Use FlashInfer's TRTLLM FP8 Blockscale GEMM (#8588)

python/sglang/srt/layers/quantization/fp8_utils.py

@@ -161,16 +161,16 @@ def flashinfer_gemm_w8a8_block_fp8_linear(
     output_shape = [*input.shape[:-1], weight.shape[0]]
 
     q_input, x_scale = sglang_per_token_group_quant_fp8(
-        input_2d, block_size[1], column_major_scales=False
+        input_2d, block_size[1], column_major_scales=True
     )
-
+    # TRTLLM requires column-major scaling factors
     output = gemm_fp8_nt_groupwise(
         q_input,
         weight,
         x_scale,
         weight_scale,
-        scale_major_mode="K",
         out_dtype=input_2d.dtype,
+        backend="trtllm",
     )
 
     if bias is not None: