From 8ab734a06342ce9a517cf2fa55af14c8a62b0a27 Mon Sep 17 00:00:00 2001 From: Scott Roy Date: Tue, 21 Jul 2026 16:27:26 -0700 Subject: [PATCH] up --- examples/models/qwen3_5_moe/test_chunked_prefill.py | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/examples/models/qwen3_5_moe/test_chunked_prefill.py b/examples/models/qwen3_5_moe/test_chunked_prefill.py index 0a3f1e7c89a..00ec55e936e 100644 --- a/examples/models/qwen3_5_moe/test_chunked_prefill.py +++ b/examples/models/qwen3_5_moe/test_chunked_prefill.py @@ -113,7 +113,10 @@ def test_chunked_prefill_matches_unchunked(self): logits_chunk.to(torch.float32), logits_full.to(torch.float32), rtol=1e-2, - atol=1e-2, + # cross-hardware MLX accumulation noise on near-zero logits (the + # greedy-token assert above is the functional guard; a real chunk- + # boundary desync would diverge by orders of magnitude more) + atol=3e-2, )