Knowledge note

Wiki Log

Chronological log of knowledge-base updates and documentation maintenance.

Updated October 10, 2026Confidence: High

Wiki Log

2026-05-29

  • Added Knowledge Base Introduction to docs/README.md as a meta page sourced from AGENTS.md. Describes the repo structure, category organization, navigation workflow, and the docs lint command.
  • Added τ²-Bench: Mechanism and Design to docs/benchmarks/agent-eval/tau2-bench-mechanism.md, sourced from raw/benchmarks/agent-eval/tau2-bench-mechanism-and-design--paper.pdf. Covers Dec-POMDP formalism, dual-control architecture, five-stage domain construction, programmatic task generation, user simulator design, ablation modes, and empirical results.

2026-05-30

  • Added τ-Voice: Full-Duplex Voice Agent Benchmark to docs/benchmarks/agent-eval/tau-voice.md, sourced from the τ-Voice preprint (arXiv:2603.13686v1, Mar 2026). Covers tick-based orchestrator design, voice user simulator pipeline (7 personas, audio environment, turn-taking policy), three domains (278 total tasks), three audio-native providers, Clean vs Realistic conditions, acoustic ablation results (accents most damaging at -10pp avg), voice interaction quality metrics, qualitative error analysis (79-90% agent errors), and key conclusions.

  • Added τ-bench: Tool-Agent-User Interaction Benchmark to docs/benchmarks/agent-eval/tau-bench.md, sourced from raw/benchmarks/agent-eval/tau-bench-tool-agent-user-interaction-benchmark--paper.pdf (arXiv:2406.12045v1). Covers benchmark architecture, two domains (τ-retail 115 tasks, τ-airline 50 tasks), three-stage construction, pass^k metric definition, full empirical results table for 12 models, method comparison, domain policy ablation, and failure analysis.

2026-06-02

  • Migrated GitHub-only agent guidance into AGENTS.md so Codex can read repository rules directly, including docs front matter, category handling, logs maintenance, and rendering constraints.
  • Copied the docs ingest agent and instruction files from .github/ into .codex/agents/ and .codex/instructions/ so the repository has a Codex-local agent setup.
  • Copied the PDF skill assets from .github/skills/pdf/ into .codex/skills/pdf/ as a Codex-facing repo-local skill mirror.

2026-06-11

  • Added NVFP4: Blackwell 4-Bit Floating Point to docs/hardware/numerics/quantization/nvfp4.md, sourced from NVIDIA’s NVFP4 inference blog and Transformer Engine NVFP4 documentation captured in raw/hardware/nvfp4-blackwell-4-bit-floating-point-references.md. Covers E2M1 encoding, FP8 E4M3 block scaling, FP32 tensor scaling, memory reduction claims, Transformer Engine recipe defaults, stochastic rounding, Random Hadamard Transform, GEMM layout, distributed-training behavior, and supported hardware.

2026-06-15

  • Added FlashAttention: IO-Aware Exact Attention to docs/algorithms/attention/flashattention/flashattention.md, sourced from raw/algorithms/attention/flashattention/flashattention-io-aware-exact-attention--arxiv-2205.14135v2.pdf. Covers the original tiled exact attention algorithm, online softmax statistics, recomputation in backward, IO complexity, block-sparse FlashAttention, empirical training speedups, long-context results, limitations, and relationship to later FlashAttention versions.
  • Added FlashAttention-2: Better Parallelism and Work Partitioning to docs/algorithms/attention/flashattention/flashattention-2.md, sourced from raw/algorithms/attention/flashattention/flashattention-2-better-parallelism-work-partitioning--arxiv-2307.08691v1.pdf. Covers reduced non-matmul overhead, logsumexp softmax bookkeeping, sequence-level thread-block parallelism, backward atomic accumulation, warp-level work partitioning, causal block skipping, A100/H100 attention-kernel results, and GPT-style training throughput.
  • Added FlashAttention-3: Hopper Asynchrony and FP8 Attention to docs/algorithms/attention/flashattention/flashattention-3.md, sourced from raw/algorithms/attention/flashattention/flashattention-3-asynchrony-low-precision--arxiv-2407.08608v2.pdf. Covers Hopper-specific warp specialization, TMA/WGMMA producer-consumer scheduling, ping-pong and two-stage GEMM-softmax overlap, FP8 layout handling, block quantization, incoherent processing, backward-pass acceleration, empirical speedups, and limitations.
  • Added SGLang: Structured Language Model Programs to docs/frameworks/serving/sglang/index.md, sourced from raw/frameworks/serving/sglang/sglang-structured-language-model-programs--arxiv-2312.07104v2.pdf. Covers the SGLang frontend primitives, interpreter/compiler execution modes, RadixAttention KV cache reuse, cache-aware scheduling, compressed finite-state-machine decoding, API speculative execution, evaluation setup, throughput/latency results, production deployment observations, and limitations.
  • Added vLLM: PagedAttention Serving Framework to docs/frameworks/serving/vllm/vllm-framework.md, sourced from raw/frameworks/serving/vllm/vllm-pagedattention-serving-framework--arxiv-2309.06180v1.pdf. Covers the PagedAttention algorithm, KV-cache block tables, virtual-memory analogy, copy-on-write sharing for parallel sampling and beam search, shared-prefix reuse, scheduling/preemption, distributed execution, implementation details, evaluation setup, throughput results, ablations, and limitations.
  • Added FlashAttention-4: Blackwell Attention Kernel Co-Design to docs/algorithms/attention/flashattention/flashattention-4.md, sourced from raw/algorithms/attention/flashattention/flashattention-4-kernel-pipelining-co-design--arxiv-2603.05451v1.pdf. Covers FA4’s Blackwell hardware motivation, forward pipeline, software-emulated exponentials, conditional online-softmax rescaling, backward TMEM pipeline, 2-CTA MMA mode, deterministic backward pass, LPT scheduling, CuTe-DSL implementation, benchmark results, and the source’s B200/B100 hardware-name inconsistency.
  • Refactored Wiki Index categories, grouping benchmark pages, inference systems/frameworks, attention/kernel algorithms, and hardware/numerics.
  • Refactored docs/ from the previous layer folders into category folders: docs/benchmarks/, docs/frameworks/, docs/algorithms/, and docs/hardware/; added category index pages, removed legacy classification front matter, and updated Jekyll navigation/configuration.

2026-06-16

  • Added FlatQuant: Fast Learnable Affine Quantization to docs/hardware/numerics/quantization/flatquant/index.md, sourced from raw/hardware/numerics/quantization/flatquant-fast-learnable-affine-quantization--arxiv-2410.09426v4.pdf. Covers FlatQuant’s flatness motivation, learnable affine transformations, Kronecker factorization, per-channel scaling, learnable clipping thresholds, Transformer integration, fused Triton kernel design, W4A4 accuracy results, latency speedups, ablations, and limitations.

2026-06-24

  • Added Matrix Exponentiation for Linear Transitions at the time under the legacy path docs/layer_0/matrix-exponentiation.md, sourced from raw/algorithms/attention/linear/matrix-exponentiation-hackerearth--web.pdf. Covers binary matrix exponentiation, row-vector transition-matrix construction, Fibonacci and general linear recurrences, prefix sums, coupled sequences, fixed linear dynamic-programming transitions, modular arithmetic, and repeated-query optimization in layer 0.

2026-07-06

  • Added DeepSWE: Long-Horizon Software Engineering Benchmark to docs/benchmarks/agent-eval/deepswe/index.md, sourced from raw/benchmarks/agent-eval/deepswe-long-horizon-software-engineering-benchmark.md. Covers benchmark motivation against contamination and verifier mismatch, corpus scope (113 tasks, 91 repositories, 5 languages), behavioral-verifier methodology, shared mini-swe-agent harness, the publication leaderboard snapshot dated 2026-05-26, qualitative failure patterns across model families, and stated limitations.
  • Added DeepSWE v1.1: Execution and Scoring Changes to docs/benchmarks/agent-eval/deepswe-v1-1/index.md, sourced from raw/benchmarks/agent-eval/deepswe-v1.1-execution-and-scoring-changes.md. Covers the v1.1 execution-and-grading update for the same 113 tasks: committed-diff isolated verification, CTRF per-test reporting, the cleaner main-branch git environment, updated July 1, 2026 leaderboard snapshot, removal of wall-clock reporting, and the reported similarity between v1 and v1.1 results.
  • Added vLLM Code Learning Path and Request Flow to docs/frameworks/serving/vllm/vllm-code-learning-path.md, sourced from the current vllm codebase entrypoint, engine, scheduler, KV-cache, executor, and worker files. Covers the current request path from /v1/chat/completions through AsyncLLM, EngineCore, Scheduler, KVCacheManager, Executor, GPUWorker, GPUModelRunner, output processing, and an achievement-driven staged plan to build a mini vLLM.
  • Updated Frameworks and Wiki Index to include the new code-oriented vLLM learning page alongside the existing paper-oriented framework notes.

2026-07-09

  • Added Pier: Coding-Agent Evaluation Harness to docs/benchmarks/agent-eval/pier/index.md, sourced from raw/benchmarks/agent-eval/pier-coding-agent-evaluation-harness.md. Covers why Pier exists as a smaller Harbor-compatible evaluation harness, its job/trial/agent/environment architecture, filtered-network installed-agent support, stricter ATIF trajectory conversion, the mini-swe-agent adapter path, the constraints of the current local-dataset-only build, and a DeepSWE-style execution recipe plus visual explainer assets.
  • Updated Benchmarks and Wiki Index to include the new Pier evaluation-harness page.
  • Updated DeepSWE: Long-Horizon Software Engineering Benchmark, sourced from raw/benchmarks/agent-eval/deepswe-long-horizon-software-engineering-benchmark.md, with a richer visual explainer section and added editable/source assets at docs/benchmarks/assets/deepswe-explainer.drawio and docs/benchmarks/assets/deepswe-explainer.svg to show why DeepSWE was introduced and how its benchmark pipeline works.
  • Updated Wiki Index to note the new DeepSWE visual explainer on the benchmark page.
  • Updated DeepSWE v1.1: Execution and Scoring Changes, sourced from raw/benchmarks/agent-eval/deepswe-v1.1-execution-and-scoring-changes.md, with a visual explainer section and added editable/source assets at docs/benchmarks/assets/deepswe-v1-1-explainer.drawio and docs/benchmarks/assets/deepswe-v1-1-explainer.svg to show the committed-diff verification flow, CTRF reporting, and why v1.1 is a measurement cleanup rather than a new benchmark corpus.
  • Updated Wiki Index to note the new DeepSWE v1.1 visual explainer on the benchmark page.
  • Added Harbor: Agent Evaluation Framework Design to docs/frameworks/infrastructure/harbor/index.md, sourced from the inspected Harbor README, docs, and runtime code copied under raw/frameworks/harbor/. Covers why Harbor exists, its task-centric design philosophy, Job -> Trial -> Agent/Environment/Verifier architecture, registry and distribution model, policy-driven sandboxing, verifier isolation, and the editable draw.io asset at docs/frameworks/assets/harbor-architecture.drawio.
  • Updated Frameworks and Wiki Index to include the new Harbor framework page.
  • Updated Harbor: Agent Evaluation Framework Design with a second, philosophy-first visual explainer and added complementary assets at docs/frameworks/assets/harbor-design-philosophy.drawio and docs/frameworks/assets/harbor-design-philosophy.svg to show Harbor’s motivating pressures, core principles, architecture consequences, and intended outcomes without replacing the existing runtime architecture visual.
  • Refactored Pier: Coding-Agent Evaluation Harness, DeepSWE: Long-Horizon Software Engineering Benchmark, DeepSWE v1.1: Execution and Scoring Changes, and Harbor: Agent Evaluation Framework Design into folder-backed index.md pages so VS Code Markdown preview can resolve relative image paths while GitHub Pages keeps pretty trailing-slash URLs.
  • Removed the repo-local Codex mirror files .codex/AGENTS.md and .codex/agents/docs-ingest.toml, and updated Knowledge Base Introduction to point back to AGENTS.md for the docs workflow source of truth.
  • Moved Matrix Exponentiation for Linear Transitions from the legacy docs/layer_0/ path into docs/algorithms/, updated Algorithms and Wiki Index to match the category-based layout, and revised AGENTS.md plus the .github docs-ingest/front-matter instructions so doc_layer now consistently represents confidence rather than folder placement.
  • Simplified the docs front matter schema to use confidence: high|medium|low directly, removed the redundant doc_layer field from pages and instructions, and updated the docs lint script so every page under docs/ now requires an explicit readable confidence value.

2026-07-14

  • Added Intrinsic Dimensionality and Language Model Fine-Tuning to docs/training/adaptation/fine-tuning/intrinsic-dimensionality-fine-tuning/index.md, sourced from raw/training/adaptation/fine-tuning/intrinsic-dimensionality-language-model-fine-tuning--arxiv-2012.13255v1.pdf. Covers DID/SAID intrinsic-dimension fine-tuning, MRPC/QQP d90 findings, pretraining as task compression, model-size trends, generalization correlations, the intrinsic-dimension compression bound, and an editable Draw.io explainer at docs/training/intrinsic-dimensionality-fine-tuning/intrinsic-dimensionality-fine-tuning.drawio.
  • Added the new Training category and updated Wiki Index, Knowledge Base Introduction, and the site landing page to include training and fine-tuning pages.

2026-07-15

  • Added DSpark: Confidence-Scheduled Speculative Decoding to docs/frameworks/speculative-decoding/dspark/index.md, sourced from raw/frameworks/speculative-decoding/dspark/dspark-confidence-scheduled-speculative-decoding--arxiv-2607.05147v1.pdf. Covers DSpark’s semi-autoregressive drafter, Markov/RNN sequential heads, calibrated confidence head, hardware-aware prefix scheduler, offline accepted-length results, DeepSeek-V4 production deployment, live traffic throughput/interactivity results, limitations, and an editable Draw.io asset at docs/frameworks/speculative-decoding/dspark/dspark-decoding-cycle.drawio.

  • Updated Frameworks and Wiki Index to include the new DSpark speculative decoding page.

  • Restyled all docs pages to comply with the updated docs-content-structure.instructions.md: converted ## Summary sections to ## TL;DR with three-sentence (What/How/The number) format, added ## The Core Idea, ## Why This Exists, ## Where It Breaks (failure mode table), ## One Thing to Remember (bold key phrase paragraph), and ## Go Deeper sections across all paper-insight pages. Updated all front matter updated dates to 2026-07-15. DSpark page already conformed; nvfp4 and vllm-code-learning-path received structural additions appropriate for their format-reference and learning-guide roles.

