# AOT ID: ['4_inference']
from ctypes import c_void_p, c_long
import torch
import math
import random
import os
import tempfile
from math import inf, nan
from torch._inductor.hooks import run_intermediate_hooks
from torch._inductor.utils import maybe_profile
from torch._inductor.codegen.memory_planning import _align as align

from torch import device, empty_strided
from torch._inductor.async_compile import AsyncCompile
from torch._inductor.select_algorithm import extern_kernels
from torch._inductor.codegen.multi_kernel import MultiKernelCall

aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
alloc_from_pool = torch.ops.inductor._alloc_from_pool
reinterpret_tensor = torch.ops.inductor._reinterpret_tensor
async_compile = AsyncCompile()


# kernel path: /tmp/torchinductor_root/na/cna5zgl6w5rl6c6llzjewozjzh2cfa4qqaedgxojdsrpbvufozpf.py
# Source Nodes: [_tgt, add, linear, tgt], Original ATen: [aten._to_copy, aten.add, aten.native_layer_norm]
# _tgt => add_2, add_3, clone, convert_element_type, mul, mul_1, rsqrt, sub, var_mean
# add => add
# linear => convert_element_type_3
# tgt => add_1
triton_red_fused__to_copy_add_native_layer_norm_0 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.reduction(
    size_hints=[8192, 256],
    reduction_hint=ReductionHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: '*fp16', 3: '*fp32', 4: '*fp32', 5: '*fp16', 6: 'i32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__to_copy_add_native_layer_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 8, 'num_reduction': 2, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
    xnumel = 4624
    rnumel = 256
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rbase = tl.arange(0, RBLOCK)[None, :]
    x0 = xindex
    tmp7_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    tmp7_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    tmp7_weight = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp0 = tl.load(in_ptr0 + (x0 + (4624*r1)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp1 = tl.load(in_ptr1 + (r1 + (256*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp2 = tl.load(in_ptr2 + (r1 + (256*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp3 = tmp1 + tmp2
        tmp4 = tmp0 + tmp3
        tmp5 = tmp4.to(tl.float32)
        tmp6 = tl.broadcast_to(tmp5, [XBLOCK, RBLOCK])
        tmp7_mean_next, tmp7_m2_next, tmp7_weight_next = triton_helpers.welford_reduce(
            tmp6, tmp7_mean, tmp7_m2, tmp7_weight, roffset == 0
        )
        tmp7_mean = tl.where(rmask & xmask, tmp7_mean_next, tmp7_mean)
        tmp7_m2 = tl.where(rmask & xmask, tmp7_m2_next, tmp7_m2)
        tmp7_weight = tl.where(rmask & xmask, tmp7_weight_next, tmp7_weight)
    tmp7_tmp, tmp8_tmp, tmp9_tmp = triton_helpers.welford(
        tmp7_mean, tmp7_m2, tmp7_weight, 1
    )
    tmp7 = tmp7_tmp[:, None]
    tmp8 = tmp8_tmp[:, None]
    tmp9 = tmp9_tmp[:, None]
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp10 = tl.load(in_ptr0 + (x0 + (4624*r1)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp11 = tl.load(in_ptr1 + (r1 + (256*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp12 = tl.load(in_ptr2 + (r1 + (256*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp23 = tl.load(in_ptr3 + (r1), rmask, eviction_policy='evict_last', other=0.0)
        tmp25 = tl.load(in_ptr4 + (r1), rmask, eviction_policy='evict_last', other=0.0)
        tmp13 = tmp11 + tmp12
        tmp14 = tmp10 + tmp13
        tmp15 = tmp14.to(tl.float32)
        tmp16 = tmp15 - tmp7
        tmp17 = 256.0
        tmp18 = tmp8 / tmp17
        tmp19 = 1e-05
        tmp20 = tmp18 + tmp19
        tmp21 = libdevice.rsqrt(tmp20)
        tmp22 = tmp16 * tmp21
        tmp24 = tmp22 * tmp23
