Change swap threshold

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

Author: Barry-Delaney

tensorrt_llm/_torch/modules/linear.py

@@ -574,7 +574,7 @@ def apply(self, module: Linear, input: torch.Tensor,
 
         if get_sm_version() == 100:
             import deep_gemm
-            if input.shape[0] < 128:
+            if input.shape[0] < 32:
                 # Swap AB
                 a, a_sf = fp8_utils.per_token_quant_and_transform(input,
                                                                   swap_ab=True)