# AOT ID: ['32_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/3o/c3oqwkgp2ohvdmnxuni6tremxihpetzns5pumbew4eabdb736ckq.py
# Source Nodes: [_tgt], Original ATen: [aten.native_layer_norm]
# _tgt => clone, var_mean
triton_red_fused_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: '*fp32', 1: '*fp32', 2: '*fp32', 3: 'i32', 4: 'i32', 5: '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, 5), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused_native_layer_norm_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 1, '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, out_ptr0, out_ptr1, ks0, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
    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
    tmp2_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    tmp2_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    tmp2_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 + (ks0*r1)), rmask & xmask, eviction_policy='evict_first', other=0.0)
        tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
        tmp2_mean_next, tmp2_m2_next, tmp2_weight_next = triton_helpers.welford_reduce(
            tmp1, tmp2_mean, tmp2_m2, tmp2_weight, roffset == 0
        )
        tmp2_mean = tl.where(rmask & xmask, tmp2_mean_next, tmp2_mean)
        tmp2_m2 = tl.where(rmask & xmask, tmp2_m2_next, tmp2_m2)
        tmp2_weight = tl.where(rmask & xmask, tmp2_weight_next, tmp2_weight)
    tmp2_tmp, tmp3_tmp, tmp4_tmp = triton_helpers.welford(
        tmp2_mean, tmp2_m2, tmp2_weight, 1
    )
    tmp2 = tmp2_tmp[:, None]
    tmp3 = tmp3_tmp[:, None]
    tmp4 = tmp4_tmp[:, None]
    tl.store(out_ptr0 + (x0), tmp2, xmask)
    tl.store(out_ptr1 + (x0), tmp3, 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/td/ctdr76h6hf6pzyvx2oh4sxjiqp6lf5z3vsoe4fm56pcedzd6ofyb.py
# Source Nodes: [_tgt], Original ATen: [aten.native_layer_norm]
# _tgt => add, add_1, clone, mul, mul_1, rsqrt, sub, var_mean
triton_poi_fused_native_layer_norm_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=[2097152], 
    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, 7), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_native_layer_norm_1', '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, ks0, xnumel, XBLOCK : tl.constexpr):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x2 = xindex
    x0 = xindex % ks0
    x1 = (xindex // ks0)
    tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last')
    tmp1 = tl.load(in_ptr1 + (x0), xmask, eviction_policy='evict_last')
    tmp3 = tl.load(in_ptr2 + (x0), xmask, eviction_policy='evict_last')
    tmp10 = tl.load(in_ptr3 + (x1), xmask, eviction_policy='evict_last')
    tmp12 = tl.load(in_ptr4 + (x1), xmask, eviction_policy='evict_last')
    tmp2 = tmp0 - tmp1
    tmp4 = 256.0
    tmp5 = tmp3 / tmp4
    tmp6 = 1e-05
    tmp7 = tmp5 + tmp6
    tmp8 = libdevice.rsqrt(tmp7)
    tmp9 = tmp2 * tmp8
    tmp11 = tmp9 * tmp10
    tmp13 = tmp11 + tmp12
    tl.store(out_ptr0 + (x2), tmp13, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/nv/cnvli6mu2mcjd6ex2fccnsefdzeslc26eayftxepoug22nm3szzk.py
# Source Nodes: [K, Q, V, q], Original ATen: [aten._to_copy, aten.add]
# K => convert_element_type_8
# Q => convert_element_type_2
# V => convert_element_type_14
# q => add_2
triton_poi_fused__to_copy_add_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.pointwise(
    size_hints=[8192, 256], tile_hint=TileHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: '*fp16', 3: '*fp16', 4: '*fp16', 5: 'i32', 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, 7), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__to_copy_add_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 2, '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, out_ptr0, out_ptr1, out_ptr2, ks0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    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 + (ks0*x1)), xmask & ymask, eviction_policy='evict_last')
