Use kernels from the kernel hub #2988
Conversation
server/text_generation_server/layers/compressed_tensors/w8a8_int.py
Outdated
Show resolved
Hide resolved
danieldk
force-pushed
the
kernel-hub
branch
4 times, most recently
from
fac14af to
3726ab7
Compare
| @@ -90,7 +90,7 @@ mkShell { | |||
|
|
|||
| postVenvCreation = '' | |||
| unset SOURCE_DATE_EPOCH | |||
| ( cd server ; python -m pip install --no-dependencies -e . ) | |||
| ( cd server ; python -m pip install --no-build-isolation --no-dependencies -e . ) | |||
There was a problem hiding this comment.
This is to avoid downloading torch as a build dependency (since without build isolation it is used from the environment).
| @@ -230,3 +232,111 @@ def _pack_weight( | |||
| moe_weight.perm[expert] = weight.perm | |||
|
|
|||
| return moe_weight | |||
|
|
|||
|
|
|||
| def fused_marlin_moe( | |||
There was a problem hiding this comment.
I'd like to keep moe on the kernel hub as close to vLLM as possible, so moved this with our own extensions here.
| @@ -146,3 +159,110 @@ def _load_expert_weights_row( | |||
| assert all_weight is not None | |||
|
|
|||
| return all_weight | |||
|
|
|||
|
|
|||
| def fused_moe( | |||
There was a problem hiding this comment.
I'd like to keep moe on the kernel hub as close to vLLM as possible, so moved this with our own extensions here.
|
Something broke on intel-CPU ,it's most likely due to a bad import failing on ipex-cpu. (The failure is in the values, because the pure CPU+transformers fallback takes over). |
Fixed. All checks pass 🎉 . |
| "filename": "build/torch26-cxx98-cu118-x86_64-linux/moe/configs/E=8,N=4096,device_name=NVIDIA_H100_80GB_HBM3,dtype=fp8_w8a8.json", | ||
| "blob_id": "cc614e635ea57327c610ce79e99ae5339614f22e" | ||
| }, | ||
|
|
There was a problem hiding this comment.
Holy guacamole. Thatś a long file....
There was a problem hiding this comment.
I think we can forego the per-file hashes later?
What does this PR do?
Use hub kernels for paged attention, MoE, and quantization (Marlin, cutlass, etc.).
Fixes # (issue)
Before submitting
Pull Request section?
to it if that's the case.
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.