Tagged With
80 posts connected to this tag.
Sep 9, 2026
CKE hardened its evidence and Whisper runtime, culminating in persistent workers that cut a controlled 42:22 transcription from 171.60 to 138.73 seconds without changing retained outputs.
Read post →
Sep 6, 2026
Qwen3.8 Flash Next brought hyper-connections, Per-Layer Embedding, recurrent DeltaNet, query-selected sparse attention and mixed packed experts into one CKE circuit. This is what changed, wh...
Read post →
Sep 5, 2026
I have been writing about how quickly C-Kernel-Engine is adding model families. Qwen3.8 had already produced useful output in CKE, so I treated that support as progress we had earned. Then i...
Read post →
Sep 5, 2026
Yesterday I recorded the second video in my kernel engineering series. Since C-Kernel-Engine already had Whisper support, I wanted to use it to transcribe my own recording. Not another small...
Read post →Sep 2, 2026
A kernel-by-kernel walkthrough of how C-Kernel-Engine runs Whisper end to end in generated FP32 C: the eight signal-processing frontend kernels that turn a WAV into 1500 tokens, the full sti...
Read post →
Aug 26, 2026
A detailed look at CKE's expanded model-family matrix, what each architecture added to the kernel portfolio, and why support must be reported with an evidence level rather than one generic c...
Read post →
Aug 25, 2026
Two measured Qwen3.8-27B runs used a real 128K prompt on one Ryzen CPU to generate complete technical SVGs, then exposed CKE's next cache, traffic and distributed-computing targets.
Read post →
Aug 23, 2026
Qwen3.8 became an explicit CKE model circuit without a runtime fork or a new arithmetic kernel. Merged Ryzen bring-up and 4,000-step X-Ray certification show what reusable numerical provider...
Read post →
Aug 22, 2026
How CKE moved from token-row prefill sharding to a weight-sharded Qwen3.5 transformer block, what reconstructed correctly, and why one Ryzen still beats the current heterogeneous pair.
Read post →
Aug 22, 2026
A receipt-backed account of CKE's founder-funded Intel and AMD CPU lab, the experiments each node unlocks, and the controlled access path planned for collaborators.
Read post →Aug 21, 2026
CKE is not inference-only. Six DSL hardening PRs taught the compiler to reject quiet boundary mismatches: map interfaces, port-vs-ABI validation, ranked providers, and faithful traces — on...
Read post →Aug 20, 2026
CKE moved a Qwen3-VL OCR path from ~281s to ~76.7s against a llama.cpp baseline of ~55.8s — not a victory lap, a map. Revisited after the harness era: corpus certification, X-Ray parity, a...
Read post →Aug 5, 2026
CKE is not inference-only. The v7 training and v8 inference lanes share one compiler-interface discipline.
Read post →Aug 4, 2026
A green CI badge is an argument, not a decoration. CKE's nightly report now captures and publishes its runner's hardware identity, renders it as an amber identity plate, and teaches operator...
Read post →
Aug 3, 2026
How LinuxUtilities connects AwesomeWM, a native GTK4 control center, screenshots, audio, storage, presentation tools, AI prompts, and CKE work — including the path-contract bug that made t...
Read post →
Aug 2, 2026
The two days after the July books closed The July recap closed its Whisper chapter at a 2.7x gap to whisper.cpp: 1.53–1.58 seconds on the controlled Base fixture against 0.567 seconds for ...
Read post →
Aug 1, 2026
A source-linked July recap of 194 PRs and 32 articles: multimodal models, long-context numerical parity, compiler and test hardening, CPU evidence, contributors, and the path toward distribu...
Read post →
Jul 31, 2026
How CKE used exact head parallelism and packed quantized kernels to accelerate Qwen3.6-27B Q4_K_M prefill 2.6x on one 5th Gen Xeon node, while remaining about 1.85x behind llama.cpp.
Read post →
Jul 29, 2026
Follow a real CKE X-Ray investigation from a wrong Qwen3.6 token to a grouped recurrent-topology mismatch, including what passes and what remains open.
Read post →
Jul 28, 2026
Read Kimi K3 from its released config, model code, and technical report: the new kernels, the role of scale and training, and what a distributed CPU runtime would need to implement.
