[Inductor] Fix combo kernels variable collision with ND tiled reductions#168946
Closed
karthickai wants to merge 17 commits intogh/karthickai/21/basefrom
Closed
[Inductor] Fix combo kernels variable collision with ND tiled reductions#168946karthickai wants to merge 17 commits intogh/karthickai/21/basefrom
karthickai wants to merge 17 commits intogh/karthickai/21/basefrom
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/168946
Note: Links to docs will display an error until the docs builds have been completed. ✅ You can merge normally! (1 Unrelated Failure)As of commit e2067b3 with merge base a208ed2 ( FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This was referenced Nov 24, 2025
…led reductions" Fixes: #168945 cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
This was referenced Nov 29, 2025
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
eellison
approved these changes
Dec 9, 2025
…led reductions" Fixes: #168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` cc voznesenskym penguinwu EikanWang jgong5 Guobing-Chen XiaobingSuper zhuhaozhe blzheng wenzhe-nrv jiayisunx ipiszy chenyang78 kadeng muchulee8 amjames chauhang aakhundov coconutruben mlazos [ghstack-poisoned]
Collaborator
Author
|
@pytorchbot merge |
Collaborator
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
tiendatngcs
pushed a commit
to tiendatngcs/pytorch-Dec25
that referenced
this pull request
Dec 10, 2025
ghstack-source-id: 64aa1c2 Pull Request resolved: pytorch/pytorch#168946
tiendatngcs
pushed a commit
to tiendatngcs/pytorch-Dec25
that referenced
this pull request
Dec 10, 2025
ghstack-source-id: 2af6539 Pull Request resolved: pytorch/pytorch#168946
tiendatngcs
pushed a commit
to tiendatngcs/pytorch-Dec25
that referenced
this pull request
Dec 10, 2025
ghstack-source-id: 2f27d82 Pull Request resolved: pytorch/pytorch#168946
tiendatngcs
pushed a commit
to tiendatngcs/pytorch-Dec25
that referenced
this pull request
Dec 10, 2025
ghstack-source-id: 1958bff Pull Request resolved: pytorch/pytorch#168946
skpark-rh
pushed a commit
to skpark-rh/pytorch
that referenced
this pull request
Dec 10, 2025
…ons (pytorch#168946) Fixes: pytorch#168945 Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed. ```python def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 RBLOCK_0: tl.constexpr = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 RBLOCK_0: tl.constexpr = 8 ^ ValueError('RBLOCK_0 is already defined. constexpr cannot be reassigned.') ``` root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel. after fix generated triton code: ```python import triton import triton.language as tl 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_heuristics.persistent_reduction( size_hints={'x': 4, 'r0_': 2, 'r1_': 8}, reduction_hint=ReductionHint.DEFAULT, filename=__file__, triton_meta={'signature': {'in_out_ptr0': '*fp32', 'in_ptr0': '*fp32', 'out_ptr0': '*fp32', 'XBLOCK': 'constexpr'}, 'device': DeviceProperties(type='cuda', index=0, multi_processor_count=132, cc=90, major=9, regs_per_multiprocessor=65536, max_threads_per_multi_processor=2048, warp_size=32), 'constants': {}, 'configs': [{(0,): [['tt.divisibility', 16]], (1,): [['tt.divisibility', 16]], (2,): [['tt.divisibility', 16]]}]}, inductor_meta={'grid_type': 'RoundRobinComboKernelGrid', 'combo_grid_meta': {'num_kernels': 1, 'min_blocks': 0, 'default_config': None, 'no_x_dim_0': None, 'xnumel_0': 4}, 'kernel_name': 'triton_per_fused_1', 'mutated_arg_names': ['in_out_ptr0'], 'backend_hash': '07C4B3116EC6B0BD20166279782DB98EA71861B79334EBBC8CCB3D36A1E5D7F2', 'assert_indirect_indexing': True, 'autotune_local_cache': True, 'autotune_pointwise': False, 'autotune_remote_cache': None, 'force_disable_caches': True, 'dynamic_scale_rblock': True, 'max_autotune': False, 'max_autotune_pointwise': False, 'min_split_scan_rblock': 256, 'spill_threshold': 16, 'store_cubin': False, 'deterministic': False, 'force_filter_reduction_configs': False, 'are_deterministic_algorithms_enabled': False} ) @triton.jit def triton_per_fused_1(in_out_ptr0, in_ptr0, out_ptr0, XBLOCK : tl.constexpr): pid = tl.program_id(0) if pid % 1 == 0: pid_offset = pid // 1 xnumel = 4 r0_numel = 2 R0_BLOCK_0: tl.constexpr = 2 r1_numel = 8 R1_BLOCK_0: tl.constexpr = 8 rnumel = r0_numel * r1_numel RBLOCK: tl.constexpr = R0_BLOCK_0*R1_BLOCK_0 xoffset = pid_offset * XBLOCK xindex = xoffset + tl.arange(0, XBLOCK)[:, None, None] xmask = xindex < xnumel r0_index = tl.arange(0, R0_BLOCK_0)[None, :, None] r0_offset = 0 r0_mask = r0_index < r0_numel r1_index = tl.arange(0, R1_BLOCK_0)[None, None, :] r1_offset = 0 r1_mask = r1_index < r1_numel roffset = r1_offset + r0_offset*r1_numel rindex = r1_index + r0_index*r1_numel r0_1 = r0_index r1_2 = r1_index x0 = xindex tmp0 = tl.load(in_ptr0 + (r1_2 + 8*x0 + 32*r0_1), r0_mask & r1_mask & xmask, other=0.0) tmp1 = tl.broadcast_to(tmp0, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp3 = tl.where(r0_mask & r1_mask & xmask, tmp1, 0) tmp4 = tl.broadcast_to(tmp1, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp6 = tl.where(r0_mask & r1_mask & xmask, tmp4, 0) tmp7 = tl.reshape(tmp6, [XBLOCK, RBLOCK]) tmp8 = tl.sum(tmp7, 1)[:, None, None].to(tl.float32) tmp9 = tl.full([1, 1, 1], 16, tl.int32) tmp10 = tmp9.to(tl.float32) tmp11 = (tmp8 / tmp10) tmp12 = tmp1 - tmp11 tmp13 = tmp12 * tmp12 tmp14 = tl.broadcast_to(tmp13, [XBLOCK, R0_BLOCK_0, R1_BLOCK_0]) tmp16 = tl.where(r0_mask & r1_mask & xmask, tmp14, 0) tmp17 = tl.reshape(tmp16, [XBLOCK, RBLOCK]) tmp18 = tl.sum(tmp17, 1)[:, None, None].to(tl.float32) tmp19 = 15.0 tmp20 = (tmp18 / tmp19) tl.debug_barrier() tl.store(in_out_ptr0 + (x0), tmp20, xmask) tl.store(out_ptr0 + (x0), tmp11, xmask) else: pass ``` Pull Request resolved: pytorch#168946 Approved by: https://github.com/eellison
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):
Fixes: #168945
Fix combo kernels crash when fusing op with multi-dim reductions (ND tiling). It caused variable name collisions in generated triton code when multiple reduction dimensions existed.
root cause: block size variable generation used sub-kernel index instead of reduction dim prefix, causing collisions when multiple reduction dims existed in the same sub-kernel.
after fix generated triton code:
cc @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @ipiszy @kadeng @muchulee8 @amjames @chauhang @aakhundov @coconutruben @jataylo @mlazos @chenyang78