[sgl-kernel] misc: update deepgemm version for sgl-kernel

.github/workflows/pr-test-sgl-kernel.yml

@@ -38,6 +38,8 @@ jobs:
         include:
           - python-version: "3.10"
             cuda-version: "12.4"
+          - python-version: "3.10"
+            cuda-version: "12.8"
           - python-version: "3.10"
             cuda-version: "12.9"
     name: Build Wheel (CUDA ${{ matrix.cuda-version }})

python/sglang/srt/layers/moe/ep_moe/layer.py

@@ -248,7 +248,6 @@ def forward_deepgemm(
             gateup_output,
             masked_m,
             expected_m,
-            recipe=(1, 128, 128) if deep_gemm_wrapper.DEEPGEMM_BLACKWELL else None,
         )
         del gateup_input
         del gateup_input_fp8
@@ -304,7 +303,6 @@ def forward_deepgemm(
             down_output,
             masked_m,
             expected_m,
-            recipe=(1, 128, 128) if deep_gemm_wrapper.DEEPGEMM_BLACKWELL else None,
         )
         del down_input
         del down_input_fp8
@@ -667,7 +665,6 @@ def forward_deepgemm_masked(
             gateup_output,
             masked_m,
             expected_m,
-            recipe=(1, 128, 128) if deep_gemm_wrapper.DEEPGEMM_BLACKWELL else None,
         )
         dispose_tensor(hidden_states_fp8[0])
 
@@ -708,9 +705,7 @@ def forward_deepgemm_masked(
             (
                 down_input_scale
                 if deep_gemm_wrapper.DEEPGEMM_SCALE_UE8M0
-                else deep_gemm_wrapper.get_col_major_tma_aligned_tensor(
-                    down_input_scale
-                )
+                else deep_gemm_wrapper.get_mn_major_tma_aligned_tensor(down_input_scale)
             ),
         )
         down_output = torch.empty(
@@ -722,7 +717,6 @@ def forward_deepgemm_masked(
             down_output,
             masked_m,
             expected_m,
-            recipe=(1, 128, 128) if deep_gemm_wrapper.DEEPGEMM_BLACKWELL else None,
         )
 
         return down_output