'tt.expand_dims' op inferred type(s) are incompatible with return type(s) of operation

[Stonepia]

Describe the bug

We witnessed such a bug during in total of 15 tests:

/tmp/torchinductor_pt-gpu/pj/cpjyiy3hzzp5fva5tt2hznb57kep2jnaxubw4yzgwc7hb7cdcjae.py:28:38: error: 'tt.expand_dims' op inferred type(s) 'tensor<1x8xi32, #ttg.linear<{register = [], lane = [[0, 1], [0, 2], [0, 4], [0, 0], [0, 0]], warp = [[0, 0]], block = []}>>' are incompatible with return type(s) of operation 'tensor<1x8xi32, #ttg.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 2], order = [0, 1]}>>'
    r0_index = tl.arange(0, R0_BLOCK)[None, :]
                                     ^
/tmp/torchinductor_pt-gpu/pj/cpjyiy3hzzp5fva5tt2hznb57kep2jnaxubw4yzgwc7hb7cdcjae.py:28:38: error: 'tt.expand_dims' op failed to infer returned types
    r0_index = tl.arange(0, R0_BLOCK)[None, :]
                                     ^
/tmp/torchinductor_pt-gpu/pj/cpjyiy3hzzp5fva5tt2hznb57kep2jnaxubw4yzgwc7hb7cdcjae.py:28:38: note: see current operation: %78 = "tt.expand_dims"(%49) <{axis = 0 : i32}> : (tensor<8xi32, #ttg.slice<{dim = 0, parent = #ttg.linear<{register = [], lane = [[0, 1], [0, 2], [0, 4], [0, 0], [0, 0]], warp = [[0, 0]], block = []}>}>>) -> tensor<1x8xi32, #ttg.blocked<{sizePerThread = [1, 1], threadsPerWarp = [1, 32], warpsPerCTA = [1, 2], order = [0, 1]}>>

The reproducer is :

import torch
from torch._inductor.async_compile import AsyncCompile
import triton
import triton.language as tl
from torch._inductor.runtime.triton_heuristics import (
    grid,
)
from torch._C import _xpu_getCurrentRawStream as get_raw_stream
from torch._C import _xpu_getCurrentRawStream as get_raw_stream

aten = torch.ops.aten
inductor_ops = torch.ops.inductor
_quantized = torch.ops._quantized
assert_size_stride = torch._C._dynamo.guards.assert_size_stride
empty_strided_cpu = torch._C._dynamo.guards._empty_strided_cpu
empty_strided_cuda = torch._C._dynamo.guards._empty_strided_cuda
empty_strided_xpu = torch._C._dynamo.guards._empty_strided_xpu
reinterpret_tensor = torch._C._dynamo.guards._reinterpret_tensor
alloc_from_pool = torch.ops.inductor._alloc_from_pool
async_compile = AsyncCompile()
empty_strided_p2p = torch._C._distributed_c10d._SymmetricMemory.empty_strided_p2p


# kernel path: /tmp/tmp0r9psgjh/pj/cpjyiy3hzzp5fva5tt2hznb57kep2jnaxubw4yzgwc7hb7cdcjae.py
# Topologically Sorted Source Nodes: [argsort], Original ATen: [aten.sort]
# Source node to ATen node mapping:
#   argsort => sort
# Graph fragment:
#   %sort : [num_users=1] = call_function[target=torch.ops.aten.sort.default](args = (%arg0_1, -3, True), kwargs = {})
triton_per_fused_sort_0 = async_compile.triton('triton_per_fused_sort_0', '''
import triton
import triton.language as tl
from triton.compiler.compiler import AttrsDescriptor

