A diagnostic for testers and for the performance work, off by default. With STRATA_MANIFEST unset every call returns at once
and nothing is printed or written. With it set, the engine writes <file> (whole file, replaced by rename) when it is ready and
again after a request if something new was recorded; the launcher (strata_control/split_engine.py) writes the all-cards split
decision to <file>.launch.json. Sorted keys, no timestamps, process ids or addresses, so two matched runs give the same text.
Content:
- engine: version, engine kind (single GPU or layer split), stages, KV format, drafter, spec window, pipeline windows, PLE I/O.
- budgets: context, KV resident cells, expert slots and cache MiB (all tiers and primary), VRAM reserve, prefill chunk cap and
stages, pool workers, memory limit, resident budget, batch slots, conversation cache.
- paths: one entry per decision with selected/declined and the reason in plain words: prompt experts per format class (MMQ,
FP16 WMMA tiles, native fused kernels, FP16 GEMM, with the gate/up and down formats and the layer count), MMQ direct and group
sort, HC_Q8, the output head (VRAM or pinned host, repacked or not), route tail skip, DMA batch mode.
- tensors: dense and head formats as held (count, bytes), the packed and HC_Q8 copies made at load.
- launch: split-engine mode, engine chosen, reasons, card architecture and count, routed-expert types, and the AMD auto rule
condition by condition (architecture, card count, expert types), so four gfx1201 with the IQ3_S set is recognised and two or
three cards are not promoted by the four-card rule.
The only text it prints is "strata: STRATA_MANIFEST: cannot write <file>" on stderr, once, when the variable is set and the file
cannot be written.
Tests: tests/core/run_manifest_test.cpp (off is free and writes nothing, matched runs in any recording order give the same text,
replace-on-rerecord, escaping, dirty tracking, unwritable target), tests/split/test_split_engine.py::LaunchManifest.
- drafting.blocked_by(named=True): only --pipeline-windows still blocks; Metal keeps the old rule (no per-request drafters there)
- apply/launcher/plan pass the choice; the all-cards split engine runs --draft dflash2 --draft-model PACK only for DFlash2 by name AND --split-engine on
- the split engine's own DFlash2 start line counts as "DFlash2 on"; a split run config with --draft dflash2 is left as it is
- app Drafting hint and Terminal --drafting help say it (owner-approved wording, English only)
- tests: terminal drafting/single/split-parallel, split engine, serve parallel, node parallel-default
split_config writes "drafting": "mtp" (what the split engine runs), so serve/server.py's drafter step, which resolves Auto to DFlash2 where the
engine has it, cannot add DFlash2 flags to the split engine's arguments; the fallback keeps the person's own drafting word (or none: Auto), so the
layer split that starts instead gets DFlash2 as usual. The rule refuses (quietly in auto) for an explicit dflash2 or off, and for Auto on a card
where the split has not been measured to decode faster than the layer split with DFlash2. Tests: all four cases, the card rule, the server's
drafter step on a split config.
strata launch --split-engine auto|on|off (settings field, saved in the preset and written to the launch config, so an off survives updates);
STRATA_SPLIT_ENGINE=0/1 for testers. auto (the default) uses the split engine where the rule says yes and says nothing where it says no; on says
why it cannot run; off never. The rule also leaves the layer split for choices the split engine does not honour (DFlash2, custom engine arguments,
replicas, a conversation memory format other than int8). The engine is looked up per card count and architecture (strata-split-N-smCC, then
strata-split-N); sm_120 joins sm_86 as a tested architecture. Tests: modes, auto/on/off, silent refusals, the flag, a saved off.
EXPERT_TYPES gains IQ2_S, IQ3_XXS, IQ3_S, IQ4_XS, IQ4_NL and Q2_0; a set with such experts may have embeddings, shared experts and an output
head in any type the loader converts to Q8_0; the memory estimate counts the Q2_0 to Q4_0 repack twice; the per-layer embedding table type is
checked. Tests for the supported sets (the IQ3_S set, UD-IQ4_XS), the plain refusal lines and the scan.
Niko1221