We read every piece of feedback, and take your input very seriously.
To see all available qualifiers, see our documentation.
1 parent fcc1020 commit f6eaba2Copy full SHA for f6eaba2
1 file changed
examples/models/qwen3_5_moe/test_chunked_prefill.py
@@ -113,7 +113,10 @@ def test_chunked_prefill_matches_unchunked(self):
113
logits_chunk.to(torch.float32),
114
logits_full.to(torch.float32),
115
rtol=1e-2,
116
- atol=1e-2,
+ # cross-hardware MLX accumulation noise on near-zero logits (the
117
+ # greedy-token assert above is the functional guard; a real chunk-
118
+ # boundary desync would diverge by orders of magnitude more)
119
+ atol=3e-2,
120
)
121
122
0 commit comments