2026-07-17

  • Added EvalScope Perf: LLM Inference Stress Testing to docs/benchmarks/serving-performance/evalscope-perf.md, sourced from the EvalScope documentation at evalscope.readthedocs.io. Covers the full stress-testing tool architecture: closed-loop vs. open-loop load generation, SLA binary-search auto-tuning with AND/OR constraint logic, multi-turn conversation benchmarking with cache-hit metrics, dataset flexibility (random, ShareGPT, OpenQA, embedding, rerank, multi-modal), comprehensive metrics coverage (TTFT/TPOT/ITL/latency with full percentile breakdowns), parameter-aligned vLLM bench comparison showing statistical consistency, warmup support, WandB/SwanLab/ClearML visualization, and documented failure modes.
  • Added AISBench Benchmark vs. EvalScope Perf, sourced from both projects’ current documentation and repositories. Compares positioning, load semantics, traffic and trace workloads, metrics, extensibility, user experience, result operations, and risks; records an AISBench SWOT, prioritized roadmap, and reproducible head-to-head experiment design at medium confidence.
  • Added an editable Draw.io competitive big-picture diagram to AISBench Benchmark vs. EvalScope Perf, showing the shared baseline, each product’s differentiated strengths, the central product gap, and AISBench’s recommended winning path.
  • Added AISBench First-Class Performance Roadmap, a presentation-ready decision document defining metric, load-model, SLA-capacity, evidence-store, and one-command UX capabilities as the five first-class priorities. Includes a detailed EvalScope feature explanation, implementation contracts, acceptance criteria, delivery phases, and an editable Draw.io roadmap.
  • Added the complete Chinese presentation edition, AISBench First-Class 性能能力路线图, preserving First-Class as the product and architecture term while retaining the five priorities, detailed EvalScope analysis, implementation contracts, acceptance criteria, architecture, delivery phases, and reciprocal language navigation.

2026-07-23

  • Added AutoJudger: Agent-Driven Efficient MLLM Benchmarking to docs/benchmarks/agent-eval/autojudger.md, sourced from raw/benchmarks/agent-eval/autojudger-agent-driven-mllm-benchmarking--arxiv-2505.21389v1.pdf (arXiv:2505.21389, May 2025). Restructured to full paper-insight template: added Landscape evolutionary tree, Deep Dive subsections with mini-template (What/Why/How/Intuition/Example/Remember), Putting It Together end-to-end trace, What This Buys You with narrative results, Where It Breaks failure modes table, One Thing to Remember, and Go Deeper. Covers IRT Rasch difficulty estimation, real-time ability tracking, max-min semantic retrieval, agent-driven question selection (Qwen2.5-VL-7B), dynamic category-aware memory, and results: 92.06% MMT-Bench ranking accuracy at 5% compression.

  • Added Socratic-SWE: Self-Evolving Coding Agents via Trace-Derived Skills to docs/training/adaptation/fine-tuning/socratic-swe/index.md, sourced from raw/training/adaptation/fine-tuning/socratic-swe-self-evolving-coding-agents--arxiv-2606.07412v1.pdf (arXiv:2606.07412v1, Jun 2026). Covers the closed-loop self-evolution framework: Agent Skill Registry (trace collection → skill extraction → deduplication), skill-guided task Generator with four-stage Verifier Gate (format/grounding/execution/semantics), gradient-aligned Generator reward ($R_G = \cos(g_\tau, G_v)$), GDPO-normalized three-component Solver reward, role-specific GRPO+GDPO training with shared weights, evolutionary landscape from static pipelines through self-play to skill-guided methods, benchmark results (50.40% SWE-bench Verified, 52.33% Lite, 47.20% Pro, 35.60% Terminal-Bench 2.0), Generator reward ablation, and an editable Draw.io big-picture diagram.

  • Expanded Megatron-LM: GPU-Cluster Training Parallelism with a new Deep Dive subsection QKV Column-Parallel Splitting in Detail. Covers the fused $d \times 3d$ QKV weight matrix, head-interleaved weight layout (why naive column split would break Q/K/V pairing), forward pass step-by-step walkthrough, and inference-time behavior (prefill vs. decode, KV cache distribution as the key decode win).

2026-08-23

  • Re-insighted FlashAttention-3 and FlashAttention-4 from their complete derived paper extractions. Reordered both pages into the required paper-insight structure, added source figures and editable landscape diagrams under docs/algorithms/attention/flashattention/assets/, sharpened worked runtime traces and symbol maps, reconciled benchmark conditions and failure modes, and recorded the FlashAttention-4 B200/B100 hardware-name caveat.

2026-07-24

  • Converted the four FlashAttention PDFs, now organized under raw/algorithms/, to markdown via MinerU precise mode, saving results to derived/pdf-markdown/algorithms/.
  • Converted raw/algorithms/foundations/softmax-function-properties-motivation-interpretation--franke-degen.pdf to markdown via MinerU precise mode, saving result to derived/pdf-markdown/algorithms/foundations/softmax-function-properties-motivation-interpretation.md.
  • Added The Softmax Function: Properties, Motivation, and Interpretation to docs/algorithms/foundations/softmax.md, sourced from raw/algorithms/foundations/softmax-function-properties-motivation-interpretation--franke-degen.pdf (Franke & Degen, tutorial). Full paper-insight structure: Big Picture Mermaid diagram, The Landscape evolutionary tree, Deep Dive subsections covering α interpretation, score-difference semantics, three conceptual interpretations (Gumbel noise-perturbed selection, maximum entropy, exploration-exploitation tradeoff), IO vs. IM model taxonomy, Putting It Together end-to-end trace, What This Buys You with proof table, Where It Breaks failure modes, One Thing to Remember, and Go Deeper.
  • Re-insighted all four FlashAttention docs pages with complete paper-derived content: added The Landscape evolutionary-tree sections with Mermaid diagrams showing the phylogenetic relationship between FlashAttention versions and related work, added The Big Picture diagrams for FlashAttention-2 and FlashAttention-3, and updated all front matter sources to include the derived pdf-markdown paths and updated dates to 2026-07-24.
  • FlashAttention v1: Added The Landscape section with Mermaid evolutionary tree linking GPU memory hierarchy, online softmax, approximate attention methods, and kernel fusion to FlashAttention’s unique combination of IO-aware tiling with exact computation.
  • FlashAttention-2: Added The Big Picture diagram comparing FA1 vs FA2 loop/warp structure, and The Landscape section showing how FA2 addresses FA1’s 25-40% utilization gap through parallelism restructuring without changing the attention formula.
  • FlashAttention-3: Added The Big Picture diagram showing the warp-specialized producer-consumer pipeline with TMA, circular SMEM buffer, and ping-pong consumer warpgroups, and The Landscape section positioning FA3 as the first version to exploit generation-specific hardware capabilities (async WGMMA, TMA, FP8).
  • FlashAttention-4: Added The Landscape section showing asymmetric Blackwell scaling (MMA 2× but exponential/SMEM unchanged) and FA4’s response through exponential emulation, conditional rescaling, TMEM pipelines, 2-CTA backward, and LPT scheduling.

2026-07-25

  • Converted raw/algorithms/attention/variants/multi-query-attention-one-write-head--paper.pdf to markdown via MinerU precise mode, saving result to derived/pdf-markdown/algorithms/attention/variants/multi-query-attention-one-write-head.md.

  • Added Multi-Query Attention: One Write-Head is All You Need to docs/algorithms/attention/variants/multi-query-attention.md, sourced from raw/algorithms/attention/variants/multi-query-attention-one-write-head--paper.pdf (Shazeer, Google, 2019). Full paper-insight structure: memory-bandwidth bottleneck analysis, MHA vs MQA einsum comparison, landscape evolutionary tree linking to GQA, empirical tables (12× speedup at −0.2 BLEU), and orthogonal composition with local attention.

  • Converted three new raw/training/ PDFs to markdown via MinerU precise mode: GPT-1, GPT-2, GPT-3, saving results to derived/pdf-markdown/training/.
  • Added GPT-1: Improving Language Understanding by Generative Pre-Training to docs/training/foundation-models/gpt-1.md, sourced from raw/training/foundation-models/gpt-1-improving-language-understanding--paper.pdf (Radford et al., OpenAI, 2018). Full paper-insight structure: architecture comparison with original Transformer, two-stage pre-train + fine-tune paradigm, task-specific input transformations, auxiliary LM objective, ablation results (pre-training matters: −14.8 pts without it), and zero-shot heuristic behaviors.
  • Added GPT-2: Language Models are Unsupervised Multitask Learners to docs/training/foundation-models/gpt-2.md, sourced from raw/training/foundation-models/gpt-2-unsupervised-multitask-learners--paper.pdf (Radford et al., OpenAI, 2019). Full paper-insight structure: WebText dataset creation, byte-level BPE tokenizer, four model sizes (117M–1.5B), zero-shot task transfer across 8+ NLP tasks, architectural changes from GPT-1 (pre-norm, initialization scaling), data contamination analysis, and scaling trend evidence.
  • Added GPT-3: Language Models are Few-Shot Learners to docs/training/foundation-models/gpt-3.md, sourced from raw/training/foundation-models/gpt-3-language-models-few-shot-learners--paper.pdf (Brown et al., OpenAI, 2020). Full paper-insight structure: 175B-parameter scaling, in-context few-shot learning emergent at scale, Common Crawl filtering pipeline, eight model sizes (125M–175B), zero/one/few-shot evaluation across 42 benchmarks, synthetic reasoning tasks, data contamination methodology, and limitations.
  • Expanded GPT-1, GPT-2, and GPT-3, sourced from the derived training Markdown and original raw PDFs, with Big Picture Mermaid diagrams, landscape trees, end-to-end traces, and additional source-backed interpretation around contiguous-text pretraining, byte-level BPE, zero-shot prompting, few-shot prompt formatting, scaling behavior, and context-only adaptation limits.
  • Converted raw/algorithms/foundations/attention-is-all-you-need--arxiv-1706.03762.pdf to markdown via MinerU precise mode, saving result to derived/pdf-markdown/algorithms/foundations/attention-is-all-you-need.md.
  • Added The Transformer: Attention Is All You Need to docs/algorithms/foundations/transformer.md, sourced from raw/algorithms/foundations/attention-is-all-you-need--arxiv-1706.03762.pdf (Vaswani et al. 2017, NeurIPS). Full paper-insight structure.
  • Converted raw/algorithms/attention/variants/collaborative-multi-head-attention--arxiv-2006.16362v2.pdf to markdown via MinerU precise mode, saving result to derived/pdf-markdown/algorithms/attention/variants/collaborative-multi-head-attention.md.
  • Added Collaborative Multi-Head Attention: Collaborate Instead of Concatenate to docs/algorithms/attention/variants/collaborative-attention.md, sourced from raw/algorithms/attention/variants/collaborative-multi-head-attention--arxiv-2006.16362v2.pdf (Cordonnier, Loukas, Jaggi, EPFL). Full paper-insight structure.
  • Converted raw/training/foundation-models/llama-open-efficient-foundation-language-models--paper.pdf to markdown via MinerU precise mode, saving result to derived/pdf-markdown/training/foundation-models/llama-open-efficient-foundation-language-models.md.
  • Added LLaMA: Open and Efficient Foundation Language Models to docs/training/foundation-models/llama.md, sourced from raw/training/foundation-models/llama-open-efficient-foundation-language-models--paper.pdf (Touvron et al., Meta AI, 2023). Covers public-data pretraining, 7B-65B architecture choices, inference-budget scaling motivation, efficient training implementation, benchmark results, instruction-tuning note, safety findings, and limitations.

2026-07-26

  • Added Grouped-Query Attention in Llama 2 to docs/algorithms/attention/variants/grouped-query-attention/index.md, sourced from raw/algorithms/attention/variants/grouped-query-attention-llama-2--paper.pdf and derived Markdown under derived/pdf-markdown/algorithms/attention/variants/grouped-query-attention-llama-2.md. Covers Llama 2’s 8-KV-group GQA choice for 34B/70B models, the 30B MHA/MQA/GQA ablation, throughput and out-of-memory behavior, tensor-parallel serving tradeoffs, and source limitations at medium confidence.
  • Converted raw/algorithms/attention/variants/deepseek-v2-multi-head-latent-attention--arxiv-2405.04434.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/algorithms/attention/variants/deepseek-v2-multi-head-latent-attention.md.
  • Added DeepSeek-V2 Multi-Head Latent Attention to docs/algorithms/attention/variants/deepseek-v2-mla.md, sourced from raw/algorithms/attention/variants/deepseek-v2-multi-head-latent-attention--arxiv-2405.04434.pdf (arXiv:2405.04434, May 2024). Covers MLA’s low-rank joint K/V compression, decoupled RoPE, query compression, DeepSeekMoE serving context, MHA/GQA/MQA comparison, ablation results, and reported 93.3% KV-cache reduction with 5.76× maximum generation throughput versus DeepSeek 67B.
  • Converted raw/algorithms/model-architectures/deepseek-v3-2/deepseek-v3.2-sparse-attention-scaled-rl-tool-use--paper.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/algorithms/model-architectures/deepseek-v3-2/deepseek-v3.2-sparse-attention-scaled-rl-tool-use.md.
  • Added DeepSeek-V3.2: Sparse Attention, Scaled RL, and Thinking in Tool-Use to docs/training/model-families/deepseek/deepseek-v3.2/index.md, sourced from raw/algorithms/model-architectures/deepseek-v3-2/deepseek-v3.2-sparse-attention-scaled-rl-tool-use--paper.pdf (DeepSeek-AI, 2026). Full paper-insight structure with Big Picture Mermaid diagram, Landscape evolutionary trees, Deep Dive sections covering DSA (lightning indexer, two-stage continued pre-training, MQA-mode instantiation under MLA), scaled GRPO (unbiased KL, off-policy sequence masking, Keep Routing, Keep Sampling Mask), thinking-in-tool-use (context management, cold-start prompting, large-scale agentic task synthesis across code/search/general/code-interpreter domains), What This Buys You with reasoning and agentic benchmark tables, Where It Breaks failure modes, and One Thing to Remember. Includes copied key figures from the original paper (DSA architecture, thinking retention, inference costs, RL curves, context management).