    tmp1 = tl.load(in_ptr1 + (x1 + (256*y0)), xmask & ymask, eviction_policy='evict_last')
    tmp2 = tmp0 + tmp1
    tmp3 = tmp2.to(tl.float32)
    tmp4 = tmp0.to(tl.float32)
    tl.store(out_ptr0 + (x1 + (256*y0)), tmp3, xmask & ymask)
    tl.store(out_ptr1 + (x1 + (256*y0)), tmp3, xmask & ymask)
    tl.store(out_ptr2 + (x1 + (256*y0)), tmp4, xmask & ymask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/kz/ckzduggze346udedgsxmdiricbabx3tfjxwxwwhusjneu3pglj4q.py
# Source Nodes: [Q], Original ATen: [aten._to_copy]
# Q => convert_element_type_1
triton_poi_fused__to_copy_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.pointwise(
    size_hints=[65536], 
    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_3', '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 = 65536
    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/d5/cd5gklptjfgmsnojexjq3gio2cjwhcypvbmcidhdrovt2bc4qqlj.py
# Source Nodes: [K], Original ATen: [aten._to_copy]
# K => convert_element_type_6
triton_poi_fused__to_copy_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.pointwise(
    size_hints=[256], 
    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_4', '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 = 256
    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/nn/cnne6gjdvkck6rhmy5kew2quih6x5nszwxn2dj2rf7icqxm2kzfn.py
# Source Nodes: [Q_1], Original ATen: [aten.div]
# Q_1 => div
triton_poi_fused_div_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=[2097152], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp32', 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_div_5', 'mutated_arg_names': ['in_out_ptr0'], 'no_x_dim': False, 'num_load': 2, '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_out_ptr0, in_ptr0, xnumel, XBLOCK : tl.constexpr):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x2 = xindex
    x0 = xindex % 256
    tmp0 = tl.load(in_out_ptr0 + (x2), xmask).to(tl.float32)
    tmp1 = tl.load(in_ptr0 + (x0), xmask, eviction_policy='evict_last')
    tmp2 = tmp1.to(tl.float32)
    tmp3 = tmp0 + tmp2
    tmp4 = 0.17677669529663687
    tmp5 = tmp3 * tmp4
    tl.store(in_out_ptr0 + (x2), tmp5, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/qq/cqqshablwml5tfeua2bypmz4kst7tikri7wcfqlb5vdoi4uy2xl5.py
# Source Nodes: [attn, matmul_1], Original ATen: [aten._softmax, aten._to_copy]
# attn => amax, convert_element_type_20, div_1, exp, sub_1, sum_1
# matmul_1 => convert_element_type_21
triton_red_fused__softmax__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.reduction(
    size_hints=[65536, 8192],
    reduction_hint=ReductionHint.INNER,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: 'i32', 3: 'i32', 4: '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), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__softmax__to_copy_6', '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, out_ptr2, ks0, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rbase = tl.arange(0, RBLOCK)[None, :]
    x0 = xindex
    _tmp3 = tl.full([XBLOCK, RBLOCK], float("-inf"), tl.float32)
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp0 = tl.load(in_ptr0 + (r1 + (ks0*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp1 = tmp0.to(tl.float32)
        tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
        tmp4 = triton_helpers.maximum(_tmp3, tmp2)
        _tmp3 = tl.where(rmask & xmask, tmp4, _tmp3)
    tmp3 = triton_helpers.max2(_tmp3, 1)[:, None]
    _tmp10 = tl.full([XBLOCK, RBLOCK], 0, tl.float32)
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp5 = tl.load(in_ptr0 + (r1 + (ks0*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp6 = tmp5.to(tl.float32)
