-
Notifications
You must be signed in to change notification settings - Fork 725
Pull requests: pytorch/torchtitan
Author
Label
Projects
Milestones
Reviews
Assignee
Sort
Pull requests list
[BE][4/n] split pipeline_llama into a separate file
CLA Signed
This label is managed by the Meta Open Source bot.
#499
by tianyu-l
was merged Aug 5, 2024
Loading…
[fix] float8 should be applied on all model_parts
CLA Signed
This label is managed by the Meta Open Source bot.
#500
by tianyu-l
was merged Aug 5, 2024
Loading…
remove PP tracer
CLA Signed
This label is managed by the Meta Open Source bot.
#555
by tianyu-l
was merged Aug 22, 2024
Loading…
Add warning to compile rmsnorm
CLA Signed
This label is managed by the Meta Open Source bot.
#505
by wanchaol
was merged Aug 6, 2024
Loading…
add float8 to README
CLA Signed
This label is managed by the Meta Open Source bot.
#509
by weifengpy
was merged Aug 7, 2024
Loading…
[405B] Add performance data for 405B model
CLA Signed
This label is managed by the Meta Open Source bot.
#554
by fduwjj
was merged Aug 23, 2024
Loading…
address TODOs as 2D recompiles is fixed
CLA Signed
This label is managed by the Meta Open Source bot.
#508
by tianyu-l
was merged Aug 7, 2024
Loading…
Implement debug mode for GarbageCollection
CLA Signed
This label is managed by the Meta Open Source bot.
#1230
by fegin
was merged May 28, 2025
Loading…
Fix: update expert_bias buffer in-place to preserve it in checkpoint
CLA Signed
This label is managed by the Meta Open Source bot.
#1226
by trestad
was merged May 27, 2025
Loading…
[ft] Skip extra quorum when using semi-sync training
CLA Signed
This label is managed by the Meta Open Source bot.
#1221
by H-Huang
was merged May 27, 2025
Loading…
[DeepSeek] Enable data parallel
CLA Signed
This label is managed by the Meta Open Source bot.
#1003
by kwen2501
was merged Mar 21, 2025
Loading…
[deepseek] fix/avoid Expert Parallel device mesh and 'duplicate gpu mapping' error
CLA Signed
This label is managed by the Meta Open Source bot.
#1229
by lessw2020
was merged May 28, 2025
Loading…
Add emulate in float8 and relative checks
CLA Signed
This label is managed by the Meta Open Source bot.
#1214
by mori360
was merged May 28, 2025
Loading…
Refine GC logging
CLA Signed
This label is managed by the Meta Open Source bot.
#1234
by fegin
was merged May 28, 2025
Loading…
Add validation and batched inference to flux
CLA Signed
This label is managed by the Meta Open Source bot.
#1205
by CarlosGomes98
was closed May 27, 2025
Loading…
[SimpleFSDP] Add CI for SimpleFSDP
CLA Signed
This label is managed by the Meta Open Source bot.
#1231
by ruisizhang123
was merged May 30, 2025
Loading…
Correct CI test name
CLA Signed
This label is managed by the Meta Open Source bot.
#1239
by ruisizhang123
was merged May 30, 2025
Loading…
[FSDP2] reshard_after_forward=False for root model
CLA Signed
This label is managed by the Meta Open Source bot.
#1252
by weifengpy
was closed Jun 2, 2025
Loading…
[deepseek] remove unset root logger from HF AutoTokenizer to avoid dupe logging
CLA Signed
This label is managed by the Meta Open Source bot.
#1249
by lessw2020
was merged Jun 1, 2025
Loading…
[Flux] Fix broken symbolic link
CLA Signed
This label is managed by the Meta Open Source bot.
#1255
by wwwjn
was merged Jun 2, 2025
Loading…
[WIP][SimpleFSDP] Add support for hsdp/ddp + tp
CLA Signed
This label is managed by the Meta Open Source bot.
#1248
by ruisizhang123
was closed May 30, 2025
•
Draft
[PP] Re-enable zero bubble tests
CLA Signed
This label is managed by the Meta Open Source bot.
#1240
by H-Huang
was merged May 30, 2025
Loading…
Add H100 GPU node for integration test
CLA Signed
This label is managed by the Meta Open Source bot.
#1235
by mori360
was merged May 30, 2025
Loading…
[Extensions] ensure extension default values from extension toml are used, not base class
CLA Signed
This label is managed by the Meta Open Source bot.
#1244
by lessw2020
was merged May 30, 2025
Loading…
[NO REVIEW] dsa test
CLA Signed
This label is managed by the Meta Open Source bot.
#2327
by RenfeiChen-FB
was closed Feb 5, 2026
•
Draft
Previous Next
ProTip!
Follow long discussions with comments:>50.