        tmp26 = tmp24 + tmp25
        tmp27 = tmp26.to(tl.float32)
        tl.store(out_ptr2 + (r1 + (256*x0)), tmp27, rmask & xmask)
''', device_str='cuda')

import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import grid, split_scan_grid, start_graph, end_graph
from torch._C import _cuda_getCurrentRawStream as get_raw_stream


# kernel path: /tmp/torchinductor_root/as/caslnaz5ovc2w4ytfnlr2jibrkfgmcp4e7e42dxcuygano3oelsl.py
# Source Nodes: [linear], Original ATen: [aten._to_copy]
# linear => convert_element_type_2
triton_poi_fused__to_copy_1 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[262144], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp16', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_1', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 262144
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (x0), None)
    tmp1 = tmp0.to(tl.float32)
    tl.store(out_ptr0 + (x0), tmp1, None)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/vb/cvbmvuwb6uxqkrrldzge7qnzlzdqk2l4gdp337mybwl44sp4h43k.py
# Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
# x_1 => var_mean_1
triton_per_fused_native_group_norm_2 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.persistent_reduction(
    size_hints=[32768, 256],
    reduction_hint=ReductionHint.INNER,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_native_group_norm_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, 'num_reduction': 3, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_(in_ptr0, in_ptr1, out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr):
    xnumel = 29088
    rnumel = 163
    RBLOCK: tl.constexpr = 256
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rindex = tl.arange(0, RBLOCK)[None, :]
    roffset = 0
    rmask = rindex < rnumel
    r3 = rindex
    x0 = xindex % 101
    x1 = (xindex // 101) % 9
    x2 = (xindex // 909)
    x5 = xindex
    tmp0 = r3 + (163*x0)
    tmp1 = tl.full([1, 1], 16441, tl.int32)
    tmp2 = tmp0 < tmp1
    tmp3 = r3 + (163*x0) + (16441*x1)
    tmp4 = tl.full([1, 1], 147968, tl.int32)
    tmp5 = tmp3 < tmp4
    tmp6 = tmp5 & tmp2
    tmp7 = tl.load(in_ptr0 + ((32*x2) + (1024*((r3 + (163*x0) + (16441*x1)) % 4624)) + (((r3 + (163*x0) + (16441*x1)) // 4624) % 32)), rmask & tmp6 & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
    tmp8 = tl.load(in_ptr1 + ((32*x2) + (((r3 + (163*x0) + (16441*x1)) // 4624) % 32)), rmask & tmp6 & xmask, eviction_policy='evict_last', other=0.0)
    tmp9 = tmp8.to(tl.float32)
    tmp10 = tmp7 + tmp9
    tmp11 = tmp10.to(tl.float32)
    tmp12 = tl.full(tmp11.shape, 0, tmp11.dtype)
    tmp13 = tl.where(tmp6, tmp11, tmp12)
    tmp14 = tl.full(tmp13.shape, 0, tmp13.dtype)
    tmp15 = tl.where(tmp2, tmp13, tmp14)
    tmp16 = 0.0
    tmp17 = tl.full(tmp16.shape, 0, tmp16.dtype)
    tmp18 = tl.where(tmp6, tmp16, tmp17)
    tmp19 = tl.full(tmp18.shape, 0, tmp18.dtype)
    tmp20 = tl.where(tmp2, tmp18, tmp19)
    tmp21 = 1.0
    tmp22 = tl.full(tmp21.shape, 0, tmp21.dtype)
    tmp23 = tl.where(tmp6, tmp21, tmp22)
    tmp24 = tl.full(tmp23.shape, 0, tmp23.dtype)
    tmp25 = tl.where(tmp2, tmp23, tmp24)
    tmp26 = tl.broadcast_to(tmp15, [XBLOCK, RBLOCK])
    tmp27 = tl.broadcast_to(tmp20, [XBLOCK, RBLOCK])
    tmp28 = tl.broadcast_to(tmp25, [XBLOCK, RBLOCK])
    tmp30 = tl.where(rmask & xmask, tmp26, 0)
    tmp31 = tl.where(rmask & xmask, tmp27, 0)
    tmp32 = tl.where(rmask & xmask, tmp28, 0)
    tmp33, tmp34, tmp35 = triton_helpers.welford(tmp30, tmp31, tmp32, 1)
    tmp36 = tmp33[:, None]
    tmp37 = tmp34[:, None]
    tmp38 = tmp35[:, None]
    tl.store(out_ptr0 + (x5), tmp36, xmask)
    tl.store(out_ptr1 + (x5), tmp37, xmask)
