Skip to content

Fetch from nvidia Megatron-LM #5

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Open
RaymondLi0 wants to merge 6,406 commits into
base: load-iter
Choose a base branch
from
Open

Fetch from nvidia Megatron-LM #5

RaymondLi0 wants to merge 6,406 commits into from
+655,069 −33,039

Conversation

@RaymondLi0
Copy link

@RaymondLi0 RaymondLi0 commented Aug 3, 2022

No description provided.

dimapihtar and others added 30 commits December 11, 2025 08:29
@dimapihtar
Remove flattened_range code paths for distributed optimizer checkpoin…
5ab481c 
…ting (#2126)

Signed-off-by: dimapihtar <[email protected]>
@dimapihtar
update commit (#2631)
5a24ff3 
Signed-off-by: dimapihtar <[email protected]>
@ko3n1g
tests: Disable grads test
f67b7bd 
Signed-off-by: oliver könig <[email protected]>
@AAnoosheh
Create separate teacher Layer Spec in KD mode (#2429)
44899aa 
Signed-off-by: Asha Anoosheh <[email protected]>
@lmcafee-nvidia
Dynamic context | Re-add max_requests arg. (#2488)
6b186c1
@lmcafee-nvidia
Inference | Fix entangled request generations. (#2584)
2ab9253
@dimapihtar
fix gpt3_mcore_reruns_resume_check_grads (#2646)
3a9f086 
Signed-off-by: dimapihtar <[email protected]>
@cuichenx @ericharper
Nemotron nano v2 vl changes for Megatron Bridge (#2078)
f5daa16 
Signed-off-by: Chen Cui <[email protected]>
Co-authored-by: Eric Harper <[email protected]>
@Phlip79
[docs] Migrate docs to new Sphinx (#2489)
4f700f7
@tdene
Add option to only log inference every N steps (#2637)
1c6f6eb
@Phlip79
[docs] Use autodoc2 and remove automodule (#2542)
0a59bea
@dimapihtar
add backward compatibility support for loading mcore 0.15 checkpoints (…
845617a 
…#2648)

Signed-off-by: dimapihtar <[email protected]>
@yeyu-nvidia
add offline eagle3 instructions to readme (#2246)
12b4406
@sidsingh-nvidia
Only initialize symmetric memory when needed (#2665)
fe7fb73
@ananthsub @chtruong814
Simplify parameter sync for checkpoint save (#2344)
e869218 
Signed-off-by: Ananth Subramaniam <[email protected]>
Co-authored-by: Charlie Truong <[email protected]>
@Phlip79
Update docstrings for dataset (#2666)
4a9b4a2
@cspades @shjwudp
[Megatron-FSDP] Support both old and new DeviceMesh APIs. (#2575)
597e88a 
Signed-off-by: Cory Ye <[email protected]>
Co-authored-by: jianbinc <[email protected]>
@sidsingh-nvidia
Enable hybrid tensor + expert + data parallelism in mcore inference (#…
ff45bd4 
…2470)
@sidsingh-nvidia
Fix failing functional tests (#2679)
43a0c33
@dimapihtar @yaoyu-33
M4 + Dist Checkpoint: Replace global parallel state with explicit gro…
5f5741d 
…up parameters (#2053)

Signed-off-by: dimapihtar <[email protected]>
Co-authored-by: yaoyu-33 <[email protected]>
Co-authored-by: Mcore Bot <[email protected]>
Co-authored-by: Yu Yao <[email protected]>
@dimapihtar
fix deprecated decorator import (#2680)
4bdd7b1 
Signed-off-by: dimapihtar <[email protected]>
@frsun-nvda
Added integration for Kitchen extensions' SDPA and FA implementations (…
36a9081 
…#2232)

Signed-off-by: Frank Sun (CHEF) <[email protected]>
@lmcafee-nvidia
Inference | Add request only if no paused requests. (#2600)
bacd164
@jalbericiola @ArEsKay3 @tdene
Pipeline parallelism fix in RL and sequence packing rewriting (#2632)
e9082fd 
Co-authored-by: Robert Kirby <[email protected]>
Co-authored-by: Robert Kirby <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: Teodor-Dumitru Ene <[email protected]>
Co-authored-by: Teodor-Dumitru Ene <[email protected]>
@Phlip79
Add oncall rotation (#2622)
815d86c
@salmanmkc
Upgrade GitHub Actions to latest versions (#2678)
bdc362a
@aschilling-nv
docs: Adding documentation.md to cover building documentation. (#2683)
2a8bcf0 
Signed-off-by: Andrew Schilling <[email protected]>
@Victarry
Add moe layer perf UT. (#2673)
cf39a4d 
Signed-off-by: Dennis Liu <[email protected]>
@cspades
[Megatron-FSDP] Build default FSDP DeviceMesh, and remove model arg f…
732bb8d 
…rom fully_shard_optimizer(). (#2471)