2026-07-28

  • Added Triton: Tiled GPU Kernel Language and Compiler to docs/frameworks/compilers-kernels/triton/index.md, sourced from raw/frameworks/compilers-kernels/triton/triton-tiled-gpu-kernel-language-compiler--mapl-2019-tillet-kung-cox.pdf (Tillet, Kung, Cox, MAPL 2019). Full paper-insight structure: Big Picture with Triton overview diagram, Landscape evolutionary tree (cuBLAS→TC→TVM→PlaidML→Triton), Deep Dive covering Triton-C (tile types, broadcasting, predication), SPMD model vs CUDA, Triton-IR (retiling, PSSA), Triton-JIT passes (pre-fetching, peephole, hierarchical tiling, memory coalescing, shared memory allocation, barrier synchronization, auto-tuner), and performance results (>90% peak matmul, cuDNN parity on convolution, fused shift-conv). Converted via MinerU precise mode to derived/pdf-markdown/frameworks/compilers-kernels/triton/triton-tiled-gpu-kernel-language-compiler.md. Copied seven renamed figure assets (architecture overview, CUDA vs Triton SPMD model, hierarchical tiling, matmul performance, roofline, memory coalescing, shared memory allocation).

  • Added MiniMax Sparse Attention (MSA) to docs/training/efficient-attention/minimax-sparse-attention/index.md, sourced from raw/training/efficient-attention/minimax-sparse-attention--paper.pdf (Lai et al., MiniMax, 2026). Full paper-insight structure with Big Picture Mermaid diagram, Landscape evolutionary tree (Softmax Attention → MQA → GQA → Native Sparse → MSA), Deep Dive sections covering Index Branch architecture, KL alignment loss, training stability mechanisms (gradient detach, indexer warmup, forced local block), exp-free TopK kernel, KV-outer sparse attention, and sparse KL loss optimization. Covers 109B MoE results: MSA-PT (from scratch) and MSA-CPT (continued pretraining) matching GQA on downstream benchmarks, 28.4× FLOPs reduction, 14.2× prefill and 7.6× decode speedup at 1M context on H800. Converted via MinerU precise mode to derived/pdf-markdown/training/efficient-attention/minimax-sparse-attention.md. Normalized raw filename from raw/model/minimax-m3.pdf to raw/training/efficient-attention/minimax-sparse-attention--paper.pdf. Copied four renamed figure assets (architecture overview, training LM loss, CPT KL loss, FLOPs reduction).

  • Added a revision-aware repository-reading workflow: ignored read-only checkouts under external-repos/, immutable commit-specific raw records, full-SHA derived evidence directories, multi-page docs_paths manifest consumers, machine-readable source metadata, stricter repository integrity checks, and scripts/repositories/init_source.py scaffolding. Consolidated the canonical workflow in .github/instructions/repo-reading.instructions.md.

  • Added Triton in Practice: How vLLM and vllm-ascend Use Triton to docs/frameworks/compilers-kernels/triton/triton-in-vllm.md, sourced from commit-pinned vLLM and vllm-ascend codebase records. Codebase-driven practical guide covering: vLLM’s triton_utils import infrastructure, direct_register_custom_op registration, ~163 kernel files across 12 categories (attention, MoE, quantization, activation/norm/RoPE, sampling, Mamba, utilities), universal coding patterns (tl.constexpr, do_not_specialize, grid-stride loops, inline PTX, tl.heuristics), vllm-ascend’s CANN-backend adaptation, dual Triton+AscendC strategy, and an AWQ dequantization walkthrough faithful to the pinned implementation. Added reproducible source records and repo-analysis notes, then updated frameworks/index.md and logs/index.md.

  • Added Triton Ascend: Ascend NPU Backend for Triton to docs/frameworks/compilers-kernels/triton-ascend/index.md, sourced from the commit-pinned triton-ascend codebase at 41f499924da1d58955196c946895597e992127f0 (main, 2026-07-28). Beginner-oriented architecture page covering: five-layer design (backend registration via name.conf + monkey-patching, compilation pipeline TTIR→HIVM→LLVM→Linalg→Bisheng, C++ MLIR passes, runtime driver with ACL, CANN extension ops), three compilation modes (SIMD/SIMT-only/Unstructured-in-SIMT), Ascend-specific autotuner (CV autotune, UB tuning, DSL analysis), relationship to vllm-ascend, and a staged codebase navigation guide for newcomers. Added Mermaid compilation-pipeline diagram and repository source record at raw/frameworks/compilers-kernels/triton-ascend/triton-ascend-codebase--github-41f499924da1.md.

  • Updated Frameworks and Wiki Index to include the new triton-ascend page.

  • Re-insighted NVFP4: Blackwell 4-Bit Floating Point with fresh web captures of Transformer Engine 2.19.0.dev0 NVFP4 documentation and NVIDIA Technical Blog (Introducing NVFP4, 2025-06-24), replacing the old manually-written raw reference. Captured as immutable web sources: raw/hardware/numerics/quantization/nvfp4-te--web-2026-07-28-6625830e9e9e.html + metadata and raw/hardware/numerics/quantization/nvfp4-blog--web-2026-07-28-a2f3eb0ba3bb.html + metadata, with derived markdown at derived/web-markdown/hardware/. Full paper-insight restructure: added The Big Picture Mermaid diagram with training-only mechanisms, Why This Exists concrete example, The Landscape evolutionary tree (FP16→FP8→MXFP8/MXFP4→NVFP4), Symbol Map, Deep Dive subsections covering two-level hierarchical scaling, fractional E4M3 vs. E8M0 with MSE comparison, 16 vs. 32 micro-block size, 2D weight scaling, stochastic rounding, Random Hadamard Transform with tiled implementation, GEMM TN-only layout with columnwise transpose, distributed training (amax sync, quantized all-gather), and swizzling. Added Putting It Together 8-step training trace, What This Buys You with full DeepSeek-R1-0528 PTQ results table, Where It Breaks (8 failure modes), and updated Go Deeper/deployment ecosystem links. Updated sources.json to reflect the two ingested web sources and removed the old manual reference record.

2026-07-27

  • Updated Megatron-LM: GPU-Cluster Training Parallelism to cover both Megatron-LM papers: the original 2019 intra-layer tensor parallelism paper (f/g conjugate operators, column-parallel GEMM splitting, BERT LayerNorm rearrangement, 8.3B GPT on 512 V100 GPUs) and the 2021 PTD-P paper (pipeline+tensor+data parallelism composition, interleaved 1F1B, scatter/gather, 1T GPT on 3072 A100 GPUs). Added derived markdown from raw/training/megatron-lm-train.pdf via MinerU precise mode at derived/pdf-markdown/training/parallelism/megatron-lm-tensor-parallelism/. Updated front matter, Big Picture, Landscape, TL;DR, and Core Idea sections.

  • Converted raw/training/parallelism/megatron-lm-gpu-cluster-training-parallelism--paper.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/training/parallelism/megatron-lm-gpu-cluster-training-parallelism.md.
  • Added Megatron-LM: GPU-Cluster Training Parallelism, sourced from raw/training/parallelism/megatron-lm-gpu-cluster-training-parallelism--paper.pdf. Covers PTD-P composition of tensor, pipeline, and data parallelism; interleaved 1F1B scheduling; scatter/gather pipeline communication; microbatch and activation recomputation tradeoffs; fused Transformer kernels; trillion-parameter throughput results; and copied local figure assets.
  • Created docs/terms/ glossary directory. Added Terms Glossary, Microbatch, and Scatter/Gather term pages with definition-first structure, cross-links to paper pages, and agent auto-ingestion rules in .github/instructions/docs-terms.instructions.md. Updated AGENTS.md with term-extraction workflow trigger, and linked term occurrences from the Megatron-LM paper page.
  • Added reusable hover-preview tooltip front matter to Microbatch and Scatter/Gather, and normalized their first definition sentence to start with the canonical term title.
  • Normalized source organization across raw/ and derived/pdf-markdown/: moved legacy category folders to canonical plural categories, renamed PDFs with deterministic title/source suffixes, added kb-categories.json and sources.json, documented source organization rules, and wired source integrity checks into ./scripts/lint-docs.sh.
  • Added term-hover styling, generated terms.json, and client-side enhancement so ordinary Markdown links to term pages render as Wikipedia-style glossary previews on the built site.
  • Revised term-link placement to prefer the first meaningful in-content occurrence and removed the detached Related terms line from the Megatron-LM page.
  • Added GPipe: Micro-Batch Pipeline Parallelism to docs/training/parallelism/gpipe/index.md, sourced from raw/training/parallelism/gpipe-micro-batch-pipeline-parallelism--arxiv-1811.06965v5.pdf (Huang et al., Google, arXiv:1811.06965v5, Jul 2019). Full paper-insight structure: pipeline parallelism diagram, landscape evolutionary tree (DP→SPMD→PipeDream→GPipe→Megatron-LM), Deep Dive on partitioning, micro-batch pipelining, re-materialization, and low communication overhead; near-linear throughput scaling (6.3× on 8 TPUs), 298× model scaling (83.9B on 128 TPUv3), design trade-offs vs. PipeDream/Mesh-TensorFlow/data parallelism, and copied local figure assets. Normalized raw filename from micro-batch-pipeline-parallelism.pdf. Updated microbatch term page with GPipe cross-link and the $M \ge 4K$ bubble rule.
  • Converted raw/training/parallelism/sequence-parallelism-long-sequence-training--arxiv-2105.13120.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/training/parallelism/sequence-parallelism-long-sequence-training.md.
  • Added Sequence Parallelism: Splitting Sequences Across GPUs, sourced from raw/training/parallelism/sequence-parallelism-long-sequence-training--arxiv-2105.13120.pdf (Li et al., NUS, arXiv:2105.13120, May 2021). Full paper-insight structure: parallelism-comparison diagram and Ring Self-Attention (RSA) visual, landscape evolutionary tree (DP→PP→TP→SP→4D), Deep Dive on RSA two-stage ring communication, memory analysis showing when SP beats TP ($BL > 16AZ$), communication parity with TP, empirical results (13.7× batch size, 3.0× sequence length, 114K tokens with sparse attention), weak scaling behavior, convergence verification, and 4D parallelism vision. Copied six renamed figure assets.
  • Created All-Gather term page: NCCL collective that gathers shards across ranks into a full replicated tensor; covers the split/all-gather pipeline-boundary pattern, why sequence parallelism avoids this cost, and common confusions with all-reduce and scatter/gather. Back-linked from Megatron-LM and Sequence Parallelism insight pages; cross-linked from scatter-gather and sequence-parallelism term pages.
  • Updated .github/instructions/docs-terms.instructions.md with retroactive back-link rules for terms created outside of paper ingest, and a signal-category table to improve term recognition for communication primitives and system-level concepts.

  • Created All-Reduce term page: NCCL collective that sums tensors element-wise across ranks; covers ring algorithm, where it appears in each parallelism strategy (data, tensor, pipeline), why sequence parallelism avoids it in MLP blocks, and common confusions with all-gather and reduce-scatter. Back-linked from Megatron-LM, Sequence Parallelism, vLLM, and LLaMA pages.

  • Converted raw/training/deepseek-v4.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/training/model-families/deepseek/deepseek-v4.md with 71 files (images).
  • Added DeepSeek V4 Attention: Code Reading Map to docs/frameworks/serving/deepseek/v4-attention-code-reading.md, sourced from commit-pinned vLLM (d18ed2304a27) and vllm-ascend (8645122088f5) codebase records. Implementation-focused code-reading map covering: DeepseekV4Attention pipeline (~700-line base class), DeepseekCompressor with fused Triton kernels (3 variants for head_dim=512 FP8, head_dim=128 FP8, head_dim=128 MXFP4), DeepseekV4FlashMLABackend and DeepseekV4FlashInferMLASparseBackend (SM10x/SM12x), multi-stream overlap architecture (up to 4-way on CUDA), Ascend DSA backend (AscendDSABackend + AscendDeepseekSparseAttention NPU op), heterogeneous KV cache (SWA + compressor state + MLA), FP8/MXFP4 quantization layout dispatch, and platform-specific extension points. Added derived repo-analysis notes at derived/repo-analysis/frameworks/serving/vllm-ascend/vllm/.../deepseek-v4-attention.md and derived/repo-analysis/frameworks/serving/vllm-ascend/vllm-ascend/.../deepseek-v4-attention.md. Updated sources.json docs_paths, docs/logs/index.md, and docs/frameworks/index.md.

