From 420a8c1af55516c4f276a8fa33a2df58bf6b31b1 Mon Sep 17 00:00:00 2001
From: RichardWooSJTU <37864677+RichardWooSJTU@users.noreply.github.com>
Date: Thu, 16 Apr 2026 17:56:56 +0800
Subject: [PATCH] fix deep gemm import (#7425)

---
 fastdeploy/model_executor/models/deepseek_v3.py | 4 +---
 1 file changed, 1 insertion(+), 3 deletions(-)

diff --git a/fastdeploy/model_executor/models/deepseek_v3.py b/fastdeploy/model_executor/models/deepseek_v3.py
index 4d6e3ee1f8..c837e3f265 100644
--- a/fastdeploy/model_executor/models/deepseek_v3.py
+++ b/fastdeploy/model_executor/models/deepseek_v3.py
@@ -77,8 +77,6 @@ if current_platform.is_cuda():
         radix_topk_ragged_transform,
     )
 
-    paddle.enable_compat(scope={"deep_gemm"})
-
 
 class DeepSeekV3MLP(nn.Layer):
     """
@@ -700,7 +698,7 @@ class Indexer(nn.Layer):
         # indexer write_cache
         indexer_k_quant_and_cache(k, self.indexer_cache, slot_mapping, self.quant_block_size, self.scale_fmt)
 
-        import deep_gemm
+        from fastdeploy.model_executor.layers.quantization.fp8_utils import deep_gemm
 
         if forward_meta.max_len_tensor_cpu[1]: