# AOT ID: ['21_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/hb/chbuz277bmiaacrdrscei26m2scqf6qln6fhcia4mwr3nyjhkezc.py
# Source Nodes: [local_attn, mul_1, qk_1, qk_2], Original ATen: [aten._softmax, aten.add, aten.mul, aten.sub]
# local_attn => amax, exp, sub_1, sum_1
# mul_1 => mul
# qk_1 => add
# qk_2 => sub
triton_red_fused__softmax_add_mul_sub_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: '*fp32', 3: '*fp32', 4: '*fp32', 5: 'i32', 6: 'i32', 7: 'i32', 8: 'i32', 9: '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, 8), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_red_fused__softmax_add_mul_sub_0', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 6, '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, ks0, ks1, ks2, xnumel, rnumel, XBLOCK : tl.constexpr, RBLOCK : tl.constexpr):
    xnumel = 4624
    rnumel = 225
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    rbase = tl.arange(0, RBLOCK)[None, :]
    x0 = xindex
    _tmp9 = 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 + (x0 + (4624*r1)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp1 = tl.load(in_ptr1 + (x0 + (ks0*r1)), rmask & xmask, eviction_policy='evict_last', other=0.0).to(tl.float32)
        tmp4 = tl.load(in_ptr2 + (x0 + (ks1*ks2*r1)), rmask & xmask, eviction_policy='evict_last', other=0.0)
        tmp2 = tmp0 + tmp1
        tmp3 = tmp2.to(tl.float32)
        tmp5 = 10000.0
        tmp6 = tmp4 * tmp5
        tmp7 = tmp3 - tmp6
        tmp8 = tl.broadcast_to(tmp7, [XBLOCK, RBLOCK])
        tmp10 = triton_helpers.maximum(_tmp9, tmp8)
        _tmp9 = tl.where(rmask & xmask, tmp10, _tmp9)
    tmp9 = triton_helpers.max2(_tmp9, 1)[:, None]
    tl.store(out_ptr0 + (x0), tmp9, xmask)
    _tmp22 = tl.full([XBLOCK, RBLOCK], 0, tl.float32)
    for roffset in range(0, rnumel, RBLOCK):
        rindex = roffset + rbase
        rmask = rindex < rnumel
        r1 = rindex
        tmp11 = tl.load(in_ptr0 + (x0 + (4624*r1)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp12 = tl.load(in_ptr1 + (x0 + (ks0*r1)), rmask & xmask, eviction_policy='evict_first', other=0.0).to(tl.float32)
        tmp15 = tl.load(in_ptr2 + (x0 + (ks1*ks2*r1)), rmask & xmask, eviction_policy='evict_first', other=0.0)
        tmp13 = tmp11 + tmp12
        tmp14 = tmp13.to(tl.float32)
        tmp16 = 10000.0
        tmp17 = tmp15 * tmp16
        tmp18 = tmp14 - tmp17
        tmp19 = tmp18 - tmp9
        tmp20 = tl_math.exp(tmp19)
        tmp21 = tl.broadcast_to(tmp20, [XBLOCK, RBLOCK])
        tmp23 = _tmp22 + tmp21
        _tmp22 = tl.where(rmask & xmask, tmp23, _tmp22)
    tmp22 = tl.sum(_tmp22, 1)[:, None]
    tl.store(out_ptr1 + (x0), tmp22, 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/re/cre6k764zxofztrayvddsp7x3cd4upt4oeavgsmq4ul2mpjjvmnc.py
# Source Nodes: [agg_bias, local_attn, mul_1, qk_1, qk_2], Original ATen: [aten._softmax, aten._to_copy, aten.add, aten.mul, aten.sub]
# agg_bias => convert_element_type_2
# local_attn => div, exp, sub_1
# mul_1 => mul
# qk_1 => add
# qk_2 => sub
triton_poi_fused__softmax__to_copy_add_mul_sub_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=[1048576], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp16', 1: '*fp16', 2: '*fp32', 3: '*fp32', 4: '*fp32', 5: '*fp32', 6: '*fp16', 7: 'i32', 8: 'i32', 9: 'i32', 10: '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, 10), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused__softmax__to_copy_add_mul_sub_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, out_ptr1, ks0, ks1, ks2, xnumel, XBLOCK : tl.constexpr):
    xnumel = 1040400
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x2 = xindex
    x0 = xindex % 4624
    x1 = (xindex // 4624)
    tmp0 = tl.load(in_ptr0 + (x2), xmask).to(tl.float32)
    tmp1 = tl.load(in_ptr1 + (x0 + (ks0*x1)), xmask).to(tl.float32)
