Example

storylm0_forward.kina

Source

let storylm0_forward = fn(
  token_ids: tensor<I64, [1, 4], @host>,
  token_embedding: tensor<F32, [8, 4], @host>,
  causal_mask: tensor<F32, [1, 1, 4, 4], @host>,
  attention_scale: tensor<F32, [1, 1, 1, 1], @host>,
  wq: tensor<F32, [4, 4], @host>,
  wk: tensor<F32, [4, 4], @host>,
  wv: tensor<F32, [4, 4], @host>,
  wo: tensor<F32, [4, 4], @host>,
  lm_head: tensor<F32, [4, 8], @host>
) -> tensor<F32, [1, 4, 8], @host> {
  let x = token_embedding.gather(token_ids) in

  // Matmul never broadcasts implicitly; parameters use canonical 2D layouts.
  let wq_broadcast = wq.broadcast([1, 4, 4]) in
  let wk_broadcast = wk.broadcast([1, 4, 4]) in
  let wv_broadcast = wv.broadcast([1, 4, 4]) in
  let wo_broadcast = wo.broadcast([1, 4, 4]) in
  let lm_head_broadcast = lm_head.broadcast([1, 4, 8]) in

  // Split D=4 into H=2 heads of Dh=2.
  let q_flat = x @ wq_broadcast in
  let k_flat = x @ wk_broadcast in
  let v_flat = x @ wv_broadcast in
  let q = q_flat.reshape([1, 4, 2, 2]).transpose([0, 2, 1, 3]) in
  let k_t = k_flat.reshape([1, 4, 2, 2]).transpose([0, 2, 3, 1]) in
  let v = v_flat.reshape([1, 4, 2, 2]).transpose([0, 2, 1, 3]) in

  // Both attention matmuls are rank 4 with leading batch/head dimensions.
  let scores = q @ k_t in
  let scale_broadcast = attention_scale.broadcast([1, 2, 4, 4]) in
  let mask_broadcast = causal_mask.broadcast([1, 2, 4, 4]) in
  let masked_scores = (scores * scale_broadcast) + mask_broadcast in
  let probs = masked_scores.softmax(axis=3) in
  let context = probs @ v in

  let merged = context.transpose([0, 2, 1, 3]).reshape([1, 4, 4]) in
  let projected = merged @ wo_broadcast in
  let residual = x + projected in
  let logits = residual @ lm_head_broadcast in
  logits
} in

// Host placeholders keep the example compile-only. The numerical validator
// invokes the emitted C wrapper directly with non-uniform tensor contents.
let token_ids = tensor<I64, [1, 4], @host> in
let token_embedding = tensor<F32, [8, 4], @host> in
let causal_mask = tensor<F32, [1, 1, 4, 4], @host> in
let attention_scale = tensor<F32, [1, 1, 1, 1], @host> in
let wq = tensor<F32, [4, 4], @host> in
let wk = tensor<F32, [4, 4], @host> in
let wv = tensor<F32, [4, 4], @host> in
let wo = tensor<F32, [4, 4], @host> in
let lm_head = tensor<F32, [4, 8], @host> in
storylm0_forward(token_ids, token_embedding, causal_mask, attention_scale, wq, wk, wv, wo, lm_head)

Emitted MLIR (golden)

module {
  func.func @fn_val_0(%1: tensor<1x32x64xf32>, %2: tensor<1x1x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
      %3 = tensor.empty() : tensor<1x32x64xf32>
      %4 = linalg.mul ins(%1, %1 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%3 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %7 = arith.constant 0.000000e+00 : f32
      %5 = tensor.empty() : tensor<1x32x1xf32>
      %6 = linalg.fill ins(%7 : f32) outs(%5 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
      %8 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
        iterator_types = ["parallel", "parallel", "reduction"]
      } ins(%4 : tensor<1x32x64xf32>) outs(%6 : tensor<1x32x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %9 = arith.addf %a0, %a1 : f32
        linalg.yield %9 : f32
      } -> tensor<1x32x1xf32>
      %10 = arith.constant dense<0.000000e+00> : tensor<1x1x1xf32>
      %11 = tensor.empty() : tensor<1x32x1xf32>
      %12 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%10 : tensor<1x1x1xf32>) outs(%11 : tensor<1x32x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x1xf32>
      %13 = tensor.empty() : tensor<1x32x1xf32>