        tmp7 = tmp6 - tmp3
        tmp8 = tl_math.exp(tmp7)
        tmp9 = tl.broadcast_to(tmp8, [XBLOCK, RBLOCK])
        tmp11 = _tmp10 + tmp9
        _tmp10 = tl.where(rmask & xmask, tmp11, _tmp10)
    tmp10 = tl.sum(_tmp10, 1)[:, None]
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp12 = tl.load(in_ptr0 + (r1 + (ks0*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp13 = tmp12.to(tl.float32)
        tmp14 = tmp13 - tmp3
        tmp15 = tl_math.exp(tmp14)
        tmp16 = tmp15 / tmp10
        tmp17 = tmp16.to(tl.float32)
        tl.store(out_ptr2 + (r1 + (ks0*x0)), tmp17, rmask & xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/vt/cvtlhmmqj3paiea6czszqhvkwy24f75kmg3xqwcrl5goxwjfnpye.py
# Source Nodes: [outputs_1], Original ATen: [aten.clone]
# outputs_1 => clone_2
triton_poi_fused_clone_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=[2097152], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: 'i32', 3: 'i32', 4: 'i32', 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), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_7', '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, ks0, ks1, ks2, ks3, xnumel, XBLOCK : tl.constexpr):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex % ks0
    x1 = (xindex // ks0) % ks1
    x2 = (xindex // ks2)
    x3 = xindex
    tmp0 = tl.load(in_ptr0 + (x0 + (x2*(256 // ks1)) + (ks3*x1*(256 // ks1))), xmask, eviction_policy='evict_last').to(tl.float32)
    tl.store(out_ptr0 + (x3), tmp0, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/ri/crirp7kbh4uk7kdv6mqujtkgg2wfzkri5b4lp46dk7irwncbtebs.py
# Source Nodes: [tgt], Original ATen: [aten.add]
# tgt => add_3
triton_poi_fused_add_8 = 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=[256, 8192], tile_hint=TileHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp16', 2: '*fp32', 3: '*fp32', 4: 'i32', 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, 5), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_8', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 3, '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, out_ptr0, ks0, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    ynumel = 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 + (x1 + (ks0*y0)), xmask & ymask, eviction_policy='evict_last')
    tmp1 = tl.load(in_ptr1 + (y0 + (256*x1)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tmp2 = tl.load(in_ptr2 + (y0), ymask, eviction_policy='evict_last')
    tmp3 = tmp2.to(tl.float32)
    tmp4 = tmp1 + tmp3
    tmp5 = tmp4.to(tl.float32)
    tmp6 = tmp0 + tmp5
    tl.store(out_ptr0 + (x1 + (ks0*y0)), tmp6, xmask & ymask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/z7/cz7o4l5j7vxspld2licxa63hezaaazcgvjuxrapbcmvhzv22iatj.py
# Source Nodes: [_tgt_1, curr_QV], Original ATen: [aten._to_copy, aten.native_layer_norm]
# _tgt_1 => add_4, add_5, clone_3, mul_4, mul_5, rsqrt_1, sub_2, var_mean_1
# curr_QV => convert_element_type_31
triton_red_fused__to_copy_native_layer_norm_9 = 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: '*fp32', 1: '*fp32', 2: '*fp32', 3: '*fp32', 4: '*fp16', 5: 'i32', 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, 7), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__to_copy_native_layer_norm_9', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 4, '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_ptr2, out_ptr3, ks0, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
    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
    tmp2_mean = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    tmp2_m2 = tl.zeros([XBLOCK, RBLOCK], tl.float32)
    tmp2_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 + (ks0*r1)), rmask & xmask, eviction_policy='evict_last', other=0.0)
