Add AutoEP #7938
tohtana
merged
93 commits into
deepspeedai:master
from
tohtana:tohtana/add_autoep
add autoep
bea50ef2
add checkpointing
2c041db4
fix format
c2a89bc5
add custom patterns
fd07c93a
fix optimizer resumption
cabfebcd
autoep: fix post-dispatch local expert permutation grouping
a2ab10dd
Merge branch 'master' into tohtana/add_autoep
e79c2e85
fix(autoep): restore ep_count helper
046db040
test(autoep): make checkpoint tests cpu-safe
71a0a36a
Fix AutoEP ZeRO-2 expert gradient scaling
fae0276f
fix(autoep): preserve manual backward parity
5f7dc1e7
fix(autoep): align combine path with grouped-mm baseline
90f86c78
feat(autoep): add selectable combine implementations
b207c5ef
Merge branch 'master' into tohtana/add_autoep
0e872cc5
Merge branch 'master' into tohtana/add_autoep
6b65486b
test(autoep): update combine_from_routed calls
2c873a9b
fix(autoep): support llama4 fused experts
240495c1
fix(autoep): match llama4 moe parity contract
3d49fbf4
fix(autoep): match qwen moe shared expert parity
424e38b1
Add AutoEP license attribution notices
851c2841
Support Qwen3.5 MoE AutoEP router capture
611645d6
Merge branch 'master' into tohtana/add_autoep
8820cafc
test(autoep): include qwen3.5 preset in coverage check
3212fda2
Merge branch 'master' into tohtana/add_autoep
5af8e41d
AutoEP: remove grouped_mm_backend surface
00aecdd5
AutoEP: align supported presets and Qwen3.5 guard
71a903d9
AutoEP: document Universal Checkpointing support
b59d3d4d
AutoEP: fix Universal Checkpoint load path
1104084d
AutoEP: guard selected preset detection
6aac5936
AutoEP: document Transformers preset compatibility
bbed7169
AutoEP: fix DeepSeek forward parity
c1d76843
AutoEP: move DeepSeek bias guards into specs
a1cdc2f9
Merge pull request #8 from tohtana/codex/autoep-pr7938-remove-grouped…
9268c5ec
docs: clarify AutoEP native MoE checkpoint compatibility
376cc098
AutoEP: delegate preset-specific parser behavior
1773510f
Merge pull request #12 from tohtana/codex/autoep-pr7938-deepseek-forw…
9592ed31
docs: integrate AutoEP universal checkpointing flow
3e73c964
docs: place AutoEP requirements after universal checkpoint steps
7012927d
AutoEP: align preset compatibility guard
489d4bf9
AutoEP: map Qwen2 model type to Qwen3 preset
52d32d2a
Merge pull request #9 from tohtana/codex/autoep-pr7938-supported-pres…
c1ec90cb
Merge remote-tracking branch 'origin/codex/autoep-pr7938-staging' int…
ba6f6c26
docs: distinguish AutoEP from DeepSpeed MoE docs
8432aa3b
Merge pull request #10 from tohtana/codex/autoep-pr7938-universal-che…
a6183d99
Merge pull request #11 from tohtana/codex/autoep-pr7938-universal-loa…
83f92611
Merge AutoEP staging branch
e463cdf6
Refactor AutoEP presets into registry package
fad3058a
Disable unsupported AutoEP expert bias
e08859ca
Address AutoEP preset registry review findings
da8074d6
Merge pull request #14 from tohtana/codex/disable-autoep-load-balance…
2b4e8047
Merge AutoEP staging expert-bias guard
224ec822
Merge pull request #13 from tohtana/codex/autoep-pr7938-preset-regist…
9d920fee
AutoEP: resolve routed scaling factor
068b59e2
Merge pull request #15 from tohtana/codex/autoep-pr7938-routed-scalin…
968ff9ca
AutoEP: skip AutoEP subtrees in AutoTP traversal
5bb74827
Merge pull request #16 from tohtana/codex/autoep-pr7938-autotp-subtre…
2d17f307
AutoEP: use sequential experts in CPU tests
8d43053c
Fix AutoEP review regressions
f8ff765d
Merge pull request #17 from tohtana/codex/autoep-pr7938-review-fixes
e6ed4268
Fix AutoEP PR7938 review follow-ups
9facfd59
Fix AutoEP HF output recorder retargeting
c019afff
Merge pull request #18 from tohtana/tohtana/autoep-pr7938-final-revie…
f9b90263
Merge upstream master for AutoEP review cleanups
f75f958e
AutoEP: support DeepSeek-V3 remote-code layout
da375f41
Merge pull request #19 from tohtana/codex/autoep-deepseek-v3-remote-c…
9d8c4a57
Fix AutoEP review cleanups
85ff3cc3
Merge pull request #20 from tohtana/tohtana/autoep-pr7938-review-clea…
58460d6d
Fix AutoEP PR 7938 review findings
a2f706e3
Fix AutoEP review followups
da2f1214
Trim AutoEP PR follow-up tests
6823e31f
Merge pull request #21 from tohtana/codex/autoep-pr7938-review-findings
6cf17970
tests/moe: compact autoep regression suite
89082cda
Merge pull request #22 from tohtana/tohtana/reduce-autoep-test-code
ff4ebf1b
Remove AutoEP backward loss multiplier
a05f0921
Use active MPU for AutoEP sequence-parallel size
71077a56
Merge pull request #23 from tohtana/tohtana/remove-autoep-loss-multip…
045b0617
Merge pull request #24 from tohtana/tohtana/fix-autoep-sp-size-before…
04b8abd4
fix(autoep+muon): support Moonlight noaux routing and grouped expert NS
6fbaedde
fix(autoep+muon): auto-fill HF config fields and restore batched NS
e94a8da1
fix(muon): restore @compiler.compile() on muon_update
6012befa
fix(muon): respect ns_method in expert group path
dde07ab8
Fix AutoEP + Muon compatibility for batched expert tensors
00edd290
Merge AutoEP Muon fixes
ea73adc2
Merge branch 'master' into tohtana/add_autoep
61f3dca6
Merge branch 'master' into tohtana/add_autoep
827b633d
Preserve AutoEP requires_grad flags
71205e37
Preserve AutoEP replacement dtype and device
b2e23a82
Format AutoEP dtype device test
31c4bd47
sfc-gh-truwase
commented
on 2026-06-10
sfc-gh-truwase
commented
on 2026-06-10
sfc-gh-truwase
commented
on 2026-06-10
Address AutoEP review feedback
ae351e70
Merge branch 'master' into tohtana/add_autoep
be534f45
sfc-gh-truwase
commented
on 2026-06-11
sfc-gh-truwase
commented
on 2026-06-11
sfc-gh-truwase
approved these changes
on 2026-06-11
Update AutoEP docs and supported presets
de8da26d
Clarify DeepSpeed MoE docs intro
2938a6b9
docs: clarify DeepSpeed MoE see also link
6b9ef4b4
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub