tinygrad/tinygrad
2
0
Fork 3
mirror of https://github.com/tinygrad/tinygrad.git synced 2026-08-29 12:56:07 +00:00
Code Issues Packages Projects Releases Wiki Activity

kerenl opts test from nan in llama 8b (#17510)

Browse Source 
* all2all

* nan

* remove that

* less

* has_local

* only the nan change here

* use nice getitem syntax for INDEX

* work

* remove

* even simpler
This commit is contained in:
qazal
2026-08-13 04:07:30 +09:00
committed by GitHub
parent de04781b36
commit ed8297a102
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/opt/test_kernel_opts.py
+9
View File
@@ -239,6 +239,15 @@ class TestKernelOpts(unittest.TestCase):
helper_linearizer_opt(a.sum().exp(), [[Opt(OptOps.PADTO, 0, 32)],])
helper_linearizer_opt(a.sum(0).exp(), [[Opt(OptOps.PADTO, 1, 32)],])
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_local, "test requires locals")
@unittest.skipUnless(Device[Device.DEFAULT].renderer.has_shared, "test requires shared")
@unittest.expectedFailure
def test_padto_group_full_unroll_sum(self):
a = Tensor.ones(2, 28, 4096, dtype=dtypes.bfloat16).realize()
out = ((a * 0.5).float().square()).sum(axis=(0, 2))
opts_to_apply = [Opt(OptOps.GROUPTOP, 1, 256), Opt(OptOps.PADTO, 3, 32), Opt(OptOps.UNROLL, 2, 0), Opt(OptOps.UPCAST, 0, 7)]
helper_linearizer_opt(out, [opts_to_apply], check_default_opt=False)
def test_padto_sum(self):
N = 18
# NOTE: this setup prevents 17 * 17 contiguous merged into one dimension