perf(pt): DPA4-family performance optimizations - #6001

Open

OutisLi wants to merge 16 commits into

deepmodeling:masterfrom

OutisLi:pr/perf

Commits on Aug 26, 2026

perf(dpa4): accelerate SeZM inference on three backends
OutisLi
committed
75940e4
perf(dpa4): reduce SeZM training projection overhead
OutisLi
committed
ca53f10
feat(pt-expt): port the accelerated SeZM and DPA4C inference paths
OutisLi
committed
9360855
perf(dpa4): accelerate SeZM force-loss training
OutisLi
committed
bd5d25a
perf(pt): capture the HybridMuon step into one CUDA graph
OutisLi
committed
a0bd43e
perf(pt-expt): mirror the accelerated training paths onto the pt_expt backend
OutisLi
committed
73b2791
fix(dpa4): keep distributed training and CPU export off the accelerated seams
OutisLi
committed
d9af23a
[pre-commit.ci] auto fixes from pre-commit.com hooks
pre-commit-ci[bot]
committed
c29b293

Commits on Aug 27, 2026

perf(dpa4): optimize fused training kernels
OutisLi
committed
13ce22c

Commits on Aug 28, 2026

fix(pt): cache HybridMuon graphs by gradient owners
OutisLi
committed
618918a

Commits on Aug 29, 2026

fix(ci): stabilize DPA4 acceleration coverage
OutisLi
committed
072748b

Commits on Aug 30, 2026

fix(pt): stabilize accelerated execution paths
OutisLi
committed
8c1d7e8

Commits on Aug 31, 2026

fix(pt_expt): run cell graph search on CPU
OutisLi
committed
be614c7
fix(pt): harden fused DPA4 edge cases
OutisLi
committed
cd7466e

Commits on Sep 1, 2026

Merge remote-tracking branch 'upstream/master' into pr/perf
OutisLi
committed
179d9d7
test(common): make EMA schema fixture valid
OutisLi
committed
801ddfc