perf(dpa4): stop broadcasting weights across the node axis; fix use_amp serialization - #5960

Open

wanghan-iapcm wants to merge 11 commits into

deepmodeling:masterfrom

wanghan-iapcm:perf-dpa4-grid-contract

Commits on Aug 4, 2026

perf(dpmodel): contract the DPA4 grid-branch router with matmul
Han Wang
committed
7518a41
Revert "perf(dpmodel): contract the DPA4 grid-branch router with matmul"
Han Wang
committed
01c58e6
perf(dpmodel): stop broadcasting DPA4 so3 linear weights across nodes
Han Wang
committed
193b49e

Commits on Aug 5, 2026

perf(dpmodel): remove the remaining DPA4 broadcast-weight contractions
Han Wang
committed
7545961
fix(dpa4): serialize use_amp so a configured false is not silently ignored
Han Wang
committed
3c2b9bf
feat(pt_expt): honor enable_tf32 / DP_TF32_INFER like the pt backend
Han Wang
committed
99d33ea
perf(dpmodel): stop spelling the DPA4 grid router as a degenerate GEMM
Han Wang
committed
504bb24

Commits on Aug 6, 2026

Merge upstream/master into perf-dpa4-grid-contract
Han Wang
committed
1270762
docs: shorten the comments added by this branch
Han Wang
committed
ae72043
revert(dpmodel): restore master's GridBranch router line
Han Wang
committed
1574442
Revert "feat(pt_expt): honor enable_tf32 / DP_TF32_INFER like the pt backend"
Han Wang
committed
1e56cf6