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

@triton_heuristics.persistent_reduction(
    size_hints={'x': 32, 'r0_': 8},
    reduction_hint=ReductionHint.DEFAULT,
    filename=__file__,
    triton_meta={'signature': {'in_ptr0': '*fp32', 'out_ptr0': '*i16', 'xnumel': 'i32', 'r0_numel': 'i32'}, 'device': DeviceProperties(type='xpu', index=0, multi_processor_count=64, cc={'architecture': 13136561920, 'driver_version': '1.3.30049+10', 'gpu_eu_count': 512, 'gpu_subslice_count': 64, 'has_atomic64': True, 'has_bfloat16_conversions': True, 'has_fp16': True, 'has_fp64': True, 'has_subgroup_2d_block_io': True, 'has_subgroup_matrix_multiply_accumulate': True, 'has_subgroup_matrix_multiply_accumulate_tensor_float32': False, 'max_compute_units': 512, 'max_num_sub_groups': 64, 'max_work_group_size': 1024, 'name': 'Intel(R) Data Center GPU Max 1550', 'platform_name': 'Intel(R) oneAPI Unified Runtime over Level-Zero', 'sub_group_sizes': [16, 32], 'total_memory': 68719476736, 'type': 'gpu', 'vendor': 'Intel(R) Corporation', 'version': '12.60.7'}, major=None, regs_per_multiprocessor=None, max_threads_per_multi_processor=None, warp_size=32), 'constants': {}, 'configs': [AttrsDescriptor.from_dict({'arg_properties': {'tt.divisibility': [(0,), (1,)], 'tt.equal_to': []}, 'cls': 'AttrsDescriptor'})]},
    inductor_meta={'autotune_hints': set(), 'kernel_name': 'triton_per_fused_sort_0', 'mutated_arg_names': [], 'optimize_mem': True, 'no_x_dim': False, 'num_load': 1, 'num_reduction': 0, 'backend_hash': 'FB40EBCFCB9A06A14744E64088225D5F51204F154F792FD0F599F4DD0B53BAD5', 'are_deterministic_algorithms_enabled': False, 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, '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}
)
@triton.jit
def triton_per_fused_sort_0(in_ptr0, out_ptr0, xnumel, r0_numel, XBLOCK : tl.constexpr):
    xnumel = 25
    r0_numel = 5
    R0_BLOCK: tl.constexpr = 8
    rnumel = r0_numel
    RBLOCK: tl.constexpr = R0_BLOCK
    xoffset = tl.program_id(0) * XBLOCK
    xindex = xoffset + tl.arange(0, XBLOCK)[:, None]
    xmask = xindex < xnumel
    r0_index = tl.arange(0, R0_BLOCK)[None, :]
    r0_offset = 0
    r0_mask = r0_index < r0_numel
    roffset = r0_offset
    rindex = r0_index
    r0_1 = r0_index
    x0 = xindex
    tmp0 = tl.load(in_ptr0 + (x0 + 25*r0_1), r0_mask & xmask, other=0.0)
    tmp1 = r0_1
    tmp2 = tmp1.to(tl.int16)
    tl.static_assert(tmp2.dtype == tl.int16)
    tmp3 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK])
    tmp4 = tl.broadcast_to(tmp2, [XBLOCK, R0_BLOCK])
    tmp5, tmp6, = triton_helpers.sort_with_index(tmp3, tmp4, rnumel, 1, stable=False, descending=True)
    tl.store(out_ptr0 + (x0 + 25*r0_1), tmp6, r0_mask & xmask)
''', device_str='xpu')



async_compile.wait(globals())
del async_compile

def call(args):
    arg0_1, = args
    args.clear()
    assert_size_stride(arg0_1, (5, 5, 5), (25, 5, 1))
    with torch.xpu._DeviceGuard(0):
        torch.xpu.set_device(0)
        buf1 = empty_strided_xpu((5, 5, 5), (25, 5, 1), torch.int16)
        # Topologically Sorted Source Nodes: [argsort], Original ATen: [aten.sort]
        stream0 = get_raw_stream(0)
        triton_per_fused_sort_0.run(arg0_1, buf1, 25, 5, grid=grid(25), stream=stream0)
        del arg0_1
    return (buf1,)



def benchmark_compiled_module(times=10, repeat=10):
    from torch._dynamo.testing import rand_strided
    from torch._inductor.utils import print_performance
    arg0_1 = rand_strided((5, 5, 5), (25, 5, 1), device='xpu:0', dtype=torch.float32)
    fn = lambda: call([arg0_1])
    return print_performance(fn, times=times, repeat=repeat)


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

Environment details

Triton: latest main (b59bb9a)
PyTorch: latest main (d0f5df83a50d9bb630764c92ac63fcb2640b1f94) + triton patch
OS: Ubuntu 24.10 / Windows 11

[riverliuintel]

@vlad-penkin new bugs for PyTorch 2.7 upstream.

[anmyachev]

Hi @Stonepia!

Could you restart inductor generation with the latest patches for Pytorch? I made fixes for PyTorch in c74534d.

[anmyachev]

Hi @Stonepia! Do you have any updates?

[Stonepia]

Hi @anmyachev , Thanks for the fix!
Sorry for my late reply. I think that problem is fixed.
Meanwhile, we are going to take a round of full tests, I will close this issue once the report is out.