From 3d1f3de190b168aa6e5568adfbe23f17755c9d59 Mon Sep 17 00:00:00 2001
From: biondizzle <biondizzle@gmail.com>
Date: Mon, 11 May 2026 22:08:50 +0000
Subject: [PATCH] =?UTF-8?q?fix:=20syntax=20error=20=E2=80=94=20move=20trit?=
 =?UTF-8?q?on=20imports=20before=20docstring,=20remove=20orphan=20@triton.?=
 =?UTF-8?q?jit?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

---
 patches/deepseek_v4.py | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/patches/deepseek_v4.py b/patches/deepseek_v4.py
index 23c51d3..50b8303 100644
--- a/patches/deepseek_v4.py
+++ b/patches/deepseek_v4.py
@@ -244,16 +244,16 @@ class DeepseekV4FP8Config(Fp8Config):
         return isinstance(layer, FusedMoE) and self.expert_dtype == "fp4"
 
 
-@triton.jit
+import triton
+import triton.language as tl
+import torch
+
 """
 NVFP4 staging kernel — full FP4 (E2M1) activations + UE4M3 block16 scales.
 
 The mxf4nvf4 PTX instruction requires BOTH A and B to be FP4 (E2M1 packed).
 This kernel quantizes BF16 activations → E2M1 packed uint8 with UE4M3 scales.
 """
-import triton
-import triton.language as tl
-import torch
 
 
 @triton.jit