Module runner 

CLI generation, prompt processing and shared decode entry points.

Modules

decode 🔒

One complete verification step, resumable between shared-GPU turns.

diagnostics 🔒

Optional CPU-reference comparisons, research dumps and per-step traces.

Structs

PrefillResume 🔒

Functions

check_budget 🔒

prefill_engine 🔒

Stream equal-sized prompt chunks so a short tail does not reread the whole store.

prefill_prompt 🔒

Fill the trunk and draft caches, then return the next token and draft chain.

prefill_row_batches 🔒

Use the decode row kernels for short prompts and explicit row-path checks.

qwen4_exp_gen_once 🔒

run