        tmp1 = tl.broadcast_to(tmp0, [XBLOCK, RBLOCK])
        tmp2_mean_next, tmp2_m2_next, tmp2_weight_next = triton_helpers.welford_reduce(
            tmp1, tmp2_mean, tmp2_m2, tmp2_weight, roffset == 0
        )
        tmp2_mean = tl.where(rmask & xmask, tmp2_mean_next, tmp2_mean)
        tmp2_m2 = tl.where(rmask & xmask, tmp2_m2_next, tmp2_m2)
        tmp2_weight = tl.where(rmask & xmask, tmp2_weight_next, tmp2_weight)
    tmp2_tmp, tmp3_tmp, tmp4_tmp = triton_helpers.welford(
        tmp2_mean, tmp2_m2, tmp2_weight, 1
    )
    tmp2 = tmp2_tmp[:, None]
    tmp3 = tmp3_tmp[:, None]
    tmp4 = tmp4_tmp[:, None]
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp5 = tl.load(in_ptr0 + (x0 + (ks0*r1)), rmask & xmask, eviction_policy='evict_first', other=0.0)
        tmp13 = tl.load(in_ptr1 + (r1), rmask, eviction_policy='evict_last', other=0.0)
        tmp15 = tl.load(in_ptr2 + (r1), rmask, eviction_policy='evict_last', other=0.0)
        tmp6 = tmp5 - tmp2
        tmp7 = 256.0
        tmp8 = tmp3 / tmp7
        tmp9 = 1e-05
        tmp10 = tmp8 + tmp9
        tmp11 = libdevice.rsqrt(tmp10)
        tmp12 = tmp6 * tmp11
        tmp14 = tmp12 * tmp13
        tmp16 = tmp14 + tmp15
        tmp17 = tmp16.to(tl.float32)
        tl.store(out_ptr2 + (r1 + (256*x0)), tmp16, rmask & xmask)
        tl.store(out_ptr3 + (r1 + (256*x0)), tmp17, rmask & xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/nt/cntgi5n3ffqlwx4jgbpznwlg7443khi6sshc4tqaccctsj75suwy.py
# Source Nodes: [curr_QV], Original ATen: [aten._to_copy]
# curr_QV => convert_element_type_30
triton_poi_fused__to_copy_10 = 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=[131072], 
    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_10', '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 = 131072
    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/if/cif37kq5wuj6dvzt5dile4o44kv53ktbtaspkmma6otkg2oqkbgm.py
# Source Nodes: [curr_QV], Original ATen: [aten._to_copy]
# curr_QV => convert_element_type_29
triton_poi_fused__to_copy_11 = 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=[512], 
    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_11', '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 = 512
    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/zd/czdpuiyjbl5qxpgwks5i4yui24ux76ypw5ofxcd4sfbxtb4hw2ed.py
# Source Nodes: [tensor], Original ATen: [aten.clone]
# tensor => clone_4
triton_poi_fused_clone_12 = 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=[256, 8192], tile_hint=TileHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: 'i32', 3: 'i32', 4: 'i32', 5: '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, 4), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_clone_12', '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, ks0, ks1, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    ynumel = 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 + (512*x1)), xmask & ymask, eviction_policy='evict_last').to(tl.float32)
    tl.store(out_ptr0 + (x1 + (ks0*ks1*y0)), tmp0, xmask & ymask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/e6/ce62mocjweni7mvpjqis5chks7hh4s33ua65vega4e4ld5weoe5u.py
# Source Nodes: [Q_3], Original ATen: [aten.div]
# Q_3 => div_2
triton_poi_fused_div_13 = 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=[2097152], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 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_div_13', '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):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex % 256
    x1 = (xindex // 256)
    x2 = xindex
    tmp0 = tl.load(in_ptr0 + (x0 + (512*x1)), xmask).to(tl.float32)
    tmp1 = 0.0625
    tmp2 = tmp0 * tmp1
    tl.store(out_ptr0 + (x2), tmp2, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/tc/ctcpdlxvy63zg5gdydpegh54odqcgrj2irlgazf6lehjt6mlsfej.py