2026-07-29

  • Converted raw/algorithms/attention/linear/transformers-are-rnns-linear-attention--arxiv-2006.16236v3.pdf with MinerU precise mode and added Transformers Are RNNs: Linear Attention at high confidence. The insight covers kernel factorization, associative reordering, causal recurrent states, linear-time gradients, carefully qualified image-generation throughput, speech-recognition quality limits, and editable Mermaid diagrams; it also adds the Linear Attention and KV Cache glossary entries and links both terms from Kimi Linear.
  • Converted raw/training/model-families/kimi/kimi-linear-expressive-efficient-attention--paper.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/training/model-families/kimi/kimi-linear-expressive-efficient-attention.md.
  • Added Kimi Linear: Expressive Efficient Attention Architecture to docs/training/model-families/kimi/kimi-linear/index.md, sourced from raw/training/model-families/kimi/kimi-linear-expressive-efficient-attention--paper.pdf (Kimi Team, Moonshot AI, 2026). Full paper-insight structure with Big Picture architecture diagram and performance scatterplots, Landscape evolutionary tree (Softmax Attention→Linear Attention→Gated DeltaNet→KDA→Hybrid), Deep Dive sections covering KDA channel-wise gating, DPLR constraint for 2× kernel speedup, 3:1 hybrid ratio with NoPE on MLA layers, output gate design, and neural parameterization. Covers 48B MoE (3B active) results: 51.0 MMLU-Pro (vs. MLA 47.2), 84.3 RULER at 128k (vs. MLA 81.3), 6.3× decoding speedup at 1M context, scaling law showing ~1.16× compute efficiency, RL convergence advantage, and NoPE vs. RoPE ablation. Normalized raw filename from kimi-linear.pdf. Copied 12 renamed figure assets (architecture, perf-vs-acceleration, TPOT, kernel speed, scaling law, prefill time, decode TPOT, KDA computation, chunk state/output, RL curves). Updated sources.json, docs/training/index.md, and docs/logs/index.md.
  • Added SWAT: Sliding Window Attention Training to docs/training/efficient-attention/swat-sliding-window-attention/index.md, sourced from raw/training/efficient-attention/swat-sliding-window-attention-training--paper.pdf (Fu et al., CityU/Tencent/Westlake, 2025). Full paper-insight structure with Big Picture Mermaid diagram, sigmoid+balanced ALiBi+RoPE combination, Landscape evolutionary tree (Standard Attention→Sparse→Linear→SSM→SWAT), Deep Dive on attention sink diagnosis, sigmoid dense attention, balanced ALiBi bidirectional slopes, and RoPE training stabilization. Covers SOTA results on 8 commonsense reasoning benchmarks (46.88% avg at 340M, surpassing Titans 46.17% and Gated DeltaNet 45.42%), SWA training ablation showing $O(N\omega)$ inference with long-context perplexity matching full-attention models, and normalized raw filename from sliding-window-attention.pdf. Copied three renamed figure assets (SWA mechanism, information transmission, training loss curves). Updated sources.json, docs/training/index.md, and docs/logs/index.md.
  • Converted raw/training/efficient-attention/gated-delta-networks-improving-mamba2-with-delta-rule--arxiv-2412.06464.pdf with MinerU precise mode and added Gated Delta Networks: Improving Mamba2 with Delta Rule at high confidence. The insight covers global decay plus key-targeted correction, decay-aware chunkwise WY training, S-NIAH memory behavior, hybrid SWA architectures, 1.3B/100B-token results, limitations, local diagrams, and the new Delta Rule glossary entry.
  • Converted raw/training/model-families/kimi/k3-technical-report--paper.pdf to Markdown via MinerU precise mode, saving result to derived/pdf-markdown/training/model-families/kimi/k3-technical-report.md.
  • Added Kimi K3: Open 3T-Class Frontier Model, sourced from raw/training/model-families/kimi/k3-technical-report--paper.pdf and derived Markdown. Covers the 2.8T/104B-active native multimodal MoE architecture, hybrid KDA/MLA attention, Stable LatentMoE with SiTU-GLU and Quantile Balancing, multi-effort agentic RL, MoonEP balanced expert training, million-token rollout cache/sandbox infrastructure, reported third-party and in-house results, and copied local figure assets.
  • Added Kimi Delta Attention and Mixture of Experts glossary entries, and back-linked Kimi K3 from the existing KV Cache, Linear Attention, and Delta Rule term pages.
  • Added vLLM-Ascend Kimi K3 MoE Forward Path, sourced from the pinned clean vllm-ascend codebase at 8645122088f5cad1701205310573c5ee05c809f5. The insight traces the Kimi K3-style routed-MoE serving substrate through patched FusedMoE, AscendMoERunner, router top-k selection, token dispatch, grouped expert MLP, combine/finalize, shared-expert overlap, and dynamic EPLB. Added derived evidence notes and a local Mermaid flow asset.
  • Added vLLM Kimi K3 Code Reading Map, sourced from the newly pinned clean upstream vLLM commit a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b. The page replaces the earlier substrate-only reading with the real upstream Kimi K3 implementation: vllm/models/kimi_k3/, XTML renderer/parser integration, multimodal Kimi-K2.5 vision wrapper, KimiLinearForCausalLM, hybrid KDA/NoPE MLA layers, latent MoE with LatentMoERunner, DeepGEMM MegaMoE, optional SM100 latent-MoE tail fusion, and KimiK3MTP.

2026-07-30

  • Re-insighted vLLM-Ascend Kimi K3 MoE Forward Insight, sourced from the verified remote main vllm-ascend checkout at e3bb5f570f0b7d7fef9df3190a450052bee090cc. The refreshed page replaces the older 8645122088f5... reading with typed MoE stage contracts, routed-expert capture, MC2/Fused-MC2 details, CANN MegaMoe buffer constraints, dynamic EPLB wiring, and the current limitation that vllm-ascend has no plugin-local kimi_k3.py model file.
  • Added 线性Attention的探索：Attention必须有个Softmax吗？, sourced from Su Jianlin’s 2020 blog post captured via Wayback Machine at raw/algorithms/attention/linear/linear-attention-without-softmax--web-2026-07-30-e116218d6fdb.html. Covers the argument that softmax is the bottleneck of standard attention, three linear attention families (kernel feature maps, double-softmax, cosine-similarity Taylor approximation), autoregressive generation via RNN-style recurrence, comparison with sparse attention/Reformer/Linformer, and the historical context of this widely cited Chinese-language ML reference.
  • Added Recurrent Neural Networks: From RNN to LSTM from an accessible mirror of the user-supplied Zhihu article, after direct Zhihu capture returned HTTP 403. The medium-confidence note explains shared recurrent weights, hidden-state compression, long-term dependency failures, LSTM cell/hidden state and gates, and the exact bridge from the RNN update pattern to linear attention’s recurrent decoding state.

2026-07-31

  • Captured the CANN 8.3.RCX Ascend C basic architecture page and the Triton Ascend Vector, Cube, and CV Fusion programming guides as immutable web sources under raw/frameworks/, with derived Markdown under derived/web-markdown/frameworks/.
  • Added and expanded Triton Ascend 算子机制学习路径, connecting AI Core compute/storage/movement/synchronization mechanisms to Triton Ascend tile design, physical-core reuse, UB/L1/L0 budgeting, Vector/Cube selection, CV Fusion, profiling diagnosis, and a four-stage practice plan. Preserved seven official CANN architecture/dataflow figures locally and added editable unified-dataflow and CV-pipeline Mermaid diagrams.
  • Added Global Memory term page (category hardware, first Hardware term in the glossary) defining GM as the off-chip device memory on an accelerator (GPU HBM or Ascend GM), with aliases GM/HBM/device memory. Back-linked the term from Triton Ascend 算子机制, FlatQuant, and MiniMax Sparse Attention, and added the entry to Terms Glossary.
  • Expanded Triton Ascend 算子机制学习路径 with a dedicated “5.4 Multibuffering: Hiding the A/B Load Behind Cube Compute” subsection covering the double-buffered K-loop timeline, the separate-queue overlap model that makes it legal, the BLOCK_K/occupancy vs. buffering-depth tradeoff, the MTE-bound failure signature, and the Triton Ascend multibuffer=True / set_workspace_multibuffer compile hints; renumbered the Cube failure modes section to 5.5.

2026-08-02

  • Captured the EaseCloud vLLM throughput guide, the official vLLM launch article, and AnyScale’s continuous-batching analysis as immutable web sources, then ingested them into vLLM Continuous Batching.
  • Added a medium-confidence, code-backed V1 runtime synthesis at pinned clean vLLM commit a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b, covering token/sequence budgets, running-before-waiting scheduling, chunked prefill, paged KV admission, persistent worker-batch reconciliation, completion, and preemption; saved an editable Mermaid Landscape, three locally preserved source figures, and derived runtime evidence.
  • Created the Continuous Batching glossary entry, back-linked the existing vLLM framework page, and updated the KV Cache term with the new serving-mechanism page.
  • Updated the repository visual workflow: original paper/web figures are now the first choice and must be preserved locally with source captions; The Landscape always uses a linked local Mermaid diagram; synthesized visuals outside The Landscape are fallback-only.

2026-08-03

  • Added CANNBot Skills: Triton Ascend Development Workflow to docs/frameworks/compilers-kernels/triton-ascend/cannbot-skills-workflow.md, sourced from the commit-pinned cannbot-skills external checkout at 326a6b47210f. Covers the seven Triton-domain skills (task-extractor, designer, coding, verifier, precision-debug, latency-optimizer, simulator-optimizer), the triton-op-generator Plugin’s six-phase pipeline, the 25-point ordered optimization system, five-stage ULP isolation debugging, and reusable design patterns for triton-ascend kernel development.
  • Refactored docs/ into per-topic subcategories to remove the flat single-level layout, and updated sources.json docs_path entries, all internal links, category indexes, and the Wiki Index.
  • docs/frameworks/ is now organized per project: vllm/ (PagedAttention paper, code learning path, continuous batching, Kimi K3 map), vllm-ascend/ (Kimi K3 MoE forward), triton/ (was triton-language/), triton-ascend/, sglang/ (was sglang-framework.md), dspark/, harbor/ (was harbor-framework/), and deepseek/ (V4 attention code-reading map); new hub pages added for vllm/, vllm-ascend/, and deepseek/.
  • docs/algorithms/ is now grouped into foundations/ (Transformer, softmax, RNN-to-LSTM), flashattention/ (FA1-4), attention-variants/ (MQA, GQA, collaborative, MLA), and linear-attention/ (linear attention, without-softmax survey, matrix exponentiation); deepseek-v3.2/ stays in algorithms/ to keep its manifest category consistent.
  • docs/benchmarks/ is now grouped into agent-eval/ (DeepSWE, DeepSWE v1.1, Pier, τ-bench family, AutoJudger) and serving-perf/ (EvalScope Perf, AISBench analysis and roadmap).
  • docs/training/ is now grouped into parallelism/ (Megatron-LM, GPipe, sequence parallelism), foundation-models/ (GPT-1/2/3, LLaMA), deepseek/ (V4), kimi/ (Kimi Linear, Kimi K3), efficient-attention/ (MiniMax MSA, SWAT, Gated DeltaNet), and fine-tuning/ (intrinsic dimensionality, Socratic-SWE).
  • docs/hardware/ is now grouped into quantization/ (FlatQuant, NVFP4).
  • Added 15 subcategory hub pages, rewrote the five category index pages to group by subcategory, updated docs/README.md structure notes, and refreshed all term-page back-links and the chronological log links to the new paths.
  • Codified the subcategory workflow in AGENTS.md (new “Docs Organization” section) and in .github/instructions/docs-front-matter.instructions.md and .github/instructions/source-organization.instructions.md: subcategory folders with hub index.md pages, the sources.json docs_path category-prefix invariant that discourages cross-category moves, and a move/rename checklist (git mv, link recomputation, manifest and index updates, lint).
  • Added vLLM-Ascend Architecture: How the Ascend NPU Port Integrates with vLLM to docs/frameworks/serving/vllm-ascend/architecture.md, sourced from a fresh main-branch inspection of vllm-project/vllm-ascend at 32a59d4e349c. Covers the five integration mechanisms (plugin registration, NPUPlatform switchboard, ModelRegistry overrides, monkey-patch layer, custom backends), end-to-end execution flow with a sequence diagram, ACL graph capture as CUDA graph replacement, HCCL communication replacing NCCL, Ascend-specific attention backends (FIA/MLAPO/SFA/DSA), CaMemAllocator sleep mode, and a clear inventory of what upstream vLLM code is reused as-is versus replaced.
  • Added Monkey-Patching to docs/terms/monkey-patching.md, sourced from the captured pytest monkeypatch documentation page. Defined the term around runtime substitution of functions, attributes, modules, and environment lookups; added it to the glossary; and back-linked its first meaningful in-content mentions from DeepSWE v1.1, Triton Ascend, vLLM Ascend, and the vLLM-Ascend architecture page.
  • Rewrote All-Gather, All-Reduce, and Scatter/Gather using the captured Medium article on distributed training communication primitives. Added the article’s raw HTML, metadata, and derived markdown as sources, refreshed the glossary summaries to emphasize the collectives as many-to-many communication patterns, and preserved three original source figures locally under docs/terms/assets/ for the updated term pages.
  • Added revision-aware code links for repository-backed pages: Markdown previews open relative files from external-repos/, while rendered Jekyll pages resolve the same references to provider-correct, full-commit remote links. Applied the first link to NPUPlatform in the vLLM-Ascend architecture map and documented the convention in the repository-reading workflow.
  • Generalized repository provenance from GitHub-only to origin-derived GitHub/GitCode providers. Repository scaffolding now normalizes SSH or HTTPS origins, source records preserve clone and web URLs, integrity checks enforce provider identity, and the CANNBot Skills source was corrected from GitHub to its actual GitCode origin without changing the pinned commit.
  • Made revision-aware code links the preferred repository-analysis workflow and enabled strict repository-path lint for the CANNBot Skills workflow: new pages link important files and symbols while evidence is collected, local previews open the checkout, web builds use commit-pinned provider links, and generated or hypothetical filenames remain prose.
  • Migrated the external-repos/ repository checkouts from independent full clones to the shared partial bare cache design (external-repos/.cache/{provider}/{owner}/{repo}.git). Every pinned revision is now a detached worktree backed by the shared object database, so multi-version repos store each version as a diff over one cache instead of duplicating clones: vllm (a0c092ee72c0, d18ed2304a27) and vllm-ascend (32a59d4e349c, 8645122088f5, e3bb5f570f0b) share one cache each, and triton-ascend (41f499924da1) and cannbot-skills (326a6b47210f) each have one. Registered local_checkout paths, pinned commits, raw source records, derived evidence, and the code_repositories.json registry are unchanged and remain fully reproducible via scripts/repositories/worktree.py materialize.
  • Extended revision-aware code links to every repository-backed docs page and enabled code_links: strict on all of them: vLLM Kimi K3 Code Reading Map, vLLM Continuous Batching, vLLM-Ascend Kimi K3 MoE Forward, Triton Ascend, vLLM-Ascend Architecture, Triton in vLLM and vllm-ascend, and DeepSeek V4 Attention Code Reading. All 223 code links resolve to commit-pinned blob URLs verified against the pinned checkouts; restored the missing vllm (d18ed2304a27) and vllm-ascend (8645122088f5) checkouts; rewrote one non-existent filename reference as prose; and disabled the markdownlint MD050 strong-style rule, which misreads __init__.py code identifiers.
  • Added vLLM Block Table Management, a code-backed deep dive into the V1 block table mechanism at vLLM commit dd11df04f3b7 (new immutable revision; kb-repo-worktree.py sync decided new revision because vllm/v1/core changed — deprecated calculate_kv_scales helpers removed and same-type KV cache group merging improved, while the core block mechanism is unchanged). Covers the BlockPool free-list and refcounts, per-request logical block tables, copy-on-write sharing, chained-hash prefix caching, multi-group coordinator fixed-point cache hits, SWA/Mamba block recycling, and worker-side MultiGroupBlockTable materialization. Preserves four original PagedAttention paper figures locally and links an editable landscape Mermaid.
  • Converted raw/frameworks/serving/vllm/vllm-pagedattention-serving-framework--arxiv-2309.06180v1.pdf with MinerU precise mode to derived/pdf-markdown/frameworks/serving/vllm/vllm-pagedattention-serving-framework.md (figures under its images/ subfolder) and added the derived_path to its sources.json entry.
  • Added Block Table and PagedAttention term pages (category frameworks / algorithms), added the new page to KV Cache appears_in, and updated the glossary index.
  • Updated the vLLM hub (docs/frameworks/serving/vllm/index.md) and Wiki Index with the new page and terms.

2026-08-04

  • Made commit-pinned external code checkouts portable across agent environments: docs/_data/code_repositories.json is now documented as the single lock table, kb-repo-worktree.py status reports offline workspace readiness, materialize-all restores selected or all registered revisions, and scripts/bootstrap-external-repos.sh provides the fresh-workspace entry point without adding large codebase downloads to the normal dependency bootstrap.
  • Migrated every existing repository-backed docs page to the new repository evidence template: added Required Code Evidence tables plus Runtime Flow Evidence to the supporting derived/repo-analysis/ notes and enabled code_evidence: strict on vLLM Kimi K3 Code Reading Map, vLLM Continuous Batching, vLLM-Ascend Kimi K3 MoE Forward, vLLM-Ascend Architecture, Triton Ascend, Triton in Practice, DeepSeek V4 Attention Code Reading, and CANNBot Skills workflow. Every one of the 105 declared findings maps to a revision-aware code link whose file, start line, and symbol were verified against the pinned checkouts; check-code-links.py --local and lint-docs.sh both pass.
  • Added MiniMax GQA W4A4 Quantization Path: GPU (vLLM) and NPU (vllm-ascend) to docs/frameworks/serving/vllm/, supported by new immutable revisions vllm-codebase--github-72cd5424da80 (72cd5424da80a4a9caa3f42fd65bc0b94e61cbf0) and vllm-ascend-codebase--github-61221e9add8c (61221e9add8c717b304005bd9d48d6215d035be7). Traces the W4A4 (4-bit weight/4-bit activation) runtime path for MiniMax M2/M2.5 GQA attention (qkv_proj/o_proj) and block-sparse MoE experts: GPU scheme selection (_get_scheme_from_parts, compressed-tensors W4A4 MXFP4/NVFP4, ModelOpt NVFP4, MXFP4 MoE-only fallback) and NPU ModelSlim dispatch (quant_model_description.json, packed_modules_model_mapping minimax_m2 entry, W4A4_MXFP4 / W4A4_FLATQUANT_DYNAMIC / W4A4_DYNAMIC schemes, npu_quant_matmul), plus the MiniMax M2 fp8-disable patch on NPU. New page uses code_links: strict and code_evidence: strict with 21 declared evidence findings.
  • Added six GEMM term pages to docs/terms/ — General Matrix Multiply (GEMM), Inner Product, Outer Product, Matrix Tiling, Systolic Array, and Memory Banking — sourced from the captured Spatial GEMM tutorial (raw/hardware/kernels-dataflow/spatial-gemm--web-2026-08-04-5ae8cd1f3ba2.html + metadata → derived/web-markdown/hardware/kernels-dataflow/spatial-gemm--web-2026-08-04-5ae8cd1f3ba2.md), plus a hardware insight page Spatial GEMM: Blocked Outer-Product Matrix Multiply that hosts the manifest docs_path (marked ingested). Back-linked first occurrences in FlashAttention, Triton, Triton-Ascend operator-mechanisms and cannbot-skills-workflow, NVFP4, Megatron-LM, linear-attention, transformer, sequence-parallelism, and the global-memory and delta-rule term pages; updated the terms glossary index and wiki index.
  • Reworked MiniMax GQA W4A4 Quantization Path for beginners: added a plain-language glossary, model-layer and offline/load/runtime Mermaid visuals, clearer platform flow, hardware fallback guidance, explicit static-analysis limitations, and a code-reading appendix while preserving strict pinned evidence links.
  • Reinsighted FlatQuant: Fast Learnable Affine Quantization from raw/hardware/numerics/quantization/flatquant-fast-learnable-affine-quantization--arxiv-2410.09426v4.pdf: extracted the PDF to derived/pdf-markdown/hardware/numerics/quantization/flatquant-fast-learnable-affine-quantization.md via MinerU (precise mode) and registered it as derived_path in sources.json. Rewrote the page to the current content-structure template (Big Picture with original Figure 3, Landscape Mermaid + editable .mmd, Symbol Map, Deep-Dive subsections, Putting It Together, What This Buys You with caution notes, Where It Breaks, one-sentence takeaway) and converted it to a folder-backed page at docs/hardware/numerics/quantization/flatquant/index.md with six original paper figures under assets/ (framework, MSE landscape, prefill/decoding speedup, kernel fusion, flatness evolution). Added cross-links to the MiniMax GQA W4A4 path and vLLM-Ascend architecture pages, registered the page in the KV Cache and GEMM term glossaries, moved the manifest docs_path, updated all internal links, and regenerated the docs link graph.
  • Ingested Kronecker Product as a web source (raw/algorithms/foundations/kronecker-product--web-2026-08-04-41ef5673a67a.html + metadata → derived/web-markdown/algorithms/foundations/kronecker-product--web-2026-08-04-41ef5673a67a.md): created the reference page Kronecker Product (hosts the manifest docs_path) and the glossary term Kronecker Product (category algorithms), both citing the captured MathWorld snapshot. Back-linked the first meaningful occurrences in FlatQuant (TL;DR) and MiniMax GQA W4A4 Quantization Path (NPU left_trans/right_trans), linked the Outer Product confusion and related-terms sections, updated the algorithms category index, terms glossary index, wiki index, and regenerated the docs link graph.
  • Removed the raw/frameworks/harbor/ snapshot (README, docs, and copied runtime scripts) and its github:harbor reference entry from sources.json; the Harbor insight is now repository-backed instead of doc-snapshot-backed.
  • Registered Harbor Codebase as an immutable repository source via kb-init-repo-source.py: raw/frameworks/infrastructure/harbor/harbor-codebase--github-97e65926410b.md + derived/repo-analysis/frameworks/infrastructure/harbor/harbor/97e65926410ba6b9cef15cd5d9bf9a0c700cef26/important-files.md at commit 97e65926410ba6b9cef15cd5d9bf9a0c700cef26 (main, 2026-08-04), registered as harbor-97e65926410b in docs/_data/code_repositories.json.
  • Reinsighted Harbor: Agent Evaluation Framework (Code Reading) from the pinned checkout: rewrote the page as a code_links: strict / code_evidence: strict page covering the task packaging model (Task, TaskConfig), four dataset source kinds (DatasetConfig, TaskClient), planning/orchestration (JobPlan, Job, TrialQueue), single-step and multi-step trial runtimes (Trial, SingleStepTrial, MultiStepTrial), execution primitives (BaseAgent, BaseEnvironment, Verifier), and derived compile/exec workflows (Compiler, Executor), with 22 declared evidence findings mapped to revision-aware code links. Added runtime-flow.mmd and landscape.mmd under docs/frameworks/infrastructure/harbor/assets/, extracted editable .drawio sources from the existing .drawio.svg renders (harbor-architecture.drawio, harbor-runtime-architecture.drawio, harbor-design-philosophy.drawio), and updated the frameworks index and wiki index.

2026-08-05

  • Added DeepSeek-V4 Lightning Indexer C8 Quantization to docs/frameworks/serving/vllm-ascend/, a code_links: strict / code_evidence: strict repository-backed insight at pinned vllm-ascend revision 32a59d4e349c12c32cdbc1916436c16e39939afc (reused the existing registered revision). Clarifies that the DeepSeek-V4 Lightning Indexer is always 8-bit (“C8”): INT8 keys + FP16 per-token-head scales on 910B/A2/A3, FP8 e4m3fn + FP32 scales on A5. A 4-bit “C4” indexer quantization is not implemented and is planned as future work; “c4” in the code only names the DSV4 layer type (compress ratio 4) that owns an indexer. Traces the quant path from AscendDeepseekV4IndexerCache dtypes and Indexer construction through indexer_quantize_query/indexer_quant_scatter, the npu_vllm_quant_lightning_indexer + AICPU metadata pre-op call sites in the DSA backend, and the SFA (V3.2) contrast branch (npu_lightning_indexer_quant). Added 23 declared evidence findings to derived/repo-analysis/frameworks/serving/vllm-ascend/vllm-ascend/32a59d4e349c12c32cdbc1916436c16e39939afc/important-files.md, registered the page in the 32a59d4e manifest docs_paths, saved two editable Mermaid diagrams under docs/frameworks/serving/vllm-ascend/assets/, and linked the page from the vllm-ascend hub, frameworks index, wiki index, DeepSeek-V4 paper page, and v4-attention code-reading map.
  • Added the Lightning Indexer glossary term (category algorithms) defining the learned top-k block selector of compressed sparse attention, with cross-links from the DeepSeek-V4 paper page and the new C8 quantization page; registered it in the Terms Glossary.
  • Reinsighted vLLM-Ascend Architecture at the pinned 32a59d4e349c... revision: corrected the attention backend table (SFA is the DeepSeek-V3.2-style sparse-flash-attention backend selected when the model has index_topk without compress_ratios; DSA is the DeepSeek-V4 backend for compress_ratios models — not “SFA for V4 dense queries”), documented the (use_mla, use_sparse, use_compress) selection rule from get_attn_backend_cls, added the AscendFABackend FA3 row, and cross-linked the Lightning Indexer C8 page. Added five new evidence findings (attn-backend-selection, sfa-sparse-flag, use-sparse-flag, sfa-backend, dsa-backend) to the supporting derived note.
  • Added vLLM Architecture and Code Organization Overview to docs/frameworks/serving/vllm/, a code_links: strict / code_evidence: strict repository-backed code-reading map at the newest pinned mainline revision a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b (reused vllm-a0c092ee72c0). Start-here overview of the whole repository: a six-layer API-process → engine-process → device-process mental model (editable Mermaid at docs/frameworks/serving/vllm/assets/vllm-overview-layers.mmd), the top-level vllm/ and vllm/v1/ directory maps, where the V0 engine went (vllm/engine/llm_engine.py is now a V1 alias), component-by-component responsibilities (entry points, frontend engine, EngineCore step loop, scheduler + KV cache, executor, worker/model runner, model registry, attention backends, compilation/distributed/platforms/config), the eight-step request lifecycle, extension points, and a failure-surfaces table. Added 26 declared evidence findings in the new derived/repo-analysis/frameworks/serving/vllm/vllm/a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b/overview-module-map.md, registered the page in the a0c092ee72c0 manifest docs_paths, and linked it from the vllm hub, frameworks index, and wiki index.
  • Reinsighted vLLM-Ascend Architecture at the pinned 32a59d4e349c... revision, reframed around the new vLLM six-layer overview: a new “How vllm-ascend Maps Onto vLLM’s Six-Layer Stack” section (reuse layers 1-4 untouched, replace the hardware half of layers 5-6, with a per-layer reuse/replace table), a rewritten Big Picture overlay diagram (editable docs/frameworks/serving/vllm-ascend/assets/vllm-ascend-architecture.mmd), a new Worker Variants section (NPUWorker / NPUWorker310 / XliteWorker + CaMemAllocator), the 310P attention-backend collapse and AscendFABackend FA3 path, a “Where It Breaks” fragility table, and a “One Thing to Remember” takeaway. Added 8 new evidence findings (plugin-register, npu-worker, npu-model-runner, fa3-backend, cumem-allocator, acl-graph-wrapper, npu-worker-310, xlite-worker) to the supporting derived note and cross-linked the page from the vLLM overview.
  • Added a “Worker vs. Model Runner: Process Shell vs. Step Executor” subsection (§6.5) to vLLM Architecture and Code Organization Overview: the worker is the process-level shell and RPC surface (device init, memory profiling, pipeline-parallel send/receive, sleep/health), the model runner is the step-level executor inside it (batch build, forward, sampling, KV tensors, graph capture). Documented the exact delegation chain (Worker.execute_model → GPUModelRunner.execute_model, etc.) with 6 new evidence findings (worker-execute, worker-memory-profile, runner-execute, runner-load, runner-kv-init, runner-graph-capture), and mirrored the delegation note on the vllm-ascend side (NPUWorker.execute_model → NPUModelRunner.execute_model, 2 new evidence findings npu-worker-execute, npu-runner-execute).
  • Reorganized the flat scripts/ and tests/ directories into matching checks, wiki, ingestion, and repository domains; added shared repository-root discovery, command and test indexes, updated automation and instruction paths, and preserved stable top-level workspace, lint, and serving entry points.
  • Added DeepSeek-V4 Inference on Ascend: The DSA Serving Stack to docs/frameworks/serving/vllm-ascend/, a code_links: strict / code_evidence: strict repository-backed insight at the pinned vllm-ascend revision 32a59d4e349c12c32cdbc1916436c16e39939afc (reused the existing registered revision). Traces the full DeepSeek-V4 serving path on Ascend NPUs: register_model override to AscendDeepseekV4ForCausalLM, mHC hyper-connections via npu_hc_pre_v2/npu_hc_post at every attention/FFN boundary, the per-layer c4/c128 compressor + Lightning Indexer construction in DeepseekV4Attention, the AscendDSABackend/AscendDSAImpl prefill and decode flows (including the multi-stream _mla_prolog_multistream), the npu_sparse_attn_sharedkv sparse-attention custom op, _forward_o_proj with its A5 FP8 / OTP all-to-all / olora TP variants, the five-type heterogeneous KV cache (SWA, compressor state, compressed MLA, indexer keys/scales) with DSV4_BLOCK_SIZES paging, the AscendHybridKVCacheCoordinator patch, and the DeepSeekV4MTP draft model. Added 36 declared evidence findings to derived/repo-analysis/frameworks/serving/vllm-ascend/vllm-ascend/32a59d4e349c12c32cdbc1916436c16e39939afc/important-files.md, registered the page in the 32a59d4e manifest docs_paths, saved two editable Mermaid diagrams under docs/frameworks/serving/vllm-ascend/assets/, and linked the page from the vllm-ascend hub, frameworks index, vllm-ascend architecture, C8 indexer page, DeepSeek-V4 paper page, and the v4-attention code-reading map. Regenerated the docs link graph.