      %14 = linalg.mul ins(%8, %12 : tensor<1x32x1xf32>, tensor<1x32x1xf32>) outs(%13 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
      %15 = arith.constant dense<0.000000e+00> : tensor<1x1x1xf32>
      %16 = tensor.empty() : tensor<1x32x1xf32>
      %17 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%15 : tensor<1x1x1xf32>) outs(%16 : tensor<1x32x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x1xf32>
      %18 = tensor.empty() : tensor<1x32x1xf32>
      %19 = linalg.add ins(%14, %17 : tensor<1x32x1xf32>, tensor<1x32x1xf32>) outs(%18 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
      %23 = tensor.empty() : tensor<1x32x1xf32>
      %24 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%19 : tensor<1x32x1xf32>) outs(%23 : tensor<1x32x1xf32>) {
      ^bb0(%x: f32, %_: f32):
        %20 = math.sqrt %x : f32
        %21 = arith.constant 1.000000e+00 : f32
        %22 = arith.divf %21, %20 : f32
        linalg.yield %22 : f32
      } -> tensor<1x32x1xf32>
      %25 = tensor.empty() : tensor<1x32x64xf32>
      %26 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%24 : tensor<1x32x1xf32>) outs(%25 : tensor<1x32x64xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x64xf32>
      %27 = tensor.empty() : tensor<1x32x64xf32>
      %28 = linalg.mul ins(%1, %26 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%27 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %29 = tensor.empty() : tensor<1x32x64xf32>
      %30 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%2 : tensor<1x1x64xf32>) outs(%29 : tensor<1x32x64xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x64xf32>
      %31 = tensor.empty() : tensor<1x32x64xf32>
      %32 = linalg.mul ins(%28, %30 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%31 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
    return %32 : tensor<1x32x64xf32>
  }
  func.func @fn_val_33(%34: tensor<1x32x64xf32>, %35: tensor<1x64x128xf32>, %36: tensor<1x128x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
      %37 = tensor.empty() : tensor<1x32x128xf32>
      %38 = linalg.batch_matmul ins(%34, %35 : tensor<1x32x64xf32>, tensor<1x64x128xf32>) outs(%37 : tensor<1x32x128xf32>) -> tensor<1x32x128xf32>
      %44 = tensor.empty() : tensor<1x32x128xf32>
      %45 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%38 : tensor<1x32x128xf32>) outs(%44 : tensor<1x32x128xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %39 = arith.constant 1.000000e+00 : f32
        %40 = arith.negf %a0 : f32
        %41 = math.exp %40 : f32
        %42 = arith.addf %39, %41 : f32
        %43 = arith.divf %39, %42 : f32
        linalg.yield %43 : f32
      } -> tensor<1x32x128xf32>
      %46 = tensor.empty() : tensor<1x32x128xf32>
      %47 = linalg.mul ins(%38, %45 : tensor<1x32x128xf32>, tensor<1x32x128xf32>) outs(%46 : tensor<1x32x128xf32>) -> tensor<1x32x128xf32>
      %48 = tensor.empty() : tensor<1x32x64xf32>
      %49 = linalg.batch_matmul ins(%47, %36 : tensor<1x32x128xf32>, tensor<1x128x64xf32>) outs(%48 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
    return %49 : tensor<1x32x64xf32>
  }
  func.func @fn_val_50(%51: tensor<1x32x64xf32>, %52: tensor<1x64x64xf32>, %53: tensor<1x64x64xf32>, %54: tensor<1x64x64xf32>, %55: tensor<1x64x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
      %56 = tensor.empty() : tensor<1x32x64xf32>
      %57 = linalg.batch_matmul ins(%51, %52 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%56 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %58 = tensor.empty() : tensor<1x32x64xf32>
