estimating LLM training FLOPs, inference FLOPs, weight memory, KV cache, and training memory.

A practical note on managing agent skills: how to install, create, remove, disable, update, and evolve skills, plus which meta-skills are worth installing first.

Routing prefill and decode to separate GPU pools eliminates interference entirely, enabling independent scaling and optimal latency — at the cost of KV cache migration across machines.

When thousands of requests share the same system prompt, recomputing its KV cache each time is pure waste. Prefix caching stores and reuses those vectors, cutting TTFT by up to 97% in common deployments.

Splitting a long prefill across multiple iterations keeps decode requests from stalling, with no extra FLOPs and negligible IO overhead.

How iteration-level scheduling eliminates GPU idle time, and how prefill and decode rows can share one packed forward pass.

How vLLM borrows the OS paging idea to eliminate KV cache memory fragmentation, pushing GPU utilization from ~30% to ~96%.

how online softmax extends the fused kernel to handle rows that exceed sram capacity, using a numerically stable 2-pass tiling algorithm.

why the key-value cache avoids redundant computation in autoregressive decoding, and the memory/compute tradeoffs it introduces.

how to write a fused softmax kernel in triton that eliminates redundant memory accesses and outperforms pytorch's native implementation.