Sync upstream v2.17 - #99

Closed

Caozhou1995 wants to merge 84 commits into

flagos-ai:v2.17.0from

Caozhou1995:sync-upstream-v2.17

Commits on Oct 16, 2025

Changed VERSION to 2.9.0
ptrendx
committed
966a5b9

Commits on Oct 17, 2025

[JAX] Fix imports in test for deprecated jax.experimental.pjit (#2274)

committed
739c656

Commits on Oct 24, 2025

Wheels for cuda 13 (#2278)

ksivaman
authored and
KshitijLakhani
committed
c2a643d
[JAX] NVFP4 recipe with option to enable/disable SR, RHT, and 2D quantization (#2270)

jberchtold-nvidia
authored and
KshitijLakhani
committed
7e72d41
Include TE core headers in final build (#2291)

ksivaman
authored and
KshitijLakhani
committed
9b75db3
Overhaul the compilation for the arch-specific features (#2279)

authored and
KshitijLakhani
committed
8b9849a

Commits on Oct 28, 2025

[PyTorch] Add max_logit support for MuonClip (#2195)

authored and
KshitijLakhani
committed
c4c185d

Commits on Oct 30, 2025

[PyTorch] Fix CI failures due to deterministic attention backend (#2288)

ksivaman
authored and
KshitijLakhani
committed
fa71964
[JAX] Fix: Skip determinism tests for bprop for all sm >=100 (#2315)

KshitijLakhani
and
pre-commit-ci[bot]
committed
fe9b150
[PyTorch] Fix attention backend and tests for `sm120` (#2320)

ksivaman
authored and
KshitijLakhani
committed
0acd0e7
[PyT] Bump the min version expected to supported FP8 current scaling determinism on Blackwell (#2316)

KshitijLakhani
and
pre-commit-ci[bot]
committed
9cc089a

Commits on Oct 31, 2025

[JAX] Ensure JAX reference impl uses an accurate backend in our tests (#2322)

jberchtold-nvidia
authored and
KshitijLakhani
committed
70f5366

Commits on Dec 11, 2025

[Version] Reset to TransformerEngine v2.9 (#5)

lxd-cumt
and
zhaoyinglia
authored
bae9d3a
Fix import bugs (#6)
lxd-cumt
authored
e13e38a

Commits on Dec 17, 2025

Fix flash-attention fallback failures (#7)
lxd-cumt
authored
ef41367

Commits on Dec 29, 2025

Multi-Backend Architecture Implementation for TransformerEngine-FL (#4)

lihongyang1990
and
panpy
authored
fd5f657

Commits on Jan 4, 2026

Add missing __init__.py files and policy test suite (#9)
lihongyang1990
authored
57adff4
Polish readme (#11)
lxd-cumt
authored
ec8edfc

Commits on Jan 6, 2026

Register get_attention_backend for all backends and fix FlashAttention fallback (#14)
lihongyang1990
authored
b26b226

Commits on Jan 7, 2026

fix nv shared lib bug. (#16)
lihongyang1990
authored
a423680

Commits on Jan 12, 2026

Add a new vendor implementation named hygon (#15)
wendell
authored
fbe34bd
Update the way the gems context is invoked in the FlagOS Backend (#18)
lxd-cumt
authored
396794e
Unify the usage of the gems context (#20)
lxd-cumt
authored
3d80e63
fix: torch SDPA backend multi-batch support (#17)
lihongyang1990
authored
f101d2c

Commits on Jan 13, 2026

Remove use_gems context and call flag_gems.xxx directly (#22)
lxd-cumt
authored
832a797

Commits on Jan 16, 2026

Add new vendor backend METAX (#21)
dinghaodhd
authored
08cabba

Commits on Jan 21, 2026

Add multi_tensor_adam_param_remainder and context parallel support (#23)

lihongyang1990
and
wenone766
authored
03d1998

Commits on Jan 22, 2026

Fix enum mismatch in plugins (#25)
lxd-cumt
authored
54390c7

Commits on Jan 25, 2026

add Vendor KUNLUNXIN (#27)
ssuurrffaaccee
authored
48c8480

Commits on Jan 26, 2026

Fix the incorrect registration on Kunlunxin (#29)
lxd-cumt
authored
de00a8a
Polish available check for kunlunxin (#30)
lxd-cumt
authored
35e1809

Commits on Jan 28, 2026

Add new register op get_attention_backend for METAX (#31)
dinghaodhd
authored
8690ab4

Commits on Feb 5, 2026

[iluvatar]add vendor/iluvatar backend (#35)
DannyP0
authored
b0a5934

Commits on Feb 10, 2026

Fix: Resolve parameter mismatch between TE_FL and NVTE functions (#34)
lihongyang1990
authored
12b2077

Commits on Mar 2, 2026

[CICD] Add workflows to validate TE QA test cases (#41)

authored
f808816

Commits on Mar 3, 2026

Refactor optimizer implementations and improve multi_tensor ops (#36)
lihongyang1990
authored
47e8ee7

Commits on Mar 11, 2026

tefl musa support (#42)
jiamingwang-mt
authored
acced6d

Commits on Mar 23, 2026

Add python-level patches to supporting multiple platforms (#49)
lxd-cumt
authored
4f54860

Commits on Mar 25, 2026

Add scaled_masked_softmax_forward/backward for flagos backend (#52)
lxd-cumt
authored
7f788a3

