#loc = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1273:5)
#loc1 = loc(unknown)
#loc10 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1289:19)
#loc33 = loc("X_ptr"(#loc))
#loc34 = loc("Out1_ptr"(#loc))
#loc35 = loc("Out2_ptr"(#loc))
#loc36 = loc("Scale1_ptr"(#loc))
#loc37 = loc("Shift1_ptr"(#loc))
#loc38 = loc("Scale2_ptr"(#loc))
#loc39 = loc("Shift2_ptr"(#loc))
#loc40 = loc("N"(#loc))
#loc41 = loc("scale_n_rows"(#loc))
#loc49 = loc("mean_sq"(#loc10))
#loc66 = loc(callsite(#loc1 at #loc49))
module {
  tt.func public @_rms_norm_dual_scale_shift_kernel(%X_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("X_ptr"(#loc)), %Out1_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("Out1_ptr"(#loc)), %Out2_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("Out2_ptr"(#loc)), %Scale1_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("Scale1_ptr"(#loc)), %Shift1_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("Shift1_ptr"(#loc)), %Scale2_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("Scale2_ptr"(#loc)), %Shift2_ptr: !tt.ptr<bf16> {tt.divisibility = 16 : i32} loc("Shift2_ptr"(#loc)), %N: i32 {tt.divisibility = 16 : i32} loc("N"(#loc)), %scale_n_rows: i32 {tt.divisibility = 16 : i32} loc("scale_n_rows"(#loc))) attributes {noinline = false} {
    %cst = arith.constant dense<0.000000e+00> : tensor<2048xbf16> loc(#loc1)
    %cst_0 = arith.constant dense<1.000000e+00> : tensor<2048xf32> loc(#loc1)
    %cst_1 = arith.constant 9.99999997E-7 : f32 loc(#loc1)
    %row = tt.get_program_id x : i32 loc(#loc42)
    %cols = tt.make_range {end = 2048 : i32, start = 0 : i32} : tensor<2048xi32> loc(#loc43)
    %mask = tt.splat %N : i32 -> tensor<2048xi32> loc(#loc44)
    %mask_2 = arith.cmpi slt, %cols, %mask : tensor<2048xi32> loc(#loc44)
    %off = arith.muli %row, %N : i32 loc(#loc45)
    %off_3 = tt.splat %off : i32 -> tensor<2048xi32> loc(#loc45)
    %off_4 = arith.addi %off_3, %cols : tensor<2048xi32> loc(#loc45)
    %x = tt.splat %X_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc46)
    %x_5 = tt.addptr %x, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc46)
    %x_6 = tt.load %x_5, %mask_2, %cst : tensor<2048x!tt.ptr<bf16>> loc(#loc47)
    %x_7 = arith.extf %x_6 : tensor<2048xbf16> to tensor<2048xf32> loc(#loc47)
    %mean_sq = arith.mulf %x_7, %x_7 : tensor<2048xf32> loc(#loc48)
    %mean_sq_8 = "tt.reduce"(%mean_sq) <{axis = 0 : i32}> ({
    ^bb0(%mean_sq_30: f32 loc(callsite(#loc1 at #loc49)), %mean_sq_31: f32 loc(callsite(#loc1 at #loc49))):
      %mean_sq_32 = arith.addf %mean_sq_30, %mean_sq_31 : f32 loc(#loc67)
      tt.reduce.return %mean_sq_32 : f32 loc(#loc65)
    }) : (tensor<2048xf32>) -> f32 loc(#loc65)
    %mean_sq_9 = arith.sitofp %N : i32 to f32 loc(#loc49)
    %mean_sq_10 = arith.divf %mean_sq_8, %mean_sq_9 : f32 loc(#loc49)
    %x_norm = arith.addf %mean_sq_10, %cst_1 : f32 loc(#loc50)
    %x_norm_11 = math.rsqrt %x_norm : f32 loc(#loc51)
    %x_norm_12 = tt.splat %x_norm_11 : f32 -> tensor<2048xf32> loc(#loc52)
    %x_norm_13 = arith.mulf %x_7, %x_norm_12 : tensor<2048xf32> loc(#loc52)
    %scale1 = tt.splat %Scale1_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc53)
    %scale1_14 = tt.addptr %scale1, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc53)
    %scale1_15 = tt.load %scale1_14, %mask_2, %cst : tensor<2048x!tt.ptr<bf16>> loc(#loc54)
    %scale1_16 = arith.extf %scale1_15 : tensor<2048xbf16> to tensor<2048xf32> loc(#loc54)
    %shift1 = tt.splat %Shift1_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc55)
    %shift1_17 = tt.addptr %shift1, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc55)
    %shift1_18 = tt.load %shift1_17, %mask_2, %cst : tensor<2048x!tt.ptr<bf16>> loc(#loc56)
    %shift1_19 = arith.extf %shift1_18 : tensor<2048xbf16> to tensor<2048xf32> loc(#loc56)
    %scale2 = tt.splat %Scale2_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc57)
    %scale2_20 = tt.addptr %scale2, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc57)
    %scale2_21 = tt.load %scale2_20, %mask_2, %cst : tensor<2048x!tt.ptr<bf16>> loc(#loc58)
    %scale2_22 = arith.extf %scale2_21 : tensor<2048xbf16> to tensor<2048xf32> loc(#loc58)
    %shift2 = tt.splat %Shift2_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc59)
    %shift2_23 = tt.addptr %shift2, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc59)
    %shift2_24 = tt.load %shift2_23, %mask_2, %cst : tensor<2048x!tt.ptr<bf16>> loc(#loc60)
    %shift2_25 = arith.extf %shift2_24 : tensor<2048xbf16> to tensor<2048xf32> loc(#loc60)
    %out1 = arith.addf %scale1_16, %cst_0 : tensor<2048xf32> loc(#loc61)
    %out1_26 = arith.mulf %x_norm_13, %out1 : tensor<2048xf32> loc(#loc62)
    %out1_27 = arith.addf %out1_26, %shift1_19 : tensor<2048xf32> loc(#loc62)
    %out2 = arith.addf %scale2_22, %cst_0 : tensor<2048xf32> loc(#loc63)
    %out2_28 = arith.mulf %x_norm_13, %out2 : tensor<2048xf32> loc(#loc64)
    %out2_29 = arith.addf %out2_28, %shift2_25 : tensor<2048xf32> loc(#loc64)
    %0 = tt.splat %Out1_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc27)
    %1 = tt.addptr %0, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc27)
    %2 = arith.truncf %out1_27 : tensor<2048xf32> to tensor<2048xbf16> loc(#loc28)
    tt.store %1, %2, %mask_2 : tensor<2048x!tt.ptr<bf16>> loc(#loc29)
    %3 = tt.splat %Out2_ptr : !tt.ptr<bf16> -> tensor<2048x!tt.ptr<bf16>> loc(#loc30)
    %4 = tt.addptr %3, %off_4 : tensor<2048x!tt.ptr<bf16>>, tensor<2048xi32> loc(#loc30)
    %5 = arith.truncf %out2_29 : tensor<2048xf32> to tensor<2048xbf16> loc(#loc31)
    tt.store %4, %5, %mask_2 : tensor<2048x!tt.ptr<bf16>> loc(#loc32)
    tt.return loc(#loc)
  } loc(#loc)
} loc(#loc)
#loc2 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1283:15)
#loc3 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1284:16)
#loc4 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1285:16)
#loc5 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1286:15)
#loc6 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1288:21)
#loc7 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1288:13)
#loc8 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1289:26)
#loc9 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\python_embeded\\Lib\\site-packages\\triton\\language\\standard.py":293:12)
#loc11 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\python_embeded\\Lib\\site-packages\\triton\\language\\standard.py":263:12)
#loc12 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1290:36)
#loc13 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1290:22)
#loc14 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1290:18)
#loc15 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1298:26)
#loc16 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1298:18)
#loc17 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1299:26)
#loc18 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1299:18)
#loc19 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1300:26)
#loc20 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1300:18)
#loc21 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1301:26)
#loc22 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1301:18)
#loc23 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1303:26)
#loc24 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1303:16)
#loc25 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1304:26)
#loc26 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1304:16)
#loc27 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1307:22)
#loc28 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1307:38)
#loc29 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1307:13)
#loc30 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1308:22)
#loc31 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1308:38)
#loc32 = loc("D:\\Comfy\\App\\ComfyUI_windows_portable\\ComfyUI\\custom_nodes\\comfyui-kjnodes\\nodes\\ltxv_nodes.py":1308:13)
#loc42 = loc("row"(#loc2))
#loc43 = loc("cols"(#loc3))
#loc44 = loc("mask"(#loc4))
#loc45 = loc("off"(#loc5))
#loc46 = loc("x"(#loc6))
#loc47 = loc("x"(#loc7))
#loc48 = loc("mean_sq"(#loc8))
#loc50 = loc("x_norm"(#loc12))
#loc51 = loc("x_norm"(#loc13))
#loc52 = loc("x_norm"(#loc14))
#loc53 = loc("scale1"(#loc15))
#loc54 = loc("scale1"(#loc16))
#loc55 = loc("shift1"(#loc17))
#loc56 = loc("shift1"(#loc18))
#loc57 = loc("scale2"(#loc19))
#loc58 = loc("scale2"(#loc20))
#loc59 = loc("shift2"(#loc21))
#loc60 = loc("shift2"(#loc22))
#loc61 = loc("out1"(#loc23))
#loc62 = loc("out1"(#loc24))
#loc63 = loc("out2"(#loc25))
#loc64 = loc("out2"(#loc26))
#loc65 = loc(callsite(#loc9 at #loc49))
#loc67 = loc(callsite(#loc11 at #loc65))