# Source Nodes: [attn_2, matmul_3], Original ATen: [aten._softmax, aten._to_copy]
# attn_2 => amax_1, convert_element_type_37, div_3, exp_1, sub_3, sum_2
# matmul_3 => convert_element_type_38
triton_red_fused__softmax__to_copy_14 = 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, 16384],
    reduction_hint=ReductionHint.INNER,
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: 'i32', 3: 'i32', 4: '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), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__softmax__to_copy_14', '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, out_ptr2, ks0, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rbase = tl.arange(0, RBLOCK)[None, :]
    x0 = xindex
    _tmp3 = tl.full([XBLOCK, RBLOCK], float("-inf"), tl.float32)
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp0 = tl.load(in_ptr0 + (r1 + (ks0*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp1 = tmp0.to(tl.float32)
        tmp2 = tl.broadcast_to(tmp1, [XBLOCK, RBLOCK])
        tmp4 = triton_helpers.maximum(_tmp3, tmp2)
        _tmp3 = tl.where(rmask & xmask, tmp4, _tmp3)
    tmp3 = triton_helpers.max2(_tmp3, 1)[:, None]
    _tmp10 = tl.full([XBLOCK, RBLOCK], 0, tl.float32)
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp5 = tl.load(in_ptr0 + (r1 + (ks0*x0)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp6 = tmp5.to(tl.float32)
        tmp7 = tmp6 - tmp3
        tmp8 = tl_math.exp(tmp7)
        tmp9 = tl.broadcast_to(tmp8, [XBLOCK, RBLOCK])
        tmp11 = _tmp10 + tmp9
        _tmp10 = tl.where(rmask & xmask, tmp11, _tmp10)
    tmp10 = tl.sum(_tmp10, 1)[:, None]
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp12 = tl.load(in_ptr0 + (r1 + (ks0*x0)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp13 = tmp12.to(tl.float32)
        tmp14 = tmp13 - tmp3
        tmp15 = tl_math.exp(tmp14)
        tmp16 = tmp15 / tmp10
        tmp17 = tmp16.to(tl.float32)
        tl.store(out_ptr2 + (r1 + (ks0*x0)), tmp17, rmask & xmask)
''', 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, arg12_1, arg13_1, arg14_1, arg15_1, arg16_1, arg17_1, arg18_1, arg19_1, arg20_1, arg21_1, arg22_1, arg23_1, arg24_1, arg25_1, arg26_1 = args
    args.clear()
    s2 = arg2_1
    s6 = arg5_1
    s7 = arg6_1
    s8 = arg13_1
    s9 = arg20_1
    s10 = arg21_1
    s11 = arg22_1
    assert_size_stride(arg0_1, (256, ), (1, ))
    assert_size_stride(arg1_1, (256, ), (1, ))
    assert_size_stride(arg3_1, (s2, 1, 256), (1, 256*s2, s2))
    assert_size_stride(arg4_1, (s2, 1, 256), (256, 256, 1))
    assert_size_stride(arg7_1, (256, 256), (256, 1))
    assert_size_stride(arg8_1, (256, ), (1, ))
    assert_size_stride(arg9_1, (256, 256), (256, 1))
    assert_size_stride(arg10_1, (256, ), (1, ))
    assert_size_stride(arg11_1, (256, 256), (256, 1))
    assert_size_stride(arg12_1, (256, ), (1, ))
    assert_size_stride(arg14_1, (256, 256), (256, 1))
    assert_size_stride(arg15_1, (256, ), (1, ))
    assert_size_stride(arg16_1, (256, ), (1, ))
    assert_size_stride(arg17_1, (256, ), (1, ))
    assert_size_stride(arg18_1, (512, 256), (256, 1))
    assert_size_stride(arg19_1, (512, ), (1, ))
    assert_size_stride(arg23_1, (s11, 1, 256), (256, 256, 1))
    assert_size_stride(arg24_1, (s11, 1, 256), (256, 256, 1))
    assert_size_stride(arg25_1, (256, 256), (256, 1))
    assert_size_stride(arg26_1, (256, ), (1, ))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        buf0 = empty_strided_cuda((s2, 1, 1), (1, s2, s2), torch.float32)
        buf1 = empty_strided_cuda((s2, 1, 1), (1, s2, s2), torch.float32)
        # Source Nodes: [_tgt], Original ATen: [aten.native_layer_norm]
        stream0 = get_raw_stream(0)
        triton_red_fused_native_layer_norm_0.run(arg3_1, buf0, buf1, s2, s2, 256, grid=grid(s2), stream=stream0)
