perf(cuda): enable hardware NVFP4/MXFP4 conversion on Blackwell builds - #1938

Merged

inureyes merged 13 commits into

mainfrom

update/issue-1934-blackwell-arch-specific

Sep 21, 2026

Commits on Sep 21, 2026

chore(ci): add a guard for Blackwell CUDA architecture lists
inureyes
committed
f930d25
perf(cuda): compile the hardware fp4 converter on Blackwell builds
inureyes
committed
a80bdb3
chore(ci): scope the fp4 SASS assertion to the one translation unit
inureyes
committed
6bfa3a7
docs: say which of the two Blackwell images actually runs
inureyes
committed
2c27521
chore(ci): select only the block-float tests that actually run
inureyes
committed
d1d5b49
chore(ci): rewrap the sm_70 job comment after the list edit
inureyes
committed
abf69d2
chore(release): keep the x86_64 build ceiling and justify it
inureyes
committed
a6eb12f
fix(ci): resolve cuobjdump from the CUDA toolkit, not from PATH
inureyes
committed
01fe5bb
perf(cuda): give only fp_quantize.cu the architecture-specific image
inureyes
committed
bffbf21
docs: record the interleaved A/B that refuted the decode regression
inureyes
committed
5bd1887
fix(build): do not inject an architecture nvcc was already given
inureyes
committed
1faf87b
docs(ci): update both gate comments for the per-source mechanism
inureyes
committed
e739da0
docs: note that every width-4 arm generated identical text
inureyes
committed
4123a95