    tmp4 = tl.load(in_ptr2 + (x0 + (ks1*ks2*x1)), xmask)
    tmp8 = tl.load(in_ptr3 + (x0), xmask, eviction_policy='evict_last')
    tmp11 = tl.load(in_ptr4 + (x0), xmask, eviction_policy='evict_last')
    tmp2 = tmp0 + tmp1
    tmp3 = tmp2.to(tl.float32)
    tmp5 = 10000.0
    tmp6 = tmp4 * tmp5
    tmp7 = tmp3 - tmp6
    tmp9 = tmp7 - tmp8
    tmp10 = tl_math.exp(tmp9)
    tmp12 = tmp10 / tmp11
    tmp13 = tmp12.to(tl.float32)
    tl.store(out_ptr0 + (x0 + (4640*x1)), tmp12, xmask)
    tl.store(out_ptr1 + (x0 + (4672*x1)), tmp13, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/3i/c3ippcyax57xhlnc6p54iobdj7q634wyem4qqrujd5dr4v7cp5no.py
# Source Nodes: [global_attn], Original ATen: [aten.zeros]
# global_attn => full_default
triton_poi_fused_zeros_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=[33554432], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '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_zeros_2', 'mutated_arg_names': [], 'no_x_dim': False, 'num_load': 0, '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_(out_ptr0, xnumel, XBLOCK : tl.constexpr):
    xnumel = 31091776
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex % 6724
    x1 = (xindex // 6724)
    tmp0 = 0.0
    tl.store(out_ptr0 + (x0 + (6752*x1)), tmp0, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/aj/cajppu57utsdna2ig4xwsx7h4vg73apa4xz3o2bimjsjbjcacscv.py
# Source Nodes: [reshape], Original ATen: [aten.clone]
# reshape => clone_1
triton_poi_fused_clone_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=[8192, 256], tile_hint=TileHint.SQUARE,
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 1: '*fp32', 2: 'i32', 3: '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_clone_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, ynumel, xnumel, YBLOCK : tl.constexpr, XBLOCK : tl.constexpr):
    ynumel = 4624
    xnumel = 225
    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 + (4640*x1)), xmask & ymask, eviction_policy='evict_last')
    tl.store(out_ptr0 + (x1 + (225*y0)), tmp0, xmask & ymask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/x7/cx7o6msxbvszi6jbyynwctlif5kjwlhqpigrm2zz4ovmikdhvwom.py
# Source Nodes: [global_attn, setitem], Original ATen: [aten.index_put, aten.zeros]
# global_attn => full_default
# setitem => index_put
triton_poi_fused_index_put_zeros_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=[1048576], 
    filename=__file__,
    triton_meta={'signature': {0: '*fp32', 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), equal_to_1=())]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_index_put_zeros_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):
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (xindex % 1040400), xmask)
    tl.store(out_ptr0 + (x0), tmp0, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/7i/c7ix4vj3a5mvmeho2is7fhutiy65gpvherhzddydtkdidewuv74b.py
# Source Nodes: [agg_value], Original ATen: [aten._to_copy]
# agg_value => convert_element_type_6
triton_poi_fused__to_copy_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=[33554432], 
    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_5', '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 = 21381376
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex % 4624
    x1 = (xindex // 4624)
    tmp0 = tl.load(in_ptr0 + (581 + (82*(x0 // 68)) + (6752*x1) + (x0 % 68)), xmask)
    tmp1 = tmp0.to(tl.float32)
    tl.store(out_ptr0 + (x0 + (4672*x1)), tmp1, xmask)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/br/cbrvodbxo3uuxus5l5rzw54iumqpwsfow3foexkyib23rvh3ozek.py
# Source Nodes: [agg_bias], Original ATen: [aten._to_copy]
# agg_bias => convert_element_type_3
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=[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_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 = 57600