2026-08-06

  • Added Sarathi: Chunked Prefills for Efficient LLM Inference, sourced from raw/frameworks/serving/sarathi/sarathi-efficient-llm-inference-with-chunked-prefills-2308.16369v1--arxiv-2308.16369v1.pdf and its manually generated MinerU Markdown. Covers chunked causal prefills, decode-maximal batching, chunk-size and tile tradeoffs, pipeline-bubble reduction, reported throughput results, and extraction limitations at medium confidence.
  • Added the Chunked Prefill glossary term and linked Sarathi from the Frameworks and Wiki indexes.
  • Converted raw/benchmarks/agent-eval/agentic-hardware-design-repository-code-evolution--arxiv-2606.28279v1.pdf through the manually completed extraction workflow and normalized the derived source to derived/pdf-markdown/benchmarks/agent-eval/agentic-hardware-design-repository-code-evolution.md.
  • Added HORIZON: Agentic Hardware Design as Repository-Level Code Evolution, covering harness compilation, git-traced evaluator feedback, benchmark convergence, 209.9M token effort, coverage interpretation, reward-hacking risk, and production-flow limitations. Confidence is medium because MinerU precise conversion was unavailable in this environment and the paper evaluates controlled RTL proxies rather than full chip-design workflows.

2026-08-12

  • Re-insighted DSpark: Confidence-Scheduled Speculative Decoding from the manually completed extraction of raw/frameworks/speculative-decoding/dspark/dspark-confidence-scheduled-speculative-decoding--arxiv-2607.05147v1.pdf, normalized to derived/pdf-markdown/frameworks/speculative-decoding/dspark/dspark-confidence-scheduled-speculative-decoding.md. The page now preserves the paper’s original architecture, acceptance, calibration, production-frontier, and load-adaptive-budget figures; explains semi-autoregressive Markov/RNN heads, STS calibration, prefix-throughput scheduling, the two-step causal production delay, variable-length kernel routing, training losses, and the limits of strict-SLA comparisons.
  • Ingested raw/training/mhc-manifold-constrained-hyper-connections.pdf with MinerU precise mode to derived/pdf-markdown/training/architectures/mhc/mhc-manifold-constrained-hyper-connections.md and added mHC: Manifold-Constrained Hyper-Connections (high confidence) to docs/training/architectures/mhc/: constrains Hyper-Connections’ residual mixing matrix to the doubly stochastic manifold (Birkhoff polytope) via Sinkhorn-Knopp, restoring the identity-mapping property; 27B MoE final loss 0.021 below baseline, composite signal gain ~3000 → ≤1.6, 6.7% overhead via fused TileLang kernels, selective recomputation, and DualPipe overlap. Preserved seven original paper figures locally and linked an editable landscape Mermaid.
  • Added Hyper-Connections glossary term (category training) and registered it in the Terms Glossary; back-linked it from the new mHC page and the DeepSeek-V4 paper page.
  • Corrected the DeepSeek-V4 paper page’s mHC reference: the arXiv ID is 2512.24880 (not 2606.04201); the Go Deeper entry now links the internal mHC page.
  • Updated the training category index and Wiki Index with the new page and term.
  • Added Qwen3.5 / Qwen3.6 Inference Path on vLLM Ascend to docs/frameworks/serving/vllm-ascend/ (page renamed from qwen3.6-inference.md because Qwen3.6 and Qwen3.5 share one architecture), a code_links: strict / code_evidence: strict repository-backed insight at a new immutable vllm-ascend revision 9a52ca5fc36c1852241822863c50717bee5dc761 (upstream main, inspected 2026-08-06). Created via an explicit release-specific --force-new-revision override of the 14-day interval (previous snapshot 2026-08-03): Qwen3.6-35B-A3B weekly test configs and the Qwen3.6 support-matrix/docs updates exist only on this upstream tip, while the previously pinned 32a59d4e349c revision already contained the Qwen3.6-27B FIA test and Qwen3.6 docs. Traces how the shared qwen3_5-family models (Qwen3.5-27B / Qwen3.6-27B dense hybrid Mamba-Transformer, multimodal; Qwen3.5/Qwen3.6-35B-A3B and Qwen3.5-397B-A17B sparse MoE) run on Ascend by reusing the qwen3_5 / qwen3_5_moe model types with zero vllm-ascend model overrides: upstream vLLM qwen3_5 model classes, the patch_qwen3_5.py monkey-patches (AscendQwen3_5DecoderLayer, AscendQwen3NextAttention, GDN, MTP), AscendGatedDeltaNetAttention + qwen_gdn_attention_core + Triton FLA chunks, the FIA AscendAttentionBackend (npu_fused_infer_attention_score + full_graph_fia/full_graph_fia_v2 ACL-graph capture), ModelSlim W8A8 quantization, qwen3_5_mtp 3-token speculative decoding, and the Qwen3.6-27B multimodal FIA e2e test plus Qwen3.6-35B-A3B weekly configs (up to max-model-len 1010000). Registered the new revision in sources.json and docs/_data/code_repositories.json, added 21 declared evidence findings to derived/repo-analysis/frameworks/serving/vllm-ascend/vllm-ascend/9a52ca5fc36c1852241822863c50717bee5dc761/important-files.md, saved an editable flow Mermaid at docs/frameworks/serving/vllm-ascend/assets/qwen3.5-qwen3.6-inference-path.mmd, and linked the page from the vllm-ascend hub and wiki index.
  • Added bidirectional glossary validation: term pages remain the canonical registry, appears_in and “Where It Appears” must agree with consumer-page Markdown links, term aliases and glossary-index coverage are checked, and unlinked plain-text term mentions are reported with an optional strict mode through scripts/checks/term_links.py and the aggregate docs lint. Repaired the existing strict relationship errors and retained 160 legacy mention candidates as non-blocking cleanup warnings.
  • Added vLLM-Ascend Prefill and Decode Scheduling: Qwen3.5 GQA, backed by pinned vLLM a0c092ee72c0 and vLLM-Ascend 9a52ca5fc36c evidence. Explains the token-budget scheduler, chunked-prefill cursor, paged KV admission, Ascend attention-state classification, mixed FIA GQA layouts, and the parallel GDN recurrent path; added two evidence notes and editable runtime/landscape Mermaid sources.
  • Cleaned up all 168 legacy unlinked-term-mention warnings reported by scripts/checks/term_links.py across ~60 docs pages: added term links at the first meaningful in-content occurrence (KV Cache, PagedAttention, GEMM, MoE, Linear Attention, Lightning Indexer, All-Gather/All-Reduce/All-to-All, Sequence Parallelism, Microbatch, Scatter/Gather, Hyper-Connections, Kimi Delta Attention, Delta Rule, Inner Product, Block Table, Chunked Prefill, Continuous Batching, Global Memory, Matrix Tiling, Memory Banking, Kronecker Product), rewrote false-positive homograph matches (e.g., the verb “blocking” matching the Matrix Tiling alias) to preserve meaning, and registered every affected page in the matching term’s appears_in front matter and “Where It Appears” section so the bidirectional glossary contract holds. Re-ran the aggregate docs lint (./scripts/lint-docs.sh) with zero issues.

2026-08-07

  • Added Qwen3.5 MTP: Drafting and Target-Model Verification, backed by pinned vllm-ascend 9a52ca5fc36c1852241822863c50717bee5dc761 and vLLM a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b evidence. The page traces Qwen3.5 MTP proposal construction, target-logit row alignment, greedy equality verification, stochastic probability-ratio rejection sampling, bonus-token handling, and accepted-token state reconciliation. Added two editable Mermaid diagrams, manifest docs paths, navigation links, and repository evidence-map entries.

2026-08-09

  • Ingested Microscaling (MX) Formats: Block Floating Point for AI Hardware from the captured FPGA.org category page, preserving the raw HTML/metadata revision and derived Markdown under raw/hardware/ and derived/web-markdown/hardware/. Covers shared-scale block encoding, FPGA-era MSFP/Brainwave history, OCP MXFP4/6/8 and MXINT8 formats, implementation-defined layout and accumulation boundaries, and the mixed-format compute flow; classified the synthesis as medium confidence.
  • Re-insighted Microscaling (MX) Formats: Block Floating Point for AI Hardware from raw/hardware/numerics/quantization/microscaling-mx-formats--ocp-v1.0.pdf, preserving the precise derived extraction under derived/pdf-markdown/hardware/microscaling-mx-formats--ocp-v1.0/. Added the OCP format table, conversion and dot-product semantics, a local specification figure, and the Microscaling glossary page; retained medium confidence for the combined normative and historical synthesis.

2026-08-10

  • Ingested Context Parallelism for Scalable Million-Token Inference from raw/algorithms/distributed-inference/context-parallelism/context-parallelism-scalable-million-token-inference--arxiv-2411.01783v3.pdf using MinerU precise mode, preserving the normalized extraction under derived/pdf-markdown/algorithms/distributed-inference/context-parallelism/context-parallelism-scalable-million-token-inference.md. Added original pass-KV, pass-Q, and 128K-to-1M scaling figures, an editable landscape, and glossary terms for context parallelism, ring attention, GQA, tensor/pipeline parallelism, and FP8; classified the source-backed insight as high confidence.
  • Corrected the ACL graph section in vLLM-Ascend Architecture: the pinned implementation captures torch.npu.NPUGraph with torch.npu.graph(...) and replays it through the NPU runtime, while reusing vLLM’s CUDA-oriented graph modes and capture orchestration only as a compatibility layer. Documented the per-batch-descriptor cache and stable-input-address contract.
  • Ingested Accelerating PyTorch with CUDA Graphs from the immutable PyTorch web revision ad0c8a15d55eedc5a21ed2de60a90e0951cac920d48afc65c2413bba79610e42, preserving the raw HTML, metadata, derived Markdown, and two selected original figures locally. Added a beginner-first guide to launch-bound workloads, stream capture, warmup, stable tensor addresses, partial graphing, NCCL, benchmark interpretation, and graph failure modes; classified the historical API synthesis as medium confidence.
  • Registered TileLang commit 5a9b2a54b9a04831886d9b4dfd7a9ad758cd7ebd as immutable repository evidence and added TileLang Design and Code Learning Path (high confidence for static structure). The page traces eager/lazy Python elaboration into TIRX, target-neutral tile operations, backend context resolution, CUDA pass ordering, tile-op lowering, artifact and adapter construction, platform divergence, failure surfaces, extension points, and a staged study plan; preserved the repository’s original lowering overview locally and declared 17 revision-aware evidence findings.
  • Added TileLang-Ascend: Ascend Backend and TileLang Integration, backed by clean tilelang-ascend commit 34a048c19bd762381db0d0f2d5acfdf3527c459f on ascendc_pto and compared with pinned upstream TileLang 5a9b2a54b9a04831886d9b4dfd7a9ad758cd7ebd. The page maps Ascend C/PTO lowering, TVM source codegen, Bisheng/CANN compilation, Cython launch, package divergence, and the current upstream BackendModule integration boundary; runtime behavior remains statically inferred.
  • Ingested the NVIDIA CUDA Programming Guide introduction and Programming Model web captures, preserving immutable HTML/metadata revisions and derived Markdown. Added a hardware-grounded CUDA insight covering host/device execution, GPC/SM placement, grid/block independence, 32-thread SIMT warps, tile programming, and registers/shared/L1/L2/global/host memory; classified it as high confidence.
  • Ingested NVIDIA’s Tile IR Introduction and Programming Model web captures, preserving immutable HTML/metadata revisions and derived Markdown. Added a design-philosophy insight on tile blocks, tensor-first values, arbitrary pointer tensors versus structured tensor views, progressive GEMM specialization, and the compiler boundary for mapping tiles to threads, memory, and tensor cores.
  • Added vLLM DCP and PCP: Decode and Prefill Context Parallelism, backed by the clean pinned vLLM revision a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b and its immutable raw/derived repository evidence. The page separates PCP’s mirrored prefill batch partitioning and hidden-state restoration from DCP’s interleaved KV ownership, local sequence metadata, AG+RS/AG+AR/A2A exact LSE merges, cache block scaling, and current support boundaries; added two editable Mermaid diagrams and a glossary backlink.

2026-08-11

  • Added vLLM DCP Attention: From Local LSE to Exact Global Output, backed by the existing pinned vLLM revision a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b. The focused insight derives the stable LSE correction kernel, proves equivalence with ordinary attention, and contrasts AG+RS, AG+AR, and A2A output distribution.
  • Optimized the docs authoring workflow in AGENTS.md, the wiki synthesis workflow, and the page-content/repository-reading instructions: every rich element now requires a reader question and teaching job, representation choice follows the information relationship, runtime explanations use concrete state traces, callouts distinguish semantic roles, charts remain source-auditable, and decorative visuals are explicitly excluded.

2026-08-13

  • Added vLLM Prefill/Decode Disaggregated Deployment Path, backed by a new immutable vLLM revision 2d24355eb87b716fc1169e66731dc0386ed1a3a2. The medium-confidence static insight traces one request through router, prefill, NIXL KV handoff, decode admission, worker transfer, block leases, completion, and streaming; compares pull and push modes; preserves vLLM’s original scheduler/worker connector figure; and adds scaling, compatibility, observability, failure-policy, multi-turn, and rollout guidance.
  • Added Layer Normalization in Transformers from an immutable capture of an accessible article that explicitly cites and summarizes the requested Zhihu page; direct Zhihu capture was blocked by its verification layer. The medium-confidence insight contrasts BatchNorm and LayerNorm axes, corrects implementation ambiguities, preserves two original figures, and adds the Layer Normalization glossary term with a Transformer backlink.

2026-08-14

  • Reviewed and resolved all 56 remaining term-link warnings: added 47 meaningful glossary links across FP8, grouped-query attention, context/ring attention, KV cache, and tensor/pipeline parallelism; navigation-only labels are now recognized as existing links rather than glossary gaps.
  • Optimized glossary maintenance by making appears_in a curated teaching set, generating exhaustive consumer backlinks in terms.json, separating structural lint from explicit path-scoped mention review, and adding per-term mention_lint modes for common versus high-value concepts.
  • Expanded FlashAttention: IO-Aware Exact Attention with a Log-Sum-Exp view of online softmax: the L = m + log(l) identity, the stable logaddexp merge rule, why log space avoids overflow and is what the backward pass needs, and a cross-link to vLLM DCP attention.
  • Clarified in the same page why only the denominator L escapes the rescale (log turns ×rescale into +) while the numerator O always needs correction — as alpha in the in-loop max-rescale form or as exp(l_b - L) in the at-merge log-space form.
  • Added a derivation of the logaddexp merge form to the same page — why factoring out max(L_old, l_b) leaves a correction log(1 + e^{-d}) bounded in [0, log 2) — plus a worked merge trace.

2026-08-16

  • Added vLLM MHA Code Path: From QKV to Paged KV Cache, a medium-confidence code_links: strict / code_evidence: strict reading of decoder MHA/GQA/MQA at pinned vLLM revision 2d24355eb87b716fc1169e66731dc0386ed1a3a2. The page separates model-side QKV/RoPE/output projection, generic attention dispatch, per-iteration block/slot metadata, paged KV-cache writes, and backend kernel execution; includes one decode-token trace and 20 declared evidence findings. A scoped freshness check returned decision: defer because relevant upstream changes are not eligible for a new immutable snapshot until 2026-08-27.
  • Ingested raw/hardware/numerics/quantization/gptq-accurate-post-training-quantization--arxiv-2210.17323v2.pdf into GPTQ: Second-Order Weight Quantization at LLM Scale with a MinerU precise extraction, the paper’s original procedure figure, 175B-scale accuracy and latency evidence, and high confidence. Added glossary pages for GPTQ and Post-Training Quantization, plus a focused FlatQuant cross-link.
  • Added a slot_mapping deep-dive subsection to vLLM Block Table Management: the per-token int64 slot buffer, CpuGpuBuffer CPU/GPU/numpy layout, the Triton slot-computation kernel and slot = block * block_size + offset formula, the two distribution formats (slot_mappings_by_gid / slot_mappings_by_layer), and -1 padding — pinned to the existing vLLM revision dd11df04f3b7.

2026-08-17

  • Removed vLLM Code Learning Path and Request Flow from docs/frameworks/serving/vllm/vllm-code-learning-path.md and cleaned up its inbound links across the vLLM hub, category index, wiki index, and four glossary term pages.
  • Re-insighted vLLM DCP and PCP: Decode and Prefill Context Parallelism around the compositional ownership model: PCP partitions current-step query rows and restores token order, while DCP partitions persistent KV rows and restores exact softmax normalization. Added an ownership comparison and an evidence-linked mixed-batch state trace; a scoped freshness check deferred a new immutable vLLM revision until 2026-08-27.
  • Converted raw/frameworks/speculative-decoding/eagle-3/eagle-3-scaling-inference-acceleration--arxiv-2503.01840v3.pdf with MinerU precise mode into derived/pdf-markdown/frameworks/speculative-decoding/eagle-3/eagle-3-scaling-inference-acceleration/eagle-3-scaling-inference-acceleration.md, preserving the extracted source figures under the derived asset directory.
  • Added EAGLE-3: Training-Time Test for Speculative Decoding, the Speculative Decoding glossary term, manifest provenance, navigation links, and an editable Landscape Mermaid source. The page covers direct token prediction, multi-layer feature fusion, training-time test, dynamic draft trees, SGLang/vLLM results, and source-reported limitations.
  • Ingested raw/frameworks/speculative-decoding/dflash/dflash-block-diffusion-flash-speculative-decoding--arxiv-2602.06036v2.pdf with MinerU precise mode into derived/pdf-markdown/frameworks/speculative-decoding/dflash/dflash-block-diffusion-flash-speculative-decoding/ and added DFlash: Block Diffusion for Flash Speculative Decoding at high confidence. The page explains target-layer feature fusion, per-layer K/V injection, parallel masked-block drafting, sparse training masks, loss weighting, SGLang/vLLM results, and source-backed limitations; preserved four original figures and an editable Landscape Mermaid source, and linked the paper from DSpark and the Speculative Decoding glossary.
  • Ingested the EAGLE and EAGLE-2 PDFs with MinerU precise mode into normalized folder-backed extractions and added high-confidence EAGLE and EAGLE-2 insight pages. Preserved their mechanism, mask, and result figures locally; added editable Landscape sources, manifest provenance, family cross-links, and the Tree Attention glossary entry.

2026-08-18

  • Added GLM-5.2 on vLLM Ascend: Request-to-Backend Inference Path, a medium-confidence request round trip backed by vLLM a0c092ee72c0 and vllm-ascend 9a52ca5fc36c. The page follows OpenAI chat admission, V1 scheduling and KV allocation, the reused GlmMoeDsaForCausalLM model shell, Ascend SFA metadata and paged-cache writes, GLM-5.2 checkpoint-shared indexers, 2,048-token sparse selection, routed/shared MoE, target sampling and optional MTP, detokenization, response emission, and cleanup. Added 25 declared evidence findings, an editable sequence diagram, manifest mappings, and hub/index links. A scoped vllm-ascend freshness check found relevant changes but deferred a new immutable revision until 2026-08-20.
  • Added GLM-5.2 on vLLM: Request-to-GPU Backend Inference Path as the primary upstream analysis at new pinned vLLM revision 5c9ff5366b039a69b344773bdfead8466ed9a097. The page traces OpenAI admission, V1 scheduling and paged-cache allocation, CUDA runner metadata, the thin GlmMoeDsaForCausalLM model alias, shared sparse indexers, conditional Hopper/Blackwell sparse-MLA backends, routed/shared MoE, target sampling, detokenization, SSE emission, and cleanup. Added 35 declared evidence findings and an editable round-trip diagram; retained the vllm-ascend page as an explicitly labeled NPU companion.

2026-08-19

  • Ingested raw/benchmarks/agent-eval/rethinking-llm-evaluation-200x-less-data--arxiv-2510.10457v1.pdf with MinerU precise mode and added EssenceBench: LLM Evaluation with 200x Less Data at medium confidence. The page explains dual redundancy filtering, genetic subset search, attribution-guided refinement, HellaSwag ranking preservation, model-shift limitations, and the abstract-versus-table discrepancy in the 200x claim; added the Benchmark Compression glossary entry and an AutoJudger comparison.

2026-08-21

  • Ingested raw/hardware/numerics/quantization/quarot-outlier-free-4bit-inference-rotated-llms--arxiv-2404.00456v2.pdf (moved from raw/algorithms/2404.00456v2.pdf to the hardware category) with MinerU precise mode and added QuaRot: Outlier-Free 4-Bit Inference in Rotated LLMs at high confidence. The page covers fused Hadamard rotations via computational invariance, online FFN and attention rotations, Post-RoPE key caching, GPTQ/per-token/asymmetric quantization stages, end-to-end W4A4KV4 accuracy and performance evidence, and failure modes; preserved four original paper figures and an editable Landscape Mermaid source. Added the Hadamard Transform glossary term and cross-linked QuaRot from FlatQuant, NVFP4, Kronecker Product, and the GPTQ/PTQ term pages.

2026-08-22

  • Added Perplexity (PPL): From Next-Token Loss to Reliable Evaluation from the requested Zhihu article’s topic, using local browser-readable excerpts of the linked PyTorch explanation and Hugging Face’s official fixed-length perplexity guide after direct Zhihu capture returned HTTP 403/access-check JSON. The medium-confidence insight derives PPL from token NLL, corrects the linked companion’s missing-negative-sign shorthand, works through a three-token example, explains strided sliding-window evaluation, and records tokenizer/protocol limitations.
  • Added the Perplexity glossary term and linked it from the PPL, GPT-1, and GPT-2 pages; registered the two corroborating web sources in sources.json and updated algorithm/term navigation.
  • Ingested OpenTelemetry: Vendor-Neutral Observability Framework from the official What is OpenTelemetry? page captured at raw/frameworks/infrastructure/opentelemetry/opentelemetry--web-2026-08-22-c0e01c8e3d66.html; synthesized its instrumentation, signals, standards, OTLP, Collector, backend boundary, and extensibility model, with an editable Landscape Mermaid source and high confidence.
  • Ingested NVIDIA GPU Evolution: From Graphics to Accelerated Computing from the captured NVIDIA Technologies & Architectures page at raw/hardware/accelerators/nvidia/nvidia-technologies--web-2026-08-22-bfd1eecedea1.html; reframed the NVIDIA hardware topic around the Celsius-to-Blackwell timeline, capability branches, and a medium-confidence evolution study path with editable Mermaid sources.
  • Expanded NVIDIA GPU Evolution: From Graphics to Accelerated Computing with captured architecture sources for Volta, Ampere, Ada Lovelace, and Blackwell, plus NVIDIA’s official Hopper and Turing architecture-in-depth articles after the category links redirected; added a detailed six-generation comparison, source-status notes, and vendor-metric caveats.

2026-08-24

  • Added vLLM Mamba2 and Linear-Attention Prefix-Cache Path, a medium-confidence static trace at pinned vLLM revision 5c9ff5366b039a69b344773bdfead8466ed9a097. The page follows recurrent-state checkpoints from prefix-hash publication and its same-step visibility gate through hybrid-group hit reconciliation, align-mode allocation and state migration, Mamba2 or linear-attention suffix execution, and publication of the next checkpoint; added 20 declared evidence findings and an editable sequence diagram. A scoped freshness check returned decision: defer until 2026-09-01.
  • Ingested raw/hardware/accelerators/nvidia/nvidia-ada-lovelace-professional-gpu-architecture--nvidia.pdf and its existing MinerU extraction into NVIDIA GPU Evolution: From Graphics to Accelerated Computing, adding source-backed AD102 hierarchy, cache, RT Core, and professional-GPU details.
  • Added the dedicated NVIDIA Ada Lovelace Professional GPU Architecture insight from the precise MinerU extraction, preserving nine selected source figures, an editable landscape, full-chip versus RTX 6000 product scopes, and the FP8 cross-link.
  • Ingested Gray Code: One-Bit Transitions for Safe State Changes from the captured HandWiki Gray code article, focusing on BRGC’s unit-distance invariant, reflect-and-prefix construction, shift/XOR conversion, cyclic behavior, transition-sensitive hardware, and nearest-neighbor communication labeling; added an editable Landscape Mermaid source and recorded the limits of Gray code as non-correcting redundancy.

2026-08-25

  • Ingested the captured SemiAnalysis AgentX / InferenceX v3: Does the CUDA Moat Hold? page. Added a source-backed insight on long-context multi-turn trace replay, anonymized DAGs, KV-cache residency and routing, disaggregation, software-driven hardware comparisons, selected source figures, and an editable landscape; classified the synthesis as medium confidence because the corpus, metrics, and leaderboard are date- and workload-dependent.
  • Ingested the captured vLLM Data Parallel Deployment page and pinned vLLM revision 8fe9317f2e401aff6e13044098ac7f59e95dce97 into vLLM Data-Parallel Deployment: Internal and External Load Balancing. Added internal, hybrid, per-rank external, and multi-port external deployment strategies, current queue/KV-aware routing behavior, MoE dummy-pass synchronization, launch recipes, local source figures, and strict repository evidence.
  • Ingested raw/hardware/numerics/quantization/hif4-format-for-language-model-inference--arxiv-2602.11287v1.pdf with MinerU precise extraction into HiFloat4 (HiF4): 4-Bit Block Floating Point for LLM Inference. Added the original format, dot-product, and MSE figures, an editable Landscape Mermaid source, comparisons with MXFP4/NVFP4, the HiGPTQ/PTQ path, and glossary coverage for HiF4, block floating point, and NVFP4; classified the synthesis as medium confidence because training and native-kernel evidence remain open.
  • Re-insighted NVFP4: Blackwell 4-Bit Floating Point with the user-supplied Transformer Engine 2.18.0 NVFP4 documentation, captured as raw/hardware/numerics/quantization/nvfp4-te--web-2026-08-25-d98d1efd9b89.html plus metadata and derived Markdown. Added an explicit data-format contract for the E2M1 payload, FP8 E4M3 block scale, FP32 tensor scale, 1D/2D sharing granularity, and storage overhead; preserved the prior 2.19.0.dev0 capture because both revisions agree on the format.

2026-08-26

  • Added DeepGEMM MegaMoE: Fused Communication and Expert Compute, backed by the clean pinned deepseek-ai/DeepGEMM checkout at revision 559d79fb6994a58b8a15b4b93bf13ccc16edf247. The beginner-oriented insight traces symmetric-buffer preparation, conservative ring capacity, expert-parallel dispatch, remote token pulls, shape-specialized SM100 JIT, fused FP8xFP4 L1/L2 GEMMs with SwiGLU, optional shared experts, and top-k combine; it includes an editable Mermaid sequence diagram and strict revision-aware code evidence. Runtime behavior and performance remain static code-derived claims because no compatible GPU or multi-process CUDA run was available.

2026-08-27

  • Added Qwen3.8-FP8 on Ascend 950: Why Weights Become MXFP8 at Load Time, backed by clean pinned vllm-ascend revision 7702ccd7d8dea6b4dabdacb0118adb522dedbec7 and vLLM revision ca90b9e7d4e3ec670143e4b1822bb856ab0260cc. The medium-confidence static insight explains source block-scale resolution, Ascend 950 E4M3/E8M0 MXFP8 re-quantization, kernel layout repacking, runtime activation quantization, MoE handling, and non-950 fallbacks.

2026-08-28

  • Added Firecracker: MicroVM Sandboxes for AI Agents, backed by clean pinned revision f3f65a3425f62bc9cf5f1c81f2963f230ed89f9a. The medium-confidence code-reading insight explains KVM and Jailer containment, per-thread seccomp, VirtIO block/network/vsock boundaries, the API-to-boot round trip, and the outer agent-policy responsibilities Firecracker does not implement.

2026-08-29

  • Ingested raw/benchmarks/agent-eval/from-raw-experience-to-skill-consumption--arxiv-2605.23899v1.pdf with MinerU precise mode into derived/pdf-markdown/benchmarks/agent-eval/from-raw-experience-to-skill-consumption/ and added From Raw Experience to Skill Consumption: Model-Generated Agent Skills at medium confidence. The page covers the three-stage lifecycle, EE/TE metrics, success/failure experience composition, extractor-target asymmetry, surface-form and pairwise-judge controls, cross-model transfer, the validated failure-aware meta-skill, negative-transfer failure modes, five local source figures, and an editable Landscape Mermaid diagram; it records the paper’s prose-versus-Figure-3 guided-accuracy inconsistency.
  • Added the Agent Skill glossary page and linked it from the new SkillLens insight and Socratic-SWE as a cross-paper concept.
  • Ingested raw/benchmarks/agent-eval/skillopt-executive-strategy-for-self-evolving-agent-skills--arxiv-2605.23904v2.pdf with MinerU precise mode into derived/pdf-markdown/benchmarks/agent-eval/skillopt-executive-strategy-for-self-evolving-agent-skills/skillopt-executive-strategy-for-self-evolving-agent-skills.md and added SkillOpt: Executive Strategy for Self-Evolving Agent Skills at medium confidence. The insight explains the train/selection/test contract, bounded add/delete/replace edits, rejected-edit buffer, slow/meta update, harness transfer, 52/52 reported cell wins, compact learned artifacts, and source-backed limitations; preserved four original figures and an editable Landscape diagram, and added SkillOpt to the Agent Skill glossary examples.

2026-08-31

  • Added GLM-5.1 W4A4C8 on Ascend 950: How vLLM-Ascend Turns On Sparse C8, backed by the existing clean pinned vllm-ascend revision 7702ccd7d8dea6b4dabdacb0118adb522dedbec7. The medium-confidence static insight separates SFA main-cache C8 from Lightning-Indexer key/scale C8, explains the A5 FP8 E4M3 plus FP32 scale contract and V1/V2 cache allocation, documents the replacement of the old enable_sparse_c8 flag, and records the Ascend 950 SFA/context-parallel and P/D validation limits. Added an editable Mermaid toggle flow, 20 declared code-evidence findings, manifest/navigation links, and the official GLM-5.1 W4A4C8 validation issue as further reading.
  • Ingested raw/hardware/accelerators/ascend/ascend-950-npu-architecture-white-paper--paper.pdf with MinerU precise mode into derived/pdf-markdown/hardware/accelerators/ascend/ascend-950-npu-architecture-white-paper/ and added Ascend 950 NPU Architecture: Compute, Memory, and Unified Bus at medium confidence. The insight covers third-generation DaVinci Cube/Vector compute, HiF8 and MX formats, chiplet UMA memory, NDDMA, L2 hints, STARS2.0, DVPP, Unified Bus/URMA/CCU, IO forwarding, and supernode scaling; preserved six original figures, added an editable Landscape Mermaid source, registered the manifest and navigation, and added glossary terms for Chiplet, HiF8, SIMD/SIMT hybrid programming, NDDMA, STARS, Unified Bus, URMA, and CCU.
  • Ingested raw/benchmarks/agent-eval/autoskill-experience-driven-lifelong-learning--arxiv-2603.01145v2.pdf with MinerU precise mode into derived/pdf-markdown/benchmarks/agent-eval/autoskill-experience-driven-lifelong-learning/autoskill-experience-driven-lifelong-learning.md and added AutoSkill: Experience-Driven Lifelong Learning at medium confidence. The insight explains query rewriting, hybrid retrieval, query-only skill extraction, add/merge/discard maintenance, versioned merging, asynchronous serving, and the 1,858-skill WildChat analysis; it preserves two source figures, adds an editable Landscape diagram, links the Agent Skill glossary, and distinguishes extraction scale from unmeasured downstream utility.

2026-09-07

  • Ingested raw/algorithms/length-generalization/looped-transformers/looped-transformers-length-generalization--arxiv-2409.15647v5.pdf with precise MinerU extraction into derived/pdf-markdown/algorithms/length-generalization/looped-transformers/looped-transformers-length-generalization.md and added Looped Transformers for Length Generalization at high confidence. The insight covers n-RASP-L decomposition, shared decoder recurrence, full-output prediction, step-dependent supervision without intermediate labels, adaptive stopping, length-generalization evidence, limitations, three local source figures, and an editable Landscape Mermaid source.

2026-09-08

  • Added GLM-5.1 on vLLM: Expert-Parallel MoE Path, backed by the pinned upstream vLLM revision a0c092ee72c0dcefbb3b3e74f97ac62d842e5f4b. The static insight traces GLM-5.1’s GlmMoeDsaForCausalLM alias through EP group sizing, global-to-local expert maps, router top-k dispatch, local expert MLP execution, all-gather/reduce-scatter and specialized prepare/finalize branches, and residual return; it records the unrelated dirty-checkout edit and the unavailable multi-GPU runtime validation.

2026-09-09

  • Captured the Python 3.14 import-system reference as an immutable web source with derived Markdown, then added Python Import System: Cache, Finders, Loaders, and Module Specs. The insight explains the cache-first pipeline, ModuleSpec, meta and path finders, loader execution and rollback, regular versus namespace packages, relative imports, and __main__.__spec__; the manifest entry is now marked ingested and the editable Landscape source is saved under docs/frameworks/languages/python/assets/.
  • Captured the Python 3.14 execution-model reference as an immutable web source with derived Markdown, then added Python Execution Model: Code Blocks, Scopes, and Runtime State. The insight explains code blocks and frames, binding and name resolution, class and annotation scopes, lazy type evaluation, exception termination, and process/interpreter/thread ownership; the manifest entry is now marked ingested and the editable execution landscape is saved under docs/frameworks/languages/python/assets/.
  • Re-fetched Your Guide to the CPython Source Code through the integrated browser, preserved the rendered DOM as immutable web revision 5a5fa6d91d40, and added an insight covering the CPython 3.8 startup, parser, compiler, frame, object, standard-library, and test path with an original runtime swim-lane figure and editable Landscape.
  • Added InferenceX: Continuous Inference Benchmark Pipeline, backed by clean pinned revision 7fa1d14b6624d433bd4f72a540da739b8f3808d9. The static code-reading insight maps declarative configuration through matrix validation, fleet-specific launchers, fixed-sequence and AgentX workload branches, result normalization, collection, and downstream artifact handoff; it includes strict revision-aware code evidence and editable runtime and landscape Mermaid sources.

2026-09-10

  • Normalized raw/training/DeepSeek_V41_Tech_Report.pdf to raw/training/model-families/deepseek/deepseek-v4.1-technical-report--paper.pdf, converted it with precise MinerU extraction into derived/pdf-markdown/training/model-families/deepseek/deepseek-v4.1-technical-report.md, and added DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression. The insight covers CED prefill reduction, CSA2 Full/Reindex/Reuse modes, hierarchical indexing, FP4 main-KV caching, bounded SWA replay, Single-Pass mHC, Engram, DSpark, scalable agent post-training, reasoning-effort control, and local source figures plus an editable Landscape diagram.
  • Ingested raw/training/efficient-attention/yoco-you-only-cache-once--arxiv-2405.05254v2.pdf with precise MinerU extraction into derived/pdf-markdown/training/efficient-attention/yoco-you-only-cache-once/ and added YOCO: You Only Cache Once at high confidence. The insight explains the self-decoder/cross-decoder cache boundary, gated retention and sliding-window alternatives, prefill early exit, chunk parallelism, source-reported 1M-context efficiency, and paper-specific limitations; preserved six original figures, added an editable Landscape Mermaid source, and added the Gated Retention and Sliding-Window Attention glossary terms.
  • Registered the related raw/training/efficient-attention/universal-yoco-efficient-depth-scaling--arxiv-2604.01220v1.pdf source as captured and linked it from the YOCO: You Only Cache Once page for future synthesis; no derived extraction or second insight was created.
  • Ingested raw/training/efficient-attention/universal-yoco-efficient-depth-scaling--arxiv-2604.01220v1.pdf with precise MinerU extraction into derived/pdf-markdown/training/efficient-attention/universal-yoco-efficient-depth-scaling/ and added Universal YOCO for Efficient Depth Scaling at high confidence. The insight explains the Universal Self-Decoder loop, the T-independent global KV cache accounting against Universal Transformer and RINS costs, cross-decoder and loop-position ablations, loop, token, and parameter scaling, 300B-token MoE training and thinking-SFT results, and 8K–256K prefill, decode, and KV-memory measurements; preserved eight original figures, added an editable Landscape Mermaid source, and added the Looped Transformers glossary term. The manifest entry now records the derived extraction and ingested status.

2026-09-11

  • Normalized and ingested raw/training/efficient-attention/indexcache-accelerating-sparse-attention-cross-layer-index-reuse--arxiv-2603.12201v1.pdf with precise MinerU extraction into derived/pdf-markdown/training/efficient-attention/indexcache-accelerating-sparse-attention-cross-layer-index-reuse/ and added IndexCache: Cross-Layer Index Reuse for Sparse Attention at medium confidence. The insight covers Full/Shared layer execution, loss-guided training-free search, multi-layer distillation, 30B and preliminary GLM-5 evidence, four source figures, failure modes, and an editable Landscape Mermaid source; added the Cross-Layer Index Reuse glossary term and corrected the Lightning Indexer definition and DeepSeek-V3.2 long-context cost caveat.

2026-09-14

  • Ingested raw/frameworks/serving/kv-cache-optimization/kv-cache-optimization-strategies--arxiv-2603.20397v1.pdf with precise MinerU extraction into derived/pdf-markdown/frameworks/serving/kv-cache-optimization/kv-cache-optimization-strategies/ and added KV Cache Optimization Strategies for Scalable and Efficient LLM Inference at medium confidence. The insight organizes eviction, compression, hybrid memory, alternative attention, and combination methods; preserves three original figures, adds an editable Landscape Mermaid source, maps the methods to deployment scenarios, and links the existing KV Cache glossary entry.

2026-09-15

  • Ingested raw/frameworks/serving/kvswap/kvswap-disk-aware-kv-cache-offloading-on-device-inference--arxiv-2511.11907v2.pdf with precise MinerU extraction into derived/pdf-markdown/frameworks/serving/kvswap/kvswap-disk-aware-kv-cache-offloading-on-device-inference/ and added KVSwap: Disk-Aware KV Cache Offloading for Long-Context On-Device Inference at medium confidence. The insight explains disk-first KV-cache placement, low-rank key prediction, grouped critical-entry reads, the rolling and reuse buffers, logical cache mapping, offline tuning, device-dependent throughput/quality trade-offs, and source-backed failure modes; it preserves nine selected original source figures and an editable Landscape Mermaid diagram.
  • Added the KV Cache Offloading and Low-Rank Approximation glossary terms, linking the KVSwap insight and the existing KV-cache optimization and DeepSeek-V2 MLA pages as representative examples.
  • Captured https://isocpp.org/wiki/faq/references as immutable web revision 4aae33c31c6c9712708957020beb4c1d9912c0aa72626380ace54a13e9a68155 and added C++ References: Aliases, Binding, Lifetime, and Interface Design at medium confidence. The insight explains references as aliases, binding versus assignment, reference returns and method chaining, pointer/reference and value interface choices, null and lifetime boundaries, handles, and the historical this pointer; it adds an editable Landscape Mermaid and cross-links the forwarding-reference page.
  • Captured https://iies.in/blog/universal-references-in-c-explained/ as immutable web revision 243851bb3cafa1e33da2d384caf55d33aed2fb69378c787e15434c6fdde17ff7 and added C++ Universal References: Forwarding and Perfect Forwarding at medium confidence. The insight explains forwarding-reference terminology, value-category deduction, reference collapsing, named-parameter behavior, std::forward versus std::move, perfect forwarding, source-code extraction limitations, and failure modes; the new C++ hub and editable Landscape Mermaid are linked from the framework indexes.

2026-09-16

  • Reorganized all manifested evidence under category-relative topic_path directories across raw/, derived/, and docs/; added topic-aware ingestion and integrity checks, preserved moved page URLs with redirects, introduced intermediate topic hubs, and reduced the Wiki Index to category and topic entry points.

2026-10-09

2026-10-10

  • Model Architecture and KV Cache Paths analyzed five open-weight model architectures and cache paths against released configs and pinned vLLM 02aa10539f15 and vLLM Ascend 51b0242c0d3b revisions; Qwen3.8-Flash-Next Ascend implementation remained unverified in the pinned code. The pages were removed on 2026-10-17.
  • Added Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression from arXiv:2609.36322v1 at medium confidence. Registered the normalized PDF and local PDFKit extraction (MinerU precise mode was unavailable because no API token was configured), preserved original figures, and documented phase-conditioned retrieval gaps, phase specialization, controlled-model evidence, and theory limits.
  • Added the Chunked KV-Cache Compression glossary term and linked the phase-sensitivity insight from the attention hub and DeepSeek-V4/V4.1 pages.
  • Added MiMo-V2.6-Pro-MOPD: vLLM Inference and Deployment from the captured Xiaomi model card and vLLM recipe, documenting the vLLM 0.31.0+ MXFP4 requirement, H200 TP8 and MI355X TP4 launch profiles, optional DFlash configuration, and the stale V2.5 image reference.
  • Added MiMo-V2.6-Flash-MOPD: vLLM Inference and Deployment from the captured Xiaomi model card and vLLM recipe, documenting the 309B/15B model profile, 173 GB checkpoint and 208 GB VRAM floor, H200 TP4 and AMD backend branches, optional DFlash, and the older V2.5 image reference.

2026-10-17

  • Removed the Model Architecture and KV Cache Paths pages and their wiki links. The pinned repository source records and code reading notes remain cited by the vLLM and vLLM Ascend hubs.
  • Added DeepSeek-V4.1-Flash on NVIDIA vLLM: GPU Inference Path from the existing pinned vLLM commit 02aa10539f15c8c73aa7fd4a5ee0f691cbf88dcb. The strict code-reading map covers worker input to sampled-token output, V4.1 cache/indexer ownership, FlashMLA/FlashInfer/Mega dispatch, and bounded SWA replay; it records that the checkpoint config and GPU runtime were not inspected. Linked the page from the DeepSeek and vLLM hubs and both related DeepSeek pages.

2026-10-18

  • Deepened DeepSeek-V4.1-Flash on NVIDIA vLLM: GPU Inference Path after a reader-comprehension review: the TL;DR now motivates cache compression (cheap top-k candidate scoring versus dense bandwidth), the mental model states the SWA / compressed-KV / index-K precision hierarchy and the compressed-state coordinate space of top-k indices, the decode trace marks index units, section 4 explains the overlap rationale and the two-level candidate-block filtering, the cache-layout table notes the FP8/MXFP4 index-K record, and the bounded SWA replay section now covers the scheduler’s replay_start assignment, the PAD_SLOT_ID protection of prefix-cached KV, and the decoder replay sub-batch. All additions link the pinned vLLM commit 02aa10539f15c8c73aa7fd4a5ee0f691cbf88dcb.
  • Added an “Explain Mechanisms, Not Just Paths” section to .github/instructions/repo-reading.instructions.md, requiring motivation-first framing, explicit index coordinate spaces, state-hierarchy comparison, overlap rationale, both ends of cross-component handoffs, and staged-mechanism coverage in repository-backed docs.