        buf3 = empty_strided_cuda((s2, 1, 256), (1, 256*s2, s2), torch.float32)
        # Source Nodes: [_tgt], Original ATen: [aten.native_layer_norm]
        triton_poi_fused_native_layer_norm_1_xnumel = 256*s2
        triton_poi_fused_native_layer_norm_1.run(arg3_1, buf0, buf1, arg0_1, arg1_1, buf3, s2, triton_poi_fused_native_layer_norm_1_xnumel, grid=grid(triton_poi_fused_native_layer_norm_1_xnumel), stream=stream0)
        del arg0_1
        del arg1_1
        del buf0
        del buf1
        buf4 = empty_strided_cuda((s2, 1, 256), (256, 256, 1), torch.float16)
        buf7 = empty_strided_cuda((s2, 1, 256), (256, 256, 1), torch.float16)
        buf15 = empty_strided_cuda((s2, 1, 256), (256, 256, 1), torch.float16)
        # Source Nodes: [K, Q, V, q], Original ATen: [aten._to_copy, aten.add]
        triton_poi_fused__to_copy_add_2.run(buf3, arg4_1, buf4, buf7, buf15, s2, s2, 256, grid=grid(s2, 256), stream=stream0)
        del arg4_1
        buf5 = empty_strided_cuda((256, 256), (256, 1), torch.float16)
        # Source Nodes: [Q], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_3.run(arg7_1, buf5, 65536, grid=grid(65536), stream=stream0)
        del arg7_1
        buf6 = empty_strided_cuda((s2, 256), (256, 1), torch.float16)
        # Source Nodes: [], Original ATen: []
        extern_kernels.mm(reinterpret_tensor(buf4, (s2, 256), (256, 1), 0), reinterpret_tensor(buf5, (256, 256), (1, 256), 0), out=buf6)
        buf8 = buf5; del buf5  # reuse
        # Source Nodes: [K], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_3.run(arg9_1, buf8, 65536, grid=grid(65536), stream=stream0)
        del arg9_1
        buf9 = empty_strided_cuda((256, ), (1, ), torch.float16)
        # Source Nodes: [K], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_4.run(arg10_1, buf9, 256, grid=grid(256), stream=stream0)
        del arg10_1
        buf10 = reinterpret_tensor(buf4, (s2, 256), (256, 1), 0); del buf4  # reuse
        # Source Nodes: [K], Original ATen: [aten._to_copy, aten.addmm]
        extern_kernels.addmm(buf9, reinterpret_tensor(buf7, (s2, 256), (256, 1), 0), reinterpret_tensor(buf8, (256, 256), (1, 256), 0), alpha=1, beta=1, out=buf10)
        del buf7
        buf11 = reinterpret_tensor(buf6, (s2, 1, 256), (256, 256, 1), 0); del buf6  # reuse
        # Source Nodes: [Q_1], Original ATen: [aten.div]
        triton_poi_fused_div_5_xnumel = 256*s2
        triton_poi_fused_div_5.run(buf11, arg8_1, triton_poi_fused_div_5_xnumel, grid=grid(triton_poi_fused_div_5_xnumel), stream=stream0)
        del arg8_1
        buf12 = empty_strided_cuda((s6, s2, s2), (s2*s2, s2, 1), torch.float16)
        # Source Nodes: [QK], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf11, (s6, s2, (256 // s6)), (s8, 256, 1), 0), reinterpret_tensor(buf10, (s6, (256 // s6), s2), (s8, 1, s6*s8), 0), out=buf12)
        del buf10
        buf19 = empty_strided_cuda((1, s6, s2, s2), (s6*(s2*s2), s2*s2, s2, 1), torch.float16)
        # Source Nodes: [attn, matmul_1], Original ATen: [aten._softmax, aten._to_copy]
        triton_red_fused__softmax__to_copy_6_xnumel = s6*s2
        triton_red_fused__softmax__to_copy_6.run(buf12, buf19, s2, triton_red_fused__softmax__to_copy_6_xnumel, s2, grid=grid(triton_red_fused__softmax__to_copy_6_xnumel), stream=stream0)
        del buf12
        buf16 = buf8; del buf8  # reuse
        # Source Nodes: [V], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_3.run(arg11_1, buf16, 65536, grid=grid(65536), stream=stream0)
        del arg11_1
        buf17 = buf9; del buf9  # reuse
        # Source Nodes: [V], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_4.run(arg12_1, buf17, 256, grid=grid(256), stream=stream0)
        del arg12_1