Read post →Jul 27, 2026
Memory Planning in C-Kernel-Engine In CKE, memory planning is not a nice cleanup pass after the kernels are written. It is part of the model contract: which buffer holds the current stream, ...
Read post →
Jul 26, 2026
An earlier article explained how an audio transformer hears: waveform to frames, frames to log-Mel features, features to encoder states, states to decoded text. This article is the implement...
Read post →
Jul 24, 2026
A closed OpenAI evaluation model compromised Hugging Face, while hosted models later refused forensic evidence. The incident strengthens the case for open weights and practical local CPU AI.
Read post →
Jul 23, 2026
Numerical parity is not one BF16 tolerance. Trace how CKE uses llama.cpp, PyTorch, X-Ray, replay provenance and phase-owned kernel contracts to locate the first divergence.
Read post →Jul 22, 2026
Derive softmax and softplus, trace Laguna's learned gate projection through full attention and backpropagation, and compare where modern model families use softplus.
Read post →
Jul 21, 2026
A source-backed walkthrough of how C-Kernel-Engine brings up GLM4 through model contracts, tensor mapping, pairwise partial RoPE, generated C, parity tools, and explicit evidence boundaries.
Read post →
Jul 20, 2026
Trace Qwen3.5's hybrid circuit, see how CKE X-ray finds numerical divergence, and separate exact runtime parity from the observed quality of a 0.8B model.
Read post →
Jul 19, 2026
Follow an LLM request from tokenization through prefill, KV-cache construction and decode, then see why CPU projections become GEMM or GEMV and how CKE compares them with llama.cpp.
Read post →
Jul 18, 2026
How CKE resolves model circuits to exact C kernel providers, preserves numerical contracts through IR, and generates correct stitched code.
Read post →
Jul 16, 2026
How PMU hardware, Linux perf_event_open, Intel Xeon and AMD EPYC counters, sampling, uncore monitoring, and profiler tools turn processor activity into reproducible optimization evidence.
Read post →
Jul 15, 2026
Why Linux profiler permissions, kernel symbol visibility, PMU compatibility, and application debug symbols are separate evidence boundaries.
Read post →Jul 14, 2026
A visual, source-linked derivation of Nemotron inference, training forward, backpropagation, and quantization across Mamba2, attention, ReLU2, and routed experts.
Read post →
Jul 13, 2026
How Nemotron mixes recurrent state, attention, dense MLP, and routed experts, and why exact routing becomes a CPU memory and scheduling problem.
Read post →Jul 12, 2026
How Gemma4-VL turns image patches into decoder-width visual tokens, why its vision bridge differs from Qwen3-VL, and how CKE is preparing compatibility.
Read post →Jul 11, 2026
How C-Kernel-Engine represents Gemma4 attention and per-layer RoPE from model metadata through IR, kernel maps, and generated C.
Read post →Jul 10, 2026
How C-Kernel-Engine model-family templates declare architecture, dataflow, lowering, memory, generated C, and evidence gates on Linux CPUs.
Read post →Jul 9, 2026
GGUF: The File Format That Made Local LLMs Practical Format deep dive GGUF is not just a container for weights. It is a single-file deployment unit that carries the model architecture, token...
Read post →Jul 8, 2026
A systems-performance note connecting Amdahl's Law, Theory of Constraints, transformer inference, multi-node CPU scaling, and how C-Kernel-Engine optimizes runtime paths by measuring the act...
Read post →Jul 7, 2026
A kernel-level reading of DeepSeek-V4: mHC residual lanes, CSA/HCA compressed attention, DSA sparse selection, MoE dispatch, low precision, and what those ideas mean for C-Kernel-Engine.
Read post →Jul 6, 2026
.jacobian-matrix-post .math-block { display: block; max-width: 100%; overflow-x: auto; overflow-y: hidden; -webkit-overflow-scrolling: touch; margin: 1rem 0; padding: 0.85rem 1rem; border: 1...
Read post →Jul 5, 2026
How Audio Transformers Work: The Encoder Path, Whisper, Timestamps, And Why Audio Is Not A VLM Patch Audio transformers use the same high-level trick as language and vision transformers: con...