    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/cx/ccxwzxbgokjb4lskr6xanvjlgd2uynkoaxmdrlx53on6hedj2p5u.py
# Source Nodes: [add_3], Original ATen: [aten.add]
# add_3 => add_1
triton_poi_fused_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=[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_add_7', '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):
    xnumel = 1183744
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:]
    xmask = xindex < xnumel
    x0 = xindex
    tmp0 = tl.load(in_out_ptr0 + (x0), None).to(tl.float32)
    tmp1 = tl.load(in_ptr0 + (x0), None).to(tl.float32)
    tmp2 = tmp0 + tmp1
    tl.store(in_out_ptr0 + (x0), tmp2, None)
''', device_str='cuda')


# kernel path: /tmp/torchinductor_root/ar/carcf7uwbqtj2hgalfjzpp5bdevfi2dvum5suqasnxet5yqes4d6.py
# Source Nodes: [output_1], Original ATen: [aten._to_copy]
# output_1 => convert_element_type_10
triton_poi_fused__to_copy_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=[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_8', '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/jn/cjnrgvzsstwynldvwnljz5fk3ilroqn6fypn5dky5hgssitiob22.py
# Source Nodes: [output_1], Original ATen: [aten._to_copy]
# output_1 => convert_element_type_9
triton_poi_fused__to_copy_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.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_9', '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')


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 = args
    args.clear()
    s0 = arg0_1
    s1 = arg1_1
    s5 = arg5_1
    assert_size_stride(arg2_1, (1, 15, 15, s0, s1), (225*s0*s1, 15*s0*s1, s0*s1, s1, 1))
    assert_size_stride(arg6_1, (1, 1, 225, s0*s1), (0, 0, s5, 1))
    assert_size_stride(arg7_1, (1, 1, 225, s0*s1), (225*s0*s1, 225*s0*s1, s0*s1, 1))
    assert_size_stride(arg8_1, (1, 256, 225), (57600, 225, 1))
    assert_size_stride(arg9_1, (1, 1, 4624, 82, 82), (0, 0, 6784, 82, 1))
    assert_size_stride(arg10_1, (1, 1, 256, 4624), (1183744, 1183744, 4624, 1))
    assert_size_stride(arg11_1, (256, 256), (256, 1))
    assert_size_stride(arg12_1, (256, ), (1, ))
    with torch.cuda._DeviceGuard(0):
        torch.cuda.set_device(0)
        buf0 = empty_strided_cuda((1, 1, 1, 4624), (4640, 4640, 4640, 1), torch.float32)
        buf1 = empty_strided_cuda((1, 1, 1, 4624), (4640, 4640, 4640, 1), torch.float32)
        # Source Nodes: [local_attn, mul_1, qk_1, qk_2], Original ATen: [aten._softmax, aten.add, aten.mul, aten.sub]
        stream0 = get_raw_stream(0)
        triton_red_fused__softmax_add_mul_sub_0.run(arg2_1, arg6_1, arg7_1, buf0, buf1, s5, s0, s1, 4624, 225, grid=grid(4624), stream=stream0)
        buf2 = empty_strided_cuda((1, 1, 225, 4624), (1044000, 1044000, 4640, 1), torch.float32)
        buf10 = empty_strided_cuda((1, 1, 225, 4624), (1051200, 4672, 4672, 1), torch.float16)
        # Source Nodes: [agg_bias, local_attn, mul_1, qk_1, qk_2], Original ATen: [aten._softmax, aten._to_copy, aten.add, aten.mul, aten.sub]
        triton_poi_fused__softmax__to_copy_add_mul_sub_1.run(arg2_1, arg6_1, arg7_1, buf0, buf1, buf2, buf10, s5, s0, s1, 1040400, grid=grid(1040400), stream=stream0)
        del arg2_1
        del arg6_1
        del arg7_1
        del buf0
        del buf1
        buf3 = empty_strided_cuda((1, 1, 4624, 82, 82), (31221248, 31221248, 6752, 82, 1), torch.float32)
        # Source Nodes: [global_attn], Original ATen: [aten.zeros]
        triton_poi_fused_zeros_2.run(buf3, 31091776, grid=grid(31091776), stream=stream0)
        buf4 = empty_strided_cuda((1, 1, 4624, 225), (1040416, 225, 225, 1), torch.float32)
        # Source Nodes: [reshape], Original ATen: [aten.clone]
        triton_poi_fused_clone_3.run(buf2, buf4, 4624, 225, grid=grid(4624, 225), stream=stream0)
        buf5 = empty_strided_cuda((225*s0*s1, ), (1, ), torch.float32)
        # Source Nodes: [global_attn, setitem], Original ATen: [aten.index_put, aten.zeros]
        triton_poi_fused_index_put_zeros_4_xnumel = 225*s0*s1
        triton_poi_fused_index_put_zeros_4.run(buf4, buf5, triton_poi_fused_index_put_zeros_4_xnumel, grid=grid(triton_poi_fused_index_put_zeros_4_xnumel), stream=stream0)
        del buf4
        aten.index_put_(buf3, [reinterpret_tensor(arg9_1, (1, 1, 4624, 82, 82), (0, 0, 6784, 82, 1), 0)], buf5, False)
        del arg9_1
        del buf5
        buf8 = empty_strided_cuda((1, 1, 4624, 4624), (21603328, 4672, 4672, 1), torch.float16)
        # Source Nodes: [agg_value], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_5.run(buf3, buf8, 21381376, grid=grid(21381376), stream=stream0)
        del buf3
        buf9 = empty_strided_cuda((1, 4624, 256), (1183744, 256, 1), torch.float16)
        # Source Nodes: [agg_value], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf8, (1, 4624, 4624), (0, 4672, 1), 0), reinterpret_tensor(arg10_1, (1, 4624, 256), (0, 1, 4624), 0), out=buf9)
        del arg10_1
        del buf8
        buf11 = empty_strided_cuda((1, 256, 225), (57600, 225, 1), torch.float16)
        # Source Nodes: [agg_bias], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_6.run(arg8_1, buf11, 57600, grid=grid(57600), stream=stream0)
        del arg8_1
        buf12 = empty_strided_cuda((1, s0*s1, 256), (256*s0*s1, 256, 1), torch.float16)
        # Source Nodes: [agg_bias], Original ATen: [aten.bmm]
        extern_kernels.bmm(reinterpret_tensor(buf10, (1, s0*s1, 225), (0, 1, 4672), 0), reinterpret_tensor(buf11, (1, 225, 256), (0, 1, 225), 0), out=buf12)
        del buf10
        del buf11
        buf13 = reinterpret_tensor(buf9, (1, 1, 4624, 256), (1183744, 1, 256, 1), 0); del buf9  # reuse
        # Source Nodes: [add_3], Original ATen: [aten.add]
        triton_poi_fused_add_7.run(buf13, buf12, 1183744, grid=grid(1183744), stream=stream0)
        del buf12
        buf14 = empty_strided_cuda((256, 256), (256, 1), torch.float16)
        # Source Nodes: [output_1], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_8.run(arg11_1, buf14, 65536, grid=grid(65536), stream=stream0)
        del arg11_1
        buf15 = empty_strided_cuda((256, ), (1, ), torch.float16)
        # Source Nodes: [output_1], Original ATen: [aten._to_copy]
        triton_poi_fused__to_copy_9.run(arg12_1, buf15, 256, grid=grid(256), stream=stream0)
        del arg12_1
        buf16 = empty_strided_cuda((4624, 256), (256, 1), torch.float16)
        # Source Nodes: [output_1], Original ATen: [aten._to_copy, aten.addmm]
        extern_kernels.addmm(buf15, reinterpret_tensor(buf13, (4624, 256), (256, 1), 0), reinterpret_tensor(buf14, (256, 256), (1, 256), 0), alpha=1, beta=1, out=buf16)
        del buf13
        del buf14
        del buf15
    return (reinterpret_tensor(buf16, (4624, 1, 256), (256, 256, 1), 0), buf2, )


def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    arg0_1 = 68
    arg1_1 = 68
    arg2_1 = rand_strided((1, 15, 15, 68, 68), (1040400, 69360, 4624, 68, 1), device='cuda:0', dtype=torch.float16)
    arg3_1 = 68
    arg4_1 = 68
    arg5_1 = 4672
    arg6_1 = rand_strided((1, 1, 225, 4624), (0, 0, 4672, 1), device='cuda:0', dtype=torch.float16)
    arg7_1 = rand_strided((1, 1, 225, 4624), (1040400, 1040400, 4624, 1), device='cuda:0', dtype=torch.float32)
    arg8_1 = rand_strided((1, 256, 225), (57600, 225, 1), device='cuda:0', dtype=torch.float32)
    arg9_1 = rand_strided((1, 1, 4624, 82, 82), (0, 0, 6784, 82, 1), device='cuda:0', dtype=torch.bool)
    arg10_1 = rand_strided((1, 1, 256, 4624), (1183744, 1183744, 4624, 1), device='cuda:0', dtype=torch.float16)
    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)
    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])
    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)