    tl.store(out_ptr2 + (x5), tmp38, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/ry/cryr4surecktcuhmwmjogkgofvheizc7oycy6s35eitytxi4j6ua.py
# Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
# x_1 => var_mean_1
triton_per_fused_native_group_norm_3 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.persistent_reduction(
    size_hints=[512, 128],
    reduction_hint=ReductionHint.INNER,
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: 'i32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_native_group_norm_3', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 3, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, out_ptr2, xnumel, rnumel, XBLOCK : tl.constexpr):
    xnumel = 288
    rnumel = 101
    RBLOCK: tl.constexpr = 128
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rindex = tl.arange(0, RBLOCK)[None, :]
    roffset = 0
    rmask = rindex < rnumel
    r1 = rindex
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (r1 + (101*x0)), rmask & xmask, other=0.0)
    tmp1 = tl.load(in_ptr1 + (r1 + (101*x0)), rmask & xmask, other=0.0)
    tmp2 = tl.load(in_ptr2 + (r1 + (101*x0)), rmask & xmask, other=0.0)
    tmp3 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
    tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
    tmp5 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
    tmp7 = tl.where(rmask & xmask, tmp3, 0)
    tmp8 = tl.where(rmask & xmask, tmp4, 0)
    tmp9 = tl.where(rmask & xmask, tmp5, 0)
    tmp10, tmp11, tmp12 = triton_helpers.welford(tmp7, tmp8, tmp9, 1)
    tmp13 = tmp10[:, None]
    tmp14 = tmp11[:, None]
    tmp15 = tmp12[:, None]
    tl.store(out_ptr0 + (x0), tmp13, xmask)
    tl.store(out_ptr1 + (x0), tmp14, xmask)
    tl.store(out_ptr2 + (x0), tmp15, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/ti/ctiiksb3r3xvomwwtvvhkcylacpicxwzqr465usrd6g2ujlxocbb.py
# Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
# x_1 => var_mean_1
triton_per_fused_native_group_norm_4 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.persistent_reduction(
    size_hints=[32, 16],
    reduction_hint=ReductionHint.INNER,
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_native_group_norm_4', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, 'num_reduction': 2, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False}
)
@triton.jit
def triton_(in_ptr0, in_ptr1, in_ptr2, out_ptr0, out_ptr1, xnumel, rnumel, XBLOCK : tl.constexpr):
    xnumel = 32
    rnumel = 9
    RBLOCK: tl.constexpr = 16
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rindex = tl.arange(0, RBLOCK)[None, :]
    roffset = 0
    rmask = rindex < rnumel
    r1 = rindex
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (r1 + (9*x0)), rmask & xmask, other=0.0)
    tmp1 = tl.load(in_ptr1 + (r1 + (9*x0)), rmask & xmask, other=0.0)
    tmp2 = tl.load(in_ptr2 + (r1 + (9*x0)), rmask & xmask, other=0.0)
    tmp3 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
    tmp4 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
    tmp5 = tl.broadcast_to(tmp2, [XBLOCK, RBLOCK])
    tmp7 = tl.where(rmask & xmask, tmp3, 0)
    tmp8 = tl.where(rmask & xmask, tmp4, 0)
    tmp9 = tl.where(rmask & xmask, tmp5, 0)
    tmp10, tmp11, tmp12 = triton_helpers.welford(tmp7, tmp8, tmp9, 1)
    tmp13 = tmp10[:, None]
    tmp14 = tmp11[:, None]
    tmp15 = tmp12[:, None]
    tl.store(out_ptr0 + (x0), tmp13, xmask)
    tl.store(out_ptr1 + (x0), tmp14, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/wi/cwint6fgs7zu33a3pnijw6lclxbp7vkpkqy6yhslrgwbtyayqvnc.py
