[fix] fix DeepGEMM blackwell input quant & ut & fix style and log (#7247)

python/sglang/srt/layers/quantization/fp8_kernel.py

@@ -290,11 +290,12 @@ def sglang_per_token_group_quant_fp8(
         x_s_mn, x_s_k = x_q_mn, x_q_k // 128
         aligned_mn = align(x_s_mn, 4)
         aligned_k = align(x_s_k, 4)
-        x_s = torch.empty(
+        # TODO(FIXME): Fix cuda kernel and recover here to empty.
+        x_s = torch.zeros(
             (aligned_k // 4, aligned_mn),
             device=x.device,
             dtype=torch.int,
-        ).permute(-1, -2)[:x_s_mn, :]
+        ).transpose(0, 1)[:x_s_mn, :]
     elif column_major_scales:
         if scale_tma_aligned:
             # TODO extract "align" function
@@ -768,7 +769,7 @@ def prepare_block_fp8_matmul_inputs(
 
     if As.dtype == torch.float:
         assert triton.cdiv(A.shape[-1], block_k) == As.shape[-1]
-    elif Bs.dtype == torch.int:
+    elif As.dtype == torch.int:
         assert (
             triton.cdiv(triton.cdiv(A.shape[-1], block_k), 4) == As.shape[-1]
         ), f"{A.shape=} {As.shape=} {block_size=}"