Xantar commited on
Commit
2e0b8ac
·
verified ·
1 Parent(s): d4a3ff4

Upload folder using huggingface_hub

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .inductor_cache/25/c25l73lomz5i5p2tjttjp4w3njz7barnpiaq2fsgnq3jaafgxdsu.py +29 -0
  2. .inductor_cache/25/ee4f2760622c1bc46e4963f5f90e772637bf02dfce9d2c784a23b91a485e6046.best_config +1 -0
  3. .inductor_cache/2b/c2br4a6gmsccqokjlyz2ntpodyioqvw3ssadyruqv4t5q2h4zfkq.py +291 -0
  4. .inductor_cache/2k/0db56ce0255d17e714d6bfefd0830f67b79f31c3cd47fef86e1215e4af155123.best_config +1 -0
  5. .inductor_cache/2k/c2kidoyobarbdwts46exm2wpdobdyvsxybtugfjxnxbjmacldvms.py +29 -0
  6. .inductor_cache/2u/c2ushp4u2x3uzdiuj2hcvldtsjtcukrffzblul73u3eczxkoqeiw.py +293 -0
  7. .inductor_cache/3e/c3exb3qmzq5ev7yyxp6jzpk55mrxicwt5fjnjg72h4blbjqteq3b.py +25 -0
  8. .inductor_cache/3h/c3hjatdysmtczt36qylu5u476tmnxdvxgytr7m6z4na54xrxxoss.py +24 -0
  9. .inductor_cache/3h/c3hzz247khkgmf7sy4wb5vbaz4z24jsfnh2gf3spavzjd7bow5gg.py +27 -0
  10. .inductor_cache/3r/c3rryctepqjzosdqqttq5mb7kr5klk4hfq7vyabrmirmlhcabwmo.py +137 -0
  11. .inductor_cache/3s/c3skh52gob24udbqevuvymapfeanznl6g6yugtw6fyk24ox5ioyb.py +142 -0
  12. .inductor_cache/3u/c3u6bcbbftmiqp5kte3h55vd5gqx7jl3zql7a2czxn52qwb3z3xt.py +289 -0
  13. .inductor_cache/3z/c3ztg2btpqdyc3uz4dzswnkglhrmeaszpogpdu23omfz4c6u7zko.py +27 -0
  14. .inductor_cache/46/9ddaabf0d63fddb6e484df5802c5ab86ddbc3d9b84c46c643b006cc0acf7b995.best_config +1 -0
  15. .inductor_cache/46/c46gch5kg22oprcsfx3k73s77cdoj5zwmm4trujfwwc4zc7cs72z.py +24 -0
  16. .inductor_cache/47/c47i4hdzezcx3xndn3biber6sa6rznklzeek5cbpexpomknwaqjt.py +177 -0
  17. .inductor_cache/4b/c4bzcwru43j45mp2ej3abi6jtvpi4pqeiwedj4gy3vcydr2jkwun.py +289 -0
  18. .inductor_cache/4b/c4bzoixvavbv7tbgxpx5sajtax6yl6d6wky5k272ff4wk64t7jd7.py +25 -0
  19. .inductor_cache/4i/c4iwhfgin6ac7qvqb6fj5msgdftp5qsxjv6sjxyuqbryq4qqwzfg.py +37 -0
  20. .inductor_cache/4l/c4ljp5q37q7ktbyo35672uzle4llt4iwt4be37ikkwlit6j7fnco.py +31 -0
  21. .inductor_cache/4l/dccab99d67a1299427a9211bfd01fd3c77aaa44ad59d9ffa12a6d00f48a13c0e.best_config +1 -0
  22. .inductor_cache/4n/c4nxaenxhyi2kg2uordabgcp5ofkprcosqtxys5klrlaly24kcgm.py +293 -0
  23. .inductor_cache/4w/88c2a3ee4b3f9188d0485d1f58c19da2936d66802465a9e388f74d179d8ec54e.best_config +1 -0
  24. .inductor_cache/4w/c4wbe3nz76vrohi4tguyhvtuuwxq5ot6a6vjcubrww3al5ynrkzg.py +26 -0
  25. .inductor_cache/5d/c5dllvj5zhsrbv5yhs4lnnrfdo5t7vsfkgewb73jsq5ipppt5rbn.py +37 -0
  26. .inductor_cache/5f/c5fm757ws5k72edgygk6waomo4dh4sqhl7n4rzqsynjfqrzjugov.py +150 -0
  27. .inductor_cache/5l/c5lhmqi55bmiwxrtdbnsnvjgutkpxgsd6sl2mrliaofj5ojwc2gs.py +24 -0
  28. .inductor_cache/5p/c5pplmatftc6r3hsuhx4skgsikwfs4k4t74mko7xfk7xp2sqleis.py +46 -0
  29. .inductor_cache/5q/c5qdusbmow5b3zrnnn4oh343d7oyptsndbeaenzknk5r4mhn7org.py +148 -0
  30. .inductor_cache/5s/c5s4bm2q3gua3tgmxnj3denzqvuhpgrb5kx5rjwxxb4isc7lmqaz.py +24 -0
  31. .inductor_cache/5v/1eb21a9ba8eeb898c12aad6f0c7f6fb253af198de2a96f9596a41157018121fd.best_config +1 -0
  32. .inductor_cache/5v/c5vsrxfkmbxmqe35ujjlxyr37blswcxqhqzlv2r6jal54yzc27hf.py +31 -0
  33. .inductor_cache/5z/c5ztwyl3hjethlv2ot2kpnvlbaqjstzofdcjttbjwa7nkglx5gan.py +45 -0
  34. .inductor_cache/62/c62vhrldvf7yqvelctjx64ts5ifznauctxiqhwo4jrqxe2rtpdgj.py +37 -0
  35. .inductor_cache/65/c156cdb4f80a03134d5d44666437b38d33a91660ba036c9600b2b2dad614b347.best_config +1 -0
  36. .inductor_cache/65/c657qxcjufcctv3zrgtm2r3tsalyqcuqrzgzqct6qvtmivqg7fnt.py +237 -0
  37. .inductor_cache/65/c65lswedy4nxnsl65ku4a6ddu3r55grrzzjaxx6jvoroewo6e2dr.py +24 -0
  38. .inductor_cache/67/c67bawm35pfjbqa2hnubkrofni4tkbxzyeo5zppu6gsylvqve6eo.py +46 -0
  39. .inductor_cache/6d/c6dqripf2w5j4sf6yps4zy7tc4ifqtp2wzcncsdxelw7lhj3o7ou.py +31 -0
  40. .inductor_cache/6d/e62944c5f38dab9ccf24b5504699965ed3bab66acfacc866085a1032c963177d.best_config +1 -0
  41. .inductor_cache/6e/c6emtbesfuldnxatnjaketbq4pjgfhpuh5rpsfrb6iagv4wjnfal.py +45 -0
  42. .inductor_cache/6j/c6j7f2zfedgsifsvrtdzvprpzw6jujih2uwmzsjrvvfsrmkhcotk.py +24 -0
  43. .inductor_cache/6p/28650630bd03810b78b7cf75754dfa35da9d5ed77d198e2f0ce9f078547bb750.best_config +1 -0
  44. .inductor_cache/6p/c6ppnyt6kjnrqeeq3ra4wx4kwkpeerqiq5ftck3hdmvntgpd6wcs.py +31 -0
  45. .inductor_cache/6r/c6rthxci3676wj2mwj62e5zmpmln5m2yrpnp2cb63gsxh4u2zxhz.py +27 -0
  46. .inductor_cache/6u/c6uum37r4isrc4g6cyxmskeeex5jl3phh26m7xujcnuux4ijq3jw.py +24 -0
  47. .inductor_cache/6y/c6ygrpcdbuohowjvpdfmrangdf56jqj3p4ig3uwlgavr6t5kkb7a.py +289 -0
  48. .inductor_cache/6z/c6zyuqh3bo3nxnmo267egfmzuh75yqhd4pdyfrvq5jc5nfexlxsk.py +29 -0
  49. .inductor_cache/76/c76y7ge3o24p7dwc6aulba5id74axw62oevitux7yjjlbrbsx66u.py +24 -0
  50. .inductor_cache/7e/c7eeblfcfg2xdo34xdor3snhltmuxl4hgqzougfrenohmqtj4r3r.py +177 -0
.inductor_cache/25/c25l73lomz5i5p2tjttjp4w3njz7barnpiaq2fsgnq3jaafgxdsu.py ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 256},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'in_ptr2': 'fp64', 'out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 4096}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
25
+ tmp3 = in_ptr2
26
+ tmp2 = tmp0 * tmp1
27
+ tmp4 = tmp3.to(tl.float32)
28
+ tmp5 = tmp2 * tmp4
29
+ tl.store(out_ptr0 + (x0), tmp5, xmask)
.inductor_cache/25/ee4f2760622c1bc46e4963f5f90e772637bf02dfce9d2c784a23b91a485e6046.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 128, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 21, "triton_cache_hash": "YG74CS4JHDXVAWUXP7QZ23NAXRAHIYIZ5FUJMRPVXQLZO4XINHZA"}
.inductor_cache/2b/c2br4a6gmsccqokjlyz2ntpodyioqvw3ssadyruqv4t5q2h4zfkq.py ADDED
@@ -0,0 +1,291 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['7_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/xb/cxbfao4txftywezeygrrwtapgjmjuaow6yckdgujtj3ln7ukpbdn.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
39
+ # Source node to ATen node mapping:
40
+ # expand_scores => mul_3
41
+ # flatten => view_1
42
+ # max_1 => getitem, max_1
43
+ # unsqueeze => unsqueeze
44
+ # view => view
45
+ # Graph fragment:
46
+ # %arg2_1 : Tensor "f32[s0, 1][1, 1]cuda:0" = PlaceHolder[target=arg2_1]
47
+ # %arg4_1 : Tensor "f32[s0, 1][1, 1]cuda:0" = PlaceHolder[target=arg4_1]
48
+ # %mul_3 : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:0" = PlaceHolder[target=mul_3]
49
+ # %unsqueeze : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg2_1, 2), kwargs = {})
50
+ # %view : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%arg4_1, [-1, 1, 1]), kwargs = {})
51
+ # %mul_3 : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:0"[num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %view), kwargs = {})
52
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
53
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
54
+ # %getitem : Tensor "f32[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=operator.getitem](args = (%max_1, 0), kwargs = {})
55
+ # return %mul_3,%getitem
56
+ triton_poi_fused_max_mul_unsqueeze_view_0 = async_compile.triton('triton_poi_fused_max_mul_unsqueeze_view_0', '''
57
+ import triton
58
+ import triton.language as tl
59
+
60
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
61
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
62
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
63
+ triton_helpers.set_driver_to_gpu()
64
+
65
+ @triton_heuristics.pointwise(
66
+ size_hints={'x': 8},
67
+ filename=__file__,
68
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
69
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
70
+ min_elem_per_thread=0
71
+ )
72
+ @triton.jit
73
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
74
+ xoffset = tl.program_id(0) * XBLOCK
75
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
76
+ xmask = xindex < xnumel
77
+ x0 = xindex
78
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
79
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
80
+ tmp2 = tmp0 * tmp1
81
+ tl.store(out_ptr0 + (x0), tmp2, xmask)
82
+ tl.store(out_ptr1 + (x0), tmp2, xmask)
83
+ ''', device_str='cuda')
84
+
85
+
86
+ # kernel path: /mnt/glm5_2/.inductor_cache/jk/cjk2l5do5kfebmlafwv2ddtivblyn54wzypw3yv4cn4tgq7w7p46.py
87
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
88
+ # Source node to ATen node mapping:
89
+ # flatten => view_1
90
+ # gather => gather
91
+ # max_1 => max_1
92
+ # Graph fragment:
93
+ # %arg5_1 : Tensor "i64[s0, 1][1, 1]cuda:0" = PlaceHolder[target=arg5_1]
94
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
95
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
96
+ # %gather : Tensor "i64[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.gather.default](args = (%arg5_1, 1, %getitem_1), kwargs = {})
97
+ # return %gather
98
+ triton_poi_fused_gather_max_view_1 = async_compile.triton('triton_poi_fused_gather_max_view_1', '''
99
+ import triton
100
+ import triton.language as tl
101
+
102
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
103
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
104
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
105
+ triton_helpers.set_driver_to_gpu()
106
+
107
+ @triton_heuristics.pointwise(
108
+ size_hints={'x': 8},
109
+ filename=__file__,
110
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
111
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
112
+ min_elem_per_thread=0
113
+ )
114
+ @triton.jit
115
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
116
+ xoffset = tl.program_id(0) * XBLOCK
117
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
118
+ xmask = xindex < xnumel
119
+ x0 = xindex
120
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
121
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
122
+ ''', device_str='cuda')
123
+
124
+
125
+ # kernel path: /mnt/glm5_2/.inductor_cache/xb/cxbqyqxwwyvdsshstrz3zl343674btixfm5kbidvymejf6qmdoaf.py
126
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
127
+ # Source node to ATen node mapping:
128
+ # batch_offsets => iota
129
+ # flat_cs => view_4
130
+ # flatten => view_1
131
+ # floordiv => div
132
+ # hidden_states => index
133
+ # max_1 => max_1
134
+ # repeat_interleave => unsqueeze_1, view_5
135
+ # selected_input_index => add_47
136
+ # Graph fragment:
137
+ # %arg7_1 : Tensor "bf16[s0, s87][s87, 1]cuda:0" = PlaceHolder[target=arg7_1]
138
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
139
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
140
+ # %view_4 : Tensor "i64[s0][1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%getitem_1, [%arg3_1]), kwargs = {})
141
+ # %div : Tensor "i64[s0][1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.div.Tensor_mode](args = (%view_4, 1), kwargs = {rounding_mode: floor})
142
+ # %iota : Tensor "i64[s0][1]cuda:0"[num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (%arg3_1,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:0, requires_grad: False})
143
+ # %unsqueeze_1 : Tensor "i64[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%iota, 1), kwargs = {})
144
+ # %view_5 : Tensor "i64[s0][1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%expand, [%arg3_1]), kwargs = {})
145
+ # %add_47 : Tensor "i64[s0][1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %view_5), kwargs = {})
146
+ # %index : Tensor "bf16[s0, s87][s87, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%arg7_1, [%add_47]), kwargs = {})
147
+ # return %index
148
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2 = async_compile.triton('triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', '''
149
+ import triton
150
+ import triton.language as tl
151
+
152
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
153
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
154
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
155
+ triton_helpers.set_driver_to_gpu()
156
+
157
+ @triton_heuristics.pointwise(
158
+ size_hints={'x': 65536},
159
+ filename=__file__,
160
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
161
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 294912}},
162
+ min_elem_per_thread=0
163
+ )
164
+ @triton.jit
165
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
166
+ xoffset = tl.program_id(0) * XBLOCK
167
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
168
+ xmask = xindex < xnumel
169
+ x2 = xindex
170
+ tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last').to(tl.float32)
171
+ tl.store(out_ptr0 + (x2), tmp0, xmask)
172
+ ''', device_str='cuda')
173
+
174
+
175
+ # kernel path: /mnt/glm5_2/.inductor_cache/f5/cf5tdjh7czx76ocdvaqqmpbpt24xihjxufk72ezsjzb26rz66qx3.py
176
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
177
+ # Source node to ATen node mapping:
178
+ # add_2 => add_53
179
+ # flatten => view_1
180
+ # max_1 => max_1
181
+ # Graph fragment:
182
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
183
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
184
+ # %add_53 : Tensor "i64[s0, 1][1, 1]cuda:0"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_1, %arg8_1), kwargs = {})
185
+ # return %add_53
186
+ triton_poi_fused_add_max_view_3 = async_compile.triton('triton_poi_fused_add_max_view_3', '''
187
+ import triton
188
+ import triton.language as tl
189
+
190
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
191
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
192
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
193
+ triton_helpers.set_driver_to_gpu()
194
+
195
+ @triton_heuristics.pointwise(
196
+ size_hints={'x': 8},
197
+ filename=__file__,
198
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
199
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 128}},
200
+ min_elem_per_thread=0
201
+ )
202
+ @triton.jit
203
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
204
+ xoffset = tl.program_id(0) * XBLOCK
205
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
206
+ xmask = xindex < xnumel
207
+ x0 = xindex
208
+ tmp0 = ks0
209
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
210
+ ''', device_str='cuda')
211
+
212
+
213
+ async_compile.wait(globals())
214
+ del async_compile
215
+
216
+ class Runner:
217
+ def __init__(self, partitions):
218
+ self.partitions = partitions
219
+
220
+ def recursively_apply_fns(self, fns):
221
+ new_callables = []
222
+ for fn, c in zip(fns, self.partitions):
223
+ new_callables.append(fn(c))
224
+ self.partitions = new_callables
225
+
226
+ def call(self, args):
227
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1 = args
228
+ args.clear()
229
+ s53 = arg0_1
230
+ s68 = arg1_1
231
+ s0 = arg3_1
232
+ s87 = arg6_1
233
+ s77 = arg8_1
234
+ assert_size_stride(arg2_1, (s0, 1), (1, 1))
235
+ assert_size_stride(arg4_1, (s0, 1), (1, 1))
236
+ assert_size_stride(arg5_1, (s0, 1), (1, 1))
237
+ assert_size_stride(arg7_1, (s0, s87), (s87, 1))
238
+ with torch.cuda._DeviceGuard(0):
239
+ torch.cuda.set_device(0)
240
+ buf0 = empty_strided_cuda((s0, 1, 1), (1, 1, 1), torch.float32)
241
+ buf1 = empty_strided_cuda((s0, 1), (1, 1), torch.float32)
242
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
243
+ stream0 = get_raw_stream(0)
244
+ triton_poi_fused_max_mul_unsqueeze_view_0.run(arg2_1, arg4_1, buf0, buf1, s0, stream=stream0)
245
+ del arg2_1
246
+ del arg4_1
247
+ buf2 = empty_strided_cuda((s0, 1), (1, 1), torch.int64)
248
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
249
+ stream0 = get_raw_stream(0)
250
+ triton_poi_fused_gather_max_view_1.run(arg5_1, buf2, s0, stream=stream0)
251
+ buf3 = empty_strided_cuda((s0, s87), (s87, 1), torch.bfloat16)
252
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
253
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel = s0*s87
254
+ stream0 = get_raw_stream(0)
255
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2.run(arg7_1, buf3, triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel, stream=stream0)
256
+ del arg7_1
257
+ buf4 = empty_strided_cuda((s0, 1), (1, 1), torch.int64)
258
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
259
+ stream0 = get_raw_stream(0)
260
+ triton_poi_fused_add_max_view_3.run(buf4, s77, s0, stream=stream0)
261
+ return (reinterpret_tensor(buf2, (s0, ), (1, ), 0), buf3, buf1, buf0, arg5_1, buf4, )
262
+
263
+ runner = Runner(partitions=[])
264
+ call = runner.call
265
+ recursively_apply_fns = runner.recursively_apply_fns
266
+
267
+
268
+ def get_args():
269
+ from torch._dynamo.testing import rand_strided
270
+ arg0_1 = 1
271
+ arg1_1 = 8
272
+ arg2_1 = rand_strided((8, 1), (1, 1), device='cuda:0', dtype=torch.float32)
273
+ arg3_1 = 8
274
+ arg4_1 = rand_strided((8, 1), (1, 1), device='cuda:0', dtype=torch.float32)
275
+ arg5_1 = rand_strided((8, 1), (1, 1), device='cuda:0', dtype=torch.int64)
276
+ arg6_1 = 6144
277
+ arg7_1 = rand_strided((8, 6144), (6144, 1), device='cuda:0', dtype=torch.bfloat16)
278
+ arg8_1 = 2
279
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1]
280
+
281
+
282
+ def benchmark_compiled_module(args, times=10, repeat=10):
283
+ from torch._inductor.utils import print_performance
284
+ fn = lambda: call(list(args))
285
+ return print_performance(fn, times=times, repeat=repeat)
286
+
287
+
288
+ if __name__ == "__main__":
289
+ from torch._inductor.wrapper_benchmark import compiled_module_main
290
+ args = get_args()
291
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/2k/0db56ce0255d17e714d6bfefd0830f67b79f31c3cd47fef86e1215e4af155123.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 256, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 21, "triton_cache_hash": "5FUAZKDVQOXOEQJ4ZSRVOWBR6F6PBWI7HIJDQ3SM267VH2IXPT5A"}