      %59 = linalg.batch_matmul ins(%51, %53 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%58 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %60 = tensor.empty() : tensor<1x32x64xf32>
      %61 = linalg.batch_matmul ins(%51, %54 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%60 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %62 = tensor.empty() : tensor<1x64x32xf32>
      %63 = linalg.transpose ins(%59 : tensor<1x32x64xf32>) outs(%62 : tensor<1x64x32xf32>) permutation = [0, 2, 1]
      %64 = tensor.empty() : tensor<1x32x32xf32>
      %65 = linalg.batch_matmul ins(%57, %63 : tensor<1x32x64xf32>, tensor<1x64x32xf32>) outs(%64 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %66 = arith.constant dense<0.000000e+00> : tensor<1x1x1xf32>
      %67 = tensor.empty() : tensor<1x32x32xf32>
      %68 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, 0, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%66 : tensor<1x1x1xf32>) outs(%67 : tensor<1x32x32xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x32xf32>
      %69 = tensor.empty() : tensor<1x32x32xf32>
      %70 = linalg.mul ins(%65, %68 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%69 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %73 = arith.constant 0xFF800000 : f32
      %71 = tensor.empty() : tensor<1x32x1xf32>
      %72 = linalg.fill ins(%73 : f32) outs(%71 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
      %74 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
        iterator_types = ["parallel", "parallel", "reduction"]
      } ins(%70 : tensor<1x32x32xf32>) outs(%72 : tensor<1x32x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %75 = arith.maximumf %a0, %a1 : f32
        linalg.yield %75 : f32
      } -> tensor<1x32x1xf32>
      %76 = tensor.empty() : tensor<1x32x32xf32>
      %77 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%74 : tensor<1x32x1xf32>) outs(%76 : tensor<1x32x32xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x32xf32>
      %78 = tensor.empty() : tensor<1x32x32xf32>
      %79 = linalg.sub ins(%70, %77 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%78 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %80 = tensor.empty() : tensor<1x32x32xf32>
      %81 = linalg.exp ins(%79 : tensor<1x32x32xf32>) outs(%80 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %84 = arith.constant 0xFF800000 : f32
      %82 = tensor.empty() : tensor<1x32x1xf32>
      %83 = linalg.fill ins(%84 : f32) outs(%82 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
      %85 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
        iterator_types = ["parallel", "parallel", "reduction"]
      } ins(%70 : tensor<1x32x32xf32>) outs(%83 : tensor<1x32x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %86 = arith.maximumf %a0, %a1 : f32
        linalg.yield %86 : f32
      } -> tensor<1x32x1xf32>
      %87 = tensor.empty() : tensor<1x32x32xf32>
      %88 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%85 : tensor<1x32x1xf32>) outs(%87 : tensor<1x32x32xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x32xf32>
      %89 = tensor.empty() : tensor<1x32x32xf32>
      %90 = linalg.sub ins(%70, %88 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%89 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %91 = tensor.empty() : tensor<1x32x32xf32>
      %92 = linalg.exp ins(%90 : tensor<1x32x32xf32>) outs(%91 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %95 = arith.constant 0.000000e+00 : f32
      %93 = tensor.empty() : tensor<1x32x1xf32>
      %94 = linalg.fill ins(%95 : f32) outs(%93 : tensor<1x32x1xf32>) -> tensor<1x32x1xf32>
      %96 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, 0)>],
        iterator_types = ["parallel", "parallel", "reduction"]
      } ins(%92 : tensor<1x32x32xf32>) outs(%94 : tensor<1x32x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %97 = arith.addf %a0, %a1 : f32
        linalg.yield %97 : f32
      } -> tensor<1x32x1xf32>
      %98 = tensor.empty() : tensor<1x32x32xf32>
      %99 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (0, d1, 0)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%96 : tensor<1x32x1xf32>) outs(%98 : tensor<1x32x32xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x32x32xf32>
      %100 = tensor.empty() : tensor<1x32x32xf32>
      %101 = linalg.div ins(%81, %99 : tensor<1x32x32xf32>, tensor<1x32x32xf32>) outs(%100 : tensor<1x32x32xf32>) -> tensor<1x32x32xf32>
      %102 = tensor.empty() : tensor<1x32x64xf32>
      %103 = linalg.batch_matmul ins(%101, %61 : tensor<1x32x32xf32>, tensor<1x32x64xf32>) outs(%102 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %104 = tensor.empty() : tensor<1x32x64xf32>
      %105 = linalg.batch_matmul ins(%103, %55 : tensor<1x32x64xf32>, tensor<1x64x64xf32>) outs(%104 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
    return %105 : tensor<1x32x64xf32>
  }
  func.func @fn_val_106(%107: tensor<1x32x64xf32>, %108: tensor<1x1x64xf32>, %109: tensor<1x64x64xf32>, %110: tensor<1x64x64xf32>, %111: tensor<1x64x64xf32>, %112: tensor<1x64x64xf32>, %113: tensor<1x1x64xf32>, %114: tensor<1x64x128xf32>, %115: tensor<1x128x64xf32>) -> tensor<1x32x64xf32> attributes {llvm.emit_c_interface} {
      %116 = func.call @fn_val_0(%107, %108) : (tensor<1x32x64xf32>, tensor<1x1x64xf32>) -> tensor<1x32x64xf32>
      %117 = func.call @fn_val_50(%116, %109, %110, %111, %112) : (tensor<1x32x64xf32>, tensor<1x64x64xf32>, tensor<1x64x64xf32>, tensor<1x64x64xf32>, tensor<1x64x64xf32>) -> tensor<1x32x64xf32>
      %118 = tensor.empty() : tensor<1x32x64xf32>
      %119 = linalg.add ins(%107, %117 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%118 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
      %120 = func.call @fn_val_0(%119, %113) : (tensor<1x32x64xf32>, tensor<1x1x64xf32>) -> tensor<1x32x64xf32>
      %121 = func.call @fn_val_33(%120, %114, %115) : (tensor<1x32x64xf32>, tensor<1x64x128xf32>, tensor<1x128x64xf32>) -> tensor<1x32x64xf32>
      %122 = tensor.empty() : tensor<1x32x64xf32>
      %123 = linalg.add ins(%119, %121 : tensor<1x32x64xf32>, tensor<1x32x64xf32>) outs(%122 : tensor<1x32x64xf32>) -> tensor<1x32x64xf32>
    return %123 : tensor<1x32x64xf32>
  }
  func.func @fn_val_124(%125: tensor<1x4xi64>, %126: tensor<8x4xf32>, %127: tensor<1x1x4x4xf32>, %128: tensor<1x1x1x1xf32>, %129: tensor<4x4xf32>, %130: tensor<4x4xf32>, %131: tensor<4x4xf32>, %132: tensor<4x4xf32>, %133: tensor<4x8xf32>) -> tensor<1x4x8xf32> attributes {llvm.emit_c_interface} {
      %134 = arith.constant 0 : i64
      %135 = arith.constant 8 : i64
      %136 = tensor.empty() : tensor<1x4xi64>
      %137 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, affine_map<(d0, d1) -> (d0, d1)>],
        iterator_types = ["parallel", "parallel"]
      } ins(%125 : tensor<1x4xi64>) outs(%136 : tensor<1x4xi64>) {
      ^bb0(%candidate: i64, %_: i64):
        %138 = arith.cmpi sge, %candidate, %134 : i64
        %139 = arith.cmpi slt, %candidate, %135 : i64
        %140 = arith.andi %138, %139 : i1
        cf.assert %140, "gather index out of bounds"
        linalg.yield %candidate : i64
      } -> tensor<1x4xi64>
      %141 = tensor.empty() : tensor<1x4x4xf32>
      %142 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%137 : tensor<1x4xi64>) outs(%141 : tensor<1x4x4xf32>) {
      ^bb0(%idx: i64, %_: f32):
        %143 = arith.index_cast %idx : i64 to index
        %144 = linalg.index 2 : index
        %145 = tensor.extract %126[%143, %144] : tensor<8x4xf32>
        linalg.yield %145 : f32
      } -> tensor<1x4x4xf32>
      %146 = tensor.empty() : tensor<1x4x4xf32>
      %147 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%129 : tensor<4x4xf32>) outs(%146 : tensor<1x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x4x4xf32>
      %148 = tensor.empty() : tensor<1x4x4xf32>
      %149 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%130 : tensor<4x4xf32>) outs(%148 : tensor<1x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x4x4xf32>
      %150 = tensor.empty() : tensor<1x4x4xf32>
      %151 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%131 : tensor<4x4xf32>) outs(%150 : tensor<1x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x4x4xf32>
      %152 = tensor.empty() : tensor<1x4x4xf32>
      %153 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%132 : tensor<4x4xf32>) outs(%152 : tensor<1x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x4x4xf32>
      %154 = tensor.empty() : tensor<1x4x8xf32>
      %155 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2) -> (d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1, d2)>],
        iterator_types = ["parallel", "parallel", "parallel"]
      } ins(%133 : tensor<4x8xf32>) outs(%154 : tensor<1x4x8xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x4x8xf32>
      %156 = tensor.empty() : tensor<1x4x4xf32>
      %157 = linalg.batch_matmul ins(%142, %147 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%156 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
      %158 = tensor.empty() : tensor<1x4x4xf32>
      %159 = linalg.batch_matmul ins(%142, %149 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%158 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
      %160 = tensor.empty() : tensor<1x4x4xf32>
      %161 = linalg.batch_matmul ins(%142, %151 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%160 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
      %162 = arith.constant dense<[1, 4, 2, 2]> : tensor<4xi32>
      %163 = tensor.reshape %157(%162) : (tensor<1x4x4xf32>, tensor<4xi32>) -> tensor<1x4x2x2xf32>
      %164 = tensor.empty() : tensor<1x2x4x2xf32>
      %165 = linalg.transpose ins(%163 : tensor<1x4x2x2xf32>) outs(%164 : tensor<1x2x4x2xf32>) permutation = [0, 2, 1, 3]
      %166 = arith.constant dense<[1, 4, 2, 2]> : tensor<4xi32>
      %167 = tensor.reshape %159(%166) : (tensor<1x4x4xf32>, tensor<4xi32>) -> tensor<1x4x2x2xf32>
      %168 = tensor.empty() : tensor<1x2x2x4xf32>
      %169 = linalg.transpose ins(%167 : tensor<1x4x2x2xf32>) outs(%168 : tensor<1x2x2x4xf32>) permutation = [0, 2, 3, 1]
      %170 = arith.constant dense<[1, 4, 2, 2]> : tensor<4xi32>
      %171 = tensor.reshape %161(%170) : (tensor<1x4x4xf32>, tensor<4xi32>) -> tensor<1x4x2x2xf32>
      %172 = tensor.empty() : tensor<1x2x4x2xf32>
      %173 = linalg.transpose ins(%171 : tensor<1x4x2x2xf32>) outs(%172 : tensor<1x2x4x2xf32>) permutation = [0, 2, 1, 3]
      %174 = tensor.collapse_shape %165 [[0, 1], [2], [3]] : tensor<1x2x4x2xf32> into tensor<2x4x2xf32>
      %175 = tensor.collapse_shape %169 [[0, 1], [2], [3]] : tensor<1x2x2x4xf32> into tensor<2x2x4xf32>
      %176 = tensor.empty() : tensor<2x4x4xf32>
      %177 = linalg.batch_matmul ins(%174, %175 : tensor<2x4x2xf32>, tensor<2x2x4xf32>) outs(%176 : tensor<2x4x4xf32>) -> tensor<2x4x4xf32>
