id int64 2.74B 3.05B | title stringlengths 1 255 | user stringlengths 2 26 | state stringclasses 2 values | labels listlengths 0 24 | comments int64 0 206 | author_association stringclasses 4 values | body stringlengths 7 62.5k ⌀ | is_title bool 1 class |
|---|---|---|---|---|---|---|---|---|
2,818,126,325 | DISABLED test_return_tuple (__main__.TestScript) | pytorch-bot[bot] | closed | [
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: dynamo"
] | 1 | NONE | Platforms: dynamo
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_return_tuple&suite=TestScript&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/36340322722).
Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes.
**Debugging instructions (after clicking on the recent samples link):**
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
1. Click on the workflow logs linked above
2. Click on the Test step of the job so that it is expanded. Otherwise, the grepping will not work.
3. Grep for `test_return_tuple`
4. There should be several instances run (as flaky tests are rerun in CI) from which you can study the logs.
<details><summary>Sample error message</summary>
```
Traceback (most recent call last):
File "/var/lib/jenkins/workspace/test/test_jit.py", line 11331, in test_return_tuple
def test_return_tuple(self):
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/testing/_internal/jit_utils.py", line 483, in checkScript
source = textwrap.dedent(inspect.getsource(script))
~~~~~~~~~~~~~~~~~^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/inspect.py", line 1256, in getsource
lines, lnum = getsourcelines(object)
~~~~~~~~~~~~~~^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/inspect.py", line 1238, in getsourcelines
lines, lnum = findsource(object)
~~~~~~~~~~^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/inspect.py", line 1074, in findsource
lines = linecache.getlines(file, module.__dict__)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1400, in __call__
return self._torchdynamo_orig_callable(
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^
frame, cache_entry, self.hooks, frame_state, skip=1
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1184, in __call__
result = self._inner_convert(
frame, cache_entry, hooks, frame_state, skip=skip + 1
)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 565, in __call__
return _compile(
frame.f_code,
...<14 lines>...
skip=skip + 1,
)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 1048, in _compile
raise InternalTorchDynamoError(
f"{type(e).__qualname__}: {str(e)}"
).with_traceback(e.__traceback__) from None
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 997, in _compile
guarded_code = compile_inner(code, one_graph, hooks, transform)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_utils_internal.py", line 95, in wrapper_function
return function(*args, **kwargs)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 726, in compile_inner
return _compile_inner(code, one_graph, hooks, transform)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 760, in _compile_inner
out_code = transform_code_object(code, transform)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/bytecode_transformation.py", line 1404, in transform_code_object
transformations(instructions, code_options)
~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 236, in _fn
return fn(*args, **kwargs)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/convert_frame.py", line 680, in transform
tracer.run()
~~~~~~~~~~^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 2906, in run
super().run()
~~~~~~~~~~~^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 1078, in run
while self.step():
~~~~~~~~~^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 988, in step
self.dispatch_table[inst.opcode](self, inst)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 3087, in RETURN_VALUE
self._return(inst)
~~~~~~~~~~~~^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 3062, in _return
and not self.symbolic_locals_contain_module_class()
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/symbolic_convert.py", line 3051, in symbolic_locals_contain_module_class
if isinstance(v, UserDefinedClassVariable) and issubclass(
~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/base.py", line 191, in __instancecheck__
instance = instance.realize()
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/lazy.py", line 67, in realize
self._cache.realize()
~~~~~~~~~~~~~~~~~~~^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/lazy.py", line 33, in realize
self.vt = VariableTracker.build(tx, self.value, source)
~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/base.py", line 454, in build
return builder.VariableBuilder(tx, source)(value)
~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/builder.py", line 385, in __call__
vt = self._wrap(value)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/builder.py", line 620, in _wrap
result = dict(
build_key_value(i, k, v)
for i, (k, v) in enumerate(get_items_from_dict(value))
)
File "/opt/conda/envs/py_3.13/lib/python3.13/site-packages/torch/_dynamo/variables/builder.py", line 622, in <genexpr>
for i, (k, v) in enumerate(get_items_from_dict(value))
~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
torch._dynamo.exc.InternalTorchDynamoError: RuntimeError: dictionary changed size during iteration
from user code:
File "/opt/conda/envs/py_3.13/lib/python3.13/linecache.py", line 38, in getlines
return cache[filename][2]
Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information
You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True
To execute this test, run the following from the base repo dir:
PYTORCH_TEST_WITH_DYNAMO=1 python test/test_jit.py TestScript.test_return_tuple
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```
</details>
Test file path: `test_jit.py`
ResponseTimeoutError: Response timeout for 5000ms, GET https://raw.githubusercontent.com/pytorch/pytorch/main/test/test_jit.py -1 (connected: true, keepalive socket: false, socketHandledRequests: 1, socketHandledResponses: 0)
headers: {}
cc @clee2000 @wdvr @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @amjames | true |
2,818,020,380 | Publish pytorch RC docker images before release | bhack | closed | [
"module: binaries",
"oncall: releng",
"triaged",
"module: docker"
] | 6 | CONTRIBUTOR | ### 🚀 The feature, motivation and pitch
I think that we need at least have RC images published before we do the final image release as currently users can test only the final release:
https://hub.docker.com/r/pytorch/pytorch
### Alternatives
_No response_
### Additional context
_No response_
cc @seemethere @malfet @osalpekar @atalman | true |
2,817,692,954 | DISABLED test_dtensor_seq_par_shard_dim_0 (__main__.MicroPipelineTPTest) | pytorch-bot[bot] | open | [
"triaged",
"module: flaky-tests",
"skipped",
"module: c10d"
] | 4 | NONE | Platforms: linux
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_dtensor_seq_par_shard_dim_0&suite=MicroPipelineTPTest&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/36332698978).
Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes.
**Debugging instructions (after clicking on the recent samples link):**
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
1. Click on the workflow logs linked above
2. Click on the Test step of the job so that it is expanded. Otherwise, the grepping will not work.
3. Grep for `test_dtensor_seq_par_shard_dim_0`
4. There should be several instances run (as flaky tests are rerun in CI) from which you can study the logs.
<details><summary>Sample error message</summary>
```
Traceback (most recent call last):
File "/var/lib/jenkins/pytorch/test/distributed/tensor/parallel/test_micro_pipeline_tp.py", line 425, in test_dtensor_seq_par
self.assertIn("fused_all_gather_matmul", code)
File "/opt/conda/envs/py_3.10/lib/python3.10/unittest/case.py", line 1112, in assertIn
self.fail(self._formatMessage(msg, standardMsg))
File "/opt/conda/envs/py_3.10/lib/python3.10/unittest/case.py", line 675, in fail
raise self.failureException(msg)
AssertionError: 'fused_all_gather_matmul' not found in '# AOT ID: [\'0_forward\']\nfrom ctypes import c_void_p, c_long, c_int\nimport torch\nimport math\nimport random\nimport os\nimport tempfile\nfrom math import inf, nan\nfrom cmath import nanj\nfrom torch._inductor.hooks import run_intermediate_hooks\nfrom torch._inductor.utils import maybe_profile\nfrom torch._inductor.codegen.memory_planning import _align as align\nfrom torch import device, empty_strided\nfrom torch._inductor.async_compile import AsyncCompile\nfrom torch._inductor.select_algorithm import extern_kernels\nfrom torch._inductor.codegen.multi_kernel import MultiKernelCall\nimport triton\nimport triton.language as tl\nfrom torch._inductor.runtime.triton_heuristics import (\n grid,\n split_scan_grid,\n grid_combo_kernels,\n start_graph,\n end_graph,\n cooperative_reduction_grid,\n)\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\nfrom torch._C import _cuda_getCurrentRawStream as get_raw_stream\n\naten = torch.ops.aten\ninductor_ops = torch.ops.inductor\n_quantized = torch.ops._quantized\nassert_size_stride = torch._C._dynamo.guards.assert_size_stride\nempty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu\nempty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda\nempty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu\nreinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor\nalloc_from_pool = torch.ops.inductor._alloc_from_pool\nasync_compile = AsyncCompile()\nempty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p\n\n\n# kernel path: /tmp/tmp4rsnqzr5/sd/csdkwmy2wmyqm3e7apig34zpheymhb5ytu7whjukqo6ftyrvn4wx.py\n# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]\n# Source node to ATen node mapping:\n# linear => constant_pad_nd_default\n# Graph fragment:\n# %constant_pad_nd_default : [num_users=1] = call_function[target=torch.ops.aten.constant_pad_nd.default](args = (%wait_tensor, [0, 2, 0, 0]), kwargs = {})\ntriton_poi_fused_mm_0 = async_compile.triton(\'triton_poi_fused_mm_0\', \'\'\'\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties\ntriton_helpers.set_driver_to_gpu()\n\n@triton_heuristics.pointwise(\n size_hints={\'x\': 256}, \n filename=__file__,\n triton_meta={\'signature\': {\'in_ptr0\': \'*fp32\', \'out_ptr0\': \'*fp32\', \'xnumel\': \'i32\'}, \'device\': DeviceProperties(type=\'hip\', index=0, multi_processor_count=104, cc=\'gfx90a\', major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=64), \'constants\': {}, \'configs\': [AttrsDescriptor.from_dict({\'arg_properties\': {\'tt.divisibility\': (0, 1, 2), \'tt.equal_to\': ()}, \'cls\': \'AttrsDescriptor\'})]},\n inductor_meta={\'autotune_hints\': set(), \'kernel_name\': \'triton_poi_fused_mm_0\', \'mutated_arg_names\': [], \'optimize_mem\': False, \'no_x_dim\': False, \'num_load\': 1, \'num_reduction\': 0, \'backend_hash\': \'F6932448A4E62C51BD00A53B3A5B319A01AE443E644330D18726F033C2FD6BBE\', \'are_deterministic_algorithms_enabled\': False, \'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, \'is_hip\': True},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_mm_0(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 192\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = (xindex % 12)\n x1 = xindex // 12\n x2 = xindex\n tmp0 = x0\n tmp1 = tl.full([1], 10, tl.int64)\n tmp2 = tmp0 < tmp1\n tmp3 = tmp2.to(tl.int1)\n tmp4 = tl.load(in_ptr0 + (x0 + 10*x1), tmp3 & xmask, other=0.0)\n tl.store(out_ptr0 + (x2), tmp4, xmask)\n\'\'\', device_str=\'cuda\')\n\n\n# kernel path: /tmp/tmp4rsnqzr5/5n/c5nsbqhb7krrbf2kn6zlpa5mnpq4uvngx473ztqryvsqxkkqdv3s.py\n# Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]\n# Source node to ATen node mapping:\n# linear => constant_pad_nd_default_1\n# Graph fragment:\n# %constant_pad_nd_default_1 : [num_users=1] = call_function[target=torch.ops.aten.constant_pad_nd.default](args = (%permute, [0, 0, 0, 2]), kwargs = {})\ntriton_poi_fused_mm_1 = async_compile.triton(\'triton_poi_fused_mm_1\', \'\'\'\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties\ntriton_helpers.set_driver_to_gpu()\n\n@triton_heuristics.pointwise(\n size_hints={\'x\': 128}, \n filename=__file__,\n triton_meta={\'signature\': {\'in_ptr0\': \'*fp32\', \'out_ptr0\': \'*fp32\', \'xnumel\': \'i32\'}, \'device\': DeviceProperties(type=\'hip\', index=0, multi_processor_count=104, cc=\'gfx90a\', major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=64), \'constants\': {}, \'configs\': [AttrsDescriptor.from_dict({\'arg_properties\': {\'tt.divisibility\': (0, 1, 2), \'tt.equal_to\': ()}, \'cls\': \'AttrsDescriptor\'})]},\n inductor_meta={\'autotune_hints\': set(), \'kernel_name\': \'triton_poi_fused_mm_1\', \'mutated_arg_names\': [], \'optimize_mem\': False, \'no_x_dim\': False, \'num_load\': 1, \'num_reduction\': 0, \'backend_hash\': \'F6932448A4E62C51BD00A53B3A5B319A01AE443E644330D18726F033C2FD6BBE\', \'are_deterministic_algorithms_enabled\': False, \'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, \'is_hip\': True},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_mm_1(in_ptr0, out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 96\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = (xindex % 12)\n x1 = xindex // 12\n x2 = xindex\n tmp0 = x0\n tmp1 = tl.full([1], 10, tl.int64)\n tmp2 = tmp0 < tmp1\n tmp3 = tmp2.to(tl.int1)\n tmp4 = tl.load(in_ptr0 + (x0 + 10*x1), tmp3 & xmask, other=0.0)\n tl.store(out_ptr0 + (x2), tmp4, xmask)\n\'\'\', device_str=\'cuda\')\n\n\n# kernel path: /tmp/tmp4rsnqzr5/4v/c4vojku6k2yymfhngu74bmljpx2ib3cus653l3lcaiuxp2x7ckvo.py\n# Topologically Sorted Source Nodes: [input_tensor_2], Original ATen: [aten.relu]\n# Source node to ATen node mapping:\n# input_tensor_2 => relu\n# Graph fragment:\n# %relu : [num_users=2] = call_function[target=torch.ops.aten.relu.default](args = (%mm_default,), kwargs = {})\ntriton_poi_fused_relu_2 = async_compile.triton(\'triton_poi_fused_relu_2\', \'\'\'\nimport triton\nimport triton.language as tl\nfrom triton.compiler.compiler import AttrsDescriptor\n\nfrom torch._inductor.runtime import triton_helpers, triton_heuristics\nfrom torch._inductor.runtime.triton_helpers import libdevice, math as tl_math\nfrom torch._inductor.runtime.hints import AutotuneHint, ReductionHint, TileHint, DeviceProperties\ntriton_helpers.set_driver_to_gpu()\n\n@triton_heuristics.pointwise(\n size_hints={\'x\': 128}, \n filename=__file__,\n triton_meta={\'signature\': {\'in_out_ptr0\': \'*fp32\', \'xnumel\': \'i32\'}, \'device\': DeviceProperties(type=\'hip\', index=0, multi_processor_count=104, cc=\'gfx90a\', major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=64), \'constants\': {}, \'configs\': [AttrsDescriptor.from_dict({\'arg_properties\': {\'tt.divisibility\': (0, 1), \'tt.equal_to\': ()}, \'cls\': \'AttrsDescriptor\'})]},\n inductor_meta={\'autotune_hints\': set(), \'kernel_name\': \'triton_poi_fused_relu_2\', \'mutated_arg_names\': [\'in_out_ptr0\'], \'optimize_mem\': False, \'no_x_dim\': False, \'num_load\': 1, \'num_reduction\': 0, \'backend_hash\': \'F6932448A4E62C51BD00A53B3A5B319A01AE443E644330D18726F033C2FD6BBE\', \'are_deterministic_algorithms_enabled\': False, \'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, \'is_hip\': True},\n min_elem_per_thread=0\n)\n@triton.jit\ndef triton_poi_fused_relu_2(in_out_ptr0, xnumel, XBLOCK : tl.constexpr):\n xnumel = 128\n xoffset = tl.program_id(0) * XBLOCK\n xindex = xoffset + tl.arange(0, XBLOCK)[:]\n xmask = xindex < xnumel\n x0 = xindex\n tmp0 = tl.load(in_out_ptr0 + (x0), xmask)\n tmp1 = tl.full([1], 0, tl.int32)\n tmp2 = triton_helpers.maximum(tmp1, tmp0)\n tl.store(in_out_ptr0 + (x0), tmp2, xmask)\n\'\'\', device_str=\'cuda\')\n\n\nasync_compile.wait(globals())\ndel async_compile\n\ndef call(args):\n primals_1, primals_2, primals_3 = args\n args.clear()\n assert_size_stride(primals_1, (8, 10), (10, 1))\n assert_size_stride(primals_2, (8, 10), (10, 1))\n assert_size_stride(primals_3, (10, 8), (8, 1))\n with torch.cuda._DeviceGuard(0):\n torch.cuda.set_device(0)\n # Topologically Sorted Source Nodes: [input_tensor_1], Original ATen: [_c10d_functional.all_gather_into_tensor]\n buf0 = torch.ops._c10d_functional.all_gather_into_tensor.default(primals_1, 2, \'0\')\n assert_size_stride(buf0, (16, 10), (10, 1))\n # Topologically Sorted Source Nodes: [input_tensor_1], Original ATen: [_c10d_functional.wait_tensor]\n torch.ops._c10d_functional.wait_tensor.default(buf0)\n del primals_1\n buf3 = empty_strided_cuda((16, 12), (12, 1), torch.float32)\n # Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]\n stream0 = get_raw_stream(0)\n triton_poi_fused_mm_0.run(buf0, buf3, 192, grid=grid(192), stream=stream0)\n buf4 = empty_strided_cuda((12, 8), (1, 12), torch.float32)\n # Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]\n stream0 = get_raw_stream(0)\n triton_poi_fused_mm_1.run(primals_2, buf4, 96, grid=grid(96), stream=stream0)\n del primals_2\n buf5 = empty_strided_cuda((16, 8), (8, 1), torch.float32)\n # Topologically Sorted Source Nodes: [linear], Original ATen: [aten.mm]\n extern_kernels.mm(buf3, buf4, out=buf5)\n del buf3\n del buf4\n buf6 = buf5; del buf5 # reuse\n # Topologically Sorted Source Nodes: [input_tensor_2], Original ATen: [aten.relu]\n stream0 = get_raw_stream(0)\n triton_poi_fused_relu_2.run(buf6, 128, grid=grid(128), stream=stream0)\n # Topologically Sorted Source Nodes: [], Original ATen: []\n buf7 = torch.ops.symm_mem.fused_matmul_reduce_scatter.default(buf6, reinterpret_tensor(primals_3, (8, 10), (1, 8), 0), \'sum\', 0, \'0\')\n buf8 = buf7\n del buf7\n return (buf8, buf0, buf6, primals_3, )\n\n\ndef benchmark_compiled_module(times=10, repeat=10):\n from torch._dynamo.testing import rand_strided\n from torch._inductor.utils import print_performance\n primals_1 = rand_strided((8, 10), (10, 1), device=\'cuda:0\', dtype=torch.float32)\n primals_2 = rand_strided((8, 10), (10, 1), device=\'cuda:0\', dtype=torch.float32)\n primals_3 = rand_strided((10, 8), (8, 1), device=\'cuda:0\', dtype=torch.float32)\n fn = lambda: call([primals_1, primals_2, primals_3])\n return print_performance(fn, times=times, repeat=repeat)\n\n\nif __name__ == "__main__":\n from torch._inductor.wrapper_benchmark import compiled_module_main\n compiled_module_main(\'None\', benchmark_compiled_module)\n'
To execute this test, run the following from the base repo dir:
PYTORCH_TEST_WITH_ROCM=1 python test/distributed/tensor/parallel/test_micro_pipeline_tp.py MicroPipelineTPTest.test_dtensor_seq_par_shard_dim_0
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```
</details>
Test file path: `distributed/tensor/parallel/test_micro_pipeline_tp.py`
cc @clee2000 @wdvr | true |
2,817,589,102 | Update mi300 labels to account for multiple clusters. | saienduri | closed | [
"module: rocm",
"triaged",
"open source",
"Merged",
"Reverted",
"topic: not user facing",
"ci-no-td"
] | 11 | CONTRIBUTOR | We now have multiple Kubernetes clusters of mi300x resources, and this commit updates labels accordingly to target both clusters evenly.
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd | true |
2,817,587,352 | Update NestedInt equality to take into account all metadata | soulitzer | open | [
"module: cpu",
"release notes: fx",
"fx",
"module: dynamo",
"ciflow/inductor",
"no-stale"
] | 2 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #146172
* #146101
* __->__ #145922
* #141842
* #141841
* #146052
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @ezyang @SherlockNoMad @EikanWang @wenzhe-nrv @voznesenskym @penguinwu @Guobing-Chen @zhuhaozhe @blzheng @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,817,550,137 | Add retain-output argument | kfojcik-intel | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo"
] | 18 | CONTRIBUTOR | This PR add retain-output argument which enables appending to the already existing output file if it exists instead of deleting it and creating a new one.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,817,546,248 | Add basic Gaudi support to benchmarks/dynamo | kfojcik-intel | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"oncall: pt2",
"module: dynamo"
] | 13 | CONTRIBUTOR | This PR adds basic Gaudi support to benchmarks/dynamo
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,817,379,388 | DISABLED test_comprehensive_any_cuda_float32 (__main__.TestInductorOpInfoCUDA) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: inductor"
] | 2 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_comprehensive_any_cuda_float32&suite=TestInductorOpInfoCUDA&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/36318896504).
Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes.
**Debugging instructions (after clicking on the recent samples link):**
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
1. Click on the workflow logs linked above
2. Click on the Test step of the job so that it is expanded. Otherwise, the grepping will not work.
3. Grep for `test_comprehensive_any_cuda_float32`
4. There should be several instances run (as flaky tests are rerun in CI) from which you can study the logs.
<details><summary>Sample error message</summary>
```
Traceback (most recent call last):
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1156, in test_wrapper
return test(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1444, in only_fn
return fn(self, *args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 2262, in wrapper
fn(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1236, in dep_fn
return fn(slf, *args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1236, in dep_fn
return fn(slf, *args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1236, in dep_fn
return fn(slf, *args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 1620, in wrapper
fn(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 1542, in wrapper
fn(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/unittest/mock.py", line 1379, in patched
return func(*newargs, **newkeywargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 79, in inner
return func(*args, **kwds)
File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 79, in inner
return func(*args, **kwds)
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor_opinfo.py", line 949, in inner
raise e
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor_opinfo.py", line 941, in inner
fn(self, device, dtype, op)
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor_opinfo.py", line 1188, in test_comprehensive
raise e
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor_opinfo.py", line 1148, in test_comprehensive
self.check_model_gpu(
File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 79, in inner
return func(*args, **kwds)
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 628, in check_model_gpu
check_model(
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 510, in check_model
self.assertEqual(
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 4042, in assertEqual
raise error_metas.pop()[0].to_error( # type: ignore[index]
AssertionError: Scalars are not close!
Expected True but got True.
Absolute difference: 0 (up to 1.5e-05 allowed)
Relative difference: 0.0 (up to 1.3e-05 allowed)
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 3120, in wrapper
method(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 3120, in wrapper
method(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 454, in instantiated_test
result = test(self, **param_kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 1620, in wrapper
fn(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1168, in test_wrapper
raise e_tracked from e
Exception: Caused by sample input at index 0: SampleInput(input=Tensor[size=(), device="cuda:0", dtype=torch.float32], args=(), kwargs={}, broadcasts_input=False, name='')
To execute this test, run the following from the base repo dir:
PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=0 PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_opinfo.py TestInductorOpInfoCUDA.test_comprehensive_any_cuda_float32
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```
</details>
Test file path: `inductor/test_torchinductor_opinfo.py`
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd @clee2000 @wdvr @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,299,863 | [Customized Optimus] Add select cat aten pass | mengluy0125 | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"module: inductor",
"ciflow/inductor",
"release notes: inductor",
"inductor_pattern_match"
] | 5 | CONTRIBUTOR | Summary: This is a follow up work of D68695717, where we can further reduce the number of cat kernels in the backward by designing new aten pass in the aten level.
Test Plan:
# unit test
```
buck2 test 'fbcode//mode/dev-nosan' fbcode//caffe2/test/inductor:split_cat_fx_aten_passes -- test_select_cat_post_grad
```
Buck UI: https://www.internalfb.com/buck2/6943087f-91be-4dbd-9693-df0a11a50b73
Test UI: https://www.internalfb.com/intern/testinfra/testrun/11821949087998233
Network: Up: 101KiB Down: 132KiB (reSessionID-60e898af-f366-4247-a9f7-d8d7cd129fe0)
Analyzing targets. Remaining 0/78148
Executing actions. Remaining 0/476147
Command: test. Finished 2 local
Tests finished: Pass 3. Fail 0. Fatal 0. Skip 0. Build failure 0
# E2E
### how to add the config
```
post_grad_fusion_options: {
"normalization_aten_pass": {},
"split_cat_aten_pass": {},
"select_cat_aten_pass": {},
}
```
{F1974778773}
baseline:
aps-recgpt_ranking_1115_pt2_optimus-e52c1f277e
proposal
aps-recgpt_ranking_1115_pt2_optimus-1b0047ee0e
Differential Revision: D68803384
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,270,391 | Draft: fix: Some smaller mingw fixes | joda01 | closed | [
"open source",
"Stale"
] | 4 | NONE | Fixes #ISSUE_NUMBER
| true |
2,817,266,886 | [inductor] Add typing to common.KernelArgs | jansel | closed | [
"Merged",
"Reverted",
"topic: not user facing",
"module: inductor",
"ciflow/inductor",
"ci-no-td"
] | 7 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #146226
* #146225
* #145993
* __->__ #145916
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,198,403 | [inductor] Add typing to common.OpDecompositions | jansel | closed | [
"Merged",
"topic: not user facing",
"ciflow/mps",
"module: inductor",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #146235
* #146226
* #146225
* #145993
* #145916
* __->__ #145915
* #145914
* #145913
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,198,349 | [inductor] Combine regexp checks in OpOverrides.paren | jansel | closed | [
"Merged",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #146235
* #146226
* #146225
* #145993
* #145916
* #145915
* __->__ #145914
* #145913
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,198,218 | [inductor] Add types to DeviceOpOverrides | jansel | closed | [
"Merged",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #146235
* #146226
* #146225
* #145993
* #145916
* #145915
* #145914
* __->__ #145913
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,159,357 | Deprecate conditional view for `torch.reshape` | kurtamohler | closed | [
"module: cpu",
"open source",
"release notes: python_frontend"
] | 2 | COLLABORATOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145912
* #145911
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 | true |
2,817,159,292 | Add future lazy clone setting and deprecate `torch.reshape` view | kurtamohler | open | [
"oncall: distributed",
"module: cpu",
"open source",
"release notes: python_frontend",
"module: inductor",
"module: dynamo",
"ciflow/inductor"
] | 3 | COLLABORATOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145911
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @desertfire @chauhang @aakhundov @ColinPeppler | true |
2,817,150,655 | Fix redundant move | cyyever | closed | [
"oncall: jit",
"open source",
"NNC",
"Stale",
"release notes: jit"
] | 2 | COLLABORATOR | Fixes #ISSUE_NUMBER
cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel | true |
2,817,133,209 | [dynamo][builin-skipfiles-cleanup] Remove types | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* (to be filled)
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,817,102,438 | Debug perf regression | huydhn | closed | [
"module: dynamo",
"ciflow/inductor",
"no-runner-experiments",
"ciflow/inductor-periodic",
"test-config/inductor_torchbench_smoketest_perf"
] | 2 | CONTRIBUTOR | DEBUG, no need to review
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,817,096,807 | There may be a documentation Error in torch.nn.CrossEntropyLoss Formula | MJWade96 | open | [
"module: docs",
"module: loss",
"triaged"
] | 1 | NONE | ### 📚 The doc issue
I am not very sure if this is an error, or just a mistake of my understanding.
Documentation Location: (https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html)
In the first list point(Class indices in the range ...), it says the range is [0, C), but in the formula for l_n, the summation starts from c=1 and ends at C.
### Suggest a potential alternative/fix
_No response_
cc @svekars @brycebortree @sekyondaMeta @AlannaBurke | true |
2,817,096,461 | Use std::string_view | cyyever | closed | [
"oncall: jit",
"open source",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/periodic"
] | 9 | COLLABORATOR | Fixes #ISSUE_NUMBER
cc @EikanWang @jgong5 @wenzhe-nrv @sanchitintel | true |
2,817,090,553 | Let PYTORCH_NO_CUDA_MEMORY_CACHING has effect only when value is 1 | cyyever | closed | [
"triaged",
"open source",
"Merged",
"ciflow/trunk",
"release notes: cuda",
"topic: bug fixes"
] | 10 | COLLABORATOR | Fixes #145661
| true |
2,817,064,531 | [inductor] add size-asserts for fallback ops | shunting314 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"module: dynamo",
"ciflow/inductor",
"keep-going"
] | 26 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145904
Fix https://github.com/pytorch/pytorch/issues/144717
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,031,480 | [NJT] Add cumsum support for nested tensors | ketansingh | open | [
"fb-exported",
"Stale",
"topic: improvements",
"release notes: nested tensor"
] | 10 | NONE | Summary: - Add cumsum support for NT
Test Plan: - added unit tests
Differential Revision: D68307097
| true |
2,817,016,949 | xpu: installed pytorch is missing aten xpu ops headers (ATen/ops/cat_xpu_dispatch.h and others) | dvrogozh | open | [
"module: binaries",
"module: build",
"triaged",
"module: xpu"
] | 12 | CONTRIBUTOR | With https://github.com/pytorch/pytorch/commit/635b98fa087fa21acfdf35e95e0f2c2f56064605.
As first noted in https://github.com/pytorch/pytorch/pull/132945#discussion_r1931407357, pytorch built with XPU backend is missing a range of ATen operator headers (missing installed: they are generated, but not actually installed). Specifically, only those headers are available which are generated from torch stock sources (i.e. from `./aten/src/ATen/native/native_functions.yaml`), these are:
```
$ find torch/include/ATen/ops -name "*xpu*"
torch/include/ATen/ops/baddbmm_xpu_dispatch.h
torch/include/ATen/ops/bmm_xpu_dispatch.h
torch/include/ATen/ops/mm_xpu_dispatch.h
torch/include/ATen/ops/_addmm_activation_xpu_dispatch.h
torch/include/ATen/ops/addmm_xpu_dispatch.h
torch/include/ATen/ops/addmv_xpu_dispatch.h
torch/include/ATen/ops/addbmm_xpu_dispatch.h
```
For CUDA a lot more are generated (from the same .yaml file):
```
$ find torch/include/ATen/ops -name "*cuda*" | wc -l
604
```
For XPU however, those missing headers are generated separately from [torch-xpu-ops](https://github.com/intel/torch-xpu-ops). These are defined by https://github.com/intel/torch-xpu-ops/blob/main/yaml/native/native_functions.yaml. These XPU headers are generated, but not installed.
These missing headers might be used by some third party projects. On pytorch level these headers are used in some tests. For example, see https://github.com/pytorch/pytorch/pull/138088 which needs `ATen/ops/cat_*_dispatch.h` and `ATen/ops/norm_*_dispatch.h`.
**Can missing aten xpu headers be installed?**
I think that current challenge is that torch-xpu-ops generates not only XPU specific files from https://github.com/intel/torch-xpu-ops/blob/main/yaml/native/native_functions.yaml, but generic files as well. I.e. we get duplicated files generated from stock pytorch sources and from xpu sources. For example:
```
$ find . -name _aminmax.h
./build/aten/src/ATen/ops/_aminmax.h
./build/xpu/ATen/ops/_aminmax.h
./torch/include/ATen/ops/_aminmax.h
```
I guess there are 2 ways forward:
1. Filter out non-XPU header files and install only them
2. Move XPU headers generation to stock pytorch and drop XPU torch-xpu-ops level customization
CC: @gujinghui @EikanWang @fengyuan14 @guangyey @jgong5 @albanD
cc @seemethere @malfet @osalpekar @atalman @gujinghui @EikanWang @fengyuan14 @guangyey | true |
2,817,008,269 | [export] nested terms in nn_module_stack deserialization | pianpwk | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"ciflow/inductor",
"release notes: export"
] | 7 | CONTRIBUTOR | Summary: accounting for terms like "getattr(getattr(a[0], b), c)".
Test Plan: test_serialize
Differential Revision: D68784736
| true |
2,817,007,138 | DISABLED test_conv_backward_dynamic_shapes_cuda (__main__.DynamicShapesGPUTests) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: inductor"
] | 3 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_conv_backward_dynamic_shapes_cuda&suite=DynamicShapesGPUTests&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/36309985985).
Over the past 3 hours, it has been determined flaky in 6 workflow(s) with 6 failures and 6 successes.
**Debugging instructions (after clicking on the recent samples link):**
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
1. Click on the workflow logs linked above
2. Click on the Test step of the job so that it is expanded. Otherwise, the grepping will not work.
3. Grep for `test_conv_backward_dynamic_shapes_cuda`
4. There should be several instances run (as flaky tests are rerun in CI) from which you can study the logs.
<details><summary>Sample error message</summary>
```
Traceback (most recent call last):
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 9482, in test_conv_backward
self.common(
File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 79, in inner
return func(*args, **kwds)
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 628, in check_model_gpu
check_model(
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 417, in check_model
eager_result = model(*ref_inputs, **ref_kwargs)
File "/var/lib/jenkins/pytorch/test/inductor/test_torchinductor.py", line 9418, in fn
out1 = aten.convolution_backward(
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/_ops.py", line 1149, in __call__
return self._op(*args, **(kwargs or {}))
MemoryError: std::bad_alloc
To execute this test, run the following from the base repo dir:
PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_torchinductor_dynamic_shapes.py DynamicShapesGPUTests.test_conv_backward_dynamic_shapes_cuda
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```
</details>
Test file path: `inductor/test_torchinductor_dynamic_shapes.py`
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd @clee2000 @wdvr @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,817,003,010 | torch.compile + torch.autograd.grad: grad can be implicitly created only for scalar outputs | xmfan | open | [
"triaged",
"oncall: pt2",
"module: aotdispatch",
"module: dynamo",
"module: pt2-dispatcher"
] | 3 | MEMBER | ### 🐛 Describe the bug
```python
import torch
from torch import nn
class Model(nn.Module):
def __init__(self):
super().__init__()
self.a = nn.Linear(10, 10)
self.b = nn.Linear(10, 10)
self.c = nn.Linear(10, 10)
def forward(self, x):
ia = self.a(x)
ib = self.b(ia)
return self.c(ib)
def compute_loss(actual):
expected = torch.ones_like(actual)
return nn.MSELoss()(actual, expected)
def train(model):
x = torch.randn(10, 10)
out = model(x)
loss = compute_loss(out)
loss.backward()
def train_split(model):
x = torch.randn(10, 10)
out = model(x)
# Step 1: Compute the loss
loss = compute_loss(out)
# Step 2: Compute the gradient of the loss w.r.t. `out`
grads_out = torch.autograd.grad(loss, out, retain_graph=True, allow_unused=True)[0]
# Step 3: Compute the gradients of `out` w.r.t. the parameters of `c`
grads_c = torch.autograd.grad(
outputs=out,
inputs=[model.c.weight, model.c.bias],
grad_outputs=grads_out,
retain_graph=True
)
model.c.weight.grad, model.c.bias.grad = grads_c
# Step 4: Compute the gradients of `out` w.r.t. the parameters of `b`
grads_ib = torch.autograd.grad(
outputs=out,
inputs=[model.b.weight, model.b.bias],
grad_outputs=grads_out,
retain_graph=True
)
model.b.weight.grad, model.b.bias.grad = grads_ib
# Step 5: Compute the gradients of `out` w.r.t. the parameters of `a`
grads_ia = torch.autograd.grad(
outputs=out,
inputs=[model.a.weight, model.a.bias],
grad_outputs=grads_out,
)
model.a.weight.grad, model.a.bias.grad = grads_ia
def test(fn):
torch.manual_seed(0)
model = Model()
fn(model)
return [param.grad for param in model.parameters()]
print("Running eager, train")
grads = test(train)
print("Running eager, train_split")
grads_split = test(train_split)
assert all(torch.equal(g1, g2) for g1, g2 in zip(grads, grads_split))
print("Running compiled, train")
c_train = torch.compile(train, backend="aot_eager")
c_grads = test(c_train)
assert all(torch.equal(g1, g2) for g1, g2 in zip(grads, c_grads))
print("Running compiled, train_split")
c_train_split = torch.compile(train_split, backend="aot_eager")
c_grads_split = test(c_train_split) # <-- errors
assert all(torch.equal(g1, g2) for g1, g2 in zip(grads, c_grads_split))
```
```
Running eager, train
Running eager, train_split
Running compiled, train
Running compiled, train_split
Traceback (most recent call last):
File "/data/users/xmfan/core/b/pytorch/err_bwd.py", line 87, in <module>
c_grads_split = test(c_train_split)
File "/data/users/xmfan/core/b/pytorch/err_bwd.py", line 68, in test
fn(model)
File "/data/users/xmfan/core/b/pytorch/torch/_dynamo/eval_frame.py", line 574, in _fn
return fn(*args, **kwargs)
File "/data/users/xmfan/core/b/pytorch/err_bwd.py", line 36, in train_split
grads_out = torch.autograd.grad(loss, out, retain_graph=True, allow_unused=True)[0]
File "/data/users/xmfan/core/b/pytorch/err_bwd.py", line 39, in torch_dynamo_resume_in_train_split_at_36
grads_c = torch.autograd.grad(
File "/data/users/xmfan/core/b/pytorch/torch/autograd/__init__.py", line 475, in grad
grad_outputs_ = _make_grads(
File "/data/users/xmfan/core/b/pytorch/torch/autograd/__init__.py", line 199, in _make_grads
raise RuntimeError(
RuntimeError: grad can be implicitly created only for scalar outputs
```
### Versions
main
cc @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @amjames @zou3519 @bdhirsh @yf225 | true |
2,816,990,621 | re-use FloorDiv for RShift | ColinPeppler | closed | [
"module: cpu",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | I encountered this C++ compilation error.
```
579 | int64_t var_6 = (static_cast<int64_t>(std::floor((1.0/2.0)*u0)) | static_cast<int64_t>(std::floor((1.0/4.0)*static_cast<int64_t>(std::floor((1.0/2.0)*u0))))) | std::floor((1.0/16.0)*(static_cast<int64_t>(std::floor((1.0/2.0)*u0)) | static_cast<int64_t>(std::floor((1.0/4.0)*static_cast<int64_t>(std::floor((1.0/2.0)*u0))))));
| ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ ^ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
| | |
| int64_t {aka long int} double
```
Then, I figured out where this std::floor came from with the help of Bob's guard provenance tool. It comes from RShift which is used in `triton.next_power_of_2`.
---
Before, we used `std::floor`
```
int64_t var_6 = (
static_cast<int64_t>(std::floor((1.0/2.0)*u0)) |
static_cast<int64_t>(std::floor((1.0/4.0)*static_cast<int64_t>(std::floor((1.0/2.0)*u0)))))
| std::floor((1.0/16.0)*(static_cast<int64_t>(std::floor((1.0/2.0)*u0)) # no cast to int here.
| static_cast<int64_t>(std::floor((1.0/4.0)*static_cast<int64_t>(std::floor((1.0/2.0)*u0))))));
```
Now, we use `c10::div_floor_integer` instead
```
int64_t var_6 = (
(c10::div_floor_integer(static_cast<int64_t>(u0), static_cast<int64_t>(2L))) |
(c10::div_floor_integer(static_cast<int64_t>(u0), static_cast<int64_t>(8L)))) |
(c10::div_floor_integer(static_cast<int64_t>((c10::div_floor_integer(static_cast<int64_t>(u0), static_cast<int64_t>(2L)))
| (c10::div_floor_integer(static_cast<int64_t>(u0), static_cast<int64_t>(8L)))), static_cast<int64_t>(16L)));
```
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145898
* #145802
cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @voznesenskym @penguinwu @EikanWang @Guobing-Chen @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @amjames @desertfire @chauhang @aakhundov | true |
2,816,969,715 | [ONNX] Change constant folding and redundancies elimination to default in torch.onnx.export(..., dynamo=True) | titaiwangms | closed | [
"module: onnx",
"triaged",
"onnx-triaged"
] | 1 | COLLABORATOR | We should make the optimization a default setting, as it's more stable now.
https://github.com/pytorch/pytorch/blob/2f24f2eb46102381430777af2985b6af6e5d2cec/torch/onnx/_internal/exporter/_onnx_program.py#L119
cc @xadupre @justinchuby @gramalingam | true |
2,816,966,418 | [1/N][cp][example] flex attention in context parallel (forward pass) | XilunWu | closed | [
"oncall: distributed",
"Merged",
"topic: not user facing",
"ciflow/inductor",
"module: context parallel"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #146397
* __->__ #145896
**Description**
This is an example of how FlexAttention can be used in a context parallel fashion. Right now it's only a flex_attention call with collectives added and has no load balancer, but we're about to add the missing parts step by step:
1. backward pass
2. static load balancing for causal masking
3. dynamic load balancing for other general maskings
4. automatic collective insertion solution
5. non-intrusive context parallel APIs
**Test**
`torchrun --standalone --nnodes=1 --nproc-per-node=4 torch/distributed/tensor/examples/flex_attention_cp.py`
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,816,963,759 | [BE] reduce log spew from test_triton_kernels.py | davidberard98 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145895
One of the tests in this file was setting `self._logging.set_logs(output_code=True)` - which would cause logs to be printed for the rest of the tests in this file.
This PR puts the log-setting in a context manager so that the old behavior is restored afterwards.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,958,058 | [Inductor-CPU] Add profiling support for codegened flex attention kernels | sanchitintel | closed | [
"open source",
"Merged",
"ciflow/trunk",
"module: inductor",
"ciflow/inductor",
"release notes: inductor"
] | 6 | COLLABORATOR | ### Summary
`RECORD_FUNCTION` wasn't present in codegened Inductor-CPU Flex Attention C++ kernels, so flex attention kernels weren't present in the PyTorch profiler profiling data.
Fixes #145825 by adding `RECORD_FUNCTION` calls in the codegened flex-attention kernels.
### Caveat
#### _Before_
No corresponding results in PyTorch profiler profiling data
#### _After_
| Inductor config settings | What kernel name looks like in profiling data | Comments|
|-------------------|------------------------------------|--------------------|
| Env variable `TORCHINDUCTOR_CPP_WRAPPER=1` OR `inductor.config.cpp_wrapper=1` in python code | `graph_x_cpp_fused_y` | No way to tell from the profiling results if the kernel is a GEMM kernel or an attention kernel |
| `inductor.config.cpp.descriptive_names = "inductor_node"` but not CPP wrapper | `graph_x_kernel` | No way to tell from the profiling results if the kernel is a GEMM kernel or an attention kernel |
| Both `inductor_config.cpp.descriptive_names = "inductor_node"` & Inductor CPP Wrapper | `graph_x_cpp_fused_flex_attention_y`| Easy to interpret data |
| Neither of the two configs | `graph_x_kernel`| No way to tell from the profiling results if the kernel is a GEMM kernel or an attention kernel |
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,931,142 | [PGNCCL] Correct some ifdef's | kwen2501 | closed | [
"oncall: distributed",
"Merged",
"ciflow/trunk",
"release notes: distributed (c10d)",
"topic: not user facing"
] | 5 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145964
* __->__ #145893
`create` function supporting `ncclConfig_t` should be wrapped inside `NCCL_HAS_CONFIG` instead of `NCCL_HAS_COMM_NONBLOCKING`
cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,816,914,454 | [dynamo][builtin-skipfiles-cleanup] remove abc, enum, importlib | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* #145876
* #145909
* __->__ #145892
* #145878
* #145875
* #145856
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,909,919 | [cutlass backend] update try_import_cutlass to accomodate for pip install | henrylhtsang | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145891
The goal of this PR is to provide 3 ways for people to try out CUTLASS backend:
1. fbcode / internal
2. pip install torch (nightly) and pip install nvidia-cutlass
3. build from source
I will go into more detailed combos between building from source and downloading via pip for torch and cutlass.
repro:
```
import torch
import torch.nn as nn
import torch._inductor.config as config
config.force_disable_caches = True
config.max_autotune = True
config.max_autotune_gemm_backends = "CUTLASS"
# the following is only needed if you use a custom cutlass library
# config.cuda.cutlass_dir = "/data/users/henrylhtsang/cutlass"
class TestModule(nn.Module):
def forward(self, A, B):
return A @ B
model = TestModule().cuda()
M, K, N = 2048, 2048, 2048
A = torch.randn(M, K).cuda().half()
B = torch.randn(K, N).cuda().half()
C = torch.compile(model, fullgraph=True)(A, B)
```
## pre-requisite
Assuming you have the right cuda toolkit. Recommend 12.4. Make sure PATH, LD_LIBRARY_PATH and CUDA_NVCC_EXECUTABLE are good.
## combo 1: pip install torch + pip install nvidia-cutlass
Check https://pytorch.org/get-started/locally/ for **nightly** install command.
```
pip install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu124
pip install nvidia-cutlass
```
Then try running the script above. It should work.
## combo 2: build torch from source + pip install nvidia-cutlass
This is going to be be pretty straightforward. Just keep in mind that even though pytorch/third_party/cutlass exists, the one that will be used is the pip package, so mindful of version differences.
## combo 3: build torch from source + use pytorch/third_party/cutlass
This is how most pytorch devs would do it. Just make sure you don't have a cutlass pip package installed, i.e., make sure `import cutlass_library` would fail on its own.
## combo 4: any torch version + cutlass library from somewhere else
This is probably the only case you need to pass in cutlass_dir. Just set cutlass_dir to the cutlass repo library. The expectations is that cutlass_dir is the directory that contains include, tool, and python/cutlass_library.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,900,868 | [CD][MacOS] Don't install `libuv` from conda | malfet | closed | [
"topic: not user facing",
"ciflow/binaries_wheel"
] | 5 | CONTRIBUTOR | As one can see from build logs, it's been build as part of TensorPipe dependency:
```
2025-01-28T21:22:21.8630760Z [3444/5234] Building C object third_party/tensorpipe/tensorpipe/CMakeFiles/tensorpipe_uv.dir/__/third_party/libuv/src/random.c.o
2025-01-28T21:22:21.8731260Z [3445/5234] Building C object third_party/tensorpipe/tensorpipe/CMakeFiles/tensorpipe_uv.dir/__/third_party/libuv/src/fs-poll.c.o
2025-01-28T21:22:21.8833110Z [3446/5234] Building C object third_party/tensorpipe/tensorpipe/CMakeFiles/tensorpipe_uv.dir/__/third_party/libuv/src/idna.c.o
2025-01-28T21:22:21.8934190Z [3447/5234] Building C object third_party/tensorpipe/tensorpipe/CMakeFiles/tensorpipe_uv.dir/__/third_party/libuv/src/inet.c.o
```
Partially addresses https://github.com/pytorch/pytorch/issues/145872
Test plan: Run https://pytorch.org/tutorials/intermediate/TCPStore_libuv_backend.html | true |
2,816,896,028 | [CD] Install OpenMP from homebrew | malfet | closed | [
"Merged",
"release notes: releng",
"topic: improvements",
"ciflow/binaries_wheel"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145889
* #145870
| true |
2,816,888,043 | [pytorch] raise exception when calling dim order on sparse tensor | Gasoonjia | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145888
This diff introduces a change to the PyTorch library that raises an exception when calling the `dim_order` method on a sparse tensor.
Differential Revision: [D68797044](https://our.internmc.facebook.com/intern/diff/D68797044/) | true |
2,816,882,504 | update aotdispatcher_inference_subclass_cpu results | davidberard98 | closed | [
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 10 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145887
The regression comes from https://github.com/pytorch/pytorch/pull/145420.
I'm not sure if there's a fixed way to determine what the new value should be - but I estimated a ~0.9% increase based on the last few days of benchmark results.
<img width="953" alt="Screenshot 2025-01-28 at 2 39 39 PM" src="https://github.com/user-attachments/assets/59d8287c-6c67-4617-842e-38623ef664cf" />
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,868,263 | [export] Add distributed tests | angelayi | closed | [
"fb-exported",
"ciflow/trunk",
"topic: not user facing"
] | 6 | CONTRIBUTOR | Test Plan: CI
Differential Revision: D68799386
| true |
2,816,867,600 | Hacky solution to bad interaction between AOTAutogradcache and Triton 3.1 | jamesjwu | closed | [
"Stale",
"topic: not user facing",
"ciflow/inductor"
] | 2 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145885
| true |
2,816,864,295 | change the test wheel to release wheel when release wheel available | pytorchbot | closed | [
"open source"
] | 4 | COLLABORATOR | change the test wheel to release wheel when release wheel available
| true |
2,816,855,618 | Inductor: don't reuse buffer if it would increase peak memory | eellison | open | [
"triaged",
"oncall: pt2",
"module: inductor",
"internal ramp-up task"
] | 0 | CONTRIBUTOR | ### 🚀 The feature, motivation and pitch
Inductor has a config `allow_buffer_reuse` which will reuse a dead Tensor in memory allocation if the Tensor matches the newly allocated # of bytes. In some cases, if we are reusing a buffer during peak memory, this can increase memory usage.
We should track current allocated and peak memory during inductor codegen. and only reuse a buffer if it does not increase peak memory. We already do a similar memory tracking [here](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/memory.py). See, [buffer reuse logic](https://github.com/pytorch/pytorch/blob/af43b445a5b03ffbeab1d430d2232f48dec3053d/torch/_inductor/codegen/wrapper.py#L492-L496).
### Alternatives
master
### Additional context
_No response_
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,849,508 | bump counters for unbacked binding names | avikchaudhuri | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"ciflow/inductor",
"release notes: export"
] | 5 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145882
Instead of bumping symint counters when we process unbacked bindings during deserialization, it's better to bump them at the beginning based on what the symbols in the original shape env before serialization were. This allows symbols in unbacked bindings to have "gaps" that bumping alone would not be able to match.
Why is bumping counters important at all? It is because when the shape env coming out of deserialization is used later for propagating symints, say in run_decompositions, we don't want new names to clash with existing names (bad things happen).
Differential Revision: [D68798191](https://our.internmc.facebook.com/intern/diff/D68798191/) | true |
2,816,788,137 | [dynamo][builtin-skipfiles-cleanup] Remove threading and multithreading | anijain2305 | closed | [
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* (to be filled)
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,788,004 | [dynamo][builtin-skipfiles-cleanup] Remove operator | anijain2305 | closed | [
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* #145881
* __->__ #145880
* #145879
* #145878
* #145876
* #145875
* #145856
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,787,653 | [dynamo][bulitin-skipfiles-cleanup] Remove traceback | anijain2305 | closed | [
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* #145881
* #145880
* __->__ #145879
* #145878
* #145876
* #145875
* #145856
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,782,063 | [dynamo][builtin-skipfiles-cleanup] Remove threading, _collections_abc, _weakrefset, threading | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* #145876
* #145909
* #145892
* __->__ #145878
* #145875
* #145856
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,760,654 | [pytorch] Sprinkle in a few `template` keywords | VasuAgrawal | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 13 | CONTRIBUTOR | Summary:
These seem to be necessary to get compilation working on Windows with
CUDA 12.8. I'm not sure whether this means that all of the previous compilers
were broken, and the new one is better, or whether this is a regression in NVCC
12.8. Either way, as long as the CI passes for existing versions, this should
unblock us from CUDA 12.8 enablement on Windows.
See D68663662 for more details on the CUDA 12.8 enablement.
Test Plan: CI!
Reviewed By: akrieger
Differential Revision: D68787925
| true |
2,816,759,508 | [dynamo][builtin-skipfiles-cleanup] Remove copy | anijain2305 | closed | [
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 1 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* __->__ #145876
* #145958
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,754,999 | [dynamo][builtin-skipfiles-removal] Remove logging | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"module: dynamo",
"ciflow/inductor"
] | 4 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* #145876
* #145909
* #145892
* #145878
* __->__ #145875
* #145856
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,715,796 | Use of broadcast_shapes() errors attempting to guard on symbolic nested int | jbschlosser | closed | [
"triaged",
"module: nestedtensor",
"oncall: pt2",
"module: dynamic shapes"
] | 0 | CONTRIBUTOR | Repro:
```python
import torch
torch._dynamo.config.capture_dynamic_output_shape_ops = True
torch._dynamo.config.capture_scalar_outputs = True
nt = torch.nested.nested_tensor([
torch.randn(2),
torch.randn(3),
torch.randn(4),
], layout=torch.jagged, device="cuda")
@torch.compile(fullgraph=True)
def f(t, mask):
nt = torch.nested.masked_select(t, mask)
return torch.where(nt > 0., torch.ones_like(nt), torch.zeros_like(nt))
t = torch.randn(3, 5)
mask = torch.randint(0, 2, t.shape, dtype=torch.bool)
output = f(t, mask)
```
Problem: `torch.nested.masked_select()` constructs an NJT in-graph, which invokes a special path for constructing a new symbolic nested int. Attempting to guard on this (due to usage of `where()` -> `broadcast_shapes()` -> `Ne(s1, 1)` guard on the symbolic nested int) gives this error:
```
...
AssertionError: s0 (could be from ['<ephemeral: intermediate_offsets_or_lengths>']) not in {s0: []}. If this assert is failing, it could be due to the issue described in https://github.com/pytorch/pytorch/pull/90665
```
cc @cpuhrsch @bhosmer @drisspg @soulitzer @davidberard98 @YuqingJ @chauhang @penguinwu @ezyang @bobrenjc93 | true |
2,816,697,644 | [WIP] Allow generation of inductor backend specific tests using instantiate_device_type_tests | kundaMwiza | open | [
"open source",
"module: inductor"
] | 2 | CONTRIBUTOR | This allows the creation of inductor backend specific test classes. Since this is an extension point for out of tree backends, it also allows out of tree backends to customise test instantiation to fit their backend / device.
To maintain backwards compatibility, `only_inductor_backends` defaults to `None` so that the behaviour of test class instantiation matches the incumbent behaviour. If `only_inductor_backends` is not None, inductor backend specific test classes will be created from a test template, e.g. `TestInductorOpInfo -> TestInductorOpInfoTritonCUSTOMDEVICE, TestInductorOpInfoHalideCUSTOMDEVICE`
An illustration of the before/after changes:
```python
# in test_inductor.py
# Inductor test template
class TestInductor:
def test_comprehensive(...)
# Original
instantiate_device_type_tests(TestSuiteTemplate)
# Generates something like this:
# TestInductorCPU
# TestInductorCUDA
# After changes
instantiate_device_type_tests(TestSuiteTemplate, enable_inductor_backend_classes=True, only_inductor_backends=["cpp", "triton"])
# TestInductorCppCPU
# TestInductorCppTriton
# TestInductorTritonCUDA
# Additionally, the new test classes if a native inductor backend is used are guarded
# e.g. TestInductorCppCPU
# is equivalent to the following class definition
# @skipUnless(HAS_CPU, "Requires C++ compiler")
# @config.patch("cpu_backend", "cpp")
# class TestInductorCppCPU(CPUTestBase)
# ...
```
Fixes #ISSUE_NUMBER
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,690,663 | [binary builds] Anaconda. Remove dependency on conda libuv module in MacOS and Windows nightly builds | atalman | closed | [
"oncall: releng",
"triaged",
"topic: binaries"
] | 6 | CONTRIBUTOR | ### 🐛 Describe the bug
Related to: https://github.com/pytorch/pytorch/issues/138506
In Windows and MacOS wheel builds we use libuv as build time dependency.
In Windows we use lubuv during nightly build:
.ci/pytorch/windows/condaenv.bat
On MacOS :
.ci/wheel/build_wheel.sh
.github/requirements/conda-env-macOS-ARM64
This package is available via conda : https://anaconda.org/anaconda/libuv
And not available via pip install.
As the first step in refactoring workflow that dependent on conda we want to refactor the code to use different install method for this package. Maybe installing via download https://github.com/libuv/libuv#downloading ? If possible one should try to use same method of installation that would work in Windows and MacOS.
### Versions
2.7.0 | true |
2,816,690,584 | [CD] Install ninja and setuptools from PyPI | malfet | closed | [
"Merged",
"Reverted",
"topic: not user facing",
"ciflow/binaries_wheel",
"ci-no-td"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145889
* #145870
* __->__ #145871
As well as typing extensions, they are available from PyPI, no reason to install them from Anaconda | true |
2,816,679,092 | [CMake] Find HomeBrew OpenMP on MacOS | malfet | closed | [
"Merged",
"Reverted",
"ciflow/trunk",
"release notes: build",
"topic: improvements",
"ciflow/binaries_wheel",
"ci-no-td"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145889
* __->__ #145870
Either via `OMP_PREFIX` envvar or by searching in `/opt/homebrew/opt/libomp` folder
Modify libomp bundling logic in setup.py to change absolute path to libomp.dylib to a relative one if necessary | true |
2,816,676,523 | Compiled flex bias appears to not work at all | leijurv | open | [
"triaged",
"bug",
"oncall: pt2",
"module: higher order operators",
"module: pt2-dispatcher",
"module: flex attention"
] | 5 | NONE | ### 🐛 Describe the bug
```python
import torch
from torch.nn.attention.flex_attention import flex_attention
N_BATCH = 16
N_HEAD = 32
N_CTX = 4096
D_QK = 128
torch.set_default_device("cuda")
@torch.compile
def main():
q = torch.randn(N_BATCH, N_HEAD, N_CTX, D_QK, requires_grad=True)
k = torch.randn(N_BATCH, N_HEAD, N_CTX, D_QK, requires_grad=True)
v = torch.randn(N_BATCH, N_HEAD, N_CTX, D_QK, requires_grad=True)
# source: https://github.com/pytorch/pytorch/blob/53fc921ce2bcfd29b0adc42b72b86a982a690e30/test/inductor/test_flex_attention.py#L2978
#bias_1 = torch.randn(N_BATCH, N_HEAD, N_CTX, N_CTX, requires_grad=True)
#def score_mod(score, b, h, q_idx, kv_idx):
# return score + bias_1[b, h, q_idx, kv_idx]
# illegal memory access
# source: https://github.com/pytorch/pytorch/blob/53fc921ce2bcfd29b0adc42b72b86a982a690e30/test/inductor/test_flex_attention.py#L4714
#bias_2 = torch.randn(N_CTX, N_CTX, requires_grad=True)
#def score_mod(score, b, h, q_idx, kv_idx):
# return score + bias_2[q_idx, kv_idx]
# AssertionError: size=[16, 32, 4096, 128], stride=[4096, 1]
# source: https://github.com/pytorch/pytorch/blob/53fc921ce2bcfd29b0adc42b72b86a982a690e30/test/inductor/test_flex_attention.py#L209
#bias_3 = torch.randn(N_HEAD, requires_grad=True)
#def score_mod(score, b, h, q_idx, kv_idx):
# return score * bias_3[h]
# AssertionError: size=[16, 32, 4096, 128], stride=[1]
flex_attention(q, k, v, score_mod=score_mod).sum().backward()
main()
```
I have taken three examples of bias from the flex tests, so I expect them to work. The first one gets an illegal memory access. The second and third get deep assertion errors. This is on latest nightly.
### Versions
```
Collecting environment information...
PyTorch version: 2.7.0.dev20250128+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35
Python version: 3.10.12 (main, Nov 6 2024, 20:22:13) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.8.0-49-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.4.131
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA H100 PCIe
Nvidia driver version: 550.127.05
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 46 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 26
On-line CPU(s) list: 0-25
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8480+
CPU family: 6
Model: 143
Thread(s) per core: 1
Core(s) per socket: 1
Socket(s): 26
Stepping: 8
BogoMIPS: 4000.00
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology cpuid tsc_known_freq pni pclmulqdq vmx ssse3 fma cx16 pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves avx_vnni avx512_bf16 wbnoinvd arat vnmi avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b fsrm md_clear serialize tsxldtrk avx512_fp16 arch_capabilities
Virtualization: VT-x
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 832 KiB (26 instances)
L1i cache: 832 KiB (26 instances)
L2 cache: 104 MiB (26 instances)
L3 cache: 416 MiB (26 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-25
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Unknown: No mitigations
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI SW loop, KVM SW loop
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Mitigation; TSX disabled
Versions of relevant libraries:
[pip3] flake8==4.0.1
[pip3] numpy==1.21.5
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] optree==0.13.1
[pip3] pytorch-triton==3.2.0+gitb2684bf3
[pip3] torch==2.7.0.dev20250128+cu124
[pip3] triton==3.1.0
[conda] Could not collect
```
cc @chauhang @penguinwu @zou3519 @ydwu4 @bdhirsh @yf225 @Chillee @drisspg @yanboliang @BoyuanFeng | true |
2,816,661,559 | Fix code cache + freezing compile-time regression | masnesral | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: inductor",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145868
Summary: The current implementation introduces a compile-time regression due to overhead hashing large constants. To support freezing+caching, we consider only the tensor metadata of frozen params, but we neglect to do the same for any constants created as a result of folding frozen params. This PR Explicitly marks the constants created during freezing (and constant folding during freezing) and uses that info in the inductor cache to determine when to hash a tensor value+metadata vs. metadata only.
Test Plan: `python benchmarks/dynamo/torchbench.py --backend inductor --device cuda --only alexnet --bfloat16 --cold-start-latency --print-compilation-time --inference --performance --freezing`
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,653,998 | [triton] Update pin to tip of 3.2 release | bertmaher | closed | [
"Merged",
"Reverted",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor",
"ci-no-td"
] | 13 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145867
| true |
2,816,650,624 | [pytorch][cuda] Improve softmax backward pass native CUDA implementation | ahmadsharif1 | closed | [
"Merged",
"ciflow/trunk",
"release notes: cuda"
] | 10 | CONTRIBUTOR | This PR is similar to https://github.com/pytorch/pytorch/pull/122970, but works on the softmax backward pass.
Specifically, it uses shared memory to cache the gradOutput when it can fit in shared memory. Before this PR we were reading gradOutput twice.
On my H100 this seems to improve the softmax backward pass performance by about 5% for problem sizes that fit within shared memory. (Note that this is not the only kernel that runs when you call softmax backward pass -- there is an elementwise kernel that runs before this; optimizing that can be a separate PR).
**Important Note**: Currently the softmax backward pass consists of an [element-wise multiply operator](https://github.com/pytorch/pytorch/blob/7f65a208848205b38445423b7e2e93a2b4994e5e/aten/src/ATen/native/cuda/SoftMax.cu#L1216), followed by [this function](https://github.com/pytorch/pytorch/blob/7f65a208848205b38445423b7e2e93a2b4994e5e/aten/src/ATen/native/cuda/SoftMax.cu#L1062) which calls the `cunn_SoftMaxBackward` kernel. With my change the kernel time reduces by about 12% (see screenshot below), while the total time (including the elementwise) reduces by about 5%.
```
Baseline This PR
N size FP32 bandwidth FP16 bandwidth N size FP32 bandwidth FP16 bandwidth fp32 diff fp16 diff
0 256 134.340966 70.042039 0 256 133.70146 70.342753 -0.48% 0.43%
1 512 233.501185 129.945803 1 512 234.057145 132.933066 0.24% 2.30%
2 1024 340.667966 229.280464 2 1024 338.833265 226.441699 -0.54% -1.24%
3 2048 379.643726 337.452058 3 2048 399.559017 338.432284 5.25% 0.29%
4 4096 416.597537 383.625364 4 4096 428.252403 396.137506 2.80% 3.26%
5 6000 431.198241 384.384384 5 6000 457.744577 406.06275 6.16% 5.64%
6 8192 462.811252 427.292573 6 8192 474.791032 428.281563 2.59% 0.23%
7 10000 464.258731 429.050294 7 10000 483.7643 446.849381 4.20% 4.15%
8 10013 465.199701 429.824179 8 10013 464.904407 428.72184 -0.06% -0.26%
9 10240 477.07359 428.853737 9 10240 485.317024 444.902586 1.73% 3.74%
10 11000 473.038785 430.778663 10 11000 488.161438 453.462162 3.20% 5.27%
11 12000 474.342475 432.594814 11 12000 490.532418 458.427653 3.41% 5.97%
12 16384 487.468854 473.611576 12 16384 488.154406 476.264631 0.14% 0.56%
13 20000 482.029793 465.666186 13 20000 482.147092 483.886193 0.02% 3.91%
14 24000 478.368093 474.159464 14 24000 478.364948 491.447921 0.00% 3.65%
15 32000 476.523796 473.18868 15 32000 476.523796 474.398962 0.00% 0.26%
16 32768 476.104723 477.493634 16 32768 476.704463 477.330606 0.13% -0.03%
17 36864 477.900663 475.472787 17 36864 477.973279 475.728454 0.02% 0.05%
18 40960 477.707561 475.559064 18 40960 478.445017 476.088067 0.15% 0.11%
19 45056 479.169812 475.865134 19 45056 479.143266 475.878202 -0.01% 0.00%
20 49152 477.804907 475.382982 20 49152 477.868404 475.976377 0.01% 0.12%
21 65536 481.274125 478.171806 21 65536 481.537733 478.703926 0.05% 0.11%
22 66000 481.64652 480.095457 22 66000 481.856013 480.466388 0.04% 0.08%
23 68608 481.745774 479.034704 23 68608 481.917596 478.856209 0.04% -0.04%
24 80000 483.409361 480.356529 24 80000 483.330481 480.375277 -0.02% 0.00%
25 98304 480.736301 481.396882 25 98304 480.789858 481.320143 0.01% -0.02%
```
NCU profiler shows lower DRAM fetches with the new kernel:
NCU reports about 12% elapsed time reduction in this kernel alone compared to baseline (and because of other kernels that are run, the overall backward pass time as seen by the user gets reduced by 5%).
I compared the binary size increase by running `python setup.py develop` before and after and diffing the .so files:
libtorch_cuda.so goes from 274,752,224 bytes to 274,787,072 bytes. The increase in size is 34kB which is about 0.01%.
I measured the compilation time for incremental development:
```
touch ./aten/src/ATen/native/cuda/SoftMax.cu
time python setup.py develop
real 0m10.083s
user 0m8.197s
sys 0m3.149s
```
Note that this uses `ccache` and does a bunch of copies and is not just measuring the `nvcc` time. I measured the `nvcc` time separately by capturing the `nvcc` command shown in [1] below and running it on the baseline and modified kernels:
```
# baseline nvcc time for SoftMax.cu
real 0m35.341s
user 0m33.801s
sys 0m1.289s
# this PR's nvcc time for SoftMax.cu
real 0m36.513s
user 0m34.722s
sys 0m1.408s
```
So the `nvcc` time increases by about 1 second, or ~3% of the baseline.
[1] `nvcc` command is here:
```
# This is the nvcc command
/usr/local/cuda/bin/nvcc -forward-unknown-to-host-compiler -DAT_PER_OPERATOR_HEADERS -DFLASHATTENTION_DISABLE_ALIBI -DFMT_HEADER_ONLY=1 -DHAVE_MALLOC_USABLE_SIZE=1 -DHAVE_MMAP=1 -DHAVE_SHM_OPEN=1 -DHAVE_SHM_UNLINK=1 -DMINIZ_DISABLE_ZIP_READER_CRC32_CHECKS -DONNXIFI_ENABLE_EXT=1 -DONNX_ML=1 -DONNX_NAMESPACE=onnx_torch -DTORCH_CUDA_BUILD_MAIN_LIB -DTORCH_CUDA_USE_NVTX3 -DUSE_C10D_GLOO -DUSE_C10D_NCCL -DUSE_CUDA -DUSE_DISTRIBUTED -DUSE_EXTERNAL_MZCRC -DUSE_FLASH_ATTENTION -DUSE_MEM_EFF_ATTENTION -DUSE_NCCL -DUSE_RPC -DUSE_TENSORPIPE -D_FILE_OFFSET_BITS=64 -Dtorch_cuda_EXPORTS -I/home/ahmads/personal/pytorch/build/aten/src -I/home/ahmads/personal/pytorch/aten/src -I/home/ahmads/personal/pytorch/build -I/home/ahmads/personal/pytorch -I/home/ahmads/personal/pytorch/cmake/../third_party/benchmark/include -I/home/ahmads/personal/pytorch/third_party/onnx -I/home/ahmads/personal/pytorch/build/third_party/onnx -I/home/ahmads/personal/pytorch/nlohmann -I/home/ahmads/personal/pytorch/aten/src/THC -I/home/ahmads/personal/pytorch/aten/src/ATen/cuda -I/home/ahmads/personal/pytorch/third_party/fmt/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/include -I/home/ahmads/personal/pytorch/aten/src/ATen/../../../third_party/cutlass/tools/util/include -I/home/ahmads/personal/pytorch/build/caffe2/aten/src -I/home/ahmads/personal/pytorch/aten/src/ATen/.. -I/home/ahmads/personal/pytorch/build/nccl/include -I/home/ahmads/personal/pytorch/c10/cuda/../.. -I/home/ahmads/personal/pytorch/c10/.. -I/home/ahmads/personal/pytorch/third_party/tensorpipe -I/home/ahmads/personal/pytorch/build/third_party/tensorpipe -I/home/ahmads/personal/pytorch/third_party/tensorpipe/third_party/libnop/include -I/home/ahmads/personal/pytorch/torch/csrc/api -I/home/ahmads/personal/pytorch/torch/csrc/api/include -isystem /home/ahmads/personal/pytorch/build/third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/gloo -isystem /home/ahmads/personal/pytorch/cmake/../third_party/tensorpipe/third_party/libuv/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googlemock/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/googletest/googletest/include -isystem /home/ahmads/personal/pytorch/third_party/protobuf/src -isystem /home/ahmads/personal/pytorch/third_party/XNNPACK/include -isystem /home/ahmads/personal/pytorch/third_party/ittapi/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/eigen -isystem /usr/local/cuda/include -isystem /home/ahmads/personal/pytorch/torch/include -isystem /home/ahmads/personal/pytorch/third_party/ideep/include -isystem /home/ahmads/personal/pytorch/torch/include/oneapi/dnnl -isystem /home/ahmads/personal/pytorch/INTERFACE -isystem /home/ahmads/personal/pytorch/third_party/nlohmann/include -isystem /home/ahmads/personal/pytorch/third_party/NVTX/c/include -isystem /home/ahmads/personal/pytorch/cmake/../third_party/cudnn_frontend/include -DLIBCUDACXX_ENABLE_SIMPLIFIED_COMPLEX_OPERATIONS -D_GLIBCXX_USE_CXX11_ABI=1 -Xfatbin -compress-all -DONNX_NAMESPACE=onnx_torch -gencode arch=compute_90,code=sm_90 -Xcudafe --diag_suppress=cc_clobber_ignored,--diag_suppress=field_without_dll_interface,--diag_suppress=base_class_has_different_dll_interface,--diag_suppress=dll_interface_conflict_none_assumed,--diag_suppress=dll_interface_conflict_dllexport_assumed,--diag_suppress=bad_friend_decl --expt-relaxed-constexpr --expt-extended-lambda -Wno-deprecated-gpu-targets --expt-extended-lambda -DCUB_WRAPPED_NAMESPACE=at_cuda_detail -DCUDA_HAS_FP16=1 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -O3 -DNDEBUG -std=c++17 -Xcompiler=-fPIC -DTORCH_USE_LIBUV -DCAFFE2_USE_GLOO -Xcompiler -Wall -Wextra -Wdeprecated -Wno-unused-parameter -Wno-missing-field-initializers -Wno-array-bounds -Wno-unknown-pragmas -Wno-strict-overflow -Wno-strict-aliasing -Wunused-function -Wunused-variable -Wunused-but-set-variable -Wno-maybe-uninitialized -MD -MT caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/SoftMax.cu.o -MF caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/SoftMax.cu.o.d -x cu -c /home/ahmads/personal/pytorch/aten/src/ATen/native/cuda/SoftMax.cu -o caffe2/CMakeFiles/torch_cuda.dir/__/aten/src/ATen/native/cuda/SoftMax.cu.o
```
| true |
2,816,641,402 | [WIP] Add test_torchinductor_opinfo.py to triton-cpu tests | davidberard98 | closed | [
"Stale",
"ciflow/inductor",
"keep-going"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145865
I'm guessing this isn't going to pass, but I want to see how long the CI takes. | true |
2,816,546,277 | Segmentation fault at _rowwise_prune with convert_hf_to_gguf from llama.cpp | atmaranto | open | [
"module: crash",
"module: windows",
"triaged",
"module: 64-bit"
] | 2 | NONE | ### 🐛 Describe the bug
When converting the DeepSeek R1 LLAMA-70B distilled model to a GGUF file, the program [convert_hf_to_gguf.py](https://github.com/ggerganov/llama.cpp/blob/master/convert_hf_to_gguf.py) silently crashes. Attaching to WinDbg, it appears torch_cpu causes a segfault:
```
(cfd4.d19c): Access violation - code c0000005 (first chance)
First chance exceptions are reported before any exception handling.
This exception may be expected and handled.
torch_cpu!at::native::_rowwise_prune+0x30b:
00007ffc`ba8da7bb 0fb610 movzx edx,byte ptr [rax] ds:00000000`5e001424=??
(some omitted for brevity)
0:000> K
# Child-SP RetAddr Call Site
00 000000a3`7c7ed990 00007ffc`ba8dafbf torch_cpu!at::native::_rowwise_prune+0x30b
01 000000a3`7c7eda60 00007ffc`bb496ace torch_cpu!at::native::_local_scalar_dense_cpu+0x3f
02 000000a3`7c7edab0 00007ffc`bb46e001 torch_cpu!at::cpu::where_outf+0x1a7e
03 000000a3`7c7edae0 00007ffc`bb0617f4 torch_cpu!at::cpu::bucketize_outf+0x41271
04 000000a3`7c7edb10 00007ffc`ba8db14f torch_cpu!at::_ops::_local_scalar_dense::call+0x154
05 000000a3`7c7edc20 00007ffc`bb78fe8e torch_cpu!at::native::item+0x17f
06 000000a3`7c7edcb0 00007ffc`bb76fe31 torch_cpu!at::compositeimplicitautograd::where+0x3cce
07 000000a3`7c7edce0 00007ffc`baf1d954 torch_cpu!at::compositeimplicitautograd::broadcast_to_symint+0x33e01
08 000000a3`7c7edd10 00007ffc`bbbd1cd3 torch_cpu!at::_ops::item::call+0x154
09 000000a3`7c7ede20 00007ffd`62ad8507 torch_cpu!at::Tensor::item<unsigned char>+0x13
0a 000000a3`7c7ede70 00007ffd`626cef39 torch_python!pybind11::detail::type_caster<at::Tensor,void>::load+0x2b7
0b 000000a3`7c7edee0 00007ffd`b5174e86 torch_python!isMainPyInterpreter+0x3ae9
0c 000000a3`7c7ee060 00007ffd`b5296fc3 python312!_PyObject_Call+0xd6 [\Objects\call.c @ 369]
0d 000000a3`7c7ee0a0 00007ffd`b5175034 python312!_PyEval_EvalFrameDefault+0x6bd3 [\PCbuild\Python\bytecodes.c @ 3263]
0e 000000a3`7c7ee290 00007ffd`b51dd37f python312!_PyFunction_Vectorcall+0x54 [\Objects\call.c @ 424]
0f 000000a3`7c7ee2d0 00007ffd`b51e1802 python312!_PyObject_VectorcallTstate+0x4f [\Include\internal\pycore_call.h @ 92]
10 (Inline Function) --------`-------- python312!vectorcall_unbound+0x26 [\Objects\typeobject.c @ 2236]
11 000000a3`7c7ee310 00007ffd`b51ec939 python312!vectorcall_method+0xd2 [\Objects\typeobject.c @ 2268]
12 000000a3`7c7ee370 00007ffd`62b20371 python312!slot_sq_item+0x49 [\Objects\typeobject.c @ 8493]
13 000000a3`7c7ee3b0 00007ffd`62b2158d torch_python!torch::utils::getTHPMemoryFormat+0x2891
14 000000a3`7c7ee530 00007ffd`62b1f8ea torch_python!torch::utils::getTHPMemoryFormat+0x3aad
15 000000a3`7c7ee800 00007ffd`62757515 torch_python!torch::utils::getTHPMemoryFormat+0x1e0a
16 000000a3`7c7eea30 00007ffd`b51c1a93 torch_python!THPPointer<THPStorage>::THPPointer<THPStorage>+0x2a095
17 000000a3`7c7eec60 00007ffd`b5174e86 python312!cfunction_call+0x63 [\Objects\methodobject.c @ 540]
18 000000a3`7c7eec90 00007ffd`da95aa0d python312!_PyObject_Call+0xd6 [\Objects\call.c @ 369]
19 000000a3`7c7eecd0 00007ffd`da931283 _safetensors_rust_cp312_win_amd64!PyInit__safetensors_rust+0x1626d
1a 000000a3`7c7eed40 00007ffd`da933be9 _safetensors_rust_cp312_win_amd64+0x11283
1b 000000a3`7c7ef080 00007ffd`b52914e0 _safetensors_rust_cp312_win_amd64+0x13be9
```
To reproduce the issue, download [this model](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) and run within the llama.cpp main directory:
```bash
python convert_hf_to_gguf.py model_path
```
### Versions
```
PyTorch version: 2.5.1
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A
OS: Microsoft Windows 10 Pro (10.0.19045 64-bit)
GCC version: Could not collect
Clang version: Could not collect
CMake version: version 3.28.0-rc5
Libc version: N/A
Python version: 3.12.8 | packaged by conda-forge | (main, Dec 5 2024, 14:06:27) [MSC v.1942 64 bit (AMD64)] (64-bit runtime)
Python platform: Windows-10-10.0.19045-SP0
Is CUDA available: True
CUDA runtime version: 12.6.85
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090
Nvidia driver version: 566.36
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Name: 13th Gen Intel(R) Core(TM) i5-13600K
Manufacturer: GenuineIntel
Family: 205
Architecture: 9
ProcessorType: 3
DeviceID: CPU0
CurrentClockSpeed: 3500
MaxClockSpeed: 3500
L2CacheSize: 8192
L2CacheSpeed: None
Revision: None
Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] torch==2.5.1
[pip3] torchaudio==2.5.1
[pip3] torchvision==0.20.1
[conda] blas 1.0 mkl
[conda] cuda-cudart 12.4.127 0 nvidia
[conda] cuda-cudart-dev 12.4.127 0 nvidia
[conda] cuda-cupti 12.4.127 0 nvidia
[conda] cuda-libraries 12.4.1 0 nvidia
[conda] cuda-libraries-dev 12.4.1 0 nvidia
[conda] cuda-nvrtc 12.4.127 0 nvidia
[conda] cuda-nvrtc-dev 12.4.127 0 nvidia
[conda] cuda-nvtx 12.4.127 0 nvidia
[conda] cuda-opencl 12.6.77 0 nvidia
[conda] cuda-opencl-dev 12.6.77 0 nvidia
[conda] cuda-runtime 12.4.1 0 nvidia
[conda] libcublas 12.4.5.8 0 nvidia
[conda] libcublas-dev 12.4.5.8 0 nvidia
[conda] libcufft 11.2.1.3 0 nvidia
[conda] libcufft-dev 11.2.1.3 0 nvidia
[conda] libcurand 10.3.7.77 0 nvidia
[conda] libcurand-dev 10.3.7.77 0 nvidia
[conda] libcusolver 11.6.1.9 0 nvidia
[conda] libcusolver-dev 11.6.1.9 0 nvidia
[conda] libcusparse 12.3.1.170 0 nvidia
[conda] libcusparse-dev 12.3.1.170 0 nvidia
[conda] libnvjitlink 12.4.127 0 nvidia
[conda] libnvjitlink-dev 12.4.127 0 nvidia
[conda] mkl 2023.1.0 h6b88ed4_46358
[conda] mkl-service 2.4.0 py312h2bbff1b_1
[conda] mkl_fft 1.3.11 py312h827c3e9_0
[conda] mkl_random 1.2.8 py312h0158946_0
[conda] numpy 1.26.4 py312hfd52020_0
[conda] numpy-base 1.26.4 py312h4dde369_0
[conda] pytorch 2.5.1 py3.12_cuda12.4_cudnn9_0 pytorch
[conda] pytorch-cuda 12.4 h3fd98bf_7 pytorch
[conda] pytorch-mutex 1.0 cuda pytorch
[conda] torchaudio 2.5.1 pypi_0 pypi
[conda] torchvision 0.20.1 pypi_0 pypi
```
cc @peterjc123 @mszhanyi @skyline75489 @nbcsm @iremyux @Blackhex | true |
2,816,540,796 | Cleanup VS 2019 refs in pytorch | Camyll | closed | [
"Merged",
"ciflow/binaries",
"ciflow/trunk",
"release notes: releng",
"test-config/default"
] | 10 | CONTRIBUTOR | Related to: https://github.com/pytorch/pytorch/issues/128835
Follow up on PR: https://github.com/pytorch/pytorch/pull/145319 | true |
2,816,533,217 | Extend abi-stable nitpick message to all the c stable files | albanD | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 4 | COLLABORATOR | null | true |
2,816,532,662 | CPU Model compile not working for flexattention | tpopok | closed | [
"oncall: pt2",
"oncall: cpu inductor"
] | 6 | NONE | ### 🐛 Describe the bug
Use CPU model (i.e., CUDA_VISIBLE_DEVICES=-1) to run the following code will throw error.
```
from torch.utils.data import DataLoader, Dataset
import torch
from torch import nn
from torch.nn.attention.flex_attention import flex_attention, create_block_mask, _create_empty_block_mask, BlockMask
# flex_attention = torch.compile(flex_attention)
# Create a simple custom dataset
class MyDataset(Dataset):
def __init__(self):
self.data = torch.randn(100, 32)
self.labels = torch.randint(0, 2, (100,))
def __len__(self):
return 10000000
def __getitem__(self, idx):
return self.data[idx % 100], self.labels[idx % 100]
class MyModel(nn.Module):
def __init__(self, *args, **kwargs):
super().__init__(*args, **kwargs)
self.loss = torch.nn.MSELoss()
self.q_proj = nn.LazyLinear(128)
self.k_proj = nn.LazyLinear(128)
self.v_proj = nn.LazyLinear(128)
def forward(self, x, y):
x = x[None, None, :, :]
r = x
q = self.q_proj(x)
k = self.k_proj(x)
v = self.v_proj(x)
r = flex_attention(q, k, v)
return self.loss(torch.sum(r[0, 0, :, :], dim=-1), y)
dataset = MyDataset()
data_loader = DataLoader(dataset, batch_size=3, shuffle=True)
model = MyModel()
model.compile()
for x, y in data_loader:
print(model(x, y))
break
```
Error trace:
```
No CUDA runtime is found, using CUDA_HOME='/usr/local/cuda'
Traceback (most recent call last):
File "/data/mizhou/workspace/torch_p13n_embedding/py/tmp/test_flex_attn.py", line 44, in <module>
print(model(x, y))
^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1737, in _wrapped_call_impl
return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/eval_frame.py", line 574, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/nn/modules/module.py", line 1750, in _call_impl
return forward_call(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 1380, in __call__
return self._torchdynamo_orig_callable(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 1164, in __call__
result = self._inner_convert(
^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 547, in __call__
return _compile(
^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 986, in _compile
guarded_code = compile_inner(code, one_graph, hooks, transform)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 715, in compile_inner
return _compile_inner(code, one_graph, hooks, transform)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_utils_internal.py", line 95, in wrapper_function
return function(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 750, in _compile_inner
out_code = transform_code_object(code, transform)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/bytecode_transformation.py", line 1361, in transform_code_object
transformations(instructions, code_options)
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 231, in _fn
return fn(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/convert_frame.py", line 662, in transform
tracer.run()
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py", line 2868, in run
super().run()
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py", line 1052, in run
while self.step():
^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py", line 962, in step
self.dispatch_table[inst.opcode](self, inst)
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py", line 3048, in RETURN_VALUE
self._return(inst)
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/symbolic_convert.py", line 3033, in _return
self.output.compile_subgraph(
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/output_graph.py", line 1101, in compile_subgraph
self.compile_and_call_fx_graph(
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/output_graph.py", line 1382, in compile_and_call_fx_graph
compiled_fn = self.call_user_compiler(gm)
^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/output_graph.py", line 1432, in call_user_compiler
return self._call_user_compiler(gm)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/output_graph.py", line 1483, in _call_user_compiler
raise BackendCompilerFailed(self.compiler_fn, e).with_traceback(
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/output_graph.py", line 1462, in _call_user_compiler
compiled_fn = compiler_fn(gm, self.example_inputs())
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/repro/after_dynamo.py", line 130, in __call__
compiled_gm = compiler_fn(gm, example_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/__init__.py", line 2340, in __call__
return compile_fx(model_, inputs_, config_patches=self.config)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1863, in compile_fx
return aot_autograd(
^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/backends/common.py", line 83, in __call__
cg = aot_module_simplified(gm, example_inputs, **self.kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_functorch/aot_autograd.py", line 1155, in aot_module_simplified
compiled_fn = dispatch_and_compile()
^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_functorch/aot_autograd.py", line 1131, in dispatch_and_compile
compiled_fn, _ = create_aot_dispatcher_function(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_functorch/aot_autograd.py", line 580, in create_aot_dispatcher_function
return _create_aot_dispatcher_function(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_functorch/aot_autograd.py", line 830, in _create_aot_dispatcher_function
compiled_fn, fw_metadata = compiler_fn(
^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 678, in aot_dispatch_autograd
compiled_fw_func = aot_config.fw_compiler(fw_module, adjusted_flat_args)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_functorch/aot_autograd.py", line 489, in __call__
return self.compiler_fn(gm, example_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1741, in fw_compiler_base
return inner_compile(
^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 569, in compile_fx_inner
return wrap_compiler_debug(_compile_fx_inner, compiler_name="inductor")(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_dynamo/repro/after_aot.py", line 102, in debug_wrapper
inner_compiled_fn = compiler_fn(gm, example_inputs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 685, in _compile_fx_inner
mb_compiled_graph = fx_codegen_and_compile(
^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1129, in fx_codegen_and_compile
return scheme.codegen_and_compile(gm, example_inputs, inputs_to_check, graph_kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 979, in codegen_and_compile
graph.run(*example_inputs)
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/graph.py", line 855, in run
return super().run(*args)
^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/fx/interpreter.py", line 167, in run
self.env[node] = self.run_node(node)
^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/graph.py", line 1496, in run_node
result = super().run_node(n)
^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/fx/interpreter.py", line 230, in run_node
return getattr(self, n.op)(n.target, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/_inductor/graph.py", line 1074, in call_function
return super().call_function(target, args, kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/data/mizhou/workspace/torch_p13n_embedding/.venv_torch26/lib/python3.12/site-packages/torch/fx/interpreter.py", line 310, in call_function
return target(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^
torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised:
IndexError: tuple index out of range
Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information
You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True
```
### Versions
PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A
OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 13.2.0-23ubuntu4) 13.2.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.39
Python version: 3.12.3 (main, Jan 17 2025, 18:03:48) [GCC 13.3.0] (64-bit runtime)
Python platform: Linux-5.10.226-214.879.amzn2.x86_64-x86_64-with-glibc2.39
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA A100-SXM4-40GB
GPU 1: NVIDIA A100-SXM4-40GB
Nvidia driver version: 550.90.12
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.1.1
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.1.1
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 48 bits virtual
CPU(s): 96
On-line CPU(s) list: 0-95
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Platinum 8275CL CPU @ 3.00GHz
Stepping: 7
CPU MHz: 3582.060
BogoMIPS: 5999.99
Hypervisor vendor: KVM
Virtualization type: full
L1d cache: 1.5 MiB
L1i cache: 1.5 MiB
L2 cache: 48 MiB
L3 cache: 71.5 MiB
NUMA node0 CPU(s): 0-23,48-71
NUMA node1 CPU(s): 24-47,72-95
Vulnerability Gather data sampling: Unknown: Dependent on hypervisor status
Vulnerability Itlb multihit: KVM: Mitigation: VMX unsupported
Vulnerability L1tf: Mitigation; PTE Inversion
Vulnerability Mds: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Meltdown: Mitigation; PTI
Vulnerability Mmio stale data: Vulnerable: Clear CPU buffers attempted, no microcode; SMT Host state unknown
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Vulnerable
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Vulnerable
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines, STIBP disabled, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single pti fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid mpx avx512f avx512dq rdseed adx smap clflushopt clwb avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves ida arat pku ospke
Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==2.1.3
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] torch==2.6.0+cu124
[pip3] torchmetrics==1.0.3
[pip3] torchrec==1.1.0+cu124
[pip3] triton==3.2.0
[conda] numpy 1.23.5 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.1.3.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.1.105 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.1.105 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.1.105 pypi_0 pypi
[conda] nvidia-cudnn-cu12 8.9.2.26 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.0.2.54 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.2.106 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.4.5.107 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.1.0.106 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.20.5 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.6.85 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.1.105 pypi_0 pypi
[conda] torch 2.3.1 pypi_0 pypi
[conda] torchaudio 2.3.1 pypi_0 pypi
[conda] torchvision 0.18.1 pypi_0 pypi
[conda] triton 2.3.1 pypi_0 pypi
cc @chauhang @penguinwu | true |
2,816,490,704 | [ONNX] Support subgraphs with 1+ outputs | justinchuby | closed | [
"module: onnx",
"open source",
"Merged",
"ciflow/trunk",
"release notes: onnx",
"topic: bug fixes"
] | 14 | COLLABORATOR | Fixed a bug in _handle_output_node where additional output values were not added as graph outputs
Fixes #145734 | true |
2,816,456,758 | [be][pytorch] Fix backend in autocast | nautsimon | closed | [
"fb-exported",
"module: amp (automated mixed precision)",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 4 | MEMBER | Summary: fixing backend typo (BAKCNEDS -> BACKENDS)
Test Plan: ci
Differential Revision: D68573324
cc @mcarilli @ptrblck @leslie-fang-intel @jgong5 | true |
2,816,409,441 | Copy model before benchmark warmup runs | angelayi | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 5 | CONTRIBUTOR | Fixes https://github.com/pytorch/pytorch/issues/144772
The eager warmup runs causes the model to change state so that later when we export it, the model is different than when we export it directly out of box. For some reason exporting the model with the changed state causes issues but exporting the inital model is ok. This is the reason why the accuracy checks pass but the performance check fails when exporting.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,401,864 | [export] don't always print GM in serdes logging | pianpwk | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"ciflow/inductor",
"release notes: export"
] | 4 | CONTRIBUTOR | Summary: Didn't realize print_readable() would also print and not just return string
Test Plan: .
Differential Revision: D68781525
| true |
2,816,393,274 | [dynamo][builtin-skipfiles-cleanup] Remove some builtins | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 7 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145804
* #145876
* #145909
* #145892
* #145878
* #145875
* __->__ #145856
[dynamo][builtin-skipfiles-cleanup] Remove more builtins
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,816,339,309 | [c10d][ez] Remove goto in PGNCCL and make linter happy for PGNCCL and NCCLUtils | fduwjj | closed | [
"oncall: distributed",
"Merged",
"ciflow/trunk",
"release notes: distributed (c10d)"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145855
While working on PGNCCL I found that the code triggers some lint warnings so this PR is to address them or add lint suppressor.
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,816,289,764 | [c10d][UCC] Support coalesing in `c10d::ProcessGroupUCC` through alltoallv | samnordmann | closed | [
"oncall: distributed",
"triaged",
"open source",
"Stale",
"release notes: distributed (c10d)"
] | 3 | CONTRIBUTOR | # What
Add support for coalescing communication in `ProcessGroupUCC`, by implementing the methods `startCoalescing` and `endCoalescing`.
We need to impose several restrictions to the coalesced group:
1) we can only coalesce `send` and `recv` ops
2) we can only coalesce one `send` and one `recv` maximum per pair of ranks
3) all ranks must participate in the `startCoalescing` and `endCoalescing` calls, even if the group is empty.
4) we do not support tags for coalesced groups.
# Why
Despite the above restrictions, we cover a number of useful data patterns, such as ring p2p, allgather, broadcast, alltoall, etc. Those data patterns or other custom ones are conveniently written in terms of coalesced send/recv calls, which this patch makes possible.
Recall that for a p2p bidirectional transfer, the send and recv need to be coalesced to enjoy full bidirectional bandwidth
# How
Since UCC does not natively support Coalescing, we implement it at the ProcessGroup level. The implementation relies on calling UCC's alltoallv, setting the count to `0` and displacement to `nullptr` for ranks that do not exchange data.
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,816,265,454 | cmake: fix detection logic when using system XNNPACK | Harry-Chen | open | [
"triaged",
"open source",
"Stale",
"topic: not user facing"
] | 6 | NONE | This commit makes the following improvements:
* The "elseif" branch now only runs when USE_XNNPACK is set.
* Use "REQUIRED" to enforce the existence of XNNPACK libraries, and remove the erroneous if statement ('or' should be 'OR').
* libmicrokernels-prod is built statically in XNNPACK [1], change in pytorch side accordingly.
[1]: https://github.com/google/XNNPACK/blob/d7f398ee5e135ef4f7045802eea973cc6cb26c6c/CMakeLists.txt#L819
| true |
2,816,243,859 | RecursionError: maximum recursion depth exceeded in comparison | vpandya-quic | open | [
"triaged",
"oncall: pt2",
"module: dynamic shapes",
"module: inductor"
] | 3 | NONE | ### 🐛 Describe the bug
Running following small tests results in recursion
```py
def test_a():
def process_tensors(
in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, out_ptr1
):
for x0 in range(32):
for x1 in range(6):
for x2 in range(64):
tmp0 = in_ptr0[x0, x1, x2]
tmp1 = in_ptr1[x1, x2]
tmp6 = in_ptr2[x0, x1, x2]
tmp13 = in_ptr3[x0 // 8, x1, x2]
tmp15 = in_ptr4[x0 // 8, x1, x2]
tmp2 = torch.cos(tmp1)
tmp3 = 1.0
tmp4 = tmp2 * tmp3
tmp5 = tmp0 * tmp4
tmp7 = torch.sin(tmp1)
tmp8 = tmp7 * tmp3
tmp9 = tmp6 * tmp8
tmp10 = tmp5 + tmp9
tmp11 = 0.3535533905932738
tmp12 = tmp10 * tmp11
tmp14 = tmp13 * tmp4
tmp16 = tmp15 * tmp8
tmp17 = tmp14 + tmp16
tmp18 = tmp17 * tmp11
out_ptr0[x0, x1, x2] = tmp12
out_ptr1[x0, x1, x2] = tmp18
# Example usage:
with torch.no_grad():
in_ptr0 = torch.randn(32, 6, 64)
in_ptr1 = torch.randn(6, 64)
in_ptr2 = torch.randn(32, 6, 64)
in_ptr3 = torch.randn(4, 6, 64)
in_ptr4 = torch.randn(4, 6, 64)
out_ptr0 = torch.zeros(32, 6, 64)
out_ptr1 = torch.zeros(32, 6, 64)
out_ptr0_ = torch.zeros(32, 6, 64)
out_ptr1_ = torch.zeros(32, 6, 64)
compiled_fn = torch.compile(
process_tensors,
backend="inductor",
fullgraph=True,
)
process_tensors(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0, out_ptr1)
compiled_fn(in_ptr0, in_ptr1, in_ptr2, in_ptr3, in_ptr4, out_ptr0_, out_ptr1_)
```
### Error logs
[log_.txt](https://github.com/user-attachments/files/18576776/log_.txt)
### Versions
PyTorch version: 2.4.1+cpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: 19.1.5 (++20241125104649+086d8e6bb5da-1~exp1~20241125104703.66)
CMake version: version 3.29.0
Libc version: glibc-2.35
Python version: 3.10.12 (main, Nov 6 2024, 20:22:13) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-122-generic-x86_64-with-glibc2.35
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 52 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: AuthenticAMD
Model name: AMD EPYC 9124 16-Core Processor
CPU family: 25
Model: 17
Thread(s) per core: 1
Core(s) per socket: 16
Socket(s): 2
Stepping: 1
Frequency boost: disabled
CPU max MHz: 3711.9141
CPU min MHz: 1500.0000
BogoMIPS: 5991.11
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp ibrs_enhanced vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d
Virtualization: AMD-V
L1d cache: 1 MiB (32 instances)
L1i cache: 1 MiB (32 instances)
L2 cache: 32 MiB (32 instances)
L3 cache: 128 MiB (8 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-15
NUMA node1 CPU(s): 16-31
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Mitigation; safe RET
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] onnx==1.16.0
[pip3] onnxruntime==1.16.3
[pip3] onnxscript==0.1.0.dev20240327
[pip3] torch==2.4.1+cpu
[pip3] torch_geometric==2.5.2
[pip3] torch_qaic==0.1.0
[pip3] torch-tb-profiler==0.4.3
[conda] Could not collect
cc @chauhang @penguinwu @ezyang @bobrenjc93 @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @aakhundov | true |
2,816,210,050 | FIRE Relative Positional Encodings | kaddu341 | open | [
"module: nn",
"triaged"
] | 2 | NONE | ### 🚀 The feature, motivation and pitch
Hi,
I'm currently working on the length generalization capabilities of transformers. As shown by Zhou et al. (https://arxiv.org/abs/2402.09371), FIRE positional encodings are excellent for this purpose as they can yield generalization results up to 2.5x the input length (in combination with other techniques).
FIRE, which stands for Functional Interpolation for Relative Positional Encodings, was introduced by Li et al. (https://arxiv.org/pdf/2310.04418).
I am planning to implement the algorithm from this paper myself, but I thought it would be useful if I could turn it into a PyTorch module so that others can benefit too. Therefore, I am proposing to add this feature to the PyTorch library.
Please let me know what you think!
### Alternatives
There are many other positional encoding types (sinusoidal, RoPE, learned, etc.), but for the specific task of length generalization, FIRE seems to be the most suitable based on several papers, which is why I am proposing this feature addition.
### Additional context
Like other relative attention mechanisms, FIRE introduces positional information in the attention layers rather than adding it to the input (so the layer would subclass nn.Module and basically function as a drop-in replacement for MultiHeadAttention).
Here is a screenshot of some evaluation results for FIRE from the original paper (Li et al., 2024):
<img width="888" alt="Image" src="https://github.com/user-attachments/assets/debfb83d-54e9-4be9-a2d5-6f8c2bf1f473" />
cc @albanD @mruberry @jbschlosser @walterddr @mikaylagawarecki | true |
2,816,207,018 | Skip search for MKL on ARM cpus | malfet | closed | [
"Merged",
"Stale",
"ciflow/trunk",
"topic: not user facing"
] | 4 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145850
It will not find it anyway and makes a bit easier parsing thru CMake log on non-x86 systems | true |
2,816,206,849 | [BE] Include CheckFunctionExists in `FindBLAS.cmake` | malfet | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145850
* __->__ #145849
It's used in the script, so it must be included | true |
2,816,179,114 | Integrate sympy expression provenance logging with structured logs | bobrenjc93 | closed | [
"Merged",
"ciflow/trunk",
"release notes: fx",
"topic: not user facing",
"fx",
"ciflow/inductor"
] | 6 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* __->__ #145848
cc @ezyang @SherlockNoMad @EikanWang @jgong5 @wenzhe-nrv | true |
2,816,159,985 | Guard the CPU cpp wrapper tests on having a cpp wrapper | charlie-wt | open | [
"triaged",
"open source",
"topic: not user facing",
"module: inductor"
] | 12 | CONTRIBUTOR | Since they're the CPU CPP wrapper tests, they should only run if the CPU backend we're using has a CPP wrapper.
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,816,086,780 | [async-TP] Fix scheduling in matmul+reduce-scatter for 2 ranks | lw | closed | [
"oncall: distributed",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 4 | CONTRIBUTOR | There's a sleep that is issued in order to "nudge" CUDA to do the right scheduling decision, but this is issued on iteration number 2. However, when the world size is 2, we never reach that iteration, which led to a suboptimal scheduling.
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,816,074,248 | [OSS] Add kwargs to fsspec reader/writer | ankitageorge | closed | [
"oncall: distributed",
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 8 | CONTRIBUTOR | Summary: Add kwargs to fsspec reader/writer. This will be used when reading/writing from huggingface because it needs a token to access the repositories
Test Plan: https://fburl.com/anp/agkrlas1 ability to read write to hf with fsspec
Differential Revision: D68738777
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o @LucasLLC @MeetVadakkanchery @mhorowitz @pradeepfn @ekr0 | true |
2,816,067,692 | Move trunk windows builds to CUDA-12.4 | atalman | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 6 | CONTRIBUTOR | Same as : https://github.com/pytorch/pytorch/pull/130446
That should catch build regressions that were previously only detectable during the nightly builds for 12.4
| true |
2,816,065,231 | WIP OpenVINOQuantizer | daniil-lyakhov | closed | [
"open source",
"release notes: quantization",
"release notes: AO frontend"
] | 1 | CONTRIBUTOR | Fixes #ISSUE_NUMBER
| true |
2,815,970,732 | [MTIA][FSDP2] Enable MTIA device in FSDP2 library code | jvandebon | closed | [
"oncall: distributed",
"fb-exported",
"Merged",
"ciflow/trunk",
"release notes: distributed (fsdp)",
"ciflow/inductor"
] | 13 | CONTRIBUTOR | Differential Revision: D68560256
cc @H-Huang @awgu @kwen2501 @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k @c-p-i-o | true |
2,815,909,429 | clamp_min did not work | moghadas76 | closed | [
"needs reproduction"
] | 2 | NONE | ### 🐛 Describe the bug
```python
import numpy as np
import torch
file = np.load("tensor.zip")
tensor = torch.from_numpy(file.f.arr_0)
print(tensor)
print(tensor.shape)
print(torch.ge(tensor.clamp_min(1e-6), 0.0).all())
print(torch.ge(tensor.abs()+0.01, 0.0).all())
# print(torch.ge(tensor.abs()+0.01, 0.0).all())
```
return
```
False
False
```
Why?
[tensor.zip](https://github.com/user-attachments/files/18575204/tensor.zip)
### Versions
PyTorch version: 2.3.0+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.4 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35
Python version: 3.11.5 (main, Sep 11 2023, 13:54:46) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-6.8.0-51-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 11.5.119
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA GeForce RTX 4090
GPU 1: NVIDIA GeForce RTX 4090
Nvidia driver version: 555.42.02
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 39 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: GenuineIntel
Model name: 13th Gen Intel(R) Core(TM) i9-13900F
CPU family: 6
Model: 183
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 1
Stepping: 1
CPU max MHz: 5600.0000
CPU min MHz: 800.0000
BogoMIPS: 3993.60
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect user_shstk avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi vnmi umip pku ospke waitpkg gfni vaes vpclmulqdq rdpid movdiri movdir64b fsrm md_clear serialize arch_lbr ibt flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 896 KiB (24 instances)
L1i cache: 1.3 MiB (24 instances)
L2 cache: 32 MiB (12 instances)
L3 cache: 36 MiB (1 instance)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Mitigation; Clear Register File
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] easy-torch==1.3.2
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.24.4
[pip3] nvidia-cublas-cu12==12.1.3.1
[pip3] nvidia-cuda-cupti-cu12==12.1.105
[pip3] nvidia-cuda-nvrtc-cu12==12.1.105
[pip3] nvidia-cuda-runtime-cu12==12.1.105
[pip3] nvidia-cudnn-cu12==8.9.2.26
[pip3] nvidia-cufft-cu12==11.0.2.54
[pip3] nvidia-curand-cu12==10.3.2.106
[pip3] nvidia-cusolver-cu12==11.4.5.107
[pip3] nvidia-cusparse-cu12==12.1.0.106
[pip3] nvidia-nccl-cu12==2.20.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.1.105
[pip3] onnx==1.15.0
[pip3] pytorch-forecasting==1.2.0
[pip3] pytorch-lightning==2.2.0
[pip3] torch==2.3.0
[pip3] torch_cluster==1.6.3+pt23cu121
[pip3] torch_geometric==2.4.0
[pip3] torch_scatter==2.1.2+pt23cu121
[pip3] torch_sparse==0.6.18+pt23cu121
[pip3] torch_spline_conv==1.2.2+pt23cu121
[pip3] torch-summary==1.4.5
[pip3] torchaudio==2.3.0
[pip3] torchinfo==1.8.0
[pip3] torchmetrics==1.3.0.post0
[pip3] torchsummary==1.5.1
[pip3] torchvision==0.18.0
[pip3] triton==2.3.0
[conda] blas 1.0 mkl
[conda] cuda-cudart 12.1.105 0 nvidia
[conda] cuda-cupti 12.1.105 0 nvidia
[conda] cuda-libraries 12.1.0 0 nvidia
[conda] cuda-nvrtc 12.1.105 0 nvidia
[conda] cuda-nvtx 12.1.105 0 nvidia
[conda] cuda-opencl 12.3.52 0 nvidia
[conda] cuda-runtime 12.1.0 0 nvidia
[conda] easy-torch 1.3.2 pypi_0 pypi
[conda] ffmpeg 4.3 hf484d3e_0 pytorch
[conda] libcublas 12.1.0.26 0 nvidia
[conda] libcufft 11.0.2.4 0 nvidia
[conda] libcurand 10.3.4.52 0 nvidia
[conda] libcusolver 11.4.4.55 0 nvidia
[conda] libcusparse 12.0.2.55 0 nvidia
[conda] libjpeg-turbo 2.0.0 h9bf148f_0 pytorch
[conda] libnvjitlink 12.1.105 0 nvidia
[conda] mkl 2023.1.0 h213fc3f_46343
[conda] mkl-service 2.4.0 py311h5eee18b_1
[conda] mkl_fft 1.3.8 py311h5eee18b_0
[conda] mkl_random 1.2.4 py311hdb19cb5_0
[conda] numpy 1.24.4 pypi_0 pypi
[conda] nvidia-cublas-cu12 12.1.3.1 pypi_0 pypi
[conda] nvidia-cuda-cupti-cu12 12.1.105 pypi_0 pypi
[conda] nvidia-cuda-nvrtc-cu12 12.1.105 pypi_0 pypi
[conda] nvidia-cuda-runtime-cu12 12.1.105 pypi_0 pypi
[conda] nvidia-cudnn-cu12 8.9.2.26 pypi_0 pypi
[conda] nvidia-cufft-cu12 11.0.2.54 pypi_0 pypi
[conda] nvidia-curand-cu12 10.3.2.106 pypi_0 pypi
[conda] nvidia-cusolver-cu12 11.4.5.107 pypi_0 pypi
[conda] nvidia-cusparse-cu12 12.1.0.106 pypi_0 pypi
[conda] nvidia-nccl-cu12 2.20.5 pypi_0 pypi
[conda] nvidia-nvjitlink-cu12 12.4.127 pypi_0 pypi
[conda] nvidia-nvtx-cu12 12.1.105 pypi_0 pypi
[conda] pytorch-cuda 12.1 ha16c6d3_5 pytorch
[conda] pytorch-forecasting 1.2.0 pypi_0 pypi
[conda] pytorch-lightning 2.2.0 pypi_0 pypi
[conda] pytorch-mutex 1.0 cuda pytorch
[conda] torch 2.3.0 pypi_0 pypi
[conda] torch-cluster 1.6.3+pt23cu121 pypi_0 pypi
[conda] torch-geometric 2.4.0 pypi_0 pypi
[conda] torch-scatter 2.1.2+pt23cu121 pypi_0 pypi
[conda] torch-sparse 0.6.18+pt23cu121 pypi_0 pypi
[conda] torch-spline-conv 1.2.2+pt23cu121 pypi_0 pypi
[conda] torch-summary 1.4.5 pypi_0 pypi
[conda] torchaudio 2.3.0 pypi_0 pypi
[conda] torchinfo 1.8.0 pypi_0 pypi
[conda] torchmetrics 1.3.0.post0 pypi_0 pypi
[conda] torchsummary 1.5.1 pypi_0 pypi
[conda] torchvision 0.18.0 pypi_0 pypi
[conda] triton 2.3.0 pypi_0 pypi | true |
2,815,877,292 | s390x ci: ensure CI starts correctly if token pipe is not removed | AlekseiNikiforovIBM | closed | [
"triaged",
"open source",
"Merged",
"topic: not user facing"
] | 3 | COLLABORATOR | Mark stop actions as "may fail".
Container is expected to stop on it's own in normal case.
Remove "may fail" mark from token generation steps. | true |
2,815,772,955 | [ATen] Implement exception handling for hipsolver APIs | danzimm | closed | [
"fb-exported",
"Merged",
"ciflow/trunk",
"topic: not user facing"
] | 5 | CONTRIBUTOR | Summary: TSA
Test Plan: CI
Differential Revision: D68741194
| true |
2,815,706,878 | [typing] Not all operators correctly infer Tensor result. | randolf-scholz | closed | [
"module: typing",
"triaged",
"better-engineering",
"needs design"
] | 2 | CONTRIBUTOR | ### 🐛 Describe the bug
<details> <summary> Static Typing Test suite </summary>
```python
# mypy: enable-error-code="unused-ignore"
import operator
import torch
from torch import Tensor
from typing_extensions import assert_type, reveal_type
x = torch.randn(3)
assert_type(x, Tensor)
i64: int = 2
f64: float = 3.14
# op(Tensor, Tensor)
assert_type(x + x, Tensor)
assert_type(x - x, Tensor)
assert_type(x * x, Tensor)
assert_type(x / x, Tensor)
assert_type(x % x, Tensor)
assert_type(x // x, Tensor) # type: ignore[assert-type]
assert_type(x**x, Tensor) # type: ignore[assert-type]
# comparisons
assert_type(x < x, Tensor)
assert_type(x > x, Tensor)
assert_type(x <= x, Tensor)
assert_type(x >= x, Tensor)
assert_type(x == x, Tensor)
assert_type(x != x, Tensor)
# op(Tensor, int)
assert_type(x + i64, Tensor)
assert_type(x - i64, Tensor)
assert_type(x * i64, Tensor)
assert_type(x / i64, Tensor)
assert_type(x % i64, Tensor)
assert_type(x // i64, Tensor) # type: ignore[assert-type]
assert_type(x**i64, Tensor) # type: ignore[assert-type]
assert_type(x < i64, Tensor)
assert_type(x > i64, Tensor)
assert_type(x <= i64, Tensor)
assert_type(x >= i64, Tensor)
assert_type(x == i64, Tensor)
assert_type(x != i64, Tensor)
# op(Tensor, float)
assert_type(x + f64, Tensor)
assert_type(x - f64, Tensor)
assert_type(x * f64, Tensor)
assert_type(x / f64, Tensor)
assert_type(x % f64, Tensor)
assert_type(x // f64, Tensor) # type: ignore[assert-type]
assert_type(x**f64, Tensor) # type: ignore[assert-type]
assert_type(x < f64, Tensor)
assert_type(x > f64, Tensor)
assert_type(x <= f64, Tensor)
assert_type(x >= f64, Tensor)
assert_type(x == f64, Tensor)
assert_type(x != f64, Tensor)
# op(int, Tensor)
assert_type(i64 + x, Tensor)
assert_type(i64 - x, Tensor) # type: ignore[assert-type]
assert_type(i64 * x, Tensor)
assert_type(i64 / x, Tensor) # type: ignore[assert-type]
assert_type(i64 % x, Tensor) # type: ignore[assert-type]
assert_type(i64 // x, Tensor) # type: ignore[assert-type]
assert_type(i64**x, Tensor) # type: ignore[assert-type]
assert_type(i64 < x, Tensor)
assert_type(i64 > x, Tensor)
assert_type(i64 <= x, Tensor)
assert_type(i64 >= x, Tensor)
assert_type(i64 == x, Tensor) # type: ignore[assert-type]
assert_type(i64 != x, Tensor) # type: ignore[assert-type]
# op(float, Tensor)
assert_type(f64 + x, Tensor)
assert_type(f64 - x, Tensor) # type: ignore[assert-type]
assert_type(f64 * x, Tensor)
assert_type(f64 / x, Tensor) # type: ignore[assert-type]
assert_type(f64 % x, Tensor) # type: ignore[assert-type]
assert_type(f64 // x, Tensor) # type: ignore[assert-type]
assert_type(f64**x, Tensor) # type: ignore[assert-type]
assert_type(f64 < x, Tensor)
assert_type(f64 > x, Tensor)
assert_type(f64 <= x, Tensor)
assert_type(f64 >= x, Tensor)
assert_type(f64 == x, Tensor) # type: ignore[assert-type]
assert_type(f64 != x, Tensor) # type: ignore[assert-type]
OPS = [
operator.add, # +
operator.sub, # -
operator.mul, # *
operator.truediv, # /
operator.mod, # %
operator.floordiv, # //
operator.pow, # **
operator.le, # <
operator.gt, # >
operator.lt, # <=
operator.ge, # >=
operator.eq, # ==
operator.ne, # !=
]
for rhs in [x, i64, f64]:
for op in OPS:
assert isinstance(op(x, rhs), Tensor)
for lhs in [x, i64, f64]:
for op in OPS:
assert isinstance(op(lhs, x), Tensor)
```
</details>
## Results
Arithmetic, should be fixable by typing `_handle_torch_function_and_wrap_type_error_to_not_implemented` decorator used in `torch/_tensor.py`.
- [ ] `Tensor // Tensor` inferred as `Any`, not `Tensor`
- [ ] `Tensor ** Tensor` inferred as `Any`, not `Tensor`
- [ ] `Tensor // Number` inferred as `Any`, not `Tensor`
- [ ] `Tensor ** Number` inferred as `Any`, not `Tensor`
- [ ] `Number - Tensor` inferred as `Any`, not `Tensor`
- [ ] `Number / Tensor` inferred as `Any`, not `Tensor`
- [ ] `Number % Tensor` inferred as `Any`, not `Tensor`
- [ ] `Number // Tensor` inferred as `Any`, not `Tensor`
- [ ] `Number ** Tensor` inferred as `Any`, not `Tensor`
Comparisons, possibly unfixable currently
- [ ] `Number == Tensor` inferred as `bool`, not `Tensor`
- [ ] `Number != Tensor` inferred as `bool`, not `Tensor`
### Versions
Tested with both 2.5.1 and 2.7.0.dev20250128+cu124 nightly
cc @ezyang @malfet @xuzhao9 @gramster | true |
2,815,599,877 | torch.nested.narrow() or torch.nested.to_padded_tensor() breaks backwards pass - invalid gradient | kkj15dk | open | [
"triaged",
"module: nestedtensor"
] | 4 | NONE | ### 🐛 Describe the bug
I am converting a jagged nested tensor to a padded tensor, then adding Rotary positional embeddings, then converting it back to a jagged tensor. The forward pass is just fine, but the backwards pass breaks. It is probably because the offsets object changes throughout the forward pass, but I cannot see how to fix this.
If there was a function to convert padded tensor -> jagged tensor, while preserving the original offsets object of the jagged tensor before padding, I think that would be a workaround.
Some code to showcase the bug:
```python
import torch
import torch.nn as nn
def padded_from_jagged(tensor, pad_value=0.0):
offsets = tensor.offsets()
padded = torch.nested.to_padded_tensor(tensor, padding=pad_value)
return padded, offsets
def jagged_from_padded(tensor, offsets, contiguous=True):
seq_lens = offsets.diff()
jagged = torch.nested.narrow(tensor, dim=1, start=0, length=seq_lens, layout=torch.jagged)
if contiguous:
jagged = jagged.contiguous()
return jagged
class test_model(nn.Module):
def __init__(self, dim, max_len):
super().__init__()
self.pos_emb = nn.Parameter(torch.randn(1, max_len, dim), requires_grad=True)
def forward(self, x): # x is a ragged tensor (batch_size=4, j, dim=64), c is a regular tensor (batch_size=4, dim=64)
for i in range(10):
x_padded, offsets = padded_from_jagged(x)
x_padded = x_padded * self.pos_emb
x = jagged_from_padded(x_padded, offsets, contiguous=True)
return x
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
batch_size = 4
max_len = 4096
dim = 64
model = test_model(dim, max_len).to(device)
batch = torch.nested.nested_tensor([torch.randn(max_len - i, dim) for i in range(batch_size)], device=device, layout=torch.jagged) # batch_size=4, j=jagged, dim=64
output = model(batch)
loss = output.mean()
loss.backward()
```
Error message:
```
Traceback (most recent call last):
File "/home/kkj/axolotl/playground/padded_bug.py", line 43, in <module>
loss.backward()
File "/home/kkj/axolotl/.venv/lib/python3.10/site-packages/torch/_tensor.py", line 648, in backward
torch.autograd.backward(
File "/home/kkj/axolotl/.venv/lib/python3.10/site-packages/torch/autograd/__init__.py", line 353, in backward
_engine_run_backward(
File "/home/kkj/axolotl/.venv/lib/python3.10/site-packages/torch/autograd/graph.py", line 814, in _engine_run_backward
return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass
RuntimeError: Function CloneBackward0 returned an invalid gradient at index 0 - got [4, j21, 64] but expected shape compatible with [4, j20, 64]
```
### Versions
--2025-01-28 13:51:32-- https://raw.githubusercontent.com/pytorch/pytorch/main/torch/utils/collect_env.py
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 185.199.111.133, 185.199.110.133, 185.199.108.133, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|185.199.111.133|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 24353 (24K) [text/plain]
Saving to: ‘collect_env.py’
collect_env.py 100%[======================================================================================================================================>] 23.78K --.-KB/s in 0s
2025-01-28 13:51:32 (88.7 MB/s) - ‘collect_env.py’ saved [24353/24353]
Collecting environment information...
PyTorch version: 2.7.0.dev20250122+cu126
Is debug build: False
CUDA used to build PyTorch: 12.6
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35
Python version: 3.10.12 (main, Jan 17 2025, 14:35:34) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.167.4-microsoft-standard-WSL2-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4050 Laptop GPU
Nvidia driver version: 561.19
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 39 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 16
On-line CPU(s) list: 0-15
Vendor ID: GenuineIntel
Model name: 13th Gen Intel(R) Core(TM) i5-13500H
CPU family: 6
Model: 186
Thread(s) per core: 2
Core(s) per socket: 8
Socket(s): 1
Stepping: 2
BogoMIPS: 6374.39
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ss ht syscall nx pdpe1gb rdtscp lm constant_tsc rep_good nopl xtopology tsc_reliable nonstop_tsc cpuid pni pclmulqdq vmx ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch invpcid_single ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves avx_vnni umip waitpkg gfni vaes vpclmulqdq rdpid movdiri movdir64b fsrm md_clear serialize flush_l1d arch_capabilities
Virtualization: VT-x
Hypervisor vendor: Microsoft
Virtualization type: full
L1d cache: 384 KiB (8 instances)
L1i cache: 256 KiB (8 instances)
L2 cache: 10 MiB (8 instances)
L3 cache: 18 MiB (1 instance)
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Reg file data sampling: Mitigation; Clear Register File
Vulnerability Retbleed: Mitigation; Enhanced IBRS
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] numpy==2.1.2
[pip3] nvidia-cublas-cu12==12.6.4.1
[pip3] nvidia-cuda-cupti-cu12==12.6.80
[pip3] nvidia-cuda-nvrtc-cu12==12.6.77
[pip3] nvidia-cuda-runtime-cu12==12.6.77
[pip3] nvidia-cudnn-cu12==9.5.1.17
[pip3] nvidia-cufft-cu12==11.3.0.4
[pip3] nvidia-curand-cu12==10.3.7.77
[pip3] nvidia-cusolver-cu12==11.7.1.2
[pip3] nvidia-cusparse-cu12==12.5.4.2
[pip3] nvidia-cusparselt-cu12==0.6.3
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.6.85
[pip3] nvidia-nvtx-cu12==12.6.77
[pip3] pytorch-triton==3.2.0+git0d4682f0
[pip3] torch==2.7.0.dev20250122+cu126
[pip3] torchaudio==2.6.0.dev20250122+cu126
[pip3] torchvision==0.22.0.dev20250122+cu126
[conda] Could not collect
cc @cpuhrsch @jbschlosser @bhosmer @drisspg @soulitzer @davidberard98 @YuqingJ | true |
2,815,574,245 | DISABLED test_op_dtype_propagation_bitwise_xor_cuda_int64 (__main__.TestCaseCUDA) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: inductor"
] | 2 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_op_dtype_propagation_bitwise_xor_cuda_int64&suite=TestCaseCUDA&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/36265413881).
Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 3 failures and 3 successes.
**Debugging instructions (after clicking on the recent samples link):**
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
1. Click on the workflow logs linked above
2. Click on the Test step of the job so that it is expanded. Otherwise, the grepping will not work.
3. Grep for `test_op_dtype_propagation_bitwise_xor_cuda_int64`
4. There should be several instances run (as flaky tests are rerun in CI) from which you can study the logs.
<details><summary>Sample error message</summary>
```
Traceback (most recent call last):
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1156, in test_wrapper
return test(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 79, in inner
return func(*args, **kwds)
File "/opt/conda/envs/py_3.10/lib/python3.10/contextlib.py", line 79, in inner
return func(*args, **kwds)
File "/var/lib/jenkins/pytorch/test/inductor/test_op_dtype_prop.py", line 81, in test_op_dtype_propagation
self.assertEqual(out, out_c)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 4042, in assertEqual
raise error_metas.pop()[0].to_error( # type: ignore[index]
AssertionError: Scalars are not equal!
Expected 7 but got 7.
Absolute difference: 0
Relative difference: 0.0
The above exception was the direct cause of the following exception:
Traceback (most recent call last):
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 3120, in wrapper
method(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 3120, in wrapper
method(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 454, in instantiated_test
result = test(self, **param_kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_utils.py", line 1620, in wrapper
fn(*args, **kwargs)
File "/opt/conda/envs/py_3.10/lib/python3.10/site-packages/torch/testing/_internal/common_device_type.py", line 1168, in test_wrapper
raise e_tracked from e
Exception: Caused by sample input at index 0: SampleInput(input=Tensor[size=(), device="cuda:0", dtype=torch.int64], args=TensorList[Tensor[size=(), device="cuda:0", dtype=torch.int64]], kwargs={}, broadcasts_input=False, name='')
To execute this test, run the following from the base repo dir:
PYTORCH_OPINFO_SAMPLE_INPUT_INDEX=0 PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_op_dtype_prop.py TestCaseCUDA.test_op_dtype_propagation_bitwise_xor_cuda_int64
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```
</details>
Test file path: `inductor/test_op_dtype_prop.py`
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd @clee2000 @wdvr @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,815,512,350 | s390x: disable test_model_exports_to_core_aten.py test | AlekseiNikiforovIBM | closed | [
"triaged",
"open source",
"Merged",
"topic: not user facing",
"ciflow/s390"
] | 3 | COLLABORATOR | It often gets killed by OOM.
Disable it while investigating.
| true |
2,815,457,152 | [inductor] Add features to docstring_linter (see #142496) | rec | closed | [
"module: lint",
"open source",
"better-engineering",
"Merged",
"ciflow/trunk",
"topic: not user facing",
"suppress-api-compatibility-check",
"suppress-bc-linter"
] | 11 | COLLABORATOR | ## Improvements to `docstring_linter`
* Add a "grandfather list" of existing undocumented classes and functions (`--grandfather`, `--grandfather-tolerance`, `--no-grandfather`, `--write-grandfather`)
* In classes, now just one of the class itself or its `__init__()` method needs to be documented (`--lint-init` turns the old behavior back on)
* Now classes and functions defined local to other functions do not need to be documented (`--lint-local` turns the old behavior back on)
* New `--report` flag produces a compact report of long, undocumented classes or function definitions: see attached example run over all pytorch: [pytorch-docs.json](https://github.com/user-attachments/files/18455981/pytorch-docs.json)
## Help text
```
$ python tools/linter/adapters/docstring_linter.py --help
usage: docstring_linter.py [-h] [-l] [-v] [--grandfather GRANDFATHER] [--grandfather-tolerance GRANDFATHER_TOLERANCE] [--lint-init]
[--lint-local] [--lint-protected] [--max-class MAX_CLASS] [--max-def MAX_DEF]
[--min-docstring MIN_DOCSTRING] [--no-grandfather] [--report] [--write-grandfather]
[files ...]
`docstring_linter` reports on long functions, methods or classes without docstrings
positional arguments:
files A list of files or directories to lint
optional arguments:
-h, --help show this help message and exit
-l, --lintrunner Run for lintrunner and print LintMessages which aren't edits
-v, --verbose Print more debug info
--grandfather GRANDFATHER, -g GRANDFATHER
Set the grandfather list
--grandfather-tolerance GRANDFATHER_TOLERANCE, -t GRANDFATHER_TOLERANCE
Tolerance for grandfather sizes, in percent
--lint-init, -i Lint __init__ and class separately
--lint-local, -o Lint definitions inside other functions
--lint-protected, -p Lint functions, methods and classes that start with _
--max-class MAX_CLASS, -c MAX_CLASS
Maximum number of lines for an undocumented class
--max-def MAX_DEF, -d MAX_DEF
Maximum number of lines for an undocumented function
--min-docstring MIN_DOCSTRING, -s MIN_DOCSTRING
Minimum number of characters for a docstring
--no-grandfather, -n Disable the grandfather list
--report, -r Print a report on all classes and defs
--write-grandfather, -w
Rewrite the grandfather list
```
---
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #148959
* #144622
* #144621
* __->__ #145834
| true |
2,815,247,667 | Implement KL for studentT | moghadas76 | closed | [
"triaged",
"open source",
"Stale"
] | 3 | NONE | Fixes #145729
| true |
2,815,213,433 | [DO NOT MERGE] [TESTING] [ROCm] Triton cherry-picks for AMD backend perf optimisation | jataylo | closed | [
"module: rocm",
"open source",
"ciflow/trunk",
"topic: not user facing",
"ciflow/periodic",
"ciflow/inductor",
"ciflow/rocm",
"ciflow/inductor-rocm"
] | 7 | COLLABORATOR | Testing for rc/3.2.x PR
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @hongxiayang @naromero77amd | true |
2,815,123,002 | Update ET pin to 41e7ffa | GregoryComer | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"ciflow/inductor"
] | 9 | MEMBER | ExecuTorch pin is failing to update due to a change in the executorch install scripts. The previous install_requirements.sh now only installs dependencies and does not build ET. There is a new script - install_executorch.sh, which both installs dependencies and builds the framework.
This PR updates the relevant CI logic to use install_executorch.sh and bumps the pin forward. This should fix the stuck ET pin. | true |
2,815,109,705 | DISABLED test_cat_max_autotune_triton (__main__.TestMaxAutotune) | pytorch-bot[bot] | closed | [
"module: rocm",
"triaged",
"module: flaky-tests",
"skipped",
"oncall: pt2",
"module: inductor"
] | 8 | NONE | Platforms: rocm
This test was disabled because it is failing in CI. See [recent examples](https://hud.pytorch.org/flakytest?name=test_cat_max_autotune_triton&suite=TestMaxAutotune&limit=100) and the most recent trunk [workflow logs](https://github.com/pytorch/pytorch/runs/36261282178).
Over the past 3 hours, it has been determined flaky in 3 workflow(s) with 14 failures and 3 successes.
**Debugging instructions (after clicking on the recent samples link):**
DO NOT ASSUME THINGS ARE OKAY IF THE CI IS GREEN. We now shield flaky tests from developers so CI will thus be green but it will be harder to parse the logs.
To find relevant log snippets:
1. Click on the workflow logs linked above
2. Click on the Test step of the job so that it is expanded. Otherwise, the grepping will not work.
3. Grep for `test_cat_max_autotune_triton`
4. There should be several instances run (as flaky tests are rerun in CI) from which you can study the logs.
<details><summary>Sample error message</summary>
```
Traceback (most recent call last):
File "/var/lib/jenkins/pytorch/test/inductor/test_max_autotune.py", line 825, in test_cat_max_autotune_triton
self._test_cat_max_autotune_impl(using_triton_mm=True)
File "/var/lib/jenkins/pytorch/test/inductor/test_max_autotune.py", line 809, in _test_cat_max_autotune_impl
FileCheck().check("call(").check_count(".run", 2, exactly=True).run(code[0])
RuntimeError: Expected to not find ".run" but found it
# Topologically Sorted Source Nodes: [add], Original ATen: [aten.add]
stream0 = get_raw_stream(0)
triton_poi_fused_add_2.run(buf0, buf3, 1024, grid=grid(1024), stream=stream0)
~~~~ <--- HERE
return (buf2, buf3, )
From CHECK-NOT: .run
To execute this test, run the following from the base repo dir:
PYTORCH_TEST_WITH_ROCM=1 python test/inductor/test_max_autotune.py TestMaxAutotune.test_cat_max_autotune_triton
This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0
```
</details>
Test file path: `inductor/test_max_autotune.py`
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd @clee2000 @wdvr @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @yf225 @chenyang78 @kadeng @muchulee8 @ColinPeppler @amjames @desertfire @chauhang @aakhundov | true |
2,815,054,139 | Ensure GPU isolation for kubernetes pod MI300 runners. | saienduri | closed | [
"module: rocm",
"open source",
"Merged",
"topic: not user facing",
"rocm",
"ciflow/unstable",
"ciflow/rocm"
] | 4 | CONTRIBUTOR | Fixes the reason behind moving the tests to unstable initially. (https://github.com/pytorch/pytorch/pull/145790)
We ensure gpu isolation for each pod within kubernetes by propagating the drivers selected for the pod from the Kubernetes layer up to the docker run in pytorch here.
Now we stick with the GPUs assigned to the pod in the first place and there is no overlap between the test runners.
cc @jeffdaily @sunway513 @jithunnair-amd @pruthvistony @ROCmSupport @dllehr-amd @jataylo @hongxiayang @naromero77amd | true |
2,814,806,838 | [dynamo][builtin-skipfiles-cleanup] Remove posixpath | anijain2305 | closed | [
"Merged",
"ciflow/trunk",
"topic: not user facing",
"module: dynamo",
"ciflow/inductor"
] | 3 | CONTRIBUTOR | Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom):
* #145559
* #145804
* __->__ #145828
* #145826
* #145753
* #145744
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @chauhang @amjames | true |
2,814,801,658 | What changed in torch 2.5.0 that cause photomaker to fail? | CaledoniaProject | open | [
"needs reproduction",
"triaged"
] | 3 | NONE | ### 🐛 Describe the bug
Hi there,
I came from this issue https://github.com/TencentARC/PhotoMaker/issues/205.
I'm playing around with PhotoMaker, and I found their IP adapter failed to work with latest torch (2.5.X). But when I downgrade the pytorch version to 2.4.1 it worked.
So I suspect something changed fundamentally in version 2.5.0. How should I start investigate this issue?
Thanks!
### Versions
Latest | true |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.