.inductor_cache/2k/c2kidoyobarbdwts46exm2wpdobdyvsxybtugfjxnxbjmacldvms.py ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 256},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'in_ptr2': 'fp64', 'out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 4096}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
25
+ tmp3 = in_ptr2
26
+ tmp2 = tmp0 * tmp1
27
+ tmp4 = tmp3.to(tl.float32)
28
+ tmp5 = tmp2 * tmp4
29
+ tl.store(out_ptr0 + (x0), tmp5, xmask)
.inductor_cache/2u/c2ushp4u2x3uzdiuj2hcvldtsjtcukrffzblul73u3eczxkoqeiw.py ADDED
@@ -0,0 +1,293 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['13_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/in/cinym3hv7fv6jpjbbmmapeljldm7umzd5htvihr3dlt7qh5h5e5p.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
39
+ # Source node to ATen node mapping:
40
+ # expand_scores => mul_3
41
+ # flatten => view_1
42
+ # max_1 => getitem, max_1
43
+ # unsqueeze => unsqueeze
44
+ # view => view
45
+ # Graph fragment:
46
+ # %arg2_1 : Tensor "f32[s0, 1][1, 1]cuda:2" = PlaceHolder[target=arg2_1]
47
+ # %arg4_1 : Tensor "f32[s0, 1][1, 1]cuda:2" = PlaceHolder[target=arg4_1]
48
+ # %mul_3 : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:2" = PlaceHolder[target=mul_3]
49
+ # %unsqueeze : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg2_1, 2), kwargs = {})
50
+ # %view : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%arg4_1, [-1, 1, 1]), kwargs = {})
51
+ # %mul_3 : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:2"[num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %view), kwargs = {})
52
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
53
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
54
+ # %getitem : Tensor "f32[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=operator.getitem](args = (%max_1, 0), kwargs = {})
55
+ # return %mul_3,%getitem
56
+ triton_poi_fused_max_mul_unsqueeze_view_0 = async_compile.triton('triton_poi_fused_max_mul_unsqueeze_view_0', '''
57
+ import triton
58
+ import triton.language as tl
59
+
60
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
61
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
62
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
63
+ triton_helpers.set_driver_to_gpu()
64
+
65
+ @triton_heuristics.pointwise(
66
+ size_hints={'x': 2},
67
+ filename=__file__,
68
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
69
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 12}},
70
+ min_elem_per_thread=0
71
+ )
72
+ @triton.jit
73
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
74
+ xoffset = tl.program_id(0) * XBLOCK
75
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
76
+ xmask = xindex < xnumel
77
+ x0 = xindex
78
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
79
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
80
+ tmp2 = tmp0 * tmp1
81
+ tl.store(out_ptr0 + (x0), tmp2, xmask)
82
+ tl.store(out_ptr1 + (x0), tmp2, xmask)
83
+ ''', device_str='cuda')
84
+
85
+
86
+ # kernel path: /mnt/glm5_2/.inductor_cache/ew/cewhk2e3t5msoqsn3m7nynvzesfjqgc746bztsvg7puyhydc7sr4.py
87
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
88
+ # Source node to ATen node mapping:
89
+ # flatten => view_1
90
+ # gather => gather
91
+ # max_1 => max_1
92
+ # Graph fragment:
93
+ # %arg6_1 : Tensor "i64[s37, 1][1, 1]cuda:2" = PlaceHolder[target=arg6_1]
94
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
95
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
96
+ # %gather : Tensor "i64[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.gather.default](args = (%arg6_1, 1, %getitem_1), kwargs = {})
97
+ # return %gather
98
+ triton_poi_fused_gather_max_view_1 = async_compile.triton('triton_poi_fused_gather_max_view_1', '''
99
+ import triton
100
+ import triton.language as tl
101
+
102
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
103
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
104
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
105
+ triton_helpers.set_driver_to_gpu()
106
+
107
+ @triton_heuristics.pointwise(
108
+ size_hints={'x': 2},
109
+ filename=__file__,
110
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
111
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 12}},
112
+ min_elem_per_thread=0
113
+ )
114
+ @triton.jit
115
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
116
+ xoffset = tl.program_id(0) * XBLOCK
117
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
118
+ xmask = xindex < xnumel
119
+ x0 = xindex
120
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
121
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
122
+ ''', device_str='cuda')
123
+
124
+
125
+ # kernel path: /mnt/glm5_2/.inductor_cache/65/c65lswedy4nxnsl65ku4a6ddu3r55grrzzjaxx6jvoroewo6e2dr.py
126
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
127
+ # Source node to ATen node mapping:
128
+ # batch_offsets => iota
129
+ # flat_cs => view_4
130
+ # flatten => view_1
131
+ # floordiv => div
132
+ # hidden_states => index
133
+ # max_1 => max_1
134
+ # repeat_interleave => unsqueeze_1, view_5
135
+ # selected_input_index => add_47
136
+ # Graph fragment:
137
+ # %arg8_1 : Tensor "bf16[s0, s87][s87, 1]cuda:2" = PlaceHolder[target=arg8_1]
138
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
139
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
140
+ # %view_4 : Tensor "i64[s0][1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%getitem_1, [%arg3_1]), kwargs = {})
141
+ # %div : Tensor "i64[s0][1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.div.Tensor_mode](args = (%view_4, 1), kwargs = {rounding_mode: floor})
142
+ # %iota : Tensor "i64[s0][1]cuda:2"[num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (%arg3_1,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:2, requires_grad: False})
143
+ # %unsqueeze_1 : Tensor "i64[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%iota, 1), kwargs = {})
144
+ # %view_5 : Tensor "i64[s0][1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%expand, [%arg3_1]), kwargs = {})
145
+ # %add_47 : Tensor "i64[s0][1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %view_5), kwargs = {})
146
+ # %index : Tensor "bf16[s0, s87][s87, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%arg8_1, [%add_47]), kwargs = {})
147
+ # return %index
148
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2 = async_compile.triton('triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', '''
149
+ import triton
150
+ import triton.language as tl
151
+
152
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
153
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
154
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
155
+ triton_helpers.set_driver_to_gpu()
156
+
157
+ @triton_heuristics.pointwise(
158
+ size_hints={'x': 16384},
159
+ filename=__file__,
160
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
161
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 73728}},
162
+ min_elem_per_thread=0
163
+ )
164
+ @triton.jit
165
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
166
+ xoffset = tl.program_id(0) * XBLOCK
167
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
168
+ xmask = xindex < xnumel
169
+ x2 = xindex
170
+ tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last').to(tl.float32)
171
+ tl.store(out_ptr0 + (x2), tmp0, xmask)
172
+ ''', device_str='cuda')
173
+
174
+
175
+ # kernel path: /mnt/glm5_2/.inductor_cache/y5/cy5huhsgetyuqpiq72l6lba4jexcwsdpngrz2iwahbpopxclyk44.py
176
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
177
+ # Source node to ATen node mapping:
178
+ # add_2 => add_53
179
+ # flatten => view_1
180
+ # max_1 => max_1
181
+ # Graph fragment:
182
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
183
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
184
+ # %add_53 : Tensor "i64[s0, 1][1, 1]cuda:2"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_1, %arg9_1), kwargs = {})
185
+ # return %add_53
186
+ triton_poi_fused_add_max_view_3 = async_compile.triton('triton_poi_fused_add_max_view_3', '''
187
+ import triton
188
+ import triton.language as tl
189
+
190
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
191
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
192
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
193
+ triton_helpers.set_driver_to_gpu()
194
+
195
+ @triton_heuristics.pointwise(
196
+ size_hints={'x': 2},
197
+ filename=__file__,
198
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
199
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 8}},
200
+ min_elem_per_thread=0
201
+ )
202
+ @triton.jit
203
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
204
+ xoffset = tl.program_id(0) * XBLOCK
205
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
206
+ xmask = xindex < xnumel
207
+ x0 = xindex
208
+ tmp0 = ks0
209
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
210
+ ''', device_str='cuda')
211
+
212
+
213
+ async_compile.wait(globals())
214
+ del async_compile
215
+
216
+ class Runner:
217
+ def __init__(self, partitions):
218
+ self.partitions = partitions
219
+
220
+ def recursively_apply_fns(self, fns):
221
+ new_callables = []
222
+ for fn, c in zip(fns, self.partitions):
223
+ new_callables.append(fn(c))
224
+ self.partitions = new_callables
225
+
226
+ def call(self, args):
227
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1, arg9_1 = args
228
+ args.clear()
229
+ s53 = arg0_1
230
+ s68 = arg1_1
231
+ s0 = arg3_1
232
+ s37 = arg5_1
233
+ s87 = arg7_1
234
+ s77 = arg9_1
235
+ assert_size_stride(arg2_1, (s0, 1), (1, 1))
236
+ assert_size_stride(arg4_1, (s0, 1), (1, 1))
237
+ assert_size_stride(arg6_1, (s37, 1), (1, 1))
238
+ assert_size_stride(arg8_1, (s0, s87), (s87, 1))
239
+ with torch.cuda._DeviceGuard(2):
240
+ torch.cuda.set_device(2)
241
+ buf0 = empty_strided_cuda((s0, 1, 1), (1, 1, 1), torch.float32)
242
+ buf1 = empty_strided_cuda((s0, 1), (1, 1), torch.float32)
243
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
244
+ stream2 = get_raw_stream(2)
245
+ triton_poi_fused_max_mul_unsqueeze_view_0.run(arg2_1, arg4_1, buf0, buf1, s0, stream=stream2)
246
+ del arg2_1
247
+ del arg4_1
248
+ buf2 = empty_strided_cuda((s0, 1), (1, 1), torch.int64)
249
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
250
+ stream2 = get_raw_stream(2)
251
+ triton_poi_fused_gather_max_view_1.run(arg6_1, buf2, s0, stream=stream2)
252
+ buf3 = empty_strided_cuda((s0, s87), (s87, 1), torch.bfloat16)
253
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
254
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel = s0*s87
255
+ stream2 = get_raw_stream(2)
256
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2.run(arg8_1, buf3, triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel, stream=stream2)
257
+ del arg8_1
258
+ buf4 = empty_strided_cuda((s0, 1), (1, 1), torch.int64)
259
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
260
+ stream2 = get_raw_stream(2)
261
+ triton_poi_fused_add_max_view_3.run(buf4, s77, s0, stream=stream2)
262
+ return (reinterpret_tensor(buf2, (s0, ), (1, ), 0), buf3, buf1, buf0, arg6_1, buf4, )
263
+
264
+ runner = Runner(partitions=[])
265
+ call = runner.call
266
+ recursively_apply_fns = runner.recursively_apply_fns
267
+
268
+
269
+ def get_args():
270
+ from torch._dynamo.testing import rand_strided
271
+ arg0_1 = 1
272
+ arg1_1 = 2
273
+ arg2_1 = rand_strided((2, 1), (1, 1), device='cuda:2', dtype=torch.float32)
274
+ arg3_1 = 2
275
+ arg4_1 = rand_strided((2, 1), (1, 1), device='cuda:2', dtype=torch.float32)
276
+ arg5_1 = 2
277
+ arg6_1 = rand_strided((2, 1), (1, 1), device='cuda:2', dtype=torch.int64)
278
+ arg7_1 = 6144
279
+ arg8_1 = rand_strided((2, 6144), (6144, 1), device='cuda:2', dtype=torch.bfloat16)
280
+ arg9_1 = 1
281
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1, arg9_1]
282
+
283
+
284
+ def benchmark_compiled_module(args, times=10, repeat=10):
285
+ from torch._inductor.utils import print_performance
286
+ fn = lambda: call(list(args))
287
+ return print_performance(fn, times=times, repeat=repeat)
288
+
289
+
290
+ if __name__ == "__main__":
291
+ from torch._inductor.wrapper_benchmark import compiled_module_main
292
+ args = get_args()
293
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/3e/c3exb3qmzq5ev7yyxp6jzpk55mrxicwt5fjnjg72h4blbjqteq3b.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp1, None)
.inductor_cache/3h/c3hjatdysmtczt36qylu5u476tmnxdvxgytr7m6z4na54xrxxoss.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 8},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
.inductor_cache/3h/c3hzz247khkgmf7sy4wb5vbaz4z24jsfnh2gf3spavzjd7bow5gg.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 32},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 384}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_0(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 32
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = xindex < xnumel
23
+ x0 = xindex
24
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
25
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
26
+ tmp2 = tmp0 * tmp1
27
+ tl.store(in_out_ptr0 + (x0), tmp2, xmask)
.inductor_cache/3r/c3rryctepqjzosdqqttq5mb7kr5klk4hfq7vyabrmirmlhcabwmo.py ADDED
@@ -0,0 +1,137 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['10_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/r6/cr6a7je7lgurvwxlurfkudvnwwl26htjtvwwjxnv75jffqdxvgwj.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, mul], Original ATen: [aten.unsqueeze, aten.mul]
39
+ # Source node to ATen node mapping:
40
+ # mul => mul_2
41
+ # unsqueeze => unsqueeze
42
+ # Graph fragment:
43
+ # %arg1_1 : Tensor "f32[1, s3][s3, 1]cuda:5" = PlaceHolder[target=arg1_1]
44
+ # %arg2_1 : Tensor "f32[1, s3, 1][s3, 1, 1]cuda:5" = PlaceHolder[target=arg2_1]
45
+ # %arg3_1 : Tensor "f64[][]cpu" = PlaceHolder[target=arg3_1]
46
+ # %unsqueeze : Tensor "f32[1, s3, 1][s3, 1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg1_1, -1), kwargs = {})
47
+ # %mul_2 : Tensor "f32[1, s3, 1][s3, 1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %arg2_1), kwargs = {})
48
+ # %convert_element_type_default_1 : Tensor "f32[][]cpu"[num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%arg3_1, torch.float32), kwargs = {})
49
+ # %mul_tensor : Tensor "f32[1, s3, 1][s3, 1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_2, %convert_element_type_default_1), kwargs = {})
50
+ # return %mul_tensor
51
+ triton_poi_fused_mul_unsqueeze_0 = async_compile.triton('triton_poi_fused_mul_unsqueeze_0', '''
52
+ import triton
53
+ import triton.language as tl
54
+
55
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
56
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
57
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
58
+ triton_helpers.set_driver_to_gpu()
59
+
60
+ @triton_heuristics.pointwise(
61
+ size_hints={'x': 32},
62
+ filename=__file__,
63
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'in_ptr2': 'fp64', 'out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
64
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 512}},
65
+ min_elem_per_thread=0
66
+ )
67
+ @triton.jit
68
+ def triton_poi_fused_mul_unsqueeze_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
69
+ xoffset = tl.program_id(0) * XBLOCK
70
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
71
+ xmask = xindex < xnumel
72
+ x0 = xindex
73
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
74
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
75
+ tmp3 = in_ptr2
76
+ tmp2 = tmp0 * tmp1
77
+ tmp4 = tmp3.to(tl.float32)
78
+ tmp5 = tmp2 * tmp4
79
+ tl.store(out_ptr0 + (x0), tmp5, xmask)
80
+ ''', device_str='cuda')
81
+
82
+
83
+ async_compile.wait(globals())
84
+ del async_compile
85
+
86
+ class Runner:
87
+ def __init__(self, partitions):
88
+ self.partitions = partitions
89
+
90
+ def recursively_apply_fns(self, fns):
91
+ new_callables = []
92
+ for fn, c in zip(fns, self.partitions):
93
+ new_callables.append(fn(c))
94
+ self.partitions = new_callables
95
+
96
+ def call(self, args):
97
+ arg0_1, arg1_1, arg2_1, arg3_1 = args
98
+ args.clear()
99
+ s3 = arg0_1
100
+ assert_size_stride(arg1_1, (1, s3), (s3, 1))
101
+ assert_size_stride(arg2_1, (1, s3, 1), (s3, 1, 1))
102
+ assert_size_stride(arg3_1, (), ())
103
+ with torch.cuda._DeviceGuard(5):
104
+ torch.cuda.set_device(5)
105
+ buf0 = empty_strided_cuda((1, s3, 1), (s3, 1, 1), torch.float32)
106
+ # Topologically Sorted Source Nodes: [unsqueeze, mul], Original ATen: [aten.unsqueeze, aten.mul]
107
+ stream5 = get_raw_stream(5)
108
+ triton_poi_fused_mul_unsqueeze_0.run(arg1_1, arg2_1, arg3_1.item(), buf0, s3, stream=stream5)
109
+ del arg1_1
110
+ del arg2_1
111
+ del arg3_1
112
+ return (buf0, )
113
+
114
+ runner = Runner(partitions=[])
115
+ call = runner.call
116
+ recursively_apply_fns = runner.recursively_apply_fns
117
+
118
+
119
+ def get_args():
120
+ from torch._dynamo.testing import rand_strided
121
+ arg0_1 = 32
122
+ arg1_1 = rand_strided((1, 32), (32, 1), device='cuda:5', dtype=torch.float32)
123
+ arg2_1 = rand_strided((1, 32, 1), (32, 1, 1), device='cuda:5', dtype=torch.float32)
124
+ arg3_1 = rand_strided((), (), device='cpu', dtype=torch.float64)
125
+ return [arg0_1, arg1_1, arg2_1, arg3_1]
126
+
127
+
128
+ def benchmark_compiled_module(args, times=10, repeat=10):
129
+ from torch._inductor.utils import print_performance
130
+ fn = lambda: call(list(args))
131
