E0011
Splitting one answer across two nodes
This page reports the record as entered. It is not an independent validation badge.
Research question
Can two Raspberry Pis cooperate on one deliverable, generating separate parts of a contact-form page in parallel, and return a valid result faster than one node?
Methodology
An opt-in contact-form experiment, outside the normal chat path, assigned markup and styling to separate child jobs, combined them locally and checked the result against a validation contract. It was not a model-sharding system.
Baseline
The single-node complete-page comparison in Part 6: the 1,024-token stylish page finished after 882 output tokens, with about 148.9 s of compute on Node 01 and 145.7 s on Node 02.
Interpretation
The scheduler achieved some parallelism, but the assembled page failed its validation contract: the expected .contact-submit CSS selector was missing, and a repair job did not fix it. Splitting a task adds interface contracts, assembly rules and repair overhead. Here it made the result slower and still invalid. Parallel independent inference was demonstrated; splitting a single answer was not shown to be better.
Limitations
A single attempt on one task. Not a model-sharding or distributed-decoding system, and not a training experiment.
Reproducibility
No repository artifact or raw log is attached. The figures are copied from the published journal entry.
Measured results
| Label | Value | Note |
|---|---|---|
| Child-job overlap | 27.9 s | The two child jobs ran at the same time for this long |
| Repair job | 111 s | Added time; did not produce a valid result |
| End to end | 153 s | Slower than the single-node complete page |