Signed-off-by: Cory Ye <[email protected]>
@Phlip79
[docs] Add ability to disable autodoc2 for local builds (#2669)
ae774fe
santhnm2 and others added 30 commits January 21, 2026 09:27
@santhnm2
Add a wrapper function for FA3 _flash_attn_forward call (#2933)
a3615d7 
Signed-off-by: Keshav Santhanam <[email protected]>
@ko3n1g
ci(hotfix): Fix unit tests on main
a1ba844 
Signed-off-by: oliver könig <[email protected]>
@ko3n1g
ci(hotfix): Tests on main
e4b18f7 
Signed-off-by: oliver könig <[email protected]>
@ko3n1g
chore: Set umask 0002 (#3027)
ba876ef 
Signed-off-by: oliver könig <[email protected]>
@mathemakitten
Make attn mask inversion in-place instead of allocating it again (#3019)
28c7221
@cspades
[Megatron-FSDP] Fix incorrect gradient scaling target. (#3023)
ba456fd 
Signed-off-by: Cory Ye <[email protected]>
@santhnm2
Ensure that last prefill chunk is handled correctly by Mamba models (#…
096dbeb 
…2897)

Signed-off-by: Keshav Santhanam <[email protected]>
@nschank
Replaces ModuleSpec with Protocols for some of the inputs to SelfAtte…
90e685b 
…ntion/CrossAttention (#2761)
@Phlip79
Update oncall schedule (#3017)
dd72aee
@jiemingz @buptzyb
Various CUDA graph improvements on capture time, replay time, memory …
8baf014 
…footprint (#2572)

Signed-off-by: Jimmy Zhang <[email protected]>
Signed-off-by: Jieming Zhang <[email protected]>
Co-authored-by: Robin Zhang <[email protected]>
@github-actions
chore: rotate oncall schedule
bbbedbb
@jon-barker
Add script for batch running CI tests across distinct nodes (#3047)
8c768ea
@ko3n1g
Revert "Various CUDA graph improvements on capture time, replay time,…
c8caeb2 
… memory footprint (#2572)"

This reverts commit 8baf014.
@wdykas @jon-barker
Refit EP support (#2972)
db7a8a8 
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: William Dykas <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: Jon Barker <[email protected]>
@tdene
Catch case of negative tokens to generate (#2985)
91ebe29
@ko3n1g
ci: Disable broken GRPO tests
27168e2 
Signed-off-by: oliver könig <[email protected]>
@Phlip79
Sync GitHub and Slack teams (#3037)
f85ad03
@nschank @yashaswikarnati
Support custom Router implementations in MoELayer (#2891)
7857383 
Co-authored-by: Yashaswi Karnati <[email protected]>
@chtruong814
ci: Remove Github transition comment from CI (#2881)
cfa240a 
Signed-off-by: Charlie Truong <[email protected]>
@ko3n1g
ci: Override N_REPEAT (#3051)
8db8d08 
Signed-off-by: oliver könig <[email protected]>
@shanmugamr1992
Supporting inference when called within an asyncio loop (#2816)
0136876
@JavaZeroo @yaox12 @Skylion007
Update type hints and doc strings for moe_utils.py (#2821)
03e0915 
Co-authored-by: Xin Yao <[email protected]>
Co-authored-by: Aaron Gokaslan <[email protected]>
Co-authored-by: Xin Yao <[email protected]>
@HaochenYuan @Phlip79
Remove calculation of padding token in moe routing loss (#2142)
10c6f01 
Co-authored-by: Philip Petrakian <[email protected]>
@jon-barker
Bug fix with --no-use-tokenizer-from-checkpoint-args (#3049)
029f48f 
Co-authored-by: Jon Barker <[email protected]>
@thomasdhc
Revert "Bug fix with --no-use-tokenizer-from-checkpoint-args (#3049)" (…
0683679 
…#3057)

Signed-off-by: Dong Hyuk Chang <[email protected]>
@santhnm2
Add health endpoint to dynamic text gen server (#3009)
93567e8 
Signed-off-by: Keshav Santhanam <[email protected]>
@github-actions
Update copy-pr-bot.yaml [skip ci]
3593301
@thomasdhc
ci: Skip test_precision_aware_optimizer (#3062)
30dea5d 
Signed-off-by: Dong Hyuk Chang <[email protected]>
@yaoyu-33 @shifangx
Support multimodule communication (#2031)
485ed18 
Co-authored-by: shifangx <[email protected]>
Co-authored-by: Mcore Bot <[email protected]>
@ko3n1g
Revert "Support multimodule communication (#2031)" (#3068)
0972f02
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

No reviews

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.