+ return print_performance(fn, times=times, repeat=repeat)
132
+
133
+
134
+ if __name__ == "__main__":
135
+ from torch._inductor.wrapper_benchmark import compiled_module_main
136
+ args = get_args()
137
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/3s/c3skh52gob24udbqevuvymapfeanznl6g6yugtw6fyk24ox5ioyb.py ADDED
@@ -0,0 +1,142 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['5_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/xd/cxd3xbystwp4gocdmghs5wj2gqmc5bgvjwrd3qc6rcbnrrgpjvtn.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, mul], Original ATen: [aten.unsqueeze, aten.mul]
39
+ # Source node to ATen node mapping:
40
+ # mul => mul_2
41
+ # unsqueeze => unsqueeze
42
+ # Graph fragment:
43
+ # %arg2_1 : Tensor "f32[s0, s3][s3, 1]cuda:6" = PlaceHolder[target=arg2_1]
44
+ # %arg3_1 : Tensor "f32[s0, s3, 1][s3, 1, 1]cuda:6" = PlaceHolder[target=arg3_1]
45
+ # %arg4_1 : Tensor "f64[][]cpu" = PlaceHolder[target=arg4_1]
46
+ # %unsqueeze : Tensor "f32[s0, s3, 1][s3, 1, 1]cuda:6"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg2_1, -1), kwargs = {})
47
+ # %mul_2 : Tensor "f32[s0, s3, 1][s3, 1, 1]cuda:6"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %arg3_1), kwargs = {})
48
+ # %convert_element_type_default_1 : Tensor "f32[][]cpu"[num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%arg4_1, torch.float32), kwargs = {})
49
+ # %mul_tensor : Tensor "f32[s0, s3, 1][s3, 1, 1]cuda:6"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_2, %convert_element_type_default_1), kwargs = {})
50
+ # return %mul_tensor
51
+ triton_poi_fused_mul_unsqueeze_0 = async_compile.triton('triton_poi_fused_mul_unsqueeze_0', '''
52
+ import triton
53
+ import triton.language as tl
54
+
55
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
56
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
57
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
58
+ triton_helpers.set_driver_to_gpu()
59
+
60
+ @triton_heuristics.pointwise(
61
+ size_hints={'x': 256},
62
+ filename=__file__,
63
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'in_ptr2': 'fp64', 'out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=6, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
64
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 4096}},
65
+ min_elem_per_thread=0
66
+ )
67
+ @triton.jit
68
+ def triton_poi_fused_mul_unsqueeze_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
69
+ xoffset = tl.program_id(0) * XBLOCK
70
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
71
+ xmask = xindex < xnumel
72
+ x0 = xindex
73
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
74
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
75
+ tmp3 = in_ptr2
76
+ tmp2 = tmp0 * tmp1
77
+ tmp4 = tmp3.to(tl.float32)
78
+ tmp5 = tmp2 * tmp4
79
+ tl.store(out_ptr0 + (x0), tmp5, xmask)
80
+ ''', device_str='cuda')
81
+
82
+
83
+ async_compile.wait(globals())
84
+ del async_compile
85
+
86
+ class Runner:
87
+ def __init__(self, partitions):
88
+ self.partitions = partitions
89
+
90
+ def recursively_apply_fns(self, fns):
91
+ new_callables = []
92
+ for fn, c in zip(fns, self.partitions):
93
+ new_callables.append(fn(c))
94
+ self.partitions = new_callables
95
+
96
+ def call(self, args):
97
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1 = args
98
+ args.clear()
99
+ s7 = arg0_1
100
+ s3 = arg1_1
101
+ arg2_1_size = arg2_1.size()
102
+ s0 = arg2_1_size[0]
103
+ assert_size_stride(arg2_1, (s0, s3), (s3, 1))
104
+ assert_size_stride(arg3_1, (s0, s3, 1), (s3, 1, 1))
105
+ assert_size_stride(arg4_1, (), ())
106
+ with torch.cuda._DeviceGuard(6):
107
+ torch.cuda.set_device(6)
108
+ buf0 = empty_strided_cuda((s0, s3, 1), (s3, 1, 1), torch.float32)
109
+ # Topologically Sorted Source Nodes: [unsqueeze, mul], Original ATen: [aten.unsqueeze, aten.mul]
110
+ triton_poi_fused_mul_unsqueeze_0_xnumel = s0*s3
111
+ stream6 = get_raw_stream(6)
112
+ triton_poi_fused_mul_unsqueeze_0.run(arg2_1, arg3_1, arg4_1.item(), buf0, triton_poi_fused_mul_unsqueeze_0_xnumel, stream=stream6)
113
+ del arg2_1
114
+ del arg3_1
115
+ del arg4_1
116
+ return (buf0, )
117
+
118
+ runner = Runner(partitions=[])
119
+ call = runner.call
120
+ recursively_apply_fns = runner.recursively_apply_fns
121
+
122
+
123
+ def get_args():
124
+ from torch._dynamo.testing import rand_strided
125
+ arg0_1 = 8
126
+ arg1_1 = 32
127
+ arg2_1 = rand_strided((8, 32), (32, 1), device='cuda:6', dtype=torch.float32)
128
+ arg3_1 = rand_strided((8, 32, 1), (32, 1, 1), device='cuda:6', dtype=torch.float32)
129
+ arg4_1 = rand_strided((), (), device='cpu', dtype=torch.float64)
130
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1]
131
+
132
+
133
+ def benchmark_compiled_module(args, times=10, repeat=10):
134
+ from torch._inductor.utils import print_performance
135
+ fn = lambda: call(list(args))
136
+ return print_performance(fn, times=times, repeat=repeat)
137
+
138
+
139
+ if __name__ == "__main__":
140
+ from torch._inductor.wrapper_benchmark import compiled_module_main
141
+ args = get_args()
142
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/3u/c3u6bcbbftmiqp5kte3h55vd5gqx7jl3zql7a2czxn52qwb3z3xt.py ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['11_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/vx/cvx3kcgx73mdufoqquzcxowf7ugclz7xreckm42p6weslmelfw3l.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
39
+ # Source node to ATen node mapping:
40
+ # expand_scores => mul_1
41
+ # flatten => view_1
42
+ # max_1 => getitem, max_1
43
+ # unsqueeze => unsqueeze
44
+ # view => view
45
+ # Graph fragment:
46
+ # %arg1_1 : Tensor "f32[1, 1][1, 1]cuda:4" = PlaceHolder[target=arg1_1]
47
+ # %arg2_1 : Tensor "f32[1, 1][1, 1]cuda:4" = PlaceHolder[target=arg2_1]
48
+ # %mul_1 : Tensor "f32[1, 1, 1][1, 1, 1]cuda:4" = PlaceHolder[target=mul_1]
49
+ # %unsqueeze : Tensor "f32[1, 1, 1][1, 1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg1_1, 2), kwargs = {})
50
+ # %view : Tensor "f32[1, 1, 1][1, 1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%arg2_1, [-1, 1, 1]), kwargs = {})
51
+ # %mul_1 : Tensor "f32[1, 1, 1][1, 1, 1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %view), kwargs = {})
52
+ # %view_1 : Tensor "f32[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_1, [1, 1]), kwargs = {})
53
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
54
+ # %getitem : Tensor "f32[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=operator.getitem](args = (%max_1, 0), kwargs = {})
55
+ # return %mul_1,%getitem
56
+ triton_poi_fused_max_mul_unsqueeze_view_0 = async_compile.triton('triton_poi_fused_max_mul_unsqueeze_view_0', '''
57
+ import triton
58
+ import triton.language as tl
59
+
60
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
61
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
62
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
63
+ triton_helpers.set_driver_to_gpu()
64
+
65
+ @triton_heuristics.pointwise(
66
+ size_hints={'x': 1},
67
+ filename=__file__,
68
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
69
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
70
+ min_elem_per_thread=0
71
+ )
72
+ @triton.jit
73
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
74
+ xnumel = 1
75
+ xoffset = tl.program_id(0) * XBLOCK
76
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
77
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
78
+ tmp0 = tl.load(in_ptr0 + (0))
79
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
80
+ tmp2 = tl.load(in_ptr1 + (0))
81
+ tmp3 = tl.broadcast_to(tmp2, [XBLOCK])
82
+ tmp4 = tmp1 * tmp3
83
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp4, None)
84
+ tl.store(out_ptr1 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp4, None)
85
+ ''', device_str='cuda')
86
+
87
+
88
+ # kernel path: /mnt/glm5_2/.inductor_cache/gp/cgpokyocc7tdinfnb6opabr746t6knz5cgybj5fxyufsucoeojra.py
89
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
90
+ # Source node to ATen node mapping:
91
+ # flatten => view_1
92
+ # gather => gather
93
+ # max_1 => max_1
94
+ # Graph fragment:
95
+ # %arg3_1 : Tensor "i64[1, 1][1, 1]cuda:4" = PlaceHolder[target=arg3_1]
96
+ # %view_1 : Tensor "f32[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_1, [1, 1]), kwargs = {})
97
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
98
+ # %gather : Tensor "i64[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.gather.default](args = (%arg3_1, 1, %getitem_1), kwargs = {})
99
+ # return %gather
100
+ triton_poi_fused_gather_max_view_1 = async_compile.triton('triton_poi_fused_gather_max_view_1', '''
101
+ import triton
102
+ import triton.language as tl
103
+
104
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
105
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
106
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
107
+ triton_helpers.set_driver_to_gpu()
108
+
109
+ @triton_heuristics.pointwise(
110
+ size_hints={'x': 1},
111
+ filename=__file__,
112
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
113
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
114
+ min_elem_per_thread=0
115
+ )
116
+ @triton.jit
117
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
118
+ xnumel = 1
119
+ xoffset = tl.program_id(0) * XBLOCK
120
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
121
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
122
+ tmp0 = tl.load(in_ptr0 + (0))
123
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
124
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp1, None)
125
+ ''', device_str='cuda')
126
+
127
+
128
+ # kernel path: /mnt/glm5_2/.inductor_cache/sl/csl35uebvdbczppjgntgmnutdjrljqewnfhmnvz33m6unz7vdiu7.py
129
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
130
+ # Source node to ATen node mapping:
131
+ # batch_offsets => iota
132
+ # flat_cs => view_4
133
+ # flatten => view_1
134
+ # floordiv => div
135
+ # hidden_states => index
136
+ # max_1 => max_1
137
+ # repeat_interleave => unsqueeze_1, view_5
138
+ # selected_input_index => add
139
+ # Graph fragment:
140
+ # %arg5_1 : Tensor "bf16[1, s87][s87, 1]cuda:4" = PlaceHolder[target=arg5_1]
141
+ # %view_1 : Tensor "f32[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_1, [1, 1]), kwargs = {})
142
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
143
+ # %view_4 : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%getitem_1, [1]), kwargs = {})
144
+ # %div : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.div.Tensor_mode](args = (%view_4, 1), kwargs = {rounding_mode: floor})
145
+ # %iota : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (1,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:4, requires_grad: False})
146
+ # %unsqueeze_1 : Tensor "i64[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%iota, 1), kwargs = {})
147
+ # %view_5 : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%expand, [1]), kwargs = {})
148
+ # %add : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %view_5), kwargs = {})
149
+ # %index : Tensor "bf16[1, s87][s87, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%arg5_1, [%add]), kwargs = {})
150
+ # return %index
151
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2 = async_compile.triton('triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', '''
152
+ import triton
153
+ import triton.language as tl
154
+
155
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
156
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
157
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
158
+ triton_helpers.set_driver_to_gpu()
159
+
160
+ @triton_heuristics.pointwise(
161
+ size_hints={'x': 8192},
162
+ filename=__file__,
163
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
164
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 36864}},
165
+ min_elem_per_thread=0
166
+ )
167
+ @triton.jit
168
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
169
+ xoffset = tl.program_id(0) * XBLOCK
170
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
171
+ xmask = xindex < xnumel
172
+ x0 = xindex
173
+ tmp0 = tl.load(in_ptr0 + (x0), xmask).to(tl.float32)
174
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
175
+ ''', device_str='cuda')
176
+
177
+
178
+ # kernel path: /mnt/glm5_2/.inductor_cache/dn/cdnvk7ketvhjg6i2ggdy5oiyemirjdox3326cagxkaqlxhmh5lpp.py
179
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
180
+ # Source node to ATen node mapping:
181
+ # add_2 => add_3
182
+ # flatten => view_1
183
+ # max_1 => max_1
184
+ # Graph fragment:
185
+ # %view_1 : Tensor "f32[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_1, [1, 1]), kwargs = {})
186
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
187
+ # %add_3 : Tensor "i64[1, 1][1, 1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_1, %arg6_1), kwargs = {})
188
+ # return %add_3
189
+ triton_poi_fused_add_max_view_3 = async_compile.triton('triton_poi_fused_add_max_view_3', '''
190
+ import triton
191
+ import triton.language as tl
192
+
193
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
194
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
195
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
196
+ triton_helpers.set_driver_to_gpu()
197
+
198
+ @triton_heuristics.pointwise(
199
+ size_hints={'x': 1},
200
+ filename=__file__,
201
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
202
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
203
+ min_elem_per_thread=0
204
+ )
205
+ @triton.jit
206
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
207
+ xnumel = 1
208
+ xoffset = tl.program_id(0) * XBLOCK
209
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
210
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
211
+ tmp0 = ks0
212
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32)), tmp0, None)
213
+ ''', device_str='cuda')
214
+
215
+
216
+ async_compile.wait(globals())
217
+ del async_compile
218
+
219
+ class Runner:
220
+ def __init__(self, partitions):
221
+ self.partitions = partitions
222
+
223
+ def recursively_apply_fns(self, fns):
224
+ new_callables = []
225
+ for fn, c in zip(fns, self.partitions):
226
+ new_callables.append(fn(c))
227
+ self.partitions = new_callables
228
+
229
+ def call(self, args):
230
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1 = args
231
+ args.clear()
232
+ s53 = arg0_1
233
+ s87 = arg4_1
234
+ s77 = arg6_1
235
+ assert_size_stride(arg1_1, (1, 1), (1, 1))
236
+ assert_size_stride(arg2_1, (1, 1), (1, 1))
237
+ assert_size_stride(arg3_1, (1, 1), (1, 1))
238
+ assert_size_stride(arg5_1, (1, s87), (s87, 1))
239
+ with torch.cuda._DeviceGuard(4):
240
+ torch.cuda.set_device(4)
241
+ buf0 = empty_strided_cuda((1, 1, 1), (1, 1, 1), torch.float32)
242
+ buf1 = empty_strided_cuda((1, 1), (1, 1), torch.float32)
243
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
244
+ stream4 = get_raw_stream(4)
245
+ triton_poi_fused_max_mul_unsqueeze_view_0.run(arg1_1, arg2_1, buf0, buf1, 1, stream=stream4)
246
+ del arg1_1
247
+ del arg2_1
248
+ buf2 = empty_strided_cuda((1, 1), (1, 1), torch.int64)
249
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
250
+ stream4 = get_raw_stream(4)
251
+ triton_poi_fused_gather_max_view_1.run(arg3_1, buf2, 1, stream=stream4)
252
+ buf3 = empty_strided_cuda((1, s87), (s87, 1), torch.bfloat16)
253
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
254
+ stream4 = get_raw_stream(4)
255
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2.run(arg5_1, buf3, s87, stream=stream4)
256
+ del arg5_1
257
+ buf4 = empty_strided_cuda((1, 1), (1, 1), torch.int64)
258
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
259
+ stream4 = get_raw_stream(4)
260
+ triton_poi_fused_add_max_view_3.run(buf4, s77, 1, stream=stream4)
261
+ return (reinterpret_tensor(buf2, (1, ), (1, ), 0), buf3, buf1, buf0, arg3_1, buf4, )
262
+
263
+ runner = Runner(partitions=[])
264
+ call = runner.call
265
+ recursively_apply_fns = runner.recursively_apply_fns
266
+
267
+
268
+ def get_args():
269
+ from torch._dynamo.testing import rand_strided
270
+ arg0_1 = 1
271
+ arg1_1 = rand_strided((1, 1), (1, 1), device='cuda:4', dtype=torch.float32)
272
+ arg2_1 = rand_strided((1, 1), (1, 1), device='cuda:4', dtype=torch.float32)
273
+ arg3_1 = rand_strided((1, 1), (1, 1), device='cuda:4', dtype=torch.int64)
274
+ arg4_1 = 6144
275
+ arg5_1 = rand_strided((1, 6144), (6144, 1), device='cuda:4', dtype=torch.bfloat16)
276
+ arg6_1 = 1
277
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1]
278
+
279
+
280
+ def benchmark_compiled_module(args, times=10, repeat=10):
281
+ from torch._inductor.utils import print_performance
282
+ fn = lambda: call(list(args))
283
+ return print_performance(fn, times=times, repeat=repeat)
284
+
285
+
286
+ if __name__ == "__main__":
287
+ from torch._inductor.wrapper_benchmark import compiled_module_main
288
+ args = get_args()
289
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/3z/c3ztg2btpqdyc3uz4dzswnkglhrmeaszpogpdu23omfz4c6u7zko.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 32},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 384}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_0(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 32
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = xindex < xnumel
23
+ x0 = xindex
24
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
25
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
26
+ tmp2 = tmp0 * tmp1
27
+ tl.store(in_out_ptr0 + (x0), tmp2, xmask)
