You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
ci: the f16 narrowing ratchet misses the vector form, so it lists 25 of the 52 bodies #2219
tools/ci/hrx_f16_narrowing.py (#2114) is the ratchet that lists every .loom body that narrows f32 to f16. Its match is the plain substrings f32 to f16 or truncf, so it misses the vector form:
%half_values = vector.fptrunc %values : vector<4xf32> to vector<4xf16>
Measured at the engine's llama.cpp pin 2b7960bae (corpus extracted with git archive, then --check --root):
Narrowing lines: 233 fptrunc lines; 101 are f32 to f16 and 132 are vector<Nxf32> to vector<Nxf16>. The corpus has no other fptrunc form.
Bodies: 52 contain a narrowing. The baseline lists 25.
So a new vector-form narrowing body can land without the ratchet failing, which is the one thing the ratchet exists to prevent. Found while reviewing #2204.
Claiming — the matcher is the fix and the baseline follows it.
The datum that decides the work is already in the issue: 233 fptrunc lines, 101 scalar (f32 to f16) and 132 vector<Nxf32> to vector<Nxf16>, with no other form in the corpus, and the baseline lists 25 of the 52 narrowing bodies. So the missing pattern is vector<…xf32> to vector<…xf16>, and the re-baseline should land at 52 bodies with the vector bodies included.
Plan: add the vector form to the matcher, regenerate the baseline, then --check --root on the extracted corpus and confirm it reports the new count with none new and none stale. A tools/ change needs no workflow scope, so it can land without waiting for an owner.
Activity
bong-water-water-bong commentedon Nov 11, 2026
Claiming — the matcher is the fix and the baseline follows it.
The datum that decides the work is already in the issue: 233
fptrunclines, 101 scalar (f32 to f16) and 132vector<Nxf32> to vector<Nxf16>, with no other form in the corpus, and the baseline lists 25 of the 52 narrowing bodies. So the missing pattern isvector<…xf32> to vector<…xf16>, and the re-baseline should land at 52 bodies with the vector bodies included.Plan: add the vector form to the matcher, regenerate the baseline, then
--check --rooton the extracted corpus and confirm it reports the new count with none new and none stale. Atools/change needs noworkflowscope, so it can land without waiting for an owner.ci: the f16 narrowing ratchet matches the vector form too (#2219)
ci: the f16 narrowing ratchet matches the vector form too (#2219)
tools: truncf is not a narrowing, and the comment said it was (engine…
tools: enumerate the three ways this corpus loses a value to half pre…
ci: the ratchet counts narrowing LINES per body, so a new one inside …
ci: the ratchet counts narrowing LINES per body, so a new one inside …
ci: the ratchet counts narrowing LINES per body, so a new one inside …