From b51d4a04f08a6a8cd1aabd2f95d4a0a9ae14abf0 Mon Sep 17 00:00:00 2001 From: rodamani Date: Thu, 1 Oct 2026 00:18:37 -0700 Subject: [PATCH] [DFlash] Keep grouped-conv taps inside each request's block so NaN/Inf cannot leak across requests (#41736) Co-authored-by: Rahul Chalamala <22563365+rchalamala@users.noreply.github.com> Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com> --- python/sglang/srt/models/dflash.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/python/sglang/srt/models/dflash.py b/python/sglang/srt/models/dflash.py index f3205f3de4..05b6f1f8d8 100644 --- a/python/sglang/srt/models/dflash.py +++ b/python/sglang/srt/models/dflash.py @@ -473,7 +473,8 @@ def _grouped_conv(hidden_states, delta, base, block_size, num_groups, group_size position = position % block_size for tap in range(1, taps): shifted = F.pad(blocks[:-tap], (0, 0, 0, 0, tap, 0)) - out = out + coefficients[:, tap] * shifted * (position >= tap).view(-1, 1, 1) + shifted = torch.where((position >= tap).view(-1, 1, 1), shifted, 0) + out = out + coefficients[:, tap] * shifted return out.flatten(-2)