# Source Nodes: [x_1, x_2, x_3], Original ATen: [aten._to_copy, aten.gelu, aten.native_group_norm]
# x_1 => add_5, mul_3
# x_2 => add_6, erf, mul_4, mul_5, mul_6
# x_3 => convert_element_type_9
triton_poi_fused__to_copy_gelu_native_group_norm_5 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[8388608], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp16', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_gelu_native_group_norm_5', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, in_ptr5, out_ptr1, xnumel, XBLOCK : tl.constexpr):
    xnumel = 4734976
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x2 = xindex
    x0 = xindex % 1024
    tmp0 = tl.load(in_ptr0 + (x2), None).to(tl.float32)
    tmp1 = tl.load(in_ptr1 + (x0), None, eviction_policy='evict_last')
    tmp5 = tl.load(in_ptr2 + ((x0 // 32)), None, eviction_policy='evict_last')
    tmp7 = tl.load(in_ptr3 + ((x0 // 32)), None, eviction_policy='evict_last')
    tmp14 = tl.load(in_ptr4 + (x0), None, eviction_policy='evict_last')
    tmp16 = tl.load(in_ptr5 + (x0), None, eviction_policy='evict_last')
    tmp2 = tmp1.to(tl.float32)
    tmp3 = tmp0 + tmp2
    tmp4 = tmp3.to(tl.float32)
    tmp6 = tmp4 - tmp5
    tmp8 = 147968.0
    tmp9 = tmp7 / tmp8
    tmp10 = 1e-05
    tmp11 = tmp9 + tmp10
    tmp12 = libdevice.rsqrt(tmp11)
    tmp13 = tmp6 * tmp12
    tmp15 = tmp13 * tmp14
    tmp17 = tmp15 + tmp16
    tmp18 = 0.5
    tmp19 = tmp17 * tmp18
    tmp20 = 0.7071067811865476
    tmp21 = tmp17 * tmp20
    tmp22 = libdevice.erf(tmp21)
    tmp23 = 1.0
    tmp24 = tmp22 + tmp23
    tmp25 = tmp19 * tmp24
    tmp26 = tmp25.to(tl.float32)
    tl.store(out_ptr1 + (x2), tmp26, None)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/yp/cypln35vatxtjpb4pa4zsjqgqtmbm56jb4xbgkykms7hlfanclpj.py
# Source Nodes: [x_3], Original ATen: [aten._to_copy]
# x_3 => convert_element_type_8
triton_poi_fused__to_copy_6 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[32768], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp16', 2: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_6', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 25600
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (x0), xmask)
    tmp1 = tmp0.to(tl.float32)
    tl.store(out_ptr0 + (x0), tmp1, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/qs/cqsphp2cldk4m2hewpws4r2xnfayv5pzgaryfkhemfzjnnrkdule.py
# Source Nodes: [add, tgt, tgt2, tgt_1], Original ATen: [aten._to_copy, aten.add]
# add => add
# tgt => add_1
# tgt2 => add_7, convert_element_type_10
# tgt_1 => add_8
triton_poi_fused__to_copy_add_7 = async_compile.triton('triton_', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

from torch._inductor.runtime import triton_helpers, triton_heuristics
from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, instance_descriptor, DeviceProperties

@triton_heuristics.pointwise(
    size_hints=[8192, 256], tile_hint=TileHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: '*fp16', 3: '*fp16', 4: '*fp32', 5: '*fp16', 6: 'i32', 7: 'i32'}, 'device': DeviceProperties(type='cuda', index=0, cc=89, major=8, regs_per_multiprocessor=65536, max_threads_per_multi_processor=1536, multi_processor_count=58), 'constants': {}, 'configs': [AttrsDescriptor(divisible_by_16=(0, 1, 2, 3, 4, 5, 6, 7), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_add_7', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 5, 'num_reduction': 0, 'backend_hash': '72c34bdb145549777ca2f0838f26abe42bb446cf528c78d229508b5a55e67a78', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': False, 'force_disable_caches': False, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False},
    min_elem_per_thread=0
)
@triton.jit
def triton_(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    ynumel = 4624
    xnumel = 256
    yoffset = (tl.program_id(1) + tl.program_id(2) * tl.num_programs(1)) * YBLOCK
    yindex = yoffset + tl.arange(0, YBLOCK)[None, :]