      %178 = tensor.expand_shape %177 [[0, 1], [2], [3]] : tensor<2x4x4xf32> into tensor<1x2x4x4xf32>
      %179 = tensor.empty() : tensor<1x2x4x4xf32>
      %180 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, 0, 0, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
        iterator_types = ["parallel", "parallel", "parallel", "parallel"]
      } ins(%128 : tensor<1x1x1x1xf32>) outs(%179 : tensor<1x2x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x2x4x4xf32>
      %181 = tensor.empty() : tensor<1x2x4x4xf32>
      %182 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, 0, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
        iterator_types = ["parallel", "parallel", "parallel", "parallel"]
      } ins(%127 : tensor<1x1x4x4xf32>) outs(%181 : tensor<1x2x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x2x4x4xf32>
      %183 = tensor.empty() : tensor<1x2x4x4xf32>
      %184 = linalg.mul ins(%178, %180 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%183 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %185 = tensor.empty() : tensor<1x2x4x4xf32>
      %186 = linalg.add ins(%184, %182 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%185 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %189 = arith.constant 0xFF800000 : f32
      %187 = tensor.empty() : tensor<1x2x4x1xf32>
      %188 = linalg.fill ins(%189 : f32) outs(%187 : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
      %190 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>],
        iterator_types = ["parallel", "parallel", "parallel", "reduction"]
      } ins(%186 : tensor<1x2x4x4xf32>) outs(%188 : tensor<1x2x4x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %191 = arith.maximumf %a0, %a1 : f32
        linalg.yield %191 : f32
      } -> tensor<1x2x4x1xf32>
      %192 = tensor.empty() : tensor<1x2x4x4xf32>
      %193 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
        iterator_types = ["parallel", "parallel", "parallel", "parallel"]
      } ins(%190 : tensor<1x2x4x1xf32>) outs(%192 : tensor<1x2x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x2x4x4xf32>
      %194 = tensor.empty() : tensor<1x2x4x4xf32>
      %195 = linalg.sub ins(%186, %193 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%194 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %196 = tensor.empty() : tensor<1x2x4x4xf32>
      %197 = linalg.exp ins(%195 : tensor<1x2x4x4xf32>) outs(%196 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %200 = arith.constant 0xFF800000 : f32
      %198 = tensor.empty() : tensor<1x2x4x1xf32>
      %199 = linalg.fill ins(%200 : f32) outs(%198 : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
      %201 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>],
        iterator_types = ["parallel", "parallel", "parallel", "reduction"]
      } ins(%186 : tensor<1x2x4x4xf32>) outs(%199 : tensor<1x2x4x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %202 = arith.maximumf %a0, %a1 : f32
        linalg.yield %202 : f32
      } -> tensor<1x2x4x1xf32>
      %203 = tensor.empty() : tensor<1x2x4x4xf32>
      %204 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
        iterator_types = ["parallel", "parallel", "parallel", "parallel"]
      } ins(%201 : tensor<1x2x4x1xf32>) outs(%203 : tensor<1x2x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x2x4x4xf32>
      %205 = tensor.empty() : tensor<1x2x4x4xf32>
      %206 = linalg.sub ins(%186, %204 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%205 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %207 = tensor.empty() : tensor<1x2x4x4xf32>
