Algorithms, data and infrastructure: project notebooks, practical guides and technical judgment.https://tech-blog-chi.vercel.app/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-prefill/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-prefill/Paired, profiled measurements of the real DeepSeek-V4.1-Flash: tensor-core sparse attention, a fused indexer and a local-row MoE combine took mini-DeepInfer from 1,085 to 4,316 input tok/s/GPU — and showed where the remaining 8.4× lives.Fri, 23 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/mini-deepinfer-runtime/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-runtime/mini-DeepInfer keeps DeepInfer's engine design — paged state, prefill/decode steps, zero-overhead scheduling and deterministic sampling — and drops the scale, so a DeepSeek-V4.1 request can be followed from the queue to its next token.Fri, 23 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/mini-deepinfer-report-b1/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-report-b1/A single fused kernel replaces the chunked einsum for indexer scoring: 12.2× faster per layer, +31% real 4-GPU prefill throughput, 3.1 GiB less peak memory.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/mini-deepinfer-report-a/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-report-a/Real V4.1 4-GPU prefill goes from 1,085 to 2,572 tok/s/GPU (2.37×) after the sparse-attention kernel moves from CUDA cores to tensor cores.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/mini-deepinfer-report-c1/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-report-c1/Combining only the local expert rows and dequantizing in place cuts routed experts from 526 to 252 ms/step and lifts real 4-GPU prefill to 4,316 tok/s/GPU.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/guides/how-to-read/https://tech-blog-chi.vercel.app/guides/how-to-read/Project-driven learning: a short path in, a working mental model, and the ability to explore independently.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/guides/moe-reading-path/https://tech-blog-chi.vercel.app/guides/moe-reading-path/A reading order through the concepts, minimal experiment, measurement method and source materials.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/guides/paired-benchmarks/https://tech-blog-chi.vercel.app/guides/paired-benchmarks/Define the boundary, validate outputs and retain the individual observations.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/guides/data-leakage/https://tech-blog-chi.vercel.app/guides/data-leakage/A small train/test overlap check with an explicit normalization rule. Python only.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/guides/expert-load-imbalance/https://tech-blog-chi.vercel.app/guides/expert-load-imbalance/Keep the total assignment count fixed and change only its distribution. Python standard library; no GPU required.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/guides/useful-work/https://tech-blog-chi.vercel.app/guides/useful-work/Three distinctions to make before interpreting an optimization result.Thu, 22 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/mini-deepinfer-design/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-design/The engineering boundaries, state invariants and verification rules mini-DeepInfer commits to, written before the implementation.Tue, 20 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/mini-deepinfer-technical-notes/https://tech-blog-chi.vercel.app/writing/mini-deepinfer-technical-notes/A code tour of how mini-DeepInfer turns one request into reproducible V4.1 inference steps: lifecycle, attention state, paged blocks, ZOS, sampling and evidence.Tue, 20 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/notes/stream-priorities/https://tech-blog-chi.vercel.app/notes/stream-priorities/CUDAThu, 15 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/notes/nsight-metrics/https://tech-blog-chi.vercel.app/notes/nsight-metrics/ProfilingSat, 10 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/cuda-graph-fallback/https://tech-blog-chi.vercel.app/writing/cuda-graph-fallback/A working analysis of CUDA Graph coverage, irregular mixture-of-experts workloads, and the cost hidden behind a seemingly simpler execution path.Thu, 01 Oct 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/notes/triton-configurations/https://tech-blog-chi.vercel.app/notes/triton-configurations/TritonMon, 28 Sep 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/notes/visualize-expert-load/https://tech-blog-chi.vercel.app/notes/visualize-expert-load/MoESun, 20 Sep 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/communication-overlap/https://tech-blog-chi.vercel.app/writing/communication-overlap/A closer look at GPU contention, NCCL, and the limits of overlap in expert-parallel inference.Fri, 18 Sep 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/irregular-moe-execution/https://tech-blog-chi.vercel.app/writing/irregular-moe-execution/Analyzing routing, token distribution, and execution behavior without reducing irregular workloads to an average case.Thu, 27 Aug 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/grouped-gemm/https://tech-blog-chi.vercel.app/writing/grouped-gemm/A measurement plan for kernel-level grouped GEMM efficiency under imbalanced workloads.Mon, 10 Aug 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/gpu-memory-long-context/https://tech-blog-chi.vercel.app/writing/gpu-memory-long-context/Questions and instrumentation for understanding memory access patterns and bandwidth utilization in long-context serving.Wed, 15 Jul 2026 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/serving-lessons/https://tech-blog-chi.vercel.app/writing/serving-lessons/A concise index of systems questions to carry from one serving runtime to the next.Fri, 12 Dec 2025 00:00:00 GMThttps://tech-blog-chi.vercel.app/writing/benchmarking-moe/https://tech-blog-chi.vercel.app/writing/benchmarking-moe/A benchmark outline that treats routing distributions and tail behavior as first-class inputs.Wed, 08 Oct 2025 00:00:00 GMT