    ymask = yindex < ynumel
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    x1 = xindex
    y0 = yindex
    tmp0 = tl.load(in_ptr0 + (y0 + (4624*x1)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp1 = tl.load(in_ptr1 + (x1 + (256*y0)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp2 = tl.load(in_ptr2 + (x1 + (256*y0)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp5 = tl.load(in_ptr3 + (x1 + (256*y0)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp6 = tl.load(in_ptr4 + (x1), xmask, eviction_policy='evict_last')
    tmp3 = tmp1 + tmp2
    tmp4 = tmp0 + tmp3
    tmp7 = tmp6.to(tl.float32)
    tmp8 = tmp5 + tmp7
    tmp9 = tmp4 + tmp8
    tl.store(out_ptr0 + (y0 + (4624*x1)), tmp9, xmask & ymask)
''', device_str='cuda')


async_compile.wait(globals())
del async_compile

def call(args):
    arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1, arg9_1, arg10_1, arg11_1 = args
    args.clear()
    assert_size_stride(arg0_1, (4624, 1, 256), (256, 256, 1))
    assert_size_stride(arg1_1, (4624, 1, 256), (256, 256, 1))
    assert_size_stride(arg2_1, (4624, 1, 256), (1, 1183744, 4624))
    assert_size_stride(arg3_1, (256, ), (1, ))
    assert_size_stride(arg4_1, (256, ), (1, ))
    assert_size_stride(arg5_1, (1024, 256), (256, 1))
    assert_size_stride(arg6_1, (1024, ), (1, ))
    assert_size_stride(arg7_1, (1024, ), (1, ))
    assert_size_stride(arg8_1, (1024, ), (1, ))
    assert_size_stride(arg9_1, (1024, 1, 5, 5), (25, 25, 5, 1))
    assert_size_stride(arg10_1, (256, 1024), (1024, 1))
    assert_size_stride(arg11_1, (256, ), (1, ))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        buf3 = empty_strided_cuda((4624, 1, 256), (256, 256, 1), torch.float16)
        # Source Nodes: [_tgt, add, linear, tgt], Original ATen: [aten._to_copy, aten.add, aten.native_layer_norm]
        stream0 = get_raw_stream(0)
        triton_red_fused__to_copy_add_native_layer_norm_0.run(arg2_1, arg1_1, arg0_1, arg3_1, arg4_1, buf3, 4624, 256, grid=grid(4624), stream=stream0)
        del arg3_1
        del arg4_1
        buf4 = empty_strided_cuda((1024, 256), (256, 1), torch.float16)
        # Source Nodes: [linear], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_1.run(arg5_1, buf4, 262144, grid=grid(262144), stream=stream0)
        del arg5_1
        buf5 = empty_strided_cuda((4624, 1024), (1024, 1), torch.float16)
        # Source Nodes: [], Original ATen: []
        extern_kernels.mm(reinterpret_tensor(buf3, (4624, 256), (256, 1), 0), reinterpret_tensor(buf4, (256, 1024), (1, 256), 0), out=buf5)
        buf6 = empty_strided_cuda((1, 32, 1, 1, 9, 101), (29088, 909, 29088, 29088, 101, 1), torch.float32)
        buf7 = empty_strided_cuda((1, 32, 1, 1, 9, 101), (29088, 909, 29088, 29088, 101, 1), torch.float32)
        buf8 = empty_strided_cuda((1, 32, 1, 1, 9, 101), (29088, 909, 29088, 29088, 101, 1), torch.float32)
        # Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
        triton_per_fused_native_group_norm_2.run(buf5, arg6_1, buf6, buf7, buf8, 29088, 163, grid=grid(29088), stream=stream0)
        buf9 = empty_strided_cuda((1, 32, 1, 1, 9), (288, 9, 288, 288, 1), torch.float32)
        buf10 = empty_strided_cuda((1, 32, 1, 1, 9), (288, 9, 288, 288, 1), torch.float32)
        buf11 = empty_strided_cuda((1, 32, 1, 1, 9), (288, 9, 288, 288, 1), torch.float32)
        # Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
        triton_per_fused_native_group_norm_3.run(buf6, buf7, buf8, buf9, buf10, buf11, 288, 101, grid=grid(288), stream=stream0)
        del buf6
        del buf7
        del buf8
        buf12 = empty_strided_cuda((1, 32, 1, 1), (32, 1, 32, 32), torch.float32)