      %208 = linalg.exp ins(%206 : tensor<1x2x4x4xf32>) outs(%207 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %211 = arith.constant 0.000000e+00 : f32
      %209 = tensor.empty() : tensor<1x2x4x1xf32>
      %210 = linalg.fill ins(%211 : f32) outs(%209 : tensor<1x2x4x1xf32>) -> tensor<1x2x4x1xf32>
      %212 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>],
        iterator_types = ["parallel", "parallel", "parallel", "reduction"]
      } ins(%208 : tensor<1x2x4x4xf32>) outs(%210 : tensor<1x2x4x1xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        %213 = arith.addf %a0, %a1 : f32
        linalg.yield %213 : f32
      } -> tensor<1x2x4x1xf32>
      %214 = tensor.empty() : tensor<1x2x4x4xf32>
      %215 = linalg.generic {
        indexing_maps = [affine_map<(d0, d1, d2, d3) -> (0, d1, d2, 0)>, affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
        iterator_types = ["parallel", "parallel", "parallel", "parallel"]
      } ins(%212 : tensor<1x2x4x1xf32>) outs(%214 : tensor<1x2x4x4xf32>) {
      ^bb0(%a0: f32, %a1: f32):
        linalg.yield %a0 : f32
      } -> tensor<1x2x4x4xf32>
      %216 = tensor.empty() : tensor<1x2x4x4xf32>
      %217 = linalg.div ins(%197, %215 : tensor<1x2x4x4xf32>, tensor<1x2x4x4xf32>) outs(%216 : tensor<1x2x4x4xf32>) -> tensor<1x2x4x4xf32>
      %218 = tensor.collapse_shape %217 [[0, 1], [2], [3]] : tensor<1x2x4x4xf32> into tensor<2x4x4xf32>
      %219 = tensor.collapse_shape %173 [[0, 1], [2], [3]] : tensor<1x2x4x2xf32> into tensor<2x4x2xf32>
      %220 = tensor.empty() : tensor<2x4x2xf32>
      %221 = linalg.batch_matmul ins(%218, %219 : tensor<2x4x4xf32>, tensor<2x4x2xf32>) outs(%220 : tensor<2x4x2xf32>) -> tensor<2x4x2xf32>
      %222 = tensor.expand_shape %221 [[0, 1], [2], [3]] : tensor<2x4x2xf32> into tensor<1x2x4x2xf32>
      %223 = tensor.empty() : tensor<1x4x2x2xf32>
      %224 = linalg.transpose ins(%222 : tensor<1x2x4x2xf32>) outs(%223 : tensor<1x4x2x2xf32>) permutation = [0, 2, 1, 3]
      %225 = arith.constant dense<[1, 4, 4]> : tensor<3xi32>
      %226 = tensor.reshape %224(%225) : (tensor<1x4x2x2xf32>, tensor<3xi32>) -> tensor<1x4x4xf32>
      %227 = tensor.empty() : tensor<1x4x4xf32>
      %228 = linalg.batch_matmul ins(%226, %153 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%227 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
      %229 = tensor.empty() : tensor<1x4x4xf32>
      %230 = linalg.add ins(%142, %228 : tensor<1x4x4xf32>, tensor<1x4x4xf32>) outs(%229 : tensor<1x4x4xf32>) -> tensor<1x4x4xf32>
      %231 = tensor.empty() : tensor<1x4x8xf32>
      %232 = linalg.batch_matmul ins(%230, %155 : tensor<1x4x4xf32>, tensor<1x4x8xf32>) outs(%231 : tensor<1x4x8xf32>) -> tensor<1x4x8xf32>
    return %232 : tensor<1x4x8xf32>
  }
  func.func @main() -> tensor<1x4x8xf32> {
    %233 = arith.constant dense<0> : tensor<1x4xi64>
    %234 = arith.constant dense<0.000000e+00> : tensor<8x4xf32>
    %235 = arith.constant dense<0.000000e+00> : tensor<1x1x4x4xf32>
    %236 = arith.constant dense<0.000000e+00> : tensor<1x1x1x1xf32>
    %237 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
    %238 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
    %239 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
    %240 = arith.constant dense<0.000000e+00> : tensor<4x4xf32>
    %241 = arith.constant dense<0.000000e+00> : tensor<4x8xf32>
    %242 = func.call @fn_val_124(%233, %234, %235, %236, %237, %238, %239, %240, %241) : (tensor<1x4xi64>, tensor<8x4xf32>, tensor<1x1x4x4xf32>, tensor<1x1x1x1xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x4xf32>, tensor<4x8xf32>) -> tensor<1x4x8xf32>
    return %242 : tensor<1x4x8xf32>
  }
}