xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-10 08:04:58 +08:00
Code Issues Packages Projects Releases Wiki Activity

[Model][AMD] ROCm support for 256 head dims for Gemma (#3972)

Browse Source
This commit is contained in:
James Whedbee 2024-04-10 10:12:00 -05:00 committed by GitHub
parent bd3c144e0b
commit 8b317c6dd0
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 2 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
vllm/attention/ops/triton_flash_attention.py
View File

@ -677,8 +677,7 @@ def check_args(
assert q.shape[-1] == k.shape[-1] and q.shape[-1] == v.shape[-1]
# TODO: Change assert if we support qkl f8 and v f16
assert q.dtype == k.dtype and q.dtype == v.dtype
# TODO: Fix assert to check head size <=256 once supported
assert head_size <= 128
assert head_size <= 256
assert o.shape == q.shape
assert (nheads_q % nheads_k) == 0
@ -729,7 +728,7 @@ class _attention(torch.autograd.Function):
o_strides = (o.stride(0), o.stride(2), o.stride(1), o.stride(3))
# Get closest power of 2 over or equal to 32.
unpadded_head_dims = {32, 64, 128}
unpadded_head_dims = {32, 64, 128, 256}
if head_size not in unpadded_head_dims:
padded_d_model = None
for i in unpadded_head_dims: