[dev] partial cuda graph support for thd format training. - #4359

Merged

HaochenYuan merged 27 commits into

HaochenYuan:thd_cuda_graph_dev

Jun 16, 2026

Commits on Jun 12, 2026

add cuda graph support for thd format training

Haochen Yuan
authored and
HaochenYuan
committed
e9265f5
add unit test
HaochenYuan
committed
81a89ed
fix & refactor pad-thd logic
HaochenYuan
committed
6e3e681
refactor
HaochenYuan
committed
00c51e1
refactor
HaochenYuan
committed
b056831
fix linting
HaochenYuan
committed
39fbb0e
fix linting
HaochenYuan
committed
258bd1e
fix linting
HaochenYuan
committed
cc928a9
fix CI
HaochenYuan
committed
4e01666
change UT cp size to avoid OOM
HaochenYuan
committed
60612ba
shorten the UT seqlen
HaochenYuan
committed
b061b75
fix UT
HaochenYuan
committed
23d8838
refactor padding
HaochenYuan
committed
685e986
skip length align when use cuda graph
HaochenYuan
committed
0b0e9fc
remove redundancy
HaochenYuan
committed
597f1d7
Apply THD CUDA graph refactor fixes
HaochenYuan
committed
6224b98
fix linting
HaochenYuan
committed
aa0cfae

Commits on Jun 16, 2026

fix graph capture slot for dynamic num_microbatch
HaochenYuan
committed
64ae18a
Add THD padding mask and aux loss fixes from Tailai Ma
HaochenYuan
committed
d98e432
fix CI
HaochenYuan
committed
0d41b56
fix linting
HaochenYuan
committed
3904a93
Merge upstream dev into thd_cuda_graph_dev
HaochenYuan
committed
67ef99f
- Use thd_max_packed_sequences to clarify dummy sequence capacity semantics
HaochenYuan
committed
5e0d88e
fix linting
HaochenYuan
committed
07a24f2
fix CI & dsv4
HaochenYuan
committed
00fa6cb
remove microbatch variance checking when using sequence packing
HaochenYuan
committed
30a727d
Merge branch 'dev' into thd_cuda_graph_dev
HaochenYuan
authored
8c012d8