module {
  tt.func public @triton__0d1d2d3d4d5d6e7de(%arg0: !tt.ptr<i64, 1> {tt.divisibility = 16 : i32}, %arg1: !tt.ptr<bf16, 1> {tt.divisibility = 16 : i32}, %arg2: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg3: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg4: !tt.ptr<i64, 1> {tt.divisibility = 16 : i32}, %arg5: !tt.ptr<f32, 1> {tt.divisibility = 16 : i32}, %arg6: i64 {tt.max_divisibility = 8 : i32}, %arg7: i64 {tt.divisibility = 16 : i32, tt.max_divisibility = 16 : i32}) attributes {noinline = false} {
    %c8_i64 = arith.constant 8 : i64
    %c7680_i64 = arith.constant 7680 : i64
    %c385973760_i64 = arith.constant 385973760 : i64
    %cst = arith.constant dense<0.000000e+00> : tensor<1x2048xbf16>
    %cst_0 = arith.constant dense<50257> : tensor<1x2048xi64>
    %cst_1 = arith.constant dense<7680> : tensor<1x2048xi64>
    %c2048_i32 = arith.constant 2048 : i32
    %c7680_i32 = arith.constant 7680 : i32
    %c0_i32 = arith.constant 0 : i32
    %cst_2 = arith.constant dense<-1> : tensor<1x2048xi64>
    %cst_3 = arith.constant dense<0.000000e+00> : tensor<1x2048xf32>
    %cst_4 = arith.constant dense<0> : tensor<1x2048xi64>
    %0 = tt.get_program_id x : i32
    %1 = arith.extsi %0 : i32 to i64
    %2 = arith.cmpi slt, %1, %c8_i64 : i64
    %3 = tt.splat %2 : (i1) -> tensor<1x1xi1>
    %4 = tt.make_range {end = 2048 : i32, start = 0 : i32} : tensor<2048xi32>
    %5 = tt.expand_dims %4 {axis = 0 : i32} : (tensor<2048xi32>) -> tensor<1x2048xi32>
    %6 = arith.extsi %5 : tensor<1x2048xi32> to tensor<1x2048xi64>
    %7 = arith.muli %1, %c7680_i64 : i64
    %8 = tt.splat %7 : (i64) -> tensor<1x2048xi64>
    %9 = tt.splat %arg0 : (!tt.ptr<i64, 1>) -> tensor<1x2048x!tt.ptr<i64, 1>>
    %10 = tt.splat %2 : (i1) -> tensor<1x2048xi1>
    %11 = tt.splat %arg2 : (!tt.ptr<f32, 1>) -> tensor<1x2048x!tt.ptr<f32, 1>>
    %12 = tt.splat %arg3 : (!tt.ptr<f32, 1>) -> tensor<1x2048x!tt.ptr<f32, 1>>
    %13 = arith.muli %1, %c385973760_i64 : i64
    %14 = tt.splat %13 : (i64) -> tensor<1x2048xi64>
    %15 = tt.splat %arg1 : (!tt.ptr<bf16, 1>) -> tensor<1x2048x!tt.ptr<bf16, 1>>
    %16:2 = scf.for %arg8 = %c0_i32 to %c7680_i32 step %c2048_i32 iter_args(%arg9 = %cst_4, %arg10 = %cst_3) -> (tensor<1x2048xi64>, tensor<1x2048xf32>)  : i32 {
      %25 = arith.extsi %arg8 : i32 to i64
      %26 = tt.splat %25 : (i64) -> tensor<1x2048xi64>
      %27 = arith.addi %26, %6 : tensor<1x2048xi64>
      %28 = arith.cmpi slt, %27, %cst_1 : tensor<1x2048xi64>
      %29 = arith.addi %27, %8 : tensor<1x2048xi64>
      %30 = tt.addptr %9, %29 : tensor<1x2048x!tt.ptr<i64, 1>>, tensor<1x2048xi64>
      %31 = arith.andi %28, %10 : tensor<1x2048xi1>
      %32 = tt.load %30, %31, %cst_4 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xi64>
      %33 = tt.addptr %11, %29 : tensor<1x2048x!tt.ptr<f32, 1>>, tensor<1x2048xi64>
      %34 = tt.load %33, %31, %cst_3 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xf32>