.inductor_cache/46/9ddaabf0d63fddb6e484df5802c5ab86ddbc3d9b84c46c643b006cc0acf7b995.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 256, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 23, "triton_cache_hash": "GIDCKWLVJJFYH72JFJCG4FFIN3BV7CWCQPTZRHJERXHCZJAW4VIA"}
.inductor_cache/46/c46gch5kg22oprcsfx3k73s77cdoj5zwmm4trujfwwc4zc7cs72z.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 8192},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 36864}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask).to(tl.float32)
24
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
.inductor_cache/47/c47i4hdzezcx3xndn3biber6sa6rznklzeek5cbpexpomknwaqjt.py ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['0_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/jr/cjrkeuk4xjhx774cklsib5ivf4nd32eudc4tub7gpkw5nbrgpf7w.py
38
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, mul, mul_1, valid_offset, sub_3, input_, invert], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.mul, aten.add, aten.sub, aten.bitwise_not]
39
+ # Source node to ATen node mapping:
40
+ # added_vocab_mask => bitwise_and_1
41
+ # ge => ge
42
+ # ge_1 => ge_1
43
+ # input_ => mul_6
44
+ # invert => bitwise_not
45
+ # lt => lt
46
+ # lt_1 => lt_1
47
+ # mul => mul
48
+ # mul_1 => mul_2
49
+ # org_vocab_mask => bitwise_and
50
+ # sub_3 => sub_13
51
+ # valid_offset => add_16
52
+ # vocab_mask => bitwise_or
53
+ # Graph fragment:
54
+ # %arg1_1 : Tensor "i64[s15][1]cuda:7" = PlaceHolder[target=arg1_1]
55
+ # %ge : Tensor "b8[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg1_1, %arg2_1), kwargs = {})
56
+ # %lt : Tensor "b8[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg1_1, %arg3_1), kwargs = {})
57
+ # %bitwise_and : Tensor "b8[s15][1]cuda:7"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge, %lt), kwargs = {})
58
+ # %ge_1 : Tensor "b8[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg1_1, %arg4_1), kwargs = {})
59
+ # %lt_1 : Tensor "b8[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg1_1, %arg5_1), kwargs = {})
60
+ # %bitwise_and_1 : Tensor "b8[s15][1]cuda:7"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge_1, %lt_1), kwargs = {})
61
+ # %bitwise_or : Tensor "b8[s15][1]cuda:7"[num_users=2] = call_function[target=torch.ops.aten.bitwise_or.Tensor](args = (%bitwise_and, %bitwise_and_1), kwargs = {})
62
+ # %mul : Tensor "i64[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_and, %arg2_1), kwargs = {})
63
+ # %mul_2 : Tensor "i64[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_and_1, %sub_8), kwargs = {})
64
+ # %add_16 : Tensor "i64[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %mul_2), kwargs = {})
65
+ # %sub_13 : Tensor "i64[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg1_1, %add_16), kwargs = {})
66
+ # %mul_6 : Tensor "i64[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_or, %sub_13), kwargs = {})
67
+ # %bitwise_not : Tensor "b8[s15][1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.bitwise_not.default](args = (%bitwise_or,), kwargs = {})
68
+ # return %mul_6,%bitwise_not
69
+ triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0 = async_compile.triton('triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0', '''
70
+ import triton
71
+ import triton.language as tl
72
+
73
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
74
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
75
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
76
+ triton_helpers.set_driver_to_gpu()
77
+
78
+ @triton_heuristics.pointwise(
79
+ size_hints={'x': 128},
80
+ filename=__file__,
81
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'out_ptr1': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]},
82
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 2912}},
83
+ min_elem_per_thread=0
84
+ )
85
+ @triton.jit
86
+ def triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, out_ptr1, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
87
+ xoffset = tl.program_id(0) * XBLOCK
88
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
89
+ xmask = xindex < xnumel
90
+ x0 = xindex
91
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
92
+ tmp1 = ks0
93
+ tmp2 = tmp0 >= tmp1
94
+ tmp3 = ks1
95
+ tmp4 = tmp0 < tmp3
96
+ tmp5 = tmp2 & tmp4
97
+ tmp6 = ks2
98
+ tmp7 = tmp0 >= tmp6
99
+ tmp8 = ks3
100
+ tmp9 = tmp0 < tmp8
101
+ tmp10 = tmp7 & tmp9
102
+ tmp11 = tmp5 | tmp10
103
+ tmp12 = tmp11.to(tl.int64)
104
+ tmp13 = tmp5.to(tl.int64)
105
+ tmp14 = tmp13 * tmp1
106
+ tmp15 = tmp10.to(tl.int64)
107
+ tmp16 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
108
+ tmp17 = tmp15 * tmp16
109
+ tmp18 = tmp14 + tmp17
110
+ tmp19 = tmp0 - tmp18
111
+ tmp20 = tmp12 * tmp19
112
+ tmp21 = tmp11 == 0
113
+ tl.store(out_ptr0 + (x0), tmp20, xmask)
114
+ tl.store(out_ptr1 + (x0), tmp21, xmask)
115
+ ''', device_str='cuda')
116
+
117
+
118
+ async_compile.wait(globals())
119
+ del async_compile
120
+
121
+ class Runner:
122
+ def __init__(self, partitions):
123
+ self.partitions = partitions
124
+
125
+ def recursively_apply_fns(self, fns):
126
+ new_callables = []
127
+ for fn, c in zip(fns, self.partitions):
128
+ new_callables.append(fn(c))
129
+ self.partitions = new_callables
130
+
131
+ def call(self, args):
132
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1 = args
133
+ args.clear()
134
+ s15 = arg0_1
135
+ s90 = arg2_1
136
+ s87 = arg3_1
137
+ s68 = arg4_1
138
+ s53 = arg5_1
139
+ s88 = arg6_1
140
+ assert_size_stride(arg1_1, (s15, ), (1, ))
141
+ with torch.cuda._DeviceGuard(7):
142
+ torch.cuda.set_device(7)
143
+ buf0 = empty_strided_cuda((s15, ), (1, ), torch.int64)
144
+ buf1 = empty_strided_cuda((s15, ), (1, ), torch.bool)
145
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, mul, mul_1, valid_offset, sub_3, input_, invert], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.mul, aten.add, aten.sub, aten.bitwise_not]
146
+ stream7 = get_raw_stream(7)
147
+ triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0.run(arg1_1, buf0, buf1, s90, s87, s68, s53, s88, s15, stream=stream7)
148
+ del arg1_1
149
+ return (buf0, buf1, )
150
+
151
+ runner = Runner(partitions=[])
152
+ call = runner.call
153
+ recursively_apply_fns = runner.recursively_apply_fns
154
+
155
+
156
+ def get_args():
157
+ from torch._dynamo.testing import rand_strided
158
+ arg0_1 = 112
159
+ arg1_1 = rand_strided((112, ), (1, ), device='cuda:7', dtype=torch.int64)
160
+ arg2_1 = 135520
161
+ arg3_1 = 154880
162
+ arg4_1 = 154880
163
+ arg5_1 = 154880
164
+ arg6_1 = 0
165
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1]
166
+
167
+
168
+ def benchmark_compiled_module(args, times=10, repeat=10):
169
+ from torch._inductor.utils import print_performance
170
+ fn = lambda: call(list(args))
171
+ return print_performance(fn, times=times, repeat=repeat)
172
+
173
+
174
+ if __name__ == "__main__":
175
+ from torch._inductor.wrapper_benchmark import compiled_module_main
176
+ args = get_args()
177
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/4b/c4bzcwru43j45mp2ej3abi6jtvpi4pqeiwedj4gy3vcydr2jkwun.py ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['6_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/pq/cpqfafvbzdc4yv5hpjbitcl5zsctcve7hn6ahg66aj4blq3vprf6.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
39
+ # Source node to ATen node mapping:
40
+ # expand_scores => mul_3
41
+ # flatten => view_1
42
+ # max_1 => getitem, max_1
43
+ # unsqueeze => unsqueeze
44
+ # view => view
45
+ # Graph fragment:
46
+ # %arg2_1 : Tensor "f32[s97, 1][1, 1]cuda:5" = PlaceHolder[target=arg2_1]
47
+ # %arg3_1 : Tensor "f32[s97, 1][1, 1]cuda:5" = PlaceHolder[target=arg3_1]
48
+ # %mul_3 : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:5" = PlaceHolder[target=mul_3]
49
+ # %unsqueeze : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg2_1, 2), kwargs = {})
50
+ # %view : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%arg3_1, [-1, 1, 1]), kwargs = {})
51
+ # %mul_3 : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:5"[num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %view), kwargs = {})
52
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
53
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
54
+ # %getitem : Tensor "f32[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=operator.getitem](args = (%max_1, 0), kwargs = {})
55
+ # return %mul_3,%getitem
56
+ triton_poi_fused_max_mul_unsqueeze_view_0 = async_compile.triton('triton_poi_fused_max_mul_unsqueeze_view_0', '''
57
+ import triton
58
+ import triton.language as tl
59
+
60
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
61
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
62
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
63
+ triton_helpers.set_driver_to_gpu()
64
+
65
+ @triton_heuristics.pointwise(
66
+ size_hints={'x': 8},
67
+ filename=__file__,
68
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
69
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
70
+ min_elem_per_thread=0
71
+ )
72
+ @triton.jit
73
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
74
+ xoffset = tl.program_id(0) * XBLOCK
75
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
76
+ xmask = xindex < xnumel
77
+ x0 = xindex
78
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
79
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
80
+ tmp2 = tmp0 * tmp1
81
+ tl.store(out_ptr0 + (x0), tmp2, xmask)
82
+ tl.store(out_ptr1 + (x0), tmp2, xmask)
83
+ ''', device_str='cuda')
84
+
85
+
86
+ # kernel path: /mnt/glm5_2/.inductor_cache/zb/czbs2fm46j6qhtd2znnofgeavkuhyl72vdcahggz7dhvxp5zvnma.py
87
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
88
+ # Source node to ATen node mapping:
89
+ # flatten => view_1
90
+ # gather => gather
91
+ # max_1 => max_1
92
+ # Graph fragment:
93
+ # %arg4_1 : Tensor "i64[s97, 1][1, 1]cuda:5" = PlaceHolder[target=arg4_1]
94
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
95
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
96
+ # %gather : Tensor "i64[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.gather.default](args = (%arg4_1, 1, %getitem_1), kwargs = {})
97
+ # return %gather
98
+ triton_poi_fused_gather_max_view_1 = async_compile.triton('triton_poi_fused_gather_max_view_1', '''
99
+ import triton
100
+ import triton.language as tl
101
+
102
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
103
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
104
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
105
+ triton_helpers.set_driver_to_gpu()
106
+
107
+ @triton_heuristics.pointwise(
108
+ size_hints={'x': 8},
109
+ filename=__file__,
110
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
111
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
112
+ min_elem_per_thread=0
113
+ )
114
+ @triton.jit
115
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
116
+ xoffset = tl.program_id(0) * XBLOCK
117
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
118
+ xmask = xindex < xnumel
119
+ x0 = xindex
120
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
121
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
122
+ ''', device_str='cuda')
123
+
124
+
125
+ # kernel path: /mnt/glm5_2/.inductor_cache/zm/czmqgy6soqztyipifu6p2jfwi4je2rv5dgpxsqvvy3vdykavyemd.py
126
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
127
+ # Source node to ATen node mapping:
128
+ # batch_offsets => iota
129
+ # flat_cs => view_4
130
+ # flatten => view_1
131
+ # floordiv => div
132
+ # hidden_states => index
133
+ # max_1 => max_1
134
+ # repeat_interleave => unsqueeze_1, view_5
135
+ # selected_input_index => add_47
136
+ # Graph fragment:
137
+ # %arg6_1 : Tensor "bf16[s97, s87][s87, 1]cuda:5" = PlaceHolder[target=arg6_1]
138
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
139
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
140
+ # %view_4 : Tensor "i64[s97][1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%getitem_1, [%arg1_1]), kwargs = {})
141
+ # %div : Tensor "i64[s97][1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.div.Tensor_mode](args = (%view_4, 1), kwargs = {rounding_mode: floor})
142
+ # %iota : Tensor "i64[s97][1]cuda:5"[num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (%arg1_1,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:5, requires_grad: False})
143
+ # %unsqueeze_1 : Tensor "i64[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%iota, 1), kwargs = {})
144
+ # %view_5 : Tensor "i64[s97][1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%expand, [%arg1_1]), kwargs = {})
145
+ # %add_47 : Tensor "i64[s97][1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %view_5), kwargs = {})
146
+ # %index : Tensor "bf16[s97, s87][s87, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%arg6_1, [%add_47]), kwargs = {})
147
+ # return %index
148
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2 = async_compile.triton('triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', '''
149
+ import triton
150
+ import triton.language as tl
151
+
152
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
153
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
154
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
155
+ triton_helpers.set_driver_to_gpu()
156
+
157
+ @triton_heuristics.pointwise(
158
+ size_hints={'x': 65536},
159
+ filename=__file__,
160
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
161
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 294912}},
162
+ min_elem_per_thread=0
163
+ )
164
+ @triton.jit
165
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
166
+ xoffset = tl.program_id(0) * XBLOCK
167
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
168
+ xmask = xindex < xnumel
169
+ x2 = xindex
170
+ tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last').to(tl.float32)
171
+ tl.store(out_ptr0 + (x2), tmp0, xmask)
172
+ ''', device_str='cuda')
173
+
174
+
175
+ # kernel path: /mnt/glm5_2/.inductor_cache/mz/cmzyybpwh6qmtcq7gwcqgwbieukiskiqslaso7ig7lcotivioxzy.py
176
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
177
+ # Source node to ATen node mapping:
178
+ # add_2 => add_53
179
+ # flatten => view_1
180
+ # max_1 => max_1
181
+ # Graph fragment:
182
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
183
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
184
+ # %add_53 : Tensor "i64[s97, 1][1, 1]cuda:5"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_1, %arg7_1), kwargs = {})
185
+ # return %add_53
186
+ triton_poi_fused_add_max_view_3 = async_compile.triton('triton_poi_fused_add_max_view_3', '''
187
+ import triton
188
+ import triton.language as tl
189
+
190
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
191
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
192
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
193
+ triton_helpers.set_driver_to_gpu()
194
+
195
+ @triton_heuristics.pointwise(
196
+ size_hints={'x': 8},
197
+ filename=__file__,
198
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
199
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 128}},
200
+ min_elem_per_thread=0
201
+ )
202
+ @triton.jit
203
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
204
+ xoffset = tl.program_id(0) * XBLOCK
205
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
206
+ xmask = xindex < xnumel
207
+ x0 = xindex
208
+ tmp0 = ks0
209
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
210
+ ''', device_str='cuda')
211
+
212
+
213
+ async_compile.wait(globals())
214
+ del async_compile
215
+
216
+ class Runner:
217
+ def __init__(self, partitions):
218
+ self.partitions = partitions
219
+
220
+ def recursively_apply_fns(self, fns):
221
+ new_callables = []
222
+ for fn, c in zip(fns, self.partitions):
223
+ new_callables.append(fn(c))
224
+ self.partitions = new_callables
225
+
226
+ def call(self, args):
227
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1 = args
228
+ args.clear()
229
+ s53 = arg0_1
230
+ s97 = arg1_1
231
+ s87 = arg5_1
232
+ s77 = arg7_1
233
+ assert_size_stride(arg2_1, (s97, 1), (1, 1))
234
+ assert_size_stride(arg3_1, (s97, 1), (1, 1))
235
+ assert_size_stride(arg4_1, (s97, 1), (1, 1))
236
+ assert_size_stride(arg6_1, (s97, s87), (s87, 1))
237
+ with torch.cuda._DeviceGuard(5):
238
+ torch.cuda.set_device(5)
239
+ buf0 = empty_strided_cuda((s97, 1, 1), (1, 1, 1), torch.float32)
240
+ buf1 = empty_strided_cuda((s97, 1), (1, 1), torch.float32)
241
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
242
+ stream5 = get_raw_stream(5)
243
+ triton_poi_fused_max_mul_unsqueeze_view_0.run(arg2_1, arg3_1, buf0, buf1, s97, stream=stream5)
244
+ del arg2_1
245
+ del arg3_1
246
+ buf2 = empty_strided_cuda((s97, 1), (1, 1), torch.int64)
247
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
248
+ stream5 = get_raw_stream(5)
249
+ triton_poi_fused_gather_max_view_1.run(arg4_1, buf2, s97, stream=stream5)
250
+ buf3 = empty_strided_cuda((s97, s87), (s87, 1), torch.bfloat16)
251
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
252
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel = s87*s97
253
+ stream5 = get_raw_stream(5)
254
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2.run(arg6_1, buf3, triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel, stream=stream5)
255
+ del arg6_1
256
+ buf4 = empty_strided_cuda((s97, 1), (1, 1), torch.int64)
257
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
258
+ stream5 = get_raw_stream(5)
259
+ triton_poi_fused_add_max_view_3.run(buf4, s77, s97, stream=stream5)
260
+ return (reinterpret_tensor(buf2, (s97, ), (1, ), 0), buf3, buf1, buf0, arg4_1, buf4, )
261
+
262
+ runner = Runner(partitions=[])
263
+ call = runner.call
264
+ recursively_apply_fns = runner.recursively_apply_fns
265
+
266
+
267
+ def get_args():
268
+ from torch._dynamo.testing import rand_strided
269
+ arg0_1 = 1
270
+ arg1_1 = 8
271
+ arg2_1 = rand_strided((8, 1), (1, 1), device='cuda:5', dtype=torch.float32)
272
+ arg3_1 = rand_strided((8, 1), (1, 1), device='cuda:5', dtype=torch.float32)
273
+ arg4_1 = rand_strided((8, 1), (1, 1), device='cuda:5', dtype=torch.int64)
274
+ arg5_1 = 6144
275
+ arg6_1 = rand_strided((8, 6144), (6144, 1), device='cuda:5', dtype=torch.bfloat16)
276
+ arg7_1 = 1
277
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1]
278
+
279
+
280
+ def benchmark_compiled_module(args, times=10, repeat=10):
281
+ from torch._inductor.utils import print_performance
282
+ fn = lambda: call(list(args))
283
+ return print_performance(fn, times=times, repeat=repeat)
284
+
285
+
286
+ if __name__ == "__main__":
287
+ from torch._inductor.wrapper_benchmark import compiled_module_main
288
+ args = get_args()
289
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/4b/c4bzoixvavbv7tbgxpx5sajtax6yl6d6wky5k272ff4wk64t7jd7.py ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp1, None)
.inductor_cache/4i/c4iwhfgin6ac7qvqb6fj5msgdftp5qsxjv6sjxyuqbryq4qqwzfg.py ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=6, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tmp2 = ks0
26
+ tmp3 = tmp1 >= tmp2
27
+ tmp4 = ks1
28
+ tmp5 = tmp1 < tmp4
29
+ tmp6 = tmp3 & tmp5
30
+ tmp7 = ks2
31
+ tmp8 = tmp1 >= tmp7
32
+ tmp9 = ks3
33
+ tmp10 = tmp1 < tmp9
34
+ tmp11 = tmp8 & tmp10
35
+ tmp12 = tmp6 | tmp11
36
+ tmp13 = tmp12 == 0
37
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp13, None)
.inductor_cache/4l/c4ljp5q37q7ktbyo35672uzle4llt4iwt4be37ikkwlit6j7fnco.py ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 2048},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'in_ptr1': 'fp64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 28672}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
24
+ tmp3 = tl.load(in_ptr0 + (x0), xmask)
25
+ tmp5 = in_ptr1
26
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
27
+ tmp2 = tmp0 * tmp1
28
+ tmp4 = tmp2 * tmp3
29
+ tmp6 = tmp5.to(tl.float32)
30
+ tmp7 = tmp4 * tmp6
31
+ tl.store(in_out_ptr0 + (x0), tmp7, xmask)
.inductor_cache/4l/dccab99d67a1299427a9211bfd01fd3c77aaa44ad59d9ffa12a6d00f48a13c0e.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 256, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 48, "triton_cache_hash": "KGGY5WAIFCN23ENKRQFNT2OZB6CJLZGZIIMD5KHX6XAEJ3DEY7VA"}
.inductor_cache/4n/c4nxaenxhyi2kg2uordabgcp5ofkprcosqtxys5klrlaly24kcgm.py ADDED
@@ -0,0 +1,293 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['13_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/qd/cqdull3kwipr33yws7qnj6z4jsiuvue3qmvlqyhxb6q2tawpraqz.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
39
+ # Source node to ATen node mapping:
40
+ # expand_scores => mul_3
41
+ # flatten => view_1
42
+ # max_1 => getitem, max_1
43
+ # unsqueeze => unsqueeze
44
+ # view => view
45
+ # Graph fragment:
46
+ # %arg2_1 : Tensor "f32[s0, 1][1, 1]cuda:3" = PlaceHolder[target=arg2_1]
47
+ # %arg4_1 : Tensor "f32[s0, 1][1, 1]cuda:3" = PlaceHolder[target=arg4_1]
48
+ # %mul_3 : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:3" = PlaceHolder[target=mul_3]
49
+ # %unsqueeze : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg2_1, 2), kwargs = {})
50
+ # %view : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%arg4_1, [-1, 1, 1]), kwargs = {})
51
+ # %mul_3 : Tensor "f32[s0, 1, 1][1, 1, 1]cuda:3"[num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %view), kwargs = {})
52
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
53
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
54
+ # %getitem : Tensor "f32[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=operator.getitem](args = (%max_1, 0), kwargs = {})
55
+ # return %mul_3,%getitem
56
+ triton_poi_fused_max_mul_unsqueeze_view_0 = async_compile.triton('triton_poi_fused_max_mul_unsqueeze_view_0', '''
57
+ import triton
58
+ import triton.language as tl
59
+
60
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
61
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
62
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
63
+ triton_helpers.set_driver_to_gpu()
64
+
65
+ @triton_heuristics.pointwise(
66
+ size_hints={'x': 2},
67
+ filename=__file__,
68
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
69
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 12}},
70
+ min_elem_per_thread=0
71
+ )
72
+ @triton.jit
73
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
74
+ xoffset = tl.program_id(0) * XBLOCK
75
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
76
+ xmask = xindex < xnumel
77
+ x0 = xindex
78
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
79
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
80
+ tmp2 = tmp0 * tmp1
81
+ tl.store(out_ptr0 + (x0), tmp2, xmask)
82
+ tl.store(out_ptr1 + (x0), tmp2, xmask)
83
+ ''', device_str='cuda')
84
+
85
+
86
+ # kernel path: /mnt/glm5_2/.inductor_cache/sm/csmhkmklmfo5h4gcgl6wazfjapbn6edilhx43by42rcvlwmhrmz6.py
87
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
88
+ # Source node to ATen node mapping:
89
+ # flatten => view_1
90
+ # gather => gather
91
+ # max_1 => max_1
92
+ # Graph fragment:
93
+ # %arg6_1 : Tensor "i64[s37, 1][1, 1]cuda:3" = PlaceHolder[target=arg6_1]
94
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
95
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
96
+ # %gather : Tensor "i64[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.gather.default](args = (%arg6_1, 1, %getitem_1), kwargs = {})
97
+ # return %gather
98
+ triton_poi_fused_gather_max_view_1 = async_compile.triton('triton_poi_fused_gather_max_view_1', '''
99
+ import triton
100
+ import triton.language as tl
101
+
102
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
103
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
104
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
105
+ triton_helpers.set_driver_to_gpu()
106
+
107
+ @triton_heuristics.pointwise(
108
+ size_hints={'x': 2},
109
+ filename=__file__,
110
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
111
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 12}},
112
+ min_elem_per_thread=0
113
+ )
114
+ @triton.jit
115
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
116
+ xoffset = tl.program_id(0) * XBLOCK
117
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
118
+ xmask = xindex < xnumel
119
+ x0 = xindex
120
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
121
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
122
+ ''', device_str='cuda')
123
+
124
+
125
+ # kernel path: /mnt/glm5_2/.inductor_cache/bu/cburtb54dxftz4ovx6wqbvnuvsq4klzrlfgkq2xyomwqr45boyxv.py
126
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
127
+ # Source node to ATen node mapping:
128
+ # batch_offsets => iota
129
+ # flat_cs => view_4
130
+ # flatten => view_1
131
+ # floordiv => div
132
+ # hidden_states => index
133
+ # max_1 => max_1
134
+ # repeat_interleave => unsqueeze_1, view_5
135
+ # selected_input_index => add_47
136
+ # Graph fragment:
137
+ # %arg8_1 : Tensor "bf16[s0, s87][s87, 1]cuda:3" = PlaceHolder[target=arg8_1]
138
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
139
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
140
+ # %view_4 : Tensor "i64[s0][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%getitem_1, [%arg3_1]), kwargs = {})
141
+ # %div : Tensor "i64[s0][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.div.Tensor_mode](args = (%view_4, 1), kwargs = {rounding_mode: floor})
142
+ # %iota : Tensor "i64[s0][1]cuda:3"[num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (%arg3_1,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:3, requires_grad: False})
143
+ # %unsqueeze_1 : Tensor "i64[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%iota, 1), kwargs = {})
144
+ # %view_5 : Tensor "i64[s0][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%expand, [%arg3_1]), kwargs = {})
145
+ # %add_47 : Tensor "i64[s0][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %view_5), kwargs = {})
146
+ # %index : Tensor "bf16[s0, s87][s87, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%arg8_1, [%add_47]), kwargs = {})
147
+ # return %index
148
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2 = async_compile.triton('triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', '''
149
+ import triton
150
+ import triton.language as tl
151
+
152
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
153
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
154
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
155
+ triton_helpers.set_driver_to_gpu()
156
+
157
+ @triton_heuristics.pointwise(
158
+ size_hints={'x': 16384},
159
+ filename=__file__,
160
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
161
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 73728}},
162
+ min_elem_per_thread=0
163
+ )
164
+ @triton.jit
165
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
166
+ xoffset = tl.program_id(0) * XBLOCK
167
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
168
+ xmask = xindex < xnumel
169
+ x2 = xindex
170
+ tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last').to(tl.float32)
171
+ tl.store(out_ptr0 + (x2), tmp0, xmask)
172
+ ''', device_str='cuda')
173
+
174
+
175
+ # kernel path: /mnt/glm5_2/.inductor_cache/ub/cubrga4p6slgmvpkerkbtbebckik3kdln7y7xyo466mbcclx3vly.py
176
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
177
+ # Source node to ATen node mapping:
178
+ # add_2 => add_53
179
+ # flatten => view_1
180
+ # max_1 => max_1
181
+ # Graph fragment:
182
+ # %view_1 : Tensor "f32[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg3_1, 1]), kwargs = {})
183
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
184
+ # %add_53 : Tensor "i64[s0, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_1, %arg9_1), kwargs = {})
185
+ # return %add_53
186
+ triton_poi_fused_add_max_view_3 = async_compile.triton('triton_poi_fused_add_max_view_3', '''
187
+ import triton
188
+ import triton.language as tl
189
+
190
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
191
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
192
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
193
+ triton_helpers.set_driver_to_gpu()
194
+
195
+ @triton_heuristics.pointwise(
196
+ size_hints={'x': 2},
197
+ filename=__file__,
198
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
199
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 8}},
200
+ min_elem_per_thread=0
201
+ )
202
+ @triton.jit
203
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
204
+ xoffset = tl.program_id(0) * XBLOCK
205
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
206
+ xmask = xindex < xnumel
207
+ x0 = xindex
208
+ tmp0 = ks0
209
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
210
+ ''', device_str='cuda')
211
+
212
+
213
+ async_compile.wait(globals())
214
+ del async_compile
215
+
216
+ class Runner:
217
+ def __init__(self, partitions):
218
+ self.partitions = partitions
219
+
220
+ def recursively_apply_fns(self, fns):
221
+ new_callables = []
222
+ for fn, c in zip(fns, self.partitions):
223
+ new_callables.append(fn(c))
224
+ self.partitions = new_callables
225
+
226
+ def call(self, args):
227
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1, arg9_1 = args
228
+ args.clear()
229
+ s53 = arg0_1
230
+ s68 = arg1_1
231
+ s0 = arg3_1
232
+ s37 = arg5_1
233
+ s87 = arg7_1
234
+ s77 = arg9_1
235
+ assert_size_stride(arg2_1, (s0, 1), (1, 1))
236
+ assert_size_stride(arg4_1, (s0, 1), (1, 1))
237
+ assert_size_stride(arg6_1, (s37, 1), (1, 1))
238
+ assert_size_stride(arg8_1, (s0, s87), (s87, 1))
239
+ with torch.cuda._DeviceGuard(3):
240
+ torch.cuda.set_device(3)
241
+ buf0 = empty_strided_cuda((s0, 1, 1), (1, 1, 1), torch.float32)
242
+ buf1 = empty_strided_cuda((s0, 1), (1, 1), torch.float32)
243
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
244
+ stream3 = get_raw_stream(3)
245
+ triton_poi_fused_max_mul_unsqueeze_view_0.run(arg2_1, arg4_1, buf0, buf1, s0, stream=stream3)
246
+ del arg2_1
247
+ del arg4_1
248
+ buf2 = empty_strided_cuda((s0, 1), (1, 1), torch.int64)
249
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
250
+ stream3 = get_raw_stream(3)
251
+ triton_poi_fused_gather_max_view_1.run(arg6_1, buf2, s0, stream=stream3)
252
+ buf3 = empty_strided_cuda((s0, s87), (s87, 1), torch.bfloat16)
253
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
254
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel = s0*s87
255
+ stream3 = get_raw_stream(3)
256
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2.run(arg8_1, buf3, triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel, stream=stream3)
257
+ del arg8_1
258
+ buf4 = empty_strided_cuda((s0, 1), (1, 1), torch.int64)
259
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
260
+ stream3 = get_raw_stream(3)
261
+ triton_poi_fused_add_max_view_3.run(buf4, s77, s0, stream=stream3)
262
+ return (reinterpret_tensor(buf2, (s0, ), (1, ), 0), buf3, buf1, buf0, arg6_1, buf4, )
263
+
264
+ runner = Runner(partitions=[])
265
+ call = runner.call
266
+ recursively_apply_fns = runner.recursively_apply_fns
267
+
268
+
269
+ def get_args():
270
+ from torch._dynamo.testing import rand_strided
271
+ arg0_1 = 1
272
+ arg1_1 = 2
273
+ arg2_1 = rand_strided((2, 1), (1, 1), device='cuda:3', dtype=torch.float32)
274
+ arg3_1 = 2
275
+ arg4_1 = rand_strided((2, 1), (1, 1), device='cuda:3', dtype=torch.float32)
276
+ arg5_1 = 2
277
+ arg6_1 = rand_strided((2, 1), (1, 1), device='cuda:3', dtype=torch.int64)
278
+ arg7_1 = 6144
279
+ arg8_1 = rand_strided((2, 6144), (6144, 1), device='cuda:3', dtype=torch.bfloat16)
280
+ arg9_1 = 1
281
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1, arg8_1, arg9_1]
282
+
283
+
284
+ def benchmark_compiled_module(args, times=10, repeat=10):
285
+ from torch._inductor.utils import print_performance
286
+ fn = lambda: call(list(args))
287
+ return print_performance(fn, times=times, repeat=repeat)
288
+
289
+
290
+ if __name__ == "__main__":
291
+ from torch._inductor.wrapper_benchmark import compiled_module_main
292
+ args = get_args()
293
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/4w/88c2a3ee4b3f9188d0485d1f58c19da2936d66802465a9e388f74d179d8ec54e.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 256, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 83, "triton_cache_hash": "K7P5VUQTWQ2GR5KCWBMAJUJLJ4ZOKI2HSLQM523UQ5RBJFUB5GQA"}
.inductor_cache/4w/c4wbe3nz76vrohi4tguyhvtuuwxq5ot6a6vjcubrww3al5ynrkzg.py ADDED
@@ -0,0 +1,26 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 256},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 3072}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_0(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
24
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
25
+ tmp2 = tmp0 * tmp1
26
+ tl.store(in_out_ptr0 + (x0), tmp2, xmask)
.inductor_cache/5d/c5dllvj5zhsrbv5yhs4lnnrfdo5t7vsfkgewb73jsq5ipppt5rbn.py ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tmp2 = ks0
26
+ tmp3 = tmp1 >= tmp2
27
+ tmp4 = ks1
28
+ tmp5 = tmp1 < tmp4
29
+ tmp6 = tmp3 & tmp5
30
+ tmp7 = ks2
31
+ tmp8 = tmp1 >= tmp7
32
+ tmp9 = ks3
33
+ tmp10 = tmp1 < tmp9
34
+ tmp11 = tmp8 & tmp10
35
+ tmp12 = tmp6 | tmp11
36
+ tmp13 = tmp12 == 0
37
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp13, None)
.inductor_cache/5f/c5fm757ws5k72edgygk6waomo4dh4sqhl7n4rzqsynjfqrzjugov.py ADDED
@@ -0,0 +1,150 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['3_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/6d/c6dqripf2w5j4sf6yps4zy7tc4ifqtp2wzcncsdxelw7lhj3o7ou.py
38
+ # Topologically Sorted Source Nodes: [weights_1, unsqueeze, mul_1], Original ATen: [aten.mul, aten.unsqueeze]
39
+ # Source node to ATen node mapping:
40
+ # mul_1 => mul_7
41
+ # unsqueeze => unsqueeze
42
+ # weights_1 => mul_2
43
+ # Graph fragment:
44
+ # %mm : Tensor "f32[s0, 32][32, 1]cuda:3" = PlaceHolder[target=mm]
45
+ # %arg3_1 : Tensor "f32[s0, 32, 1][32, 1, 1]cuda:3" = PlaceHolder[target=arg3_1]
46
+ # %arg4_1 : Tensor "f64[][]cpu" = PlaceHolder[target=arg4_1]
47
+ # %mul_2 : Tensor "f32[s0, 32][32, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mm, 0.1767766952966369), kwargs = {})
48
+ # %unsqueeze : Tensor "f32[s0, 32, 1][32, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%mul_2, -1), kwargs = {})
49
+ # %mul_7 : Tensor "f32[s0, 32, 1][32, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %arg3_1), kwargs = {})
50
+ # %convert_element_type_default_1 : Tensor "f32[][]cpu"[num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%arg4_1, torch.float32), kwargs = {})
51
+ # %mul_tensor : Tensor "f32[s0, 32, 1][32, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_7, %convert_element_type_default_1), kwargs = {})
52
+ # return %mul_tensor
53
+ triton_poi_fused_mul_unsqueeze_0 = async_compile.triton('triton_poi_fused_mul_unsqueeze_0', '''
54
+ import triton
55
+ import triton.language as tl
56
+
57
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
58
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
59
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
60
+ triton_helpers.set_driver_to_gpu()
61
+
62
+ @triton_heuristics.pointwise(
63
+ size_hints={'x': 2048},
64
+ filename=__file__,
65
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'in_ptr1': 'fp64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
66
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 28672}},
67
+ min_elem_per_thread=0
68
+ )
69
+ @triton.jit
70
+ def triton_poi_fused_mul_unsqueeze_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
71
+ xoffset = tl.program_id(0) * XBLOCK
72
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
73
+ xmask = xindex < xnumel
74
+ x0 = xindex
75
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
76
+ tmp3 = tl.load(in_ptr0 + (x0), xmask)
77
+ tmp5 = in_ptr1
78
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
79
+ tmp2 = tmp0 * tmp1
80
+ tmp4 = tmp2 * tmp3
81
+ tmp6 = tmp5.to(tl.float32)
82
+ tmp7 = tmp4 * tmp6
83
+ tl.store(in_out_ptr0 + (x0), tmp7, xmask)
84
+ ''', device_str='cuda')
85
+
86
+
87
+ async_compile.wait(globals())
88
+ del async_compile
89
+
90
+ class Runner:
91
+ def __init__(self, partitions):
92
+ self.partitions = partitions
93
+
94
+ def recursively_apply_fns(self, fns):
95
+ new_callables = []
96
+ for fn, c in zip(fns, self.partitions):
97
+ new_callables.append(fn(c))
98
+ self.partitions = new_callables
99
+
100
+ def call(self, args):
101
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1 = args
102
+ args.clear()
103
+ s77 = arg1_1
104
+ arg2_1_size = arg2_1.size()
105
+ s0 = arg2_1_size[0]
106
+ assert_size_stride(arg0_1, (32, 6144), (6144, 1))
107
+ assert_size_stride(arg2_1, (s0, 6144), (6144, 1))
108
+ assert_size_stride(arg3_1, (s0, 32, 1), (32, 1, 1))
109
+ assert_size_stride(arg4_1, (), ())
110
+ with torch.cuda._DeviceGuard(3):
111
+ torch.cuda.set_device(3)
112
+ buf0 = empty_strided_cuda((s0, 32), (32, 1), torch.float32)
113
+ # Topologically Sorted Source Nodes: [t, weights], Original ATen: [aten.t, aten.mm]
114
+ extern_kernels.mm_dtype(arg2_1, reinterpret_tensor(arg0_1, (6144, 32), (1, 6144), 0), out_dtype=torch.float32, out=buf0)
115
+ del arg0_1
116
+ del arg2_1
117
+ buf1 = reinterpret_tensor(buf0, (s0, 32, 1), (32, 1, 1), 0); del buf0 # reuse
118
+ # Topologically Sorted Source Nodes: [weights_1, unsqueeze, mul_1], Original ATen: [aten.mul, aten.unsqueeze]
119
+ triton_poi_fused_mul_unsqueeze_0_xnumel = 32*s0
120
+ stream3 = get_raw_stream(3)
121
+ triton_poi_fused_mul_unsqueeze_0.run(buf1, arg3_1, arg4_1.item(), triton_poi_fused_mul_unsqueeze_0_xnumel, stream=stream3)
122
+ del arg3_1
123
+ del arg4_1
124
+ return (buf1, )
125
+
126
+ runner = Runner(partitions=[])
127
+ call = runner.call
128
+ recursively_apply_fns = runner.recursively_apply_fns
129
+
130
+
131
+ def get_args():
132
+ from torch._dynamo.testing import rand_strided
133
+ arg0_1 = rand_strided((32, 6144), (6144, 1), device='cuda:3', dtype=torch.bfloat16)
134
+ arg1_1 = 56
135
+ arg2_1 = rand_strided((56, 6144), (6144, 1), device='cuda:3', dtype=torch.bfloat16)
136
+ arg3_1 = rand_strided((56, 32, 1), (32, 1, 1), device='cuda:3', dtype=torch.float32)
137
+ arg4_1 = rand_strided((), (), device='cpu', dtype=torch.float64)
138
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1]
139
+
140
+
141
+ def benchmark_compiled_module(args, times=10, repeat=10):
142
+ from torch._inductor.utils import print_performance
143
+ fn = lambda: call(list(args))
144
+ return print_performance(fn, times=times, repeat=repeat)
145
+
146
+
147
+ if __name__ == "__main__":
148
+ from torch._inductor.wrapper_benchmark import compiled_module_main
149
+ args = get_args()
150
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/5l/c5lhmqi55bmiwxrtdbnsnvjgutkpxgsd6sl2mrliaofj5ojwc2gs.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = ks0
24
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32)), tmp0, None)
.inductor_cache/5p/c5pplmatftc6r3hsuhx4skgsikwfs4k4t74mko7xfk7xp2sqleis.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 128},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'out_ptr1': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 2912}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, out_ptr1, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tmp1 = ks0
25
+ tmp2 = tmp0 >= tmp1
26
+ tmp3 = ks1
27
+ tmp4 = tmp0 < tmp3
28
+ tmp5 = tmp2 & tmp4
29
+ tmp6 = ks2
30
+ tmp7 = tmp0 >= tmp6
31
+ tmp8 = ks3
32
+ tmp9 = tmp0 < tmp8
33
+ tmp10 = tmp7 & tmp9
34
+ tmp11 = tmp5 | tmp10
35
+ tmp12 = tmp11.to(tl.int64)
36
+ tmp13 = tmp5.to(tl.int64)
37
+ tmp14 = tmp13 * tmp1
38
+ tmp15 = tmp10.to(tl.int64)
39
+ tmp16 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
40
+ tmp17 = tmp15 * tmp16
41
+ tmp18 = tmp14 + tmp17
42
+ tmp19 = tmp0 - tmp18
43
+ tmp20 = tmp12 * tmp19
44
+ tmp21 = tmp11 == 0
45
+ tl.store(out_ptr0 + (x0), tmp20, xmask)
46
+ tl.store(out_ptr1 + (x0), tmp21, xmask)
.inductor_cache/5q/c5qdusbmow5b3zrnnn4oh343d7oyptsndbeaenzknk5r4mhn7org.py ADDED
@@ -0,0 +1,148 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['1_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/bk/cbk2gyqxm7qwcf2a3uzdxgrzshzr3heruzlee42ped765ktbbc6f.py
38
+ # Topologically Sorted Source Nodes: [weights_1, unsqueeze, mul_1], Original ATen: [aten.mul, aten.unsqueeze]
39
+ # Source node to ATen node mapping:
40
+ # mul_1 => mul_7
41
+ # unsqueeze => unsqueeze
42
+ # weights_1 => mul_2
43
+ # Graph fragment:
44
+ # %mm : Tensor "f32[s77, 32][32, 1]cuda:7" = PlaceHolder[target=mm]
45
+ # %arg3_1 : Tensor "f32[s77, 32, 1][32, 1, 1]cuda:7" = PlaceHolder[target=arg3_1]
46
+ # %arg4_1 : Tensor "f64[][]cpu" = PlaceHolder[target=arg4_1]
47
+ # %mul_2 : Tensor "f32[s77, 32][32, 1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mm, 0.1767766952966369), kwargs = {})
48
+ # %unsqueeze : Tensor "f32[s77, 32, 1][32, 1, 1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%mul_2, -1), kwargs = {})
49
+ # %mul_7 : Tensor "f32[s77, 32, 1][32, 1, 1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %arg3_1), kwargs = {})
50
+ # %convert_element_type_default_1 : Tensor "f32[][]cpu"[num_users=1] = call_function[target=torch.ops.prims.convert_element_type.default](args = (%arg4_1, torch.float32), kwargs = {})
51
+ # %mul_tensor : Tensor "f32[s77, 32, 1][32, 1, 1]cuda:7"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%mul_7, %convert_element_type_default_1), kwargs = {})
52
+ # return %mul_tensor
53
+ triton_poi_fused_mul_unsqueeze_0 = async_compile.triton('triton_poi_fused_mul_unsqueeze_0', '''
54
+ import triton
55
+ import triton.language as tl
56
+
57
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
58
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
59
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
60
+ triton_helpers.set_driver_to_gpu()
61
+
62
+ @triton_heuristics.pointwise(
63
+ size_hints={'x': 4096},
64
+ filename=__file__,
65
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'in_ptr1': 'fp64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
66
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 57344}},
67
+ min_elem_per_thread=0
68
+ )
69
+ @triton.jit
70
+ def triton_poi_fused_mul_unsqueeze_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
71
+ xoffset = tl.program_id(0) * XBLOCK
72
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
73
+ xmask = xindex < xnumel
74
+ x0 = xindex
75
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
76
+ tmp3 = tl.load(in_ptr0 + (x0), xmask)
77
+ tmp5 = in_ptr1
78
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
79
+ tmp2 = tmp0 * tmp1
80
+ tmp4 = tmp2 * tmp3
81
+ tmp6 = tmp5.to(tl.float32)
82
+ tmp7 = tmp4 * tmp6
83
+ tl.store(in_out_ptr0 + (x0), tmp7, xmask)
84
+ ''', device_str='cuda')
85
+
86
+
87
+ async_compile.wait(globals())
88
+ del async_compile
89
+
90
+ class Runner:
91
+ def __init__(self, partitions):
92
+ self.partitions = partitions
93
+
94
+ def recursively_apply_fns(self, fns):
95
+ new_callables = []
96
+ for fn, c in zip(fns, self.partitions):
97
+ new_callables.append(fn(c))
98
+ self.partitions = new_callables
99
+
100
+ def call(self, args):
101
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1 = args
102
+ args.clear()
103
+ s77 = arg1_1
104
+ assert_size_stride(arg0_1, (32, 6144), (6144, 1))
105
+ assert_size_stride(arg2_1, (s77, 6144), (6144, 1))
106
+ assert_size_stride(arg3_1, (s77, 32, 1), (32, 1, 1))
107
+ assert_size_stride(arg4_1, (), ())
108
+ with torch.cuda._DeviceGuard(7):
109
+ torch.cuda.set_device(7)
110
+ buf0 = empty_strided_cuda((s77, 32), (32, 1), torch.float32)
111
+ # Topologically Sorted Source Nodes: [t, weights], Original ATen: [aten.t, aten.mm]
112
+ extern_kernels.mm_dtype(arg2_1, reinterpret_tensor(arg0_1, (6144, 32), (1, 6144), 0), out_dtype=torch.float32, out=buf0)
113
+ del arg0_1
114
+ del arg2_1
115
+ buf1 = reinterpret_tensor(buf0, (s77, 32, 1), (32, 1, 1), 0); del buf0 # reuse
116
+ # Topologically Sorted Source Nodes: [weights_1, unsqueeze, mul_1], Original ATen: [aten.mul, aten.unsqueeze]
117
+ triton_poi_fused_mul_unsqueeze_0_xnumel = 32*s77
118
+ stream7 = get_raw_stream(7)
119
+ triton_poi_fused_mul_unsqueeze_0.run(buf1, arg3_1, arg4_1.item(), triton_poi_fused_mul_unsqueeze_0_xnumel, stream=stream7)
120
+ del arg3_1
121
+ del arg4_1
122
+ return (buf1, )
123
+
124
+ runner = Runner(partitions=[])
125
+ call = runner.call
126
+ recursively_apply_fns = runner.recursively_apply_fns
127
+
128
+
129
+ def get_args():
130
+ from torch._dynamo.testing import rand_strided
131
+ arg0_1 = rand_strided((32, 6144), (6144, 1), device='cuda:7', dtype=torch.bfloat16)
132
+ arg1_1 = 112
133
+ arg2_1 = rand_strided((112, 6144), (6144, 1), device='cuda:7', dtype=torch.bfloat16)
134
+ arg3_1 = rand_strided((112, 32, 1), (32, 1, 1), device='cuda:7', dtype=torch.float32)
135
+ arg4_1 = rand_strided((), (), device='cpu', dtype=torch.float64)
136
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1]
137
+
138
+
139
+ def benchmark_compiled_module(args, times=10, repeat=10):
140
+ from torch._inductor.utils import print_performance
141
+ fn = lambda: call(list(args))
142
+ return print_performance(fn, times=times, repeat=repeat)
143
+
144
+
145
+ if __name__ == "__main__":
146
+ from torch._inductor.wrapper_benchmark import compiled_module_main
147
+ args = get_args()
148
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/5s/c5s4bm2q3gua3tgmxnj3denzqvuhpgrb5kx5rjwxxb4isc7lmqaz.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 2},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 8}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = ks0
24
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
.inductor_cache/5v/1eb21a9ba8eeb898c12aad6f0c7f6fb253af198de2a96f9596a41157018121fd.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 128, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 49, "triton_cache_hash": "XSMSLG6JQYYM2VVQRQ2PIUDTYEQTGTTF7NZPLRX5QSBQLLXPCRSA"}
.inductor_cache/5v/c5vsrxfkmbxmqe35ujjlxyr37blswcxqhqzlv2r6jal54yzc27hf.py ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 4096},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'in_ptr1': 'fp64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 57344}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
24
+ tmp3 = tl.load(in_ptr0 + (x0), xmask)
25
+ tmp5 = in_ptr1
26
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
27
+ tmp2 = tmp0 * tmp1
28
+ tmp4 = tmp2 * tmp3
29
+ tmp6 = tmp5.to(tl.float32)
30
+ tmp7 = tmp4 * tmp6
31
+ tl.store(in_out_ptr0 + (x0), tmp7, xmask)
.inductor_cache/5z/c5ztwyl3hjethlv2ot2kpnvlbaqjstzofdcjttbjwa7nkglx5gan.py ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tmp2 = ks0
26
+ tmp3 = tmp1 >= tmp2
27
+ tmp4 = ks1
28
+ tmp5 = tmp1 < tmp4
29
+ tmp6 = tmp3 & tmp5
30
+ tmp7 = ks2
31
+ tmp8 = tmp1 >= tmp7
32
+ tmp9 = ks3
33
+ tmp10 = tmp1 < tmp9
34
+ tmp11 = tmp8 & tmp10
35
+ tmp12 = tmp6 | tmp11
36
+ tmp13 = tmp12.to(tl.int64)
37
+ tmp14 = tmp6.to(tl.int64)
38
+ tmp15 = tmp14 * tmp2
39
+ tmp16 = tmp11.to(tl.int64)
40
+ tmp17 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
41
+ tmp18 = tmp16 * tmp17
42
+ tmp19 = tmp15 + tmp18
43
+ tmp20 = tmp1 - tmp19
44
+ tmp21 = tmp13 * tmp20
45
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp21, None)
.inductor_cache/62/c62vhrldvf7yqvelctjx64ts5ifznauctxiqhwo4jrqxe2rtpdgj.py ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tmp2 = ks0
26
+ tmp3 = tmp1 >= tmp2
27
+ tmp4 = ks1
28
+ tmp5 = tmp1 < tmp4
29
+ tmp6 = tmp3 & tmp5
30
+ tmp7 = ks2
31
+ tmp8 = tmp1 >= tmp7
32
+ tmp9 = ks3
33
+ tmp10 = tmp1 < tmp9
34
+ tmp11 = tmp8 & tmp10
35
+ tmp12 = tmp6 | tmp11
36
+ tmp13 = tmp12 == 0
37
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp13, None)
.inductor_cache/65/c156cdb4f80a03134d5d44666437b38d33a91660ba036c9600b2b2dad614b347.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 128, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 21, "triton_cache_hash": "Q2N5PO2YINGSELIO3M4VOOWLWHTNDJUQK2K7ETAZL7ECQK6GFLMA"}
.inductor_cache/65/c657qxcjufcctv3zrgtm2r3tsalyqcuqrzgzqct6qvtmivqg7fnt.py ADDED
@@ -0,0 +1,237 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['8_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/6e/c6emtbesfuldnxatnjaketbq4pjgfhpuh5rpsfrb6iagv4wjnfal.py
38
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, mul, mul_1, valid_offset, sub_3, input_], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.mul, aten.add, aten.sub]
39
+ # Source node to ATen node mapping:
40
+ # added_vocab_mask => bitwise_and_1
41
+ # ge => ge
42
+ # ge_1 => ge_1
43
+ # input_ => mul_2
44
+ # lt => lt
45
+ # lt_1 => lt_1
46
+ # mul => mul
47
+ # mul_1 => mul_1
48
+ # org_vocab_mask => bitwise_and
49
+ # sub_3 => sub_3
50
+ # valid_offset => add
51
+ # vocab_mask => bitwise_or
52
+ # Graph fragment:
53
+ # %arg0_1 : Tensor "i64[1][1]cuda:4" = PlaceHolder[target=arg0_1]
54
+ # %ge : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg0_1, %arg1_1), kwargs = {})
55
+ # %lt : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg0_1, %arg2_1), kwargs = {})
56
+ # %bitwise_and : Tensor "b8[1][1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge, %lt), kwargs = {})
57
+ # %ge_1 : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg0_1, %arg3_1), kwargs = {})
58
+ # %lt_1 : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg0_1, %arg4_1), kwargs = {})
59
+ # %bitwise_and_1 : Tensor "b8[1][1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge_1, %lt_1), kwargs = {})
60
+ # %bitwise_or : Tensor "b8[1][1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.bitwise_or.Tensor](args = (%bitwise_and, %bitwise_and_1), kwargs = {})
61
+ # %mul : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_and, %arg1_1), kwargs = {})
62
+ # %mul_1 : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_and_1, %sub_2), kwargs = {})
63
+ # %add : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %mul_1), kwargs = {})
64
+ # %sub_3 : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg0_1, %add), kwargs = {})
65
+ # %mul_2 : Tensor "i64[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_or, %sub_3), kwargs = {})
66
+ # return %mul_2
67
+ triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0 = async_compile.triton('triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0', '''
68
+ import triton
69
+ import triton.language as tl
70
+
71
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
72
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
73
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
74
+ triton_helpers.set_driver_to_gpu()
75
+
76
+ @triton_heuristics.pointwise(
77
+ size_hints={'x': 1},
78
+ filename=__file__,
79
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
80
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
81
+ min_elem_per_thread=0
82
+ )
83
+ @triton.jit
84
+ def triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
85
+ xnumel = 1
86
+ xoffset = tl.program_id(0) * XBLOCK
87
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
88
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
89
+ tmp0 = tl.load(in_ptr0 + (0))
90
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
91
+ tmp2 = ks0
92
+ tmp3 = tmp1 >= tmp2
93
+ tmp4 = ks1
94
+ tmp5 = tmp1 < tmp4
95
+ tmp6 = tmp3 & tmp5
96
+ tmp7 = ks2
97
+ tmp8 = tmp1 >= tmp7
98
+ tmp9 = ks3
99
+ tmp10 = tmp1 < tmp9
100
+ tmp11 = tmp8 & tmp10
101
+ tmp12 = tmp6 | tmp11
102
+ tmp13 = tmp12.to(tl.int64)
103
+ tmp14 = tmp6.to(tl.int64)
104
+ tmp15 = tmp14 * tmp2
105
+ tmp16 = tmp11.to(tl.int64)
106
+ tmp17 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
107
+ tmp18 = tmp16 * tmp17
108
+ tmp19 = tmp15 + tmp18
109
+ tmp20 = tmp1 - tmp19
110
+ tmp21 = tmp13 * tmp20
111
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp21, None)
112
+ ''', device_str='cuda')
113
+
114
+
115
+ # kernel path: /mnt/glm5_2/.inductor_cache/ck/cck7lywyewjkzmkn5tgb3zgvxscxs7idurr7myyeg2z5dguv75br.py
116
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, invert], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.bitwise_not]
117
+ # Source node to ATen node mapping:
118
+ # added_vocab_mask => bitwise_and_1
119
+ # ge => ge
120
+ # ge_1 => ge_1
121
+ # invert => bitwise_not
122
+ # lt => lt
123
+ # lt_1 => lt_1
124
+ # org_vocab_mask => bitwise_and
125
+ # vocab_mask => bitwise_or
126
+ # Graph fragment:
127
+ # %arg0_1 : Tensor "i64[1][1]cuda:4" = PlaceHolder[target=arg0_1]
128
+ # %ge : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg0_1, %arg1_1), kwargs = {})
129
+ # %lt : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg0_1, %arg2_1), kwargs = {})
130
+ # %bitwise_and : Tensor "b8[1][1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge, %lt), kwargs = {})
131
+ # %ge_1 : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg0_1, %arg3_1), kwargs = {})
132
+ # %lt_1 : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg0_1, %arg4_1), kwargs = {})
133
+ # %bitwise_and_1 : Tensor "b8[1][1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge_1, %lt_1), kwargs = {})
134
+ # %bitwise_or : Tensor "b8[1][1]cuda:4"[num_users=2] = call_function[target=torch.ops.aten.bitwise_or.Tensor](args = (%bitwise_and, %bitwise_and_1), kwargs = {})
135
+ # %bitwise_not : Tensor "b8[1][1]cuda:4"[num_users=1] = call_function[target=torch.ops.aten.bitwise_not.default](args = (%bitwise_or,), kwargs = {})
136
+ # return %bitwise_not
137
+ triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1 = async_compile.triton('triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1', '''
138
+ import triton
139
+ import triton.language as tl
140
+
141
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
142
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
143
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
144
+ triton_helpers.set_driver_to_gpu()
145
+
146
+ @triton_heuristics.pointwise(
147
+ size_hints={'x': 1},
148
+ filename=__file__,
149
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
150
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
151
+ min_elem_per_thread=0
152
+ )
153
+ @triton.jit
154
+ def triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, xnumel, XBLOCK : tl.constexpr):
155
+ xnumel = 1
156
+ xoffset = tl.program_id(0) * XBLOCK
157
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
158
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
159
+ tmp0 = tl.load(in_ptr0 + (0))
160
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
161
+ tmp2 = ks0
162
+ tmp3 = tmp1 >= tmp2
163
+ tmp4 = ks1
164
+ tmp5 = tmp1 < tmp4
165
+ tmp6 = tmp3 & tmp5
166
+ tmp7 = ks2
167
+ tmp8 = tmp1 >= tmp7
168
+ tmp9 = ks3
169
+ tmp10 = tmp1 < tmp9
170
+ tmp11 = tmp8 & tmp10
171
+ tmp12 = tmp6 | tmp11
172
+ tmp13 = tmp12 == 0
173
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp13, None)
174
+ ''', device_str='cuda')
175
+
176
+
177
+ async_compile.wait(globals())
178
+ del async_compile
179
+
180
+ class Runner:
181
+ def __init__(self, partitions):
182
+ self.partitions = partitions
183
+
184
+ def recursively_apply_fns(self, fns):
185
+ new_callables = []
186
+ for fn, c in zip(fns, self.partitions):
187
+ new_callables.append(fn(c))
188
+ self.partitions = new_callables
189
+
190
+ def call(self, args):
191
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1 = args
192
+ args.clear()
193
+ s90 = arg1_1
194
+ s87 = arg2_1
195
+ s68 = arg3_1
196
+ s53 = arg4_1
197
+ s88 = arg5_1
198
+ assert_size_stride(arg0_1, (1, ), (1, ))
199
+ with torch.cuda._DeviceGuard(4):
200
+ torch.cuda.set_device(4)
201
+ buf0 = empty_strided_cuda((1, ), (1, ), torch.int64)
202
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, mul, mul_1, valid_offset, sub_3, input_], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.mul, aten.add, aten.sub]
203
+ stream4 = get_raw_stream(4)
204
+ triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0.run(arg0_1, buf0, s90, s87, s68, s53, s88, 1, stream=stream4)
205
+ buf1 = empty_strided_cuda((1, ), (1, ), torch.bool)
206
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, invert], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.bitwise_not]
207
+ stream4 = get_raw_stream(4)
208
+ triton_poi_fused_bitwise_and_bitwise_not_bitwise_or_ge_lt_1.run(arg0_1, buf1, s90, s87, s68, s53, 1, stream=stream4)
209
+ del arg0_1
210
+ return (buf0, buf1, )
211
+
212
+ runner = Runner(partitions=[])
213
+ call = runner.call
214
+ recursively_apply_fns = runner.recursively_apply_fns
215
+
216
+
217
+ def get_args():
218
+ from torch._dynamo.testing import rand_strided
219
+ arg0_1 = rand_strided((1, ), (1, ), device='cuda:4', dtype=torch.int64)
220
+ arg1_1 = 77440
221
+ arg2_1 = 96800
222
+ arg3_1 = 154880
223
+ arg4_1 = 154880
224
+ arg5_1 = 0
225
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1]
226
+
227
+
228
+ def benchmark_compiled_module(args, times=10, repeat=10):
229
+ from torch._inductor.utils import print_performance
230
+ fn = lambda: call(list(args))
231
+ return print_performance(fn, times=times, repeat=repeat)
232
+
233
+
234
+ if __name__ == "__main__":
235
+ from torch._inductor.wrapper_benchmark import compiled_module_main
236
+ args = get_args()
237
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/65/c65lswedy4nxnsl65ku4a6ddu3r55grrzzjaxx6jvoroewo6e2dr.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 16384},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 73728}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x2 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last').to(tl.float32)
24
+ tl.store(out_ptr0 + (x2), tmp0, xmask)
.inductor_cache/67/c67bawm35pfjbqa2hnubkrofni4tkbxzyeo5zppu6gsylvqve6eo.py ADDED
@@ -0,0 +1,46 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 128},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'out_ptr1': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 2912}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, out_ptr1, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tmp1 = ks0
25
+ tmp2 = tmp0 >= tmp1
26
+ tmp3 = ks1
27
+ tmp4 = tmp0 < tmp3
28
+ tmp5 = tmp2 & tmp4
29
+ tmp6 = ks2
30
+ tmp7 = tmp0 >= tmp6
31
+ tmp8 = ks3
32
+ tmp9 = tmp0 < tmp8
33
+ tmp10 = tmp7 & tmp9
34
+ tmp11 = tmp5 | tmp10
35
+ tmp12 = tmp11.to(tl.int64)
36
+ tmp13 = tmp5.to(tl.int64)
37
+ tmp14 = tmp13 * tmp1
38
+ tmp15 = tmp10.to(tl.int64)
39
+ tmp16 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
40
+ tmp17 = tmp15 * tmp16
41
+ tmp18 = tmp14 + tmp17
42
+ tmp19 = tmp0 - tmp18
43
+ tmp20 = tmp12 * tmp19
44
+ tmp21 = tmp11 == 0
45
+ tl.store(out_ptr0 + (x0), tmp20, xmask)
46
+ tl.store(out_ptr1 + (x0), tmp21, xmask)
.inductor_cache/6d/c6dqripf2w5j4sf6yps4zy7tc4ifqtp2wzcncsdxelw7lhj3o7ou.py ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 2048},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'in_ptr1': 'fp64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 28672}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
24
+ tmp3 = tl.load(in_ptr0 + (x0), xmask)
25
+ tmp5 = in_ptr1
26
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
27
+ tmp2 = tmp0 * tmp1
28
+ tmp4 = tmp2 * tmp3
29
+ tmp6 = tmp5.to(tl.float32)
30
+ tmp7 = tmp4 * tmp6
31
+ tl.store(in_out_ptr0 + (x0), tmp7, xmask)
.inductor_cache/6d/e62944c5f38dab9ccf24b5504699965ed3bab66acfacc866085a1032c963177d.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 256, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 45, "triton_cache_hash": "KGGY5WAIFCN23ENKRQFNT2OZB6CJLZGZIIMD5KHX6XAEJ3DEY7VA"}
.inductor_cache/6e/c6emtbesfuldnxatnjaketbq4pjgfhpuh5rpsfrb6iagv4wjnfal.py ADDED
@@ -0,0 +1,45 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_bitwise_and_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = tl.load(in_ptr0 + (0))
24
+ tmp1 = tl.broadcast_to(tmp0, [XBLOCK])
25
+ tmp2 = ks0
26
+ tmp3 = tmp1 >= tmp2
27
+ tmp4 = ks1
28
+ tmp5 = tmp1 < tmp4
29
+ tmp6 = tmp3 & tmp5
30
+ tmp7 = ks2
31
+ tmp8 = tmp1 >= tmp7
32
+ tmp9 = ks3
33
+ tmp10 = tmp1 < tmp9
34
+ tmp11 = tmp8 & tmp10
35
+ tmp12 = tmp6 | tmp11
36
+ tmp13 = tmp12.to(tl.int64)
37
+ tmp14 = tmp6.to(tl.int64)
38
+ tmp15 = tmp14 * tmp2
39
+ tmp16 = tmp11.to(tl.int64)
40
+ tmp17 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
41
+ tmp18 = tmp16 * tmp17
42
+ tmp19 = tmp15 + tmp18
43
+ tmp20 = tmp1 - tmp19
44
+ tmp21 = tmp13 * tmp20
45
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32).broadcast_to(XBLOCK)), tmp21, None)
.inductor_cache/6j/c6j7f2zfedgsifsvrtdzvprpzw6jujih2uwmzsjrvvfsrmkhcotk.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 2},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=1, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 12}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
.inductor_cache/6p/28650630bd03810b78b7cf75754dfa35da9d5ed77d198e2f0ce9f078547bb750.best_config ADDED
@@ -0,0 +1 @@
 
 
1
+ {"XBLOCK": 256, "num_warps": 4, "num_stages": 1, "configs_hash": "1b2cc4dbebb9680d3ce31843331593b159e4046c056f195ca1ccf2464d5b37d1", "found_by_coordesc": false, "time_taken_ms": 43, "triton_cache_hash": "KGGY5WAIFCN23ENKRQFNT2OZB6CJLZGZIIMD5KHX6XAEJ3DEY7VA"}
.inductor_cache/6p/c6ppnyt6kjnrqeeq3ra4wx4kwkpeerqiq5ftck3hdmvntgpd6wcs.py ADDED
@@ -0,0 +1,31 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 4096},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'in_ptr1': 'fp64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=5, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': ['in_out_ptr0'], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 57344}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_out_ptr0, in_ptr0, in_ptr1, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_out_ptr0 + (x0), xmask)
24
+ tmp3 = tl.load(in_ptr0 + (x0), xmask)
25
+ tmp5 = in_ptr1
26
+ tmp1 = tl.full([1], 0.1767766952966369, tl.float32)
27
+ tmp2 = tmp0 * tmp1
28
+ tmp4 = tmp2 * tmp3
29
+ tmp6 = tmp5.to(tl.float32)
30
+ tmp7 = tmp4 * tmp6
31
+ tl.store(in_out_ptr0 + (x0), tmp7, xmask)
.inductor_cache/6r/c6rthxci3676wj2mwj62e5zmpmln5m2yrpnp2cb63gsxh4u2zxhz.py ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 2},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=4, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 12}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
25
+ tmp2 = tmp0 * tmp1
26
+ tl.store(out_ptr0 + (x0), tmp2, xmask)
27
+ tl.store(out_ptr1 + (x0), tmp2, xmask)
.inductor_cache/6u/c6uum37r4isrc4g6cyxmskeeex5jl3phh26m7xujcnuux4ijq3jw.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 1},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'constexpr', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=2, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {'xnumel': 1}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
19
+ xnumel = 1
20
+ xoffset = tl.program_id(0) * XBLOCK
21
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
22
+ xmask = tl.full([XBLOCK], True, tl.int1)[:]
23
+ tmp0 = ks0
24
+ tl.store(out_ptr0 + (tl.full([XBLOCK], 0, tl.int32)), tmp0, None)
.inductor_cache/6y/c6ygrpcdbuohowjvpdfmrangdf56jqj3p4ig3uwlgavr6t5kkb7a.py ADDED
@@ -0,0 +1,289 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['6_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/eu/ceujptddvohmrpmi3d35cfzrtagwszsxwklnkpos5z5fi3vn46sj.py
38
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
39
+ # Source node to ATen node mapping:
40
+ # expand_scores => mul_3
41
+ # flatten => view_1
42
+ # max_1 => getitem, max_1
43
+ # unsqueeze => unsqueeze
44
+ # view => view
45
+ # Graph fragment:
46
+ # %arg2_1 : Tensor "f32[s97, 1][1, 1]cuda:3" = PlaceHolder[target=arg2_1]
47
+ # %arg3_1 : Tensor "f32[s97, 1][1, 1]cuda:3" = PlaceHolder[target=arg3_1]
48
+ # %mul_3 : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:3" = PlaceHolder[target=mul_3]
49
+ # %unsqueeze : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%arg2_1, 2), kwargs = {})
50
+ # %view : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%arg3_1, [-1, 1, 1]), kwargs = {})
51
+ # %mul_3 : Tensor "f32[s97, 1, 1][1, 1, 1]cuda:3"[num_users=2] = call_function[target=torch.ops.aten.mul.Tensor](args = (%unsqueeze, %view), kwargs = {})
52
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
53
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
54
+ # %getitem : Tensor "f32[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=operator.getitem](args = (%max_1, 0), kwargs = {})
55
+ # return %mul_3,%getitem
56
+ triton_poi_fused_max_mul_unsqueeze_view_0 = async_compile.triton('triton_poi_fused_max_mul_unsqueeze_view_0', '''
57
+ import triton
58
+ import triton.language as tl
59
+
60
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
61
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
62
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
63
+ triton_helpers.set_driver_to_gpu()
64
+
65
+ @triton_heuristics.pointwise(
66
+ size_hints={'x': 8},
67
+ filename=__file__,
68
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'out_ptr0': '*fp32', 'out_ptr1': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
69
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_max_mul_unsqueeze_view_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 2, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
70
+ min_elem_per_thread=0
71
+ )
72
+ @triton.jit
73
+ def triton_poi_fused_max_mul_unsqueeze_view_0(in_ptr0, in_ptr1, out_ptr0, out_ptr1, xnumel, XBLOCK : tl.constexpr):
74
+ xoffset = tl.program_id(0) * XBLOCK
75
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
76
+ xmask = xindex < xnumel
77
+ x0 = xindex
78
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
79
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
80
+ tmp2 = tmp0 * tmp1
81
+ tl.store(out_ptr0 + (x0), tmp2, xmask)
82
+ tl.store(out_ptr1 + (x0), tmp2, xmask)
83
+ ''', device_str='cuda')
84
+
85
+
86
+ # kernel path: /mnt/glm5_2/.inductor_cache/jw/cjwnm4nhtti6dx4kp7rcgvjpxw7ab72wy3zgvwtnnqtlbtlmkzlu.py
87
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
88
+ # Source node to ATen node mapping:
89
+ # flatten => view_1
90
+ # gather => gather
91
+ # max_1 => max_1
92
+ # Graph fragment:
93
+ # %arg4_1 : Tensor "i64[s97, 1][1, 1]cuda:3" = PlaceHolder[target=arg4_1]
94
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
95
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
96
+ # %gather : Tensor "i64[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.gather.default](args = (%arg4_1, 1, %getitem_1), kwargs = {})
97
+ # return %gather
98
+ triton_poi_fused_gather_max_view_1 = async_compile.triton('triton_poi_fused_gather_max_view_1', '''
99
+ import triton
100
+ import triton.language as tl
101
+
102
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
103
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
104
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
105
+ triton_helpers.set_driver_to_gpu()
106
+
107
+ @triton_heuristics.pointwise(
108
+ size_hints={'x': 8},
109
+ filename=__file__,
110
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
111
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_gather_max_view_1', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 192}},
112
+ min_elem_per_thread=0
113
+ )
114
+ @triton.jit
115
+ def triton_poi_fused_gather_max_view_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
116
+ xoffset = tl.program_id(0) * XBLOCK
117
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
118
+ xmask = xindex < xnumel
119
+ x0 = xindex
120
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
121
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
122
+ ''', device_str='cuda')
123
+
124
+
125
+ # kernel path: /mnt/glm5_2/.inductor_cache/o6/co6hyomcl3npwceolu64q5bdvrb7h6qa4h3fycnlkntxzzmdzmue.py
126
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
127
+ # Source node to ATen node mapping:
128
+ # batch_offsets => iota
129
+ # flat_cs => view_4
130
+ # flatten => view_1
131
+ # floordiv => div
132
+ # hidden_states => index
133
+ # max_1 => max_1
134
+ # repeat_interleave => unsqueeze_1, view_5
135
+ # selected_input_index => add_47
136
+ # Graph fragment:
137
+ # %arg6_1 : Tensor "bf16[s97, s87][s87, 1]cuda:3" = PlaceHolder[target=arg6_1]
138
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
139
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
140
+ # %view_4 : Tensor "i64[s97][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%getitem_1, [%arg1_1]), kwargs = {})
141
+ # %div : Tensor "i64[s97][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.div.Tensor_mode](args = (%view_4, 1), kwargs = {rounding_mode: floor})
142
+ # %iota : Tensor "i64[s97][1]cuda:3"[num_users=1] = call_function[target=torch.ops.prims.iota.default](args = (%arg1_1,), kwargs = {start: 0, step: 1, dtype: torch.int64, device: cuda:3, requires_grad: False})
143
+ # %unsqueeze_1 : Tensor "i64[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.unsqueeze.default](args = (%iota, 1), kwargs = {})
144
+ # %view_5 : Tensor "i64[s97][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%expand, [%arg1_1]), kwargs = {})
145
+ # %add_47 : Tensor "i64[s97][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%div, %view_5), kwargs = {})
146
+ # %index : Tensor "bf16[s97, s87][s87, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.index.Tensor](args = (%arg6_1, [%add_47]), kwargs = {})
147
+ # return %index
148
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2 = async_compile.triton('triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', '''
149
+ import triton
150
+ import triton.language as tl
151
+
152
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
153
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
154
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
155
+ triton_helpers.set_driver_to_gpu()
156
+
157
+ @triton_heuristics.pointwise(
158
+ size_hints={'x': 65536},
159
+ filename=__file__,
160
+ triton_meta={'signature': {'in_ptr0': '*bf16', 'out_ptr0': '*bf16', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]]}]},
161
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 294912}},
162
+ min_elem_per_thread=0
163
+ )
164
+ @triton.jit
165
+ def triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):
166
+ xoffset = tl.program_id(0) * XBLOCK
167
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
168
+ xmask = xindex < xnumel
169
+ x2 = xindex
170
+ tmp0 = tl.load(in_ptr0 + (x2), xmask, eviction_policy='evict_last').to(tl.float32)
171
+ tl.store(out_ptr0 + (x2), tmp0, xmask)
172
+ ''', device_str='cuda')
173
+
174
+
175
+ # kernel path: /mnt/glm5_2/.inductor_cache/dm/cdmvkduxzdqrvvphvyrynpwm77ulv5m44qrmm2izq36cx2yge35l.py
176
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
177
+ # Source node to ATen node mapping:
178
+ # add_2 => add_53
179
+ # flatten => view_1
180
+ # max_1 => max_1
181
+ # Graph fragment:
182
+ # %view_1 : Tensor "f32[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.reshape.default](args = (%mul_3, [%arg1_1, 1]), kwargs = {})
183
+ # %max_1 : [num_users=2] = call_function[target=torch.ops.aten.max.dim](args = (%view_1, -1, True), kwargs = {})
184
+ # %add_53 : Tensor "i64[s97, 1][1, 1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%getitem_1, %arg7_1), kwargs = {})
185
+ # return %add_53
186
+ triton_poi_fused_add_max_view_3 = async_compile.triton('triton_poi_fused_add_max_view_3', '''
187
+ import triton
188
+ import triton.language as tl
189
+
190
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
191
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
192
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
193
+ triton_helpers.set_driver_to_gpu()
194
+
195
+ @triton_heuristics.pointwise(
196
+ size_hints={'x': 8},
197
+ filename=__file__,
198
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
199
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 128}},
200
+ min_elem_per_thread=0
201
+ )
202
+ @triton.jit
203
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
204
+ xoffset = tl.program_id(0) * XBLOCK
205
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
206
+ xmask = xindex < xnumel
207
+ x0 = xindex
208
+ tmp0 = ks0
209
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
210
+ ''', device_str='cuda')
211
+
212
+
213
+ async_compile.wait(globals())
214
+ del async_compile
215
+
216
+ class Runner:
217
+ def __init__(self, partitions):
218
+ self.partitions = partitions
219
+
220
+ def recursively_apply_fns(self, fns):
221
+ new_callables = []
222
+ for fn, c in zip(fns, self.partitions):
223
+ new_callables.append(fn(c))
224
+ self.partitions = new_callables
225
+
226
+ def call(self, args):
227
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1 = args
228
+ args.clear()
229
+ s53 = arg0_1
230
+ s97 = arg1_1
231
+ s87 = arg5_1
232
+ s77 = arg7_1
233
+ assert_size_stride(arg2_1, (s97, 1), (1, 1))
234
+ assert_size_stride(arg3_1, (s97, 1), (1, 1))
235
+ assert_size_stride(arg4_1, (s97, 1), (1, 1))
236
+ assert_size_stride(arg6_1, (s97, s87), (s87, 1))
237
+ with torch.cuda._DeviceGuard(3):
238
+ torch.cuda.set_device(3)
239
+ buf0 = empty_strided_cuda((s97, 1, 1), (1, 1, 1), torch.float32)
240
+ buf1 = empty_strided_cuda((s97, 1), (1, 1), torch.float32)
241
+ # Topologically Sorted Source Nodes: [unsqueeze, view, expand_scores, flatten, max_1], Original ATen: [aten.unsqueeze, aten.view, aten.mul, aten.max]
242
+ stream3 = get_raw_stream(3)
243
+ triton_poi_fused_max_mul_unsqueeze_view_0.run(arg2_1, arg3_1, buf0, buf1, s97, stream=stream3)
244
+ del arg2_1
245
+ del arg3_1
246
+ buf2 = empty_strided_cuda((s97, 1), (1, 1), torch.int64)
247
+ # Topologically Sorted Source Nodes: [flatten, max_1, gather], Original ATen: [aten.view, aten.max, aten.gather]
248
+ stream3 = get_raw_stream(3)
249
+ triton_poi_fused_gather_max_view_1.run(arg4_1, buf2, s97, stream=stream3)
250
+ buf3 = empty_strided_cuda((s97, s87), (s87, 1), torch.bfloat16)
251
+ # Topologically Sorted Source Nodes: [flatten, max_1, flat_cs, floordiv, batch_offsets, repeat_interleave, selected_input_index, hidden_states], Original ATen: [aten.view, aten.max, aten.floor_divide, aten.arange, aten.unsqueeze, aten.add, aten.index]
252
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel = s87*s97
253
+ stream3 = get_raw_stream(3)
254
+ triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2.run(arg6_1, buf3, triton_poi_fused_add_arange_floor_divide_index_max_unsqueeze_view_2_xnumel, stream=stream3)
255
+ del arg6_1
256
+ buf4 = empty_strided_cuda((s97, 1), (1, 1), torch.int64)
257
+ # Topologically Sorted Source Nodes: [flatten, max_1, add_2], Original ATen: [aten.view, aten.max, aten.add]
258
+ stream3 = get_raw_stream(3)
259
+ triton_poi_fused_add_max_view_3.run(buf4, s77, s97, stream=stream3)
260
+ return (reinterpret_tensor(buf2, (s97, ), (1, ), 0), buf3, buf1, buf0, arg4_1, buf4, )
261
+
262
+ runner = Runner(partitions=[])
263
+ call = runner.call
264
+ recursively_apply_fns = runner.recursively_apply_fns
265
+
266
+
267
+ def get_args():
268
+ from torch._dynamo.testing import rand_strided
269
+ arg0_1 = 1
270
+ arg1_1 = 8
271
+ arg2_1 = rand_strided((8, 1), (1, 1), device='cuda:3', dtype=torch.float32)
272
+ arg3_1 = rand_strided((8, 1), (1, 1), device='cuda:3', dtype=torch.float32)
273
+ arg4_1 = rand_strided((8, 1), (1, 1), device='cuda:3', dtype=torch.int64)
274
+ arg5_1 = 6144
275
+ arg6_1 = rand_strided((8, 6144), (6144, 1), device='cuda:3', dtype=torch.bfloat16)
276
+ arg7_1 = 1
277
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1, arg7_1]
278
+
279
+
280
+ def benchmark_compiled_module(args, times=10, repeat=10):
281
+ from torch._inductor.utils import print_performance
282
+ fn = lambda: call(list(args))
283
+ return print_performance(fn, times=times, repeat=repeat)
284
+
285
+
286
+ if __name__ == "__main__":
287
+ from torch._inductor.wrapper_benchmark import compiled_module_main
288
+ args = get_args()
289
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))
.inductor_cache/6z/c6zyuqh3bo3nxnmo267egfmzuh75yqhd4pdyfrvq5jc5nfexlxsk.py ADDED
@@ -0,0 +1,29 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 32},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'in_ptr0': '*fp32', 'in_ptr1': '*fp32', 'in_ptr2': 'fp64', 'out_ptr0': '*fp32', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=1, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (3,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_mul_unsqueeze_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 3, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 512}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_mul_unsqueeze_0(in_ptr0, in_ptr1, in_ptr2, out_ptr0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
24
+ tmp1 = tl.load(in_ptr1 + (x0), xmask)
25
+ tmp3 = in_ptr2
26
+ tmp2 = tmp0 * tmp1
27
+ tmp4 = tmp3.to(tl.float32)
28
+ tmp5 = tmp2 * tmp4
29
+ tl.store(out_ptr0 + (x0), tmp5, xmask)
.inductor_cache/76/c76y7ge3o24p7dwc6aulba5id74axw62oevitux7yjjlbrbsx66u.py ADDED
@@ -0,0 +1,24 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+
2
+ import triton
3
+ import triton.language as tl
4
+
5
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
6
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
7
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
8
+ triton_helpers.set_driver_to_gpu()
9
+
10
+ @triton_heuristics.pointwise(
11
+ size_hints={'x': 8},
12
+ filename=__file__,
13
+ triton_meta={'signature': {'out_ptr0': '*i64', 'ks0': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=7, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]]}]},
14
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_max_view_3', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 0, 'num_store': 1, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 128}},
15
+ min_elem_per_thread=0
16
+ )
17
+ @triton.jit
18
+ def triton_poi_fused_add_max_view_3(out_ptr0, ks0, xnumel, XBLOCK : tl.constexpr):
19
+ xoffset = tl.program_id(0) * XBLOCK
20
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
21
+ xmask = xindex < xnumel
22
+ x0 = xindex
23
+ tmp0 = ks0
24
+ tl.store(out_ptr0 + (x0), tmp0, xmask)
.inductor_cache/7e/c7eeblfcfg2xdo34xdor3snhltmuxl4hgqzougfrenohmqtj4r3r.py ADDED
@@ -0,0 +1,177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # AOT ID: ['0_inference']
2
+ from ctypes import c_void_p, c_long, c_int
3
+ import torch
4
+ import math
5
+ import random
6
+ import os
7
+ import tempfile
8
+ from math import inf, nan
9
+ from cmath import nanj
10
+ from torch._inductor.hooks import run_intermediate_hooks
11
+ from torch._inductor.utils import maybe_profile
12
+ from torch._inductor.codegen.memory_planning import _align as align
13
+ from torch import device, empty_strided
14
+ from torch._inductor.async_compile import AsyncCompile
15
+ from torch._inductor.select_algorithm import extern_kernels
16
+ import triton
17
+ import triton.language as tl
18
+ from torch._inductor.runtime.triton_heuristics import start_graph, end_graph
19
+ from torch._C import _cuda_getCurrentRawStream as get_raw_stream
20
+
21
+ aten = torch.ops.aten
22
+ inductor_ops = torch.ops.inductor
23
+ _quantized = torch.ops._quantized
24
+ assert_size_stride = torch._C._dynamo.guards.assert_size_stride
25
+ assert_alignment = torch._C._dynamo.guards.assert_alignment
26
+ empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
27
+ empty_strided_cpu_pinned = torch._C._dynamo.guards._empty_strided_cpu_pinned
28
+ empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
29
+ empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
30
+ empty_strided_mtia = torch._C._dynamo.guards._empty_strided_mtia
31
+ reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
32
+ alloc_from_pool = torch.ops.inductor._alloc_from_pool
33
+ async_compile = AsyncCompile()
34
+ empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p
35
+
36
+
37
+ # kernel path: /mnt/glm5_2/.inductor_cache/5p/c5pplmatftc6r3hsuhx4skgsikwfs4k4t74mko7xfk7xp2sqleis.py
38
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, mul, mul_1, valid_offset, sub_3, input_, invert], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.mul, aten.add, aten.sub, aten.bitwise_not]
39
+ # Source node to ATen node mapping:
40
+ # added_vocab_mask => bitwise_and_1
41
+ # ge => ge
42
+ # ge_1 => ge_1
43
+ # input_ => mul_6
44
+ # invert => bitwise_not
45
+ # lt => lt
46
+ # lt_1 => lt_1
47
+ # mul => mul
48
+ # mul_1 => mul_2
49
+ # org_vocab_mask => bitwise_and
50
+ # sub_3 => sub_13
51
+ # valid_offset => add_16
52
+ # vocab_mask => bitwise_or
53
+ # Graph fragment:
54
+ # %arg1_1 : Tensor "i64[s15][1]cuda:3" = PlaceHolder[target=arg1_1]
55
+ # %ge : Tensor "b8[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg1_1, %arg2_1), kwargs = {})
56
+ # %lt : Tensor "b8[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg1_1, %arg3_1), kwargs = {})
57
+ # %bitwise_and : Tensor "b8[s15][1]cuda:3"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge, %lt), kwargs = {})
58
+ # %ge_1 : Tensor "b8[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.ge.Scalar](args = (%arg1_1, %arg4_1), kwargs = {})
59
+ # %lt_1 : Tensor "b8[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.lt.Scalar](args = (%arg1_1, %arg5_1), kwargs = {})
60
+ # %bitwise_and_1 : Tensor "b8[s15][1]cuda:3"[num_users=2] = call_function[target=torch.ops.aten.bitwise_and.Tensor](args = (%ge_1, %lt_1), kwargs = {})
61
+ # %bitwise_or : Tensor "b8[s15][1]cuda:3"[num_users=2] = call_function[target=torch.ops.aten.bitwise_or.Tensor](args = (%bitwise_and, %bitwise_and_1), kwargs = {})
62
+ # %mul : Tensor "i64[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_and, %arg2_1), kwargs = {})
63
+ # %mul_2 : Tensor "i64[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_and_1, %sub_8), kwargs = {})
64
+ # %add_16 : Tensor "i64[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.add.Tensor](args = (%mul, %mul_2), kwargs = {})
65
+ # %sub_13 : Tensor "i64[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.sub.Tensor](args = (%arg1_1, %add_16), kwargs = {})
66
+ # %mul_6 : Tensor "i64[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.mul.Tensor](args = (%bitwise_or, %sub_13), kwargs = {})
67
+ # %bitwise_not : Tensor "b8[s15][1]cuda:3"[num_users=1] = call_function[target=torch.ops.aten.bitwise_not.default](args = (%bitwise_or,), kwargs = {})
68
+ # return %mul_6,%bitwise_not
69
+ triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0 = async_compile.triton('triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0', '''
70
+ import triton
71
+ import triton.language as tl
72
+
73
+ from torch._inductor.runtime import triton_helpers, triton_heuristics
74
+ from torch._inductor.runtime.triton_helpers import libdevice, math as tl_math
75
+ from torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties
76
+ triton_helpers.set_driver_to_gpu()
77
+
78
+ @triton_heuristics.pointwise(
79
+ size_hints={'x': 128},
80
+ filename=__file__,
81
+ triton_meta={'signature': {'in_ptr0': '*i64', 'out_ptr0': '*i64', 'out_ptr1': '*i1', 'ks0': 'i64', 'ks1': 'i64', 'ks2': 'i64', 'ks3': 'i64', 'ks4': 'i64', 'xnumel': 'i32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=3, multi_processor_count=148, cc=100, major=10, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, max_threads_per_block=1024, warp_size=32), 'constants': {}, 'native_matmul': False, 'enable_fp_fusion': True, 'launch_pdl': False, 'disable_ftz': False, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]},
82
+ inductor_meta={'grid_type': 'Grid1D', 'autotune_hints': set(), 'kernel_name': 'triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'atomic_add_found': False, 'num_load': 1, 'num_store': 2, 'num_reduction': 0, 'backend_hash': 'B9347ED8E82563C0DF6E4C45E8A40BBC93F818749C1452A3CD0A7365095FE2BD', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': True, 'autotune_remote_cache': None, '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, 'deterministic': False, 'force_filter_reduction_configs': False, 'mix_order_reduction_allow_multi_stages': False, 'are_deterministic_algorithms_enabled': False, 'tiling_scores': {'x': 2912}},
83
+ min_elem_per_thread=0
84
+ )
85
+ @triton.jit
86
+ def triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0(in_ptr0, out_ptr0, out_ptr1, ks0, ks1, ks2, ks3, ks4, xnumel, XBLOCK : tl.constexpr):
87
+ xoffset = tl.program_id(0) * XBLOCK
88
+ xindex = xoffset + tl.arange(0, XBLOCK)[:]
89
+ xmask = xindex < xnumel
90
+ x0 = xindex
91
+ tmp0 = tl.load(in_ptr0 + (x0), xmask)
92
+ tmp1 = ks0
93
+ tmp2 = tmp0 >= tmp1
94
+ tmp3 = ks1
95
+ tmp4 = tmp0 < tmp3
96
+ tmp5 = tmp2 & tmp4
97
+ tmp6 = ks2
98
+ tmp7 = tmp0 >= tmp6
99
+ tmp8 = ks3
100
+ tmp9 = tmp0 < tmp8
101
+ tmp10 = tmp7 & tmp9
102
+ tmp11 = tmp5 | tmp10
103
+ tmp12 = tmp11.to(tl.int64)
104
+ tmp13 = tmp5.to(tl.int64)
105
+ tmp14 = tmp13 * tmp1
106
+ tmp15 = tmp10.to(tl.int64)
107
+ tmp16 = ks0 + ks2 + ((-1)*ks1) + ((-1)*ks4)
108
+ tmp17 = tmp15 * tmp16
109
+ tmp18 = tmp14 + tmp17
110
+ tmp19 = tmp0 - tmp18
111
+ tmp20 = tmp12 * tmp19
112
+ tmp21 = tmp11 == 0
113
+ tl.store(out_ptr0 + (x0), tmp20, xmask)
114
+ tl.store(out_ptr1 + (x0), tmp21, xmask)
115
+ ''', device_str='cuda')
116
+
117
+
118
+ async_compile.wait(globals())
119
+ del async_compile
120
+
121
+ class Runner:
122
+ def __init__(self, partitions):
123
+ self.partitions = partitions
124
+
125
+ def recursively_apply_fns(self, fns):
126
+ new_callables = []
127
+ for fn, c in zip(fns, self.partitions):
128
+ new_callables.append(fn(c))
129
+ self.partitions = new_callables
130
+
131
+ def call(self, args):
132
+ arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1 = args
133
+ args.clear()
134
+ s15 = arg0_1
135
+ s90 = arg2_1
136
+ s87 = arg3_1
137
+ s68 = arg4_1
138
+ s53 = arg5_1
139
+ s88 = arg6_1
140
+ assert_size_stride(arg1_1, (s15, ), (1, ))
141
+ with torch.cuda._DeviceGuard(3):
142
+ torch.cuda.set_device(3)
143
+ buf0 = empty_strided_cuda((s15, ), (1, ), torch.int64)
144
+ buf1 = empty_strided_cuda((s15, ), (1, ), torch.bool)
145
+ # Topologically Sorted Source Nodes: [ge, lt, org_vocab_mask, ge_1, lt_1, added_vocab_mask, vocab_mask, mul, mul_1, valid_offset, sub_3, input_, invert], Original ATen: [aten.ge, aten.lt, aten.bitwise_and, aten.bitwise_or, aten.mul, aten.add, aten.sub, aten.bitwise_not]
146
+ stream3 = get_raw_stream(3)
147
+ triton_poi_fused_add_bitwise_and_bitwise_not_bitwise_or_ge_lt_mul_sub_0.run(arg1_1, buf0, buf1, s90, s87, s68, s53, s88, s15, stream=stream3)
148
+ del arg1_1
149
+ return (buf0, buf1, )
150
+
151
+ runner = Runner(partitions=[])
152
+ call = runner.call
153
+ recursively_apply_fns = runner.recursively_apply_fns
154
+
155
+
156
+ def get_args():
157
+ from torch._dynamo.testing import rand_strided
158
+ arg0_1 = 112
159
+ arg1_1 = rand_strided((112, ), (1, ), device='cuda:3', dtype=torch.int64)
160
+ arg2_1 = 58080
161
+ arg3_1 = 77440
162
+ arg4_1 = 154880
163
+ arg5_1 = 154880
164
+ arg6_1 = 0
165
+ return [arg0_1, arg1_1, arg2_1, arg3_1, arg4_1, arg5_1, arg6_1]
166
+
167
+
168
+ def benchmark_compiled_module(args, times=10, repeat=10):
169
+ from torch._inductor.utils import print_performance
170
+ fn = lambda: call(list(args))
171
+ return print_performance(fn, times=times, repeat=repeat)
172
+
173
+
174
+ if __name__ == "__main__":
175
+ from torch._inductor.wrapper_benchmark import compiled_module_main
176
+ args = get_args()
177
+ compiled_module_main('None', lambda times, repeat: benchmark_compiled_module(args, times=times, repeat=repeat))