microsoft/DeepSpeed

Pull Requests Commits

Merge branch 'master' into features/rebase-quant-fp6

loadams committed 1 year ago

Verified 4a917880

MOE gate fixes and enhancements (#5156)

mosheisland committed 1 year ago

Verified 5a2e705b

Merge branch 'master' into features/rebase-quant-fp6

loadams committed 1 year ago

Verified f34312ad

Add the missed torch import.

JamesTheZ committed 1 year ago

bed775e1

Pin pydantic to latest version

loadams committed 1 year ago

f8454a08

Update version.txt after 0.13.5 release (#5229)

mrwyattii committed 1 year ago

Verified db70c183

Fix some review comments.

JamesTheZ committed 1 year ago

7e28144d

fix pydantic import

mrwyattii committed 1 year ago

e86611fc

Update pydantic version

loadams committed 1 year ago

e1a4ce04

fix fused_qkv model accuracy issue (#5217)

Yejing-Lai committed 1 year ago

Verified bc0d2465

revert testing to fix A6000 test

mrwyattii committed 1 year ago

Verified f6bcdee0

Update requirements.txt

mrwyattii committed 1 year ago

Verified 77f3883d

Update deepspeed/inference/v2/modules/implementations/linear/quantized_linear.py

mrwyattii committed 1 year ago

Verified fb8887c9

refactor fp6 tests, fix import error

mrwyattii committed 1 year ago

c2e6ebb9

Update deepspeed/inference/v2/config_v2.py

mrwyattii committed 1 year ago

Verified b15a1a10

Update deepspeed/inference/v2/config_v2.py

mrwyattii committed 1 year ago

Verified ed00ac92

Fix the GPU architecture requirement of FP6 kernel.

JamesTheZ committed 1 year ago

63489d17

Only compiler the kernel body for SM >= 8.0.

JamesTheZ committed 1 year ago

35c82f25

Remove debug files.

JamesTheZ committed 1 year ago

0ddbfd11

Refine split-k setting.

JamesTheZ committed 1 year ago

56eb8b90

Fix at::nullopt and at::optional conflicts.

JamesTheZ committed 1 year ago

118af370

Fix the CI workflow problem for FP6 end-to-end.

JamesTheZ committed 1 year ago

62a2d495

Update according to review comments.

JamesTheZ committed 1 year ago

a6d2f2f0

Deal with the subnormal FP6 and FP16 values and refine the UT.

JamesTheZ committed 1 year ago

c43947a2

JamesTheZ committed 1 year ago

20b543ca

JamesTheZ committed 1 year ago

6ed67f77

Fix the bug of subnormal FP6 casting and the 2bit/4bit tensor allocation.

JamesTheZ committed 1 year ago

b025c5ad

Update FP6 kernels.

JamesTheZ committed 1 year ago

f8e3acfb

Fix FP16-FP6 cast problems.

JamesTheZ committed 1 year ago

90b710d2

Add unit tests for FP6 quantizer

arashb committed 1 year ago

6c45a84b

Older