cuda: enable topk-moe fusion for 288 experts (#25267)
Performance-sensitive backend path - could affect local throughput.
Independent publication // local-first AI // field reports from owned systems
Nodehome
A running publication about local models, private inference, self-hosted agents, weird hardware, research sweeps, and the builders wiring their own AI stack together.
AI is getting physical again. It shows up in terminals, racks, side projects, and ugly little workflows people actually control.
Latest
Daily Sweep - Jul 04, 2026clean up fragile code
Karpathy signal - watch for workflow shifts, repo movement, or vocabulary changes.
add inference benchmarking
Karpathy signal - watch for workflow shifts, repo movement, or vocabulary changes.
add a few tests
Karpathy signal - watch for workflow shifts, repo movement, or vocabulary changes.
Andrej Karpathy: logged GitHub activity for karpathy pushed nanochat (8 events)
Routine GitHub activity compressed for scanability.
unify the two optimizer implementations into one and prevent bugs lik...
Karpathy signal - watch for workflow shifts, repo movement, or vocabulary changes.
harden and shorten the execution sandbox
Karpathy signal - watch for workflow shifts, repo movement, or vocabulary changes.
fix deprecated commit on torchao
Karpathy signal - watch for workflow shifts, repo movement, or vocabulary changes.
Field Reports
builds, experiments, notesThree RTX 3090s, One 32B Model: A Pipeline-Parallel Canary
A current field note on why the 3x3090 serving path moved through pipeline parallelism, not tensor parallelism, for the tested 32B AWQ model.
Gemma 4 12B And The Sensory Agent Lane
A public-safe read on Gemma 4 12B as a local sensory preprocessor: useful for seeing, hearing, and structuring observations without turning into an action system.
Hardware
machines, thermals, economicsPower Caps On Three RTX 3090s: Bursts Versus Sustained Load
A measured note on 300W bursty inference, lower caps for sustained runs, and why power-cap sweet spots are workload-specific.
Parallel Agent Serving Is A Hardware Shape Now
A field-report read on 14x RTX 3090 agent serving, EXL3, FP8 KV cache, Aphrodite, and why concurrency is becoming the local hardware metric.