https://posttrainllm.com/devlog Notes from building PostTrainLLM — kernel measurements, honest negative results, decisions made while AI-pairing on a Mac-local LLM factory. en-us Fri, 31 Jul 2026 05:50:09 GMT https://posttrainllm.com/devlog#devlog-redesign https://posttrainllm.com/devlog#devlog-redesign Unified the site into one instrument: a single dark design system, the loss-curve logo as the universal mark, machine-transparent claims, and structured-data SEO. Fri, 10 Jul 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-factory-loop https://posttrainllm.com/devlog#devlog-factory-loop The north star became one canonical run: target → data → post-train → eval → package → report, ending in a documented ship/reject/retry decision. It executed end-to-end for the first time on 2026-07-04. Sat, 04 Jul 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-dpo-saga https://posttrainllm.com/devlog#devlog-dpo-saga Three DPO runs to strip SQL wrapper prose: ref-free SimPO collapsed (exec 0.860 → 0.080), reference-anchored DPO cured the collapse (0.900), but clean-SQL rate stayed 0.000 — the wrapper is a base-model prior a rank-4 adapter can't override. Sat, 11 Jul 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-frontier-parity-4b https://posttrainllm.com/devlog#devlog-frontier-parity-4b A Qwen3-4B distilled on multi-turn file-ops trajectories reached 100% on the BFCL hard gate, up from 58%. The honest cost: out-of-domain breadth dropped 59.6% → 42.3% — real catastrophic forgetting. Ships as a routed specialist. Tue, 16 Jun 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-broken-eval https://posttrainllm.com/devlog#devlog-broken-eval A fine-tune looked like it beat its baseline — until the baseline turned out to be non-reproducible and the v1 fixture set was measuring the wrong thing. The whole factory's eval discipline came out of that miss. Mon, 01 Jun 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-mac-runtime https://posttrainllm.com/devlog#devlog-mac-runtime The factory is a native Swift/MLX CLI — 100+ subcommands, one binary. It serves an OpenAI- and Ollama-compatible endpoint, runs an agent loop with FSM-constrained JSON, and packages to MLX, safetensors, and CoreML. Sat, 20 Jun 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-curriculum https://posttrainllm.com/devlog#devlog-curriculum A ground-up curriculum: ten modules from 'a model is a parameterized function' to a self-improving factory, each with a toy exercise, a repo anchor, and a mastery gate. A coverage map ties every subsystem to a learning anchor. Thu, 25 Jun 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-memory64 https://posttrainllm.com/devlog#devlog-memory64 WebAssembly's -sMEMORY64=1 + -sWASM_BIGINT flags switch the module to 64-bit pointers, lifting the cap into the tens of GB. Allocated a 473M-parameter model end-to-end in a browser tab. Wed, 20 May 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-matmul-sweep https://posttrainllm.com/devlog#devlog-matmul-sweep Workgroup-shared tiling (16×16) + thread-level register blocking (4×4) gave up to 5.18× speedup. f16-packed storage was not additive with tiling. 8×8 register block lost at every size — register spill on Apple GPUs. Fri, 15 May 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-speed-evolution https://posttrainllm.com/devlog#devlog-speed-evolution From naive WASM to tiled+blocked WebGPU to Flash Attention 2: the cumulative speedup curve across presets, with loss-drift parity checks at every step. Mon, 25 May 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-fa2 https://posttrainllm.com/devlog#devlog-fa2 FA2 on Apple GPUs: a hand-written Metal kernel with the tiled+blocked structure from the matmul sweep. Measured forward-pass speedup and the end-to-end training parity check. Fri, 05 Jun 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-whats-next https://posttrainllm.com/devlog#devlog-whats-next The honest roadmap: what's proven, what's parked, and what the next real milestone is — not aspiration, but the next gate that produces a ship/reject decision. Wed, 01 Jul 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-integration-bug https://posttrainllm.com/devlog#devlog-integration-bug An end-to-end training loss diverged 30× despite every standalone kernel benchmark passing. The bug was in the integration — the end-to-end parity test became the bar. Mon, 18 May 2026 00:00:00 GMT https://posttrainllm.com/devlog#devlog-pair-programming https://posttrainllm.com/devlog#devlog-pair-programming Lessons from building the entire project AI-paired: where the AI helps, where it doesn't, and the workflow patterns that produced real results vs. wasted cycles. Sun, 28 Jun 2026 00:00:00 GMT