        buf18 = reinterpret_tensor(buf11, (s2, 256), (256, 1), 0); del buf11  # reuse
        # Source Nodes: [V], Original ATen: [aten._to_copy, aten.addmm]
        extern_kernels.addmm(buf17, reinterpret_tensor(buf15, (s2, 256), (256, 1), 0), reinterpret_tensor(buf16, (256, 256), (1, 256), 0), alpha=1, beta=1, out=buf18)
        buf20 = empty_strided_cuda((s6, s2, (256 // s6)), (s2*(256 // s6), (256 // s6), 1), torch.float16)
        # Source Nodes: [matmul_1], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf19, (s6, s2, s2), (s2*s2, s2, 1), 0), reinterpret_tensor(buf18, (s6, s2, (256 // s6)), (s7, s6*s7, 1), 0), out=buf20)
        del buf19
        ps0 = (256 // s6)
        ps1 = s6*(256 // s6)
        buf21 = empty_strided_cuda((s2, 1, s6, (256 // s6)), (s6*(256 // s6), 1, (256 // s6), 1), torch.float16)
        # Source Nodes: [outputs_1], Original ATen: [aten.clone]
        triton_poi_fused_clone_7_xnumel = s6*s2*(256 // s6)
        triton_poi_fused_clone_7.run(buf20, buf21, ps0, s6, ps1, s2, triton_poi_fused_clone_7_xnumel, grid=grid(triton_poi_fused_clone_7_xnumel), stream=stream0)
        del buf20
        buf22 = buf16; del buf16  # reuse
        # Source Nodes: [outputs_2], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_3.run(arg14_1, buf22, 65536, grid=grid(65536), stream=stream0)
        del arg14_1
        buf23 = buf18; del buf18  # reuse
        # Source Nodes: [], Original ATen: []
        extern_kernels.mm(reinterpret_tensor(buf21, (s2, s6*(256 // s6)), (s6*(256 // s6), 1), 0), reinterpret_tensor(buf22, (256, 256), (1, 256), 0), out=buf23)
        del buf21
        buf24 = buf3; del buf3  # reuse
        # Source Nodes: [tgt], Original ATen: [aten.add]
        triton_poi_fused_add_8.run(arg3_1, buf23, arg15_1, buf24, s2, 256, s2, grid=grid(256, s2), stream=stream0)
        del arg15_1
        del arg3_1
        buf28 = empty_strided_cuda((s2, 1, 256), (256, 256, 1), torch.float32)
        buf29 = reinterpret_tensor(buf23, (s2, 1, 256), (256, 256, 1), 0); del buf23  # reuse
        # Source Nodes: [_tgt_1, curr_QV], Original ATen: [aten._to_copy, aten.native_layer_norm]
        triton_red_fused__to_copy_native_layer_norm_9.run(buf24, arg16_1, arg17_1, buf28, buf29, s2, s2, 256, grid=grid(s2), stream=stream0)
        del arg16_1
        del arg17_1
        buf30 = empty_strided_cuda((512, 256), (256, 1), torch.float16)
        # Source Nodes: [curr_QV], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_10.run(arg18_1, buf30, 131072, grid=grid(131072), stream=stream0)
        del arg18_1
        buf31 = empty_strided_cuda((512, ), (1, ), torch.float16)
        # Source Nodes: [curr_QV], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_11.run(arg19_1, buf31, 512, grid=grid(512), stream=stream0)
        del arg19_1
        buf32 = empty_strided_cuda((s2, 512), (512, 1), torch.float16)
        # Source Nodes: [curr_QV], Original ATen: [aten._to_copy, aten.addmm]
        extern_kernels.addmm(buf31, reinterpret_tensor(buf29, (s2, 256), (256, 1), 0), reinterpret_tensor(buf30, (256, 512), (1, 256), 0), alpha=1, beta=1, out=buf32)
        del buf30
        del buf31
        buf33 = empty_strided_cuda((1, 256, s9, s10), (256*s10*s9, s10*s9, s10, 1), torch.float16)
        # Source Nodes: [tensor], Original ATen: [aten.clone]
        triton_poi_fused_clone_12_xnumel = s10*s9
        triton_poi_fused_clone_12.run(buf32, buf33, s10, s9, 256, triton_poi_fused_clone_12_xnumel, grid=grid(256, triton_poi_fused_clone_12_xnumel), stream=stream0)
        buf34 = buf29; del buf29  # reuse
        # Source Nodes: [Q_3], Original ATen: [aten.div]
        triton_poi_fused_div_13_xnumel = 256*s2
        triton_poi_fused_div_13.run(buf32, buf34, triton_poi_fused_div_13_xnumel, grid=grid(triton_poi_fused_div_13_xnumel), stream=stream0)
