microsoft/DeepSpeed

Pull Requests Commits

jeffra committed 4 years ago

a708b187

Update setup.py (#1361)

anukaal committed 4 years ago

Verified 8e577c92

fix unit test failure when merging PP ckpt files (#1359)

eltonzheng committed 4 years ago

Verified 9b915fed

Remove dropout as client code can do it independently. (#1354)

awan-10 committed 4 years ago

Verified 9f5939d2

Correctness fix PP+ZeRO for gradient accumulation (#1264)

jeffra committed 4 years ago

Verified b712babe

Update website hiring link

jeffra committed 4 years ago

Verified bff6126f

update hiring link

jeffra committed 4 years ago

Verified 168fce30

Update matmul.py (#1349)

hyunwoongko committed 4 years ago

Verified eb97a424

Added 4-byte alignment on NCCL/RCCL (#1328)

amathews-amd committed 4 years ago

Verified 2c628439

[actions] revert unit-test build name

jeffra committed 4 years ago

Verified 3e7d06ad

[actions] add torch version runner label

jeffra committed 4 years ago

Verified 0a32c3e2

[actions] add master to formatting trigger

jeffra committed 4 years ago

Verified 86b948f9

[actions] update branch triggers

jeffra committed 4 years ago

Verified 8f299be5

[actions] split formatting and unit tests into two jobs (#1339)

jeffra committed 4 years ago

Verified 600db091

Update main.yml (#1338)

jeffra committed 4 years ago

Verified 74f058b9

Use mpu in zero.Init() (#1325)

tjruwase committed 4 years ago

Verified e08c239c

Support client lr schedulers that are not subclass of torch _LRScheduler (#1337)

tjruwase committed 4 years ago

Verified 8e301b6e

Added drop_last to DeepSpeedDataLoader (#1321)

hxri committed 4 years ago

Verified c0b27fb0

Remove duplicate clip grad function in deepspeed (#1333)

awan-10 committed 4 years ago

Verified ddffbae0

update for cuda-11.4 (#1329)

stas00 committed 4 years ago

Verified b9ece257

Use clone to avoid checkpoint bloat (#1326)

tjruwase committed 4 years ago

Verified 336dd089

Activation checkpointing improvements (#1254)

tjruwase committed 4 years ago

Verified 85acf14c

use scalar cpu-adam in case of exception in builder (#1259)

RezaYazdaniAminabadi committed 4 years ago

Verified 9645e7bc

Add more synchronizations and barriers for the multi-gpu inference case (#1309)

RezaYazdaniAminabadi committed 4 years ago

Verified 0ec11daa

jeffra committed 4 years ago

c1b0a4e1

Reducing the memory-overhead of creating model for multi-GPU run (#1244)

RezaYazdaniAminabadi committed 4 years ago

Verified 49b6a632

Support Callable type for client optimizer and lr_scheduler (#1316)

tjruwase committed 4 years ago

Verified 274c375c

[zero] restore fp16 params if no zero ckpts available (#1322)

jeffra committed 4 years ago

Verified aa121291

Add issue templates

jeffra committed 4 years ago

10b48405

[docs] update moe features and news post

jeffra committed 4 years ago

6cd5f87b

Newer Older