Commits on Mar 26, 2026

Fix quantizer dtype conversion errors (#54)
lxd-cumt
authored
1f98511

Commits on Mar 30, 2026

apply flagos te_groups_gemm op (#55)
chai-xiaonan
authored
2188137

Commits on Apr 2, 2026

[CICD] support Metax MACA workflow (#48)

authored
ebcfadc

Commits on Apr 9, 2026

[CICD] Upload unittest coverage report to FlagCICD platform && Access FlagCICD runner (#58)
BrianPei
authored
9d1c48a

Commits on Apr 24, 2026

[CICD] Refactor workflows, Add integration_tests, Switch to FlagCICD metax runner (#60)

authored
d7e9e7b

Commits on May 12, 2026

Add the new vendor backend ENFLAME (#61)

gongxijun
and
xijun.gong
authored
38bce13

Commits on May 13, 2026

add kunlunxin vendor op (#66)
sunge666-ui
authored
b75e354

Commits on May 15, 2026

TE-FL Upgrade: Synchronization with TE Release V2.14 (#62)

authored
badccf2

Commits on May 29, 2026

Fix op register errors when skip cuda (#69)
lxd-cumt
authored
9a686fa

Commits on Jun 5, 2026

Support bias for generic_gemm (#70)
wenqingqian
authored
3c34bb9

Commits on Jun 10, 2026

add flagos layernorm (#72)
wenqingqian
authored
25e80e9

Commits on Jun 11, 2026

add multi_tensor_compute_scale_inv_e8m0 and change call erro (#74)
sunge666-ui
authored
23b5013
add new unittest (#77)
zhaoyinglia
authored
b7f65d1

Commits on Jun 25, 2026

[fix] fix plugin te_general_grouped test bug (#81)
lihongyang1990
authored
9aa7e20

Commits on Jun 29, 2026

fix(hygon/platform): Improve Hygon library path resolution, fall back… (#82)

wangxshuai
and
wangyl166
authored
5dd31d7

Commits on Jul 1, 2026

Commits on Jul 9, 2026

Add FlagOS Triton fused RoPE kernels (#83)
lihongyang1990
authored
d83b4e7

Commits on Jul 13, 2026

hcu: implement multi_tensor_scale_tensor using multi_tensor_scale in transformer_engine_hygon 2.13 (#85)

wangxshuai
and
wangyl166
authored
3b5fbb5

Commits on Jul 23, 2026

feat(backend): support tsingmicro txda backend (#88)

authored
2cb485f

Commits on Jul 28, 2026

[Ascend] Integrate transformer_engine_npu && Fix the GEMM operator bug in the reference backend (#89)
Darryl233
authored
711887e

Commits on Jul 31, 2026

Integrate KunLunXin TE-FL backend patches (#84)
Jonahmoon
authored
5aa024d

Commits on Aug 4, 2026

[CICD] Add Ascend NPU unit test support (#91)

authored
7ce3fce

Commits on Aug 6, 2026

[CICD] Add BW1000 reference CI baseline and standardize plugin tests (#92)

AlexMa616
and
wkhylyh-debug
authored
ed37db3

Commits on Aug 7, 2026

[CICD] Add MUSA test workflow (#93)

authored
dea7cd6

Commits on Aug 10, 2026

Stage 3 complete: Merge upstream v2.17 into main
FlagScale Agent
committed
642108f
Stage 4: Add 2 utils bindings to pybind.cpp
FlagScale Agent
committed
c296127
Stage 5: Patch CUDA hardcoding to TE_DEVICE_TYPE
FlagScale Agent
committed
29c1da2
Stage 7: Fix setup.py missing InstallCommand import
FlagScale Agent
committed
a43c966
Fix 58 residual merge conflict markers from Stage 3
FlagScale Agent
committed
80778da
Remove temporary analysis files from sync process
FlagScale Agent
committed
f7b076f
Stage 4: Sync plugin API with upstream v2.17 pybind changes
FlagScale Agent
committed
5d01bea
Fix: Restore TE_DEVICE_TYPE and te_device_type() lost during merge
FlagScale Agent
committed
88a9532
fix: Restore te_device_type() for multi-backend support in DotProductAttention
FlagScale Agent
committed
a07604f
fix: Restore te_device_type() in permutation.py (partial, 10+ occurrences)
FlagScale Agent
committed
4f73cab
fix: Restore te_device_type() in triton/permutation.py and utils.py
FlagScale Agent
committed
402cc76
fix: Restore te_device_type() in distributed.py (9 locations)
FlagScale Agent
committed
45429d5
fix: Restore te_device_type() in quantization.py (7 locations)
FlagScale Agent
committed
9d632ed
fix: Restore te_device_type() in module/base.py (3 locations)
FlagScale Agent
committed
925121e
fix: Batch restore te_device_type() in 10 files (attention, jit, tensor, gemm, optimizers)
FlagScale Agent
committed
09ed9b1
fix: Replace hardcoded cuda with te_device_type() in layernorm_mlp.py
FlagScale Agent
committed
48e7727
fix: Fix syntax error in utils.py import statement
FlagScale Agent
committed
179930a
fix: Add te_device_type() to cpu_offload.py
FlagScale Agent
committed
2cb9180
fix: Add missing te_device_type() import in dot_product_attention.py
FlagScale Agent
committed
e9b40fe
fix: Correct te_device_type() import paths to match main branch convention
FlagScale Agent
committed
2bc1987