      %35 = tt.addptr %12, %29 : tensor<1x2048x!tt.ptr<f32, 1>>, tensor<1x2048xi64>
      %36 = tt.load %35, %31, %cst_3 {cache = 1 : i32, evict = 2 : i32, isVolatile = false} : tensor<1x2048xf32>
      %37 = arith.cmpi ne, %32, %cst_2 : tensor<1x2048xi64>
      %38 = arith.extui %37 : tensor<1x2048xi1> to tensor<1x2048xi64>
      %39 = arith.addi %arg9, %38 : tensor<1x2048xi64>
      %40 = arith.select %31, %39, %arg9 : tensor<1x2048xi1>, tensor<1x2048xi64>
      %41 = arith.select %37, %32, %cst_4 : tensor<1x2048xi1>, tensor<1x2048xi64>
      %42 = arith.addi %41, %cst_0 : tensor<1x2048xi64>
      %43 = arith.cmpi slt, %41, %cst_4 : tensor<1x2048xi64>
      %44 = arith.select %43, %42, %41 : tensor<1x2048xi1>, tensor<1x2048xi64>
      %45 = arith.cmpi sge, %44, %cst_4 : tensor<1x2048xi64>
      %46 = arith.cmpi slt, %44, %cst_0 : tensor<1x2048xi64>
      %47 = arith.andi %45, %46 : tensor<1x2048xi1>
      tt.assert %47, "index out of bounds: 0 <= tmp11 < 50257", "<frozen importlib._bootstrap_external>", "_call_with_frames_removed", 883 : tensor<1x2048xi1>
      %48 = arith.muli %27, %cst_0 : tensor<1x2048xi64>
      %49 = arith.addi %44, %48 : tensor<1x2048xi64>
      %50 = arith.addi %49, %14 : tensor<1x2048xi64>
      %51 = tt.addptr %15, %50 : tensor<1x2048x!tt.ptr<bf16, 1>>, tensor<1x2048xi64>
      %52 = tt.load %51, %31, %cst {cache = 1 : i32, evict = 3 : i32, isVolatile = false} : tensor<1x2048xbf16>
      %53 = arith.extf %52 : tensor<1x2048xbf16> to tensor<1x2048xf32>
      %54 = arith.subf %53, %34 : tensor<1x2048xf32>
      %55 = math.log %36 : tensor<1x2048xf32>
      %56 = arith.subf %54, %55 : tensor<1x2048xf32>
      %57 = arith.subf %cst_3, %56 : tensor<1x2048xf32>
      %58 = arith.select %37, %57, %cst_3 : tensor<1x2048xi1>, tensor<1x2048xf32>
      %59 = arith.addf %arg10, %58 : tensor<1x2048xf32>
      %60 = arith.select %31, %59, %arg10 : tensor<1x2048xi1>, tensor<1x2048xf32>
      scf.yield %40, %60 : tensor<1x2048xi64>, tensor<1x2048xf32>
    }
    %17 = "tt.reduce"(%16#0) <{axis = 1 : i32}> ({
    ^bb0(%arg8: i64, %arg9: i64):
      %25 = arith.addi %arg8, %arg9 : i64
      tt.reduce.return %25 : i64
    }) : (tensor<1x2048xi64>) -> tensor<1xi64>
    %18 = tt.expand_dims %17 {axis = 1 : i32} : (tensor<1xi64>) -> tensor<1x1xi64>
    %19 = tt.addptr %arg4, %1 : !tt.ptr<i64, 1>, i64
    %20 = tt.splat %19 : (!tt.ptr<i64, 1>) -> tensor<1x1x!tt.ptr<i64, 1>>
    tt.store %20, %18, %3 {cache = 1 : i32, evict = 1 : i32} : tensor<1x1xi64>
    %21 = "tt.reduce"(%16#1) <{axis = 1 : i32}> ({
    ^bb0(%arg8: f32, %arg9: f32):
      %25 = arith.addf %arg8, %arg9 : f32
      tt.reduce.return %25 : f32
    }) : (tensor<1x2048xf32>) -> tensor<1xf32>
    %22 = tt.expand_dims %21 {axis = 1 : i32} : (tensor<1xf32>) -> tensor<1x1xf32>
    %23 = tt.addptr %arg5, %1 : !tt.ptr<f32, 1>, i64
    %24 = tt.splat %23 : (!tt.ptr<f32, 1>) -> tensor<1x1x!tt.ptr<f32, 1>>
    tt.store %24, %22, %3 {cache = 1 : i32, evict = 1 : i32} : tensor<1x1xf32>
    tt.return
  }
}