Read post →Jul 4, 2026
C-Kernel-Engine architecture note DeltaNet and SSM/Mamba are two ways to reduce the amount of unbounded attention history a model has to carry by introducing fixed-size learned state. They a...
Read post →
Jul 3, 2026
C-Kernel-Engine architecture note Nemotron-style hybrid models are interesting to C-Kernel-Engine because they stop looking like a simple stack of attention plus MLP blocks. A runtime has to...
Read post →
Jul 2, 2026
ML systems · Quantization internals This ShivasNotes deep dive is the granular companion to the broader Quantization Deep Dive in this series. The earlier post explained why quantization ma...
Read post →Jul 1, 2026
C-Kernel-Engine runtime comparison This post is a follow-up to the earlier thread-pool deep dive. The previous post built the pthread idea from first principles. This one asks the systems qu...
Read post →
Jun 30, 2026
Distributed runtime note C-Kernel-Engine has been hardened first as a CPU-native runtime on single machines. The next phase is not just buying faster CPUs. It is learning how to make multipl...
Read post →Jun 29, 2026
Systems note This post moves from model math into the machine that actually runs the kernels: cores, caches, NUMA domains, MPI ranks, RDMA-capable fabrics, and the memory-layout discipline n...
Read post →
Jun 28, 2026
Muon vs AdamW explained from update mathematics to optimizer state, PyTorch's BF16 Newton-Schulz implementation, matrix-aware training kernels, memory cost, and a proposed C-Kernel-Engine pa...
Read post →
Jun 27, 2026
Kernel Engineers Are Worth Billions In 2026: Chris Lattner, Mojo, LLVM, and the Compiler Layer of AI Compiler systems note Chris Lattner's career is a reminder that compiler engineers and ke...
Read post →
Jun 26, 2026
Systems note A fast kernel is not just a loop. A fast kernel is a loop, a memory layout, a thread placement policy, a page-fault policy, and a measurement discipline running inside a real Li...
Read post →
Jun 25, 2026
KV cache memory for CPU-native inference This ShivasNotes deep dive is written for engineers who want to understand the single largest memory consumer in autoregressive LLM inference: the KV...
Read post →
Jun 24, 2026
v8 inference hardening This ShivasNotes deep dive is written for engineers who want to understand what happens between "here is a model on Hugging Face" and "here is a compiled .so that runs...
Read post →
Jun 22, 2026
CPU performance engineering · C-Kernel-Engine This ShivasNotes deep dive is written for CPU silicon teams and systems engineers who need to understand not just whether an AI framework is fa...
Read post →
Jun 21, 2026
ML fundamentals This ShivasNotes deep dive is written for CPU silicon teams asking a very specific question: does C-Kernel-Engine really own attention on CPU down at the kernel layer? The an...
Read post →
Jun 20, 2026
Quantization deep dive · ML fundamentals This ShivasNotes deep dive is written for CPU silicon teams evaluating whether C-Kernel-Engine really owns its quantized inference path. The answer ...
Read post →
Jun 19, 2026
ARM NEON · C-Kernel-Engine CKE has already run Qwen2 and Qwen3 on ARM Cortex-A72 hardware (TI TDA4VM). This is not a roadmap post. This is a “what we already ship” post. For ARM silicon...
Read post →
Jun 18, 2026
ML fundamentals · SIMD deep dive This ShivasNotes deep dive is written for CPU silicon teams evaluating kernel-level software competence. The subject is not “can the model run?” but “...
Read post →
Jun 17, 2026
Lab note This ShivasNotes lab note studies how a real vision-language model turns images into tokens that a language decoder can use. It connects earlier notes on activations, normalization,...
Read post →
Jun 16, 2026
Lab note This ShivasNotes lab note studies the alternative used by the Qwen3.5 / qwen3next recurrent path in C-Kernel-Engine: Gated DeltaNet, where the model stops trying to store every toke...
Read post →
Jun 15, 2026
Lab note Part of the ShivasNotes transformer-from-scratch series. Previously: dL/d(LLM): The Full Backward Pass. The full backward-pass post ended with every weight in the model holding a gr...