        buf13 = empty_strided_cuda((1, 32, 1, 1), (32, 1, 32, 32), torch.float32)
        # Source Nodes: [x_1], Original ATen: [aten.native_group_norm]
        triton_per_fused_native_group_norm_4.run(buf9, buf10, buf11, buf12, buf13, 32, 9, grid=grid(32), stream=stream0)
        del buf10
        del buf11
        del buf9
        buf16 = empty_strided_cuda((1, 1024, 68, 68), (4734976, 1, 69632, 1024), torch.float16)
        # Source Nodes: [x_1, x_2, x_3], Original ATen: [aten._to_copy, aten.gelu, aten.native_group_norm]
        triton_poi_fused__to_copy_gelu_native_group_norm_5.run(buf5, arg6_1, buf12, buf13, arg7_1, arg8_1, buf16, 4734976, grid=grid(4734976), stream=stream0)
        del arg6_1
        del arg7_1
        del arg8_1
        del buf12
        del buf13
        del buf5
        buf17 = empty_strided_cuda((1024, 1, 5, 5), (25, 25, 5, 1), torch.float16)
        # Source Nodes: [x_3], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_6.run(arg9_1, buf17, 25600, grid=grid(25600), stream=stream0)
        del arg9_1
        # Source Nodes: [x_2, x_3], Original ATen: [aten._to_copy, aten.convolution, aten.gelu]
        buf18 = extern_kernels.convolution(buf16, buf17, stride=(1, 1), padding=(2, 2), dilation=(1, 1), transposed=False, output_padding=(0, 0), groups=1024, bias=None)
        assert_size_stride(buf18, (1, 1024, 68, 68), (4734976, 1, 69632, 1024))
        del buf16
        del buf17
        buf19 = reinterpret_tensor(buf4, (256, 1024), (1024, 1), 0); del buf4  # reuse
        # Source Nodes: [tgt2], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_1.run(arg10_1, buf19, 262144, grid=grid(262144), stream=stream0)
        del arg10_1
        buf20 = buf3; del buf3  # reuse
        # Source Nodes: [tgt2], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf18, (4624, 1, 1024), (1024, 0, 1), 0), reinterpret_tensor(buf19, (4624, 1024, 256), (0, 1, 1024), 0), out=buf20)
        del buf18
        del buf19
        buf21 = empty_strided_cuda((4624, 1, 256), (1, 1183744, 4624), torch.float16)
        # Source Nodes: [add, tgt, tgt2, tgt_1], Original ATen: [aten._to_copy, aten.add]
        triton_poi_fused__to_copy_add_7.run(arg2_1, arg1_1, arg0_1, buf20, arg11_1, buf21, 4624, 256, grid=grid(4624, 256), stream=stream0)
        del arg0_1
        del arg11_1
        del arg1_1
        del arg2_1
        del buf20
    return (buf21, )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    arg0_1 = rand_strided((4624, 1, 256), (256, 256, 1), device='cuda:0', dtype=torch.float16)
    arg1_1 = rand_strided((4624, 1, 256), (256, 256, 1), device='cuda:0', dtype=torch.float16)
    arg2_1 = rand_strided((4624, 1, 256), (1, 1183744, 4624), device='cuda:0', dtype=torch.float16)
    arg3_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg4_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg5_1 = rand_strided((1024, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg6_1 = rand_strided((1024, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg7_1 = rand_strided((1024, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg8_1 = rand_strided((1024, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg9_1 = rand_strided((1024, 1, 5, 5), (25, 25, 5, 1), device='cuda:0', dtype=torch.float32)
    arg10_1 = rand_strided((256, 1024), (1024, 1), device='cuda:0', dtype=torch.float32)
    arg11_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    fn = lambda: call([arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1, arg9_1, arg10_1, arg11_1])
    return print_performance(fn, times=times, repeat=repeat)


if __name__ == "__main__":
    from torch._inductor.wrapper_benchmark import compiled_module_main
    compiled_module_main('None', benchmark_compiled_module)