iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

[ROCm][CI] Use larger atol value for INT3 in test_quick_all_reduce.py (#50450)

Browse Source 
Signed-off-by: Dino Music <Dino.Music@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
music-dino
2026-07-30 21:41:44 -05:00
committed by GitHub
co-authored by Andreas Karatzas
parent d91f7af7d0
commit 5d5f22ed7a
1 changed files with 14 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/distributed/test_quick_all_reduce.py
+14 -4
View File
@@ -38,6 +38,14 @@ def on_gfx942() -> bool:
return False
def _quickreduce_tolerance(quant_level: QuickReduceRegime) -> dict:
# With INT3 inputs of absmax 23, per-rank error peaks at 2.75,
# doubled by the TP2 sum = 5.5.
if quant_level == QuickReduceRegime.INT3:
return {"atol": 5.5, "rtol": 0.1}
return {"atol": 2.5, "rtol": 0.1}
set_random_seed(42)
_test_size_rng = random.Random(44)
# Size over 8MB is sufficient for custom quick allreduce.
@@ -280,8 +288,9 @@ def graph_quickreduce(
out2 = tensor_model_parallel_all_reduce(inp2)
dist.all_reduce(inp2, group=group)
graph.replay()
torch.testing.assert_close(out1, inp1, atol=2.5, rtol=0.1)
torch.testing.assert_close(out2, inp2, atol=2.5, rtol=0.1)
tol = _quickreduce_tolerance(fa.qr_quant_level)
torch.testing.assert_close(out1, inp1, **tol)
torch.testing.assert_close(out2, inp2, **tol)
@ray.remote(num_gpus=1, max_calls=1)
@@ -309,14 +318,15 @@ def eager_quickreduce(
)
_assert_quickreduce(fa, inp)
out = fa.quick_all_reduce(inp)
torch.testing.assert_close(out, inp * tp_size, atol=2.5, rtol=0.1)
tol = _quickreduce_tolerance(fa.qr_quant_level)
torch.testing.assert_close(out, inp * tp_size, **tol)
inp = torch.tensor(
[1.0 * ((i) % 23) for i in range(sz)], dtype=torch.bfloat16, device=device
)
_assert_quickreduce(fa, inp)
out = fa.quick_all_reduce(inp)
torch.testing.assert_close(out, inp * tp_size, atol=2.5, rtol=0.1)
torch.testing.assert_close(out, inp * tp_size, **tol)
@ray.remote(num_gpus=1, max_calls=1)