Read post →
Jun 14, 2026
AI Kernel Engineer Beginner Guide: Math, Linear Algebra, C/Linux Part 1: Beginner roadmap A viewer asked a useful question: “I know some C, some linear algebra, and some calculus. What are...
Read post →
Jun 13, 2026
Thread Pools in C: How CPU Runtimes Dispatch Work Across Cores Runtime note Companion note for the threadpool video. This post zooms into the local runtime layer: how persistent workers rece...
Read post →
Jun 12, 2026
A capstone walkthrough of the full LLM backward pass: loss to LM head, final norm, decoder layers, attention, FFN, residual splits, embedding scatter-add, training loop, AdamW, and C-Kernel-...
Read post →
Jun 11, 2026
Lab note Previously: Attention: The Core Of The Transformer. The previous post showed how attention routes information sideways across a sequence. This post is about the architectural trick ...
Read post →
Jun 10, 2026
Attention is the core transformer mechanism: Q/K/V projections, head splitting, RoPE, scaled dot-products, masking, softmax, weighted value sums, GQA, Flash Attention, and the full backward ...
Read post →
Jun 9, 2026
Lab note Companion post to the Tokenization carousel. Previously: Positional Encoding: Teaching Transformers Where To Look. Tokenization is the first place where text stops being human langu...
Read post →
Jun 8, 2026
Lab note Companion post to the Positional Encoding carousel. Previously: LayerNorm And RMSNorm: Stabilizing The Signal. Transformers are brilliant at reading an entire sequence in parallel, ...
Read post →Jun 7, 2026
Lab note Companion post to the LayerNorm & RMSNorm carousel. Watch the companion video: Why LayerNorm and RMSNorm Matter in a C LLM Runtime. Previously: Activation Functions: The Nonlinearit...
Read post →Jun 6, 2026
Lab note Companion post to the Activation Functions carousel. Previously: Softmax: The Probability Engine. The previous Softmax post ended with a strange contrast. Softmax was beautifully sm...
Read post →Jun 5, 2026
Lab note Companion post to the Softmax carousel. Previously: Matrix Wx+b: From Scalars To Transformers. The previous post was about the moment a scalar weighted sum grew into a matrix multip...
Read post →
Jun 4, 2026
A step-by-step explanation of how scalar wx+b becomes matrix Wx+b, how shape rules govern transformer layers, and why GEMM is the physical kernel underneath modern models.
Read post →
Jun 3, 2026
Lab note This is the companion post to today's backprop-wxplusb carousel. I wanted a version with more room for the pictures, the algebra, and the engineering intuition. Previous post in thi...
Read post →
May 31, 2026
Why I keep building systems from scratch: Antsand, C Kernel Engine, and my flight-controller stack all come from the same instinct to understand primitives, remove bloat, think like a compil...
Read post →May 31, 2026
A research-style report on training small C Kernel Engine models to emit SVG-oriented DSLs, showing why lower loss is not enough, how tokenizer and compiler contracts shaped the work, and wh...
Read post →
May 29, 2026
I am slowly moving more of my engineering life back into Linux. Not as a slogan. As a workflow. This is part two of my hardening series. Part one was Hardening Long Projects in the Age of A...
Read post →
May 26, 2026
Abstract / Executive Summary C Kernel Engine is an AI LLM inference and training engine optimized for CPUs. It takes model weights, model configuration, and computational graph templates, th...
Read post →
May 24, 2026
For the last few months, I have been noticing a strange shift in my work. I am still building software. I am still thinking about AI runtimes, Linux, C Kernel Engine, Antsand, content system...
Read post →May 20, 2026
Qwen3 is not just a model name inside C Kernel Engine. It is a concrete runtime target: a GGUF file, a model configuration, a set of tensor shapes, a generated C runtime, a packed weight blo...
Read post →May 15, 2026
Images generated by chatGPTI have been thinking a lot about hardening lately, especially with Antsand. Not hardening in the dramatic cybersecurity movie sense. I mean the boring kind of hard...
Read post →Apr 21, 2026
Last week, Claude Opus 4.7 "got released." And the Anthropic team is doing what Americans are genius at: marketing, marketing, marketing. The whole Claude, OpenAI, Gemini saga is the same pl...
Read post →Get my rants delivered to your inbox