use constraints-dev.txt in e2e tests
#3320
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
For instructlab, "pip install ." does not install vllm, but it does install an uncapped torch (2.7.0 currently). When we install vllm later, we compile a binary flash_attn wheel against torch 2.7.0. vllm 0.8.4 requires torch==2.6.0, so we downgrade torch, and then we use that with the incompatible flash_attn binary wheel. To resolve this, use constraints-dev.txt in the first pip install operation. This restricts torch to 2.6.0 immediately when we first install instructlab, so that we will compile flash_attn against that torch version. Signed-off-by: Ken Dreyer <[email protected]>
|
E2E (NVIDIA L40S x4) workflow launched on this PR: View run |
6 tasks
booxter
approved these changes
Apr 28, 2025
|
e2e workflow failed on this PR: View run, please investigate. |
Contributor
Author
|
This fixes the flash-attn problem. The e2e tests get further, but they still fail in NCCL timeouts. I've filed #3321 to track that separately. |
cdoern
approved these changes
Apr 28, 2025
Contributor
|
Force merging since Ken confirmed this improves situation even if it doesn't make CI green yet. |
Contributor
|
@mergify backport release-v0.26 |
Contributor
✅ Backports have been createdDetails
|
booxter
added a commit
that referenced
this pull request
Apr 30, 2025
…-3320 use `constraints-dev.txt` in e2e tests (backport #3320)
bbrowning
added a commit
to bbrowning/instructlab-sdg
that referenced
this pull request
May 5, 2025
This is a port of instructlab/instructlab#3320 over to the SDG repository. While doing so, I noticed we also were not using "-DGGML_CUDA=ON" so updated that as well, since it's the same pip install line in the file. Signed-off-by: Ben Browning <[email protected]>
mergify bot
pushed a commit
to instructlab/sdg
that referenced
this pull request
May 15, 2025
This is a port of instructlab/instructlab#3320 over to the SDG repository. While doing so, I noticed we also were not using "-DGGML_CUDA=ON" so updated that as well, since it's the same pip install line in the file. Signed-off-by: Ben Browning <[email protected]> (cherry picked from commit 225612c)
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
For instructlab,
pip install .does not installvllm, but it does install an uncappedtorch(2.7.0currently).When we install
vllmlater, we compile a binaryflash_attnwheel againsttorch 2.7.0.vllm 0.8.4requirestorch==2.6.0, so we downgradetorch, and then we use that with the incompatibleflash_attnbinary wheel.ImportErrorlooks like:To resolve this, use
constraints-dev.txtin the firstpip installoperation. This restrictstorchto2.6.0immediately when we first install instructlab, so that we will compileflash_attnagainst thattorchversion.