From 40955b28751e261dad7171b3c345c6d1fea9f0a9 Mon Sep 17 00:00:00 2001
From: Yuhao Yang <47235274+yhyang201@users.noreply.github.com>
Date: Tue, 19 May 2026 14:48:34 +0800
Subject: [PATCH] [Bug] Fix V4-Pro NaN on Blackwell by converting fp8_einsum
 input scale to ue8m0 (#25733)

---
 python/sglang/srt/models/deepseek_v4.py | 1 +
 1 file changed, 1 insertion(+)

diff --git a/python/sglang/srt/models/deepseek_v4.py b/python/sglang/srt/models/deepseek_v4.py
index b8df0c3135e2..16d42526b6b3 100644
--- a/python/sglang/srt/models/deepseek_v4.py
+++ b/python/sglang/srt/models/deepseek_v4.py
@@ -594,6 +594,7 @@ def forward(
                 o.reshape(T * G, D).contiguous(),
                 group_size=128,
             )
+            o_s = deep_gemm.ceil_to_ue8m0(o_s)
             output = torch.empty(T, G, R, device=o.device, dtype=torch.bfloat16)
             deep_gemm.fp8_einsum(
                 "bhr,hdr->bhd",