        buf35 = empty_strided_cuda((1, s2, s11), (s11*s2, s11, 1), torch.float16)
        # Source Nodes: [QK_1], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf34, (1, s2, 256), (0, 256, 1), 0), reinterpret_tensor(arg23_1, (1, 256, s11), (0, 1, 256), 0), out=buf35)
        del arg23_1
        buf38 = empty_strided_cuda((1, 1, s2, s11), (s11*s2, 1, s11, 1), torch.float16)
        # Source Nodes: [attn_2, matmul_3], Original ATen: [aten._softmax, aten._to_copy]
        triton_red_fused__softmax__to_copy_14.run(buf35, buf38, s11, s2, s11, grid=grid(s2), stream=stream0)
        del buf35
        buf39 = reinterpret_tensor(buf34, (1, s2, 256), (256*s2, 256, 1), 0); del buf34  # reuse
        # Source Nodes: [matmul_3], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf38, (1, s2, s11), (0, s11, 1), 0), reinterpret_tensor(arg24_1, (1, s11, 256), (256*s11, 256, 1), 0), out=buf39)
        del arg24_1
        del buf38
        buf40 = buf22; del buf22  # reuse
        # Source Nodes: [outputs_5], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_3.run(arg25_1, buf40, 65536, grid=grid(65536), stream=stream0)
        del arg25_1
        buf41 = buf17; del buf17  # reuse
        # Source Nodes: [outputs_5], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_4.run(arg26_1, buf41, 256, grid=grid(256), stream=stream0)
        del arg26_1
        buf42 = reinterpret_tensor(buf15, (s2, 256), (256, 1), 0); del buf15  # reuse
        # Source Nodes: [outputs_5], Original ATen: [aten._to_copy, aten.addmm]
        extern_kernels.addmm(buf41, reinterpret_tensor(buf39, (s2, 256), (256, 1), 0), reinterpret_tensor(buf40, (256, 256), (1, 256), 0), alpha=1, beta=1, out=buf42)
        del buf39
        del buf40
        del buf41
    return (buf33, buf24, buf28, reinterpret_tensor(buf42, (s2, 1, 256), (256, 256, 1), 0), reinterpret_tensor(buf32, (s2, 1, 256), (512, 512, 1), 0), reinterpret_tensor(buf32, (s2, 1, 256), (512, 512, 1), 256), )


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((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg1_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg2_1 = 4624
    arg3_1 = rand_strided((4624, 1, 256), (1, 1183744, 4624), device='cuda:0', dtype=torch.float32)
    arg4_1 = rand_strided((4624, 1, 256), (256, 256, 1), device='cuda:0', dtype=torch.float32)
    arg5_1 = 8
    arg6_1 = 32
    arg7_1 = rand_strided((256, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg8_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg9_1 = rand_strided((256, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg10_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg11_1 = rand_strided((256, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg12_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg13_1 = 32
    arg14_1 = rand_strided((256, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg15_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg16_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg17_1 = rand_strided((256, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg18_1 = rand_strided((512, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg19_1 = rand_strided((512, ), (1, ), device='cuda:0', dtype=torch.float32)
    arg20_1 = 68
    arg21_1 = 68
    arg22_1 = 9248
    arg23_1 = rand_strided((9248, 1, 256), (256, 256, 1), device='cuda:0', dtype=torch.float16)
    arg24_1 = rand_strided((9248, 1, 256), (256, 256, 1), device='cuda:0', dtype=torch.float16)
    arg25_1 = rand_strided((256, 256), (256, 1), device='cuda:0', dtype=torch.float32)
    arg26_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, arg12_1, arg13_1, arg14_1, arg15_1, arg16_1, arg17_1, arg18_1, arg19_1, arg20_1, arg21_1, arg22_1, arg23_1, arg24_1, arg25_1, arg26_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)