Support DeepGEMM swap-AB on sm100

Signed-off-by: Barry Kang <[email protected]>

Author: Barry-Delaney

.gitmodules

@@ -28,4 +28,5 @@
 	url = https://github.com/zeromq/cppzmq.git
 [submodule "3rdparty/DeepGEMM"]
 	path = 3rdparty/DeepGEMM
-	url = https://github.com/deepseek-ai/DeepGEMM.git
+	url = https://github.com/ruoqianguo/DeepGEMM.git
+	branch = dev/ruoqiang/swapab_sm100

tensorrt_llm/_torch/custom_ops/torch_custom_ops.py

@@ -927,7 +927,7 @@ def get_valid_tactics(
         profile: OptimizationProfile,
     ) -> List[int]:
         # Encode swap_ab as False (0) and True (1). Currently only add one tactic here.
-        return [0]
+        return [0, 1]
 
     def forward(
         self,
@@ -941,14 +941,22 @@ def forward(
             device=input.device,
             dtype=self.output_dtype,
         )
-        # TODO: add swap_ab=tactic == 0 to detemrmine the swap_ab value
-        # Treat the default tactic=-1 as swap_ab=False
-        deep_gemm.fp8_gemm_nt(
-            (a, a_sf),
-            (weight, weight_scale),
-            output,
-            disable_ue8m0_cast=self.disable_ue8m0_cast,
-        )
+
+        swap_ab = tactic == 1
+        if swap_ab:
+            deep_gemm.fp8_gemm_ntt(
+                (a, a_sf),
+                (weight, weight_scale),
+                output,
+                disable_ue8m0_cast=self.disable_ue8m0_cast,
+            )
+        else:
+            deep_gemm.fp8_gemm_nt(
+                (a, a_sf),
+                (weight, weight_scale),
+                output,
+                disable_ue8m0_cast=self.disable_ue8m0_cast,
+            )
         return output