Closed Eta0 closed 8 months ago
@wbrown Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/6962388870
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-a628650-nccl-cuda11.8.0-nccl2.16.2-1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn1.0.9
@wbrown Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/6962388870
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-a628650-nccl-cuda11.8.0-nccl2.16.2-1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn1.0.9
@Eta0 Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/7132227698
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-7ade069-nccl-cuda11.8.0-nccl2.16.2-1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn2.3.6
@Eta0 Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/7132227698
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-7ade069-base-cuda12.1.1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn2.3.6
@Eta0 Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/7132227698
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-7ade069-nccl-cuda12.0.1-nccl2.18.5-1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn2.3.6
@Eta0 Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/7132227698
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-7ade069-nccl-cuda12.2.2-nccl2.18.5-1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn2.3.6
@Eta0 Build complete, success: https://github.com/coreweave/ml-containers/actions/runs/7132227698
Image: ghcr.io/coreweave/ml-containers/torch-extras:es-limit-concurrency-7ade069-nccl-cuda12.1.1-nccl2.18.3-1-torch2.0.1-vision0.15.2-audio2.0.2-flash_attn2.3.6
Fix
xformers
& Limit Concurrency ontorch-extras
BuildsThis change updates
flash-attention
to v2.3.6, and removes theflash-attention
v1.0.9 builds.xformers
's bundledflash-attention
is disabled so that it uses the manually built one, becausexformers
's ownflash-attention
build currently segfaults on CUDA 12.1.This change also adds hard limits on the number of threads used during
torch-extras
compilation jobs to avoid excessive RAM use.