Issue Station
Sign In
Tools hereToolbox

pytorch - 💡(How to fix) Fix [aot_autograd] RuntimeError: derivative for aten::_scaled_dot_product_flash_attention_for_cpu_backward is not implemented [1 participants]

Official PRs (…)
ON THIS PAGE

Recommended Tools

×6

Utilities matched from this issue’s tags and category — try them while you read without losing context.

GitHub issue graph ai analysis

Paste a GitHub issue URL. We fetch that issue, discover linked issues from bodies/comments/timeline, collect linked pull requests, and produce a structured English report.

The report is written in English Markdown for sharing and archival.

Helpful · Quick feedback

Loading…
GitHub stats
pytorch/pytorch#181177•Fetched 2026-04-23 07:22:08
View on GitHub
Comments
0
Participants
1
Timeline
109
Reactions
0
Author
jasminetrailjasminetrail
Participants
jasminetrailjasminetrail
Timeline (top)
mentioned ×50subscribed ×50labeled ×9

Error Message

Error logs

/home/.venv/lib/python3.10/site-packages/torch/autograd/graph.py:869: UserWarning: Error detected in torch::autograd::NotImplemented. Traceback of forward call that caused the error: Previous calculation was induced by ScaledDotProductFlashAttentionForCpuBackward0. Traceback of forward call that induced the previous calculation: Traceback (most recent call last):

Fix Action

Fix / Workaround

Previous calculation was induced by ScaledDotProductFlashAttentionForCpuBackward0. Traceback of forward call that induced the previous calculation: File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/apis.py", line 433, in wrapper return eager_transforms.grad_impl(func, argnums, has_aux, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1506, in grad_impl results = grad_and_value_impl(func, argnums, has_aux, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/vmap.py", line 56, in fn return f(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1457, in grad_and_value_impl output = func(*args, **kwargs) File "/home/bugs/crash_29f934d5_reporting.py", line 15, in _fn t338 = F.scaled_dot_product_attention(t332, t332, t332) (Triggered internally at /pytorch/torch/csrc/autograd/python_anomaly_mode.cpp:129.) return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass Traceback (most recent call last): File "/home/bugs/crash_29f934d5_reporting.py", line 22, in <module> _fgrad_compiled = torch.compile(_torch_func.grad(_fn))(t330) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1038, in compile_wrapper raise e.remove_dynamo_frames() from None # see TORCHDYNAMO_VERBOSE=1 File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1024, in compile_wrapper return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 2316, in call result = self._torchdynamo_orig_backend( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 2052, in call result = self._inner_convert( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 729, in call result = _compile( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1827, in _compile guarded_code, tracer_output = compile_inner(code, one_graph, hooks) File "/home/.venv/lib/python3.10/site-packages/torch/_utils_internal.py", line 96, in wrapper_function return function(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1500, in compile_inner return _compile_inner(code, one_graph, hooks) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1534, in _compile_inner dynamo_output = compile_frame( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1408, in compile_frame bytecode, tracer_output = transform_code_object(code, transform) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/bytecode_transformation.py", line 1608, in transform_code_object tracer_output = transformations(instructions, code_options) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1380, in transform tracer_output = trace_frame( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 341, in _fn return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 863, in trace_frame run_tracer() File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 844, in run_tracer tracer.run() File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1794, in run while self.step(): File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1459, in step self.dispatch_table[inst.opcode](self, inst) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 5570, in RETURN_VALUE self._return(inst) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 5552, in _return all_stack_locals_metadata = self.output.compile_subgraph( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 1889, in compile_subgraph self.compile_and_call_fx_graph(tx, pass2.graph_output_vars(), root) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2460, in compile_and_call_fx_graph compiled_fn = self.call_user_compiler(gm, self.example_inputs()) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2613, in call_user_compiler return self._call_user_compiler(gm, example_inputs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2689, in _call_user_compiler raise BackendCompilerFailed( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2664, in call_user_compiler compiled_fn = compiler_fn(gm, example_inputs) File "/home/.venv/lib/python3.10/site-packages/torch/dynamo/repro/after_dynamo.py", line 156, in call compiled_gm = compiler_fn(gm, example_inputs) File "/home/.venv/lib/python3.10/site-packages/torch/init.py", line 2461, in call return compile_fx(model, inputs, config_patches=self.config) File "/home/.venv/lib/python3.10/site-packages/torch/_inductor/compile_fx.py", line 2578, in compile_fx return _maybe_wrap_and_compile_fx_main( File "/home/.venv/lib/python3.10/site-packages/torch/_inductor/compile_fx.py", line 2655, in _maybe_wrap_and_compile_fx_main return _compile_fx_main( File "/home/.venv/lib/python3.10/site-packages/torch/_inductor/compile_fx.py", line 2850, in _compile_fx_main return aot_autograd( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/backends/common.py", line 124, in call cg = aot_module_simplified(gm, example_inputs, **self.kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py", line 1160, in aot_module_simplified aot_graph_capture = aot_stage1_graph_capture(aot_state, functional_call) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_compile.py", line 223, in aot_stage1_graph_capture ) = aot_dispatch_autograd_graph( File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture.py", line 482, in aot_dispatch_autograd_graph fx_g = _create_graph( File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture.py", line 96, in _create_graph fx_g = make_fx( File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 2824, in wrapped return make_fx_tracer.trace(f, *args) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 2725, in trace return self._trace_inner(f, *args) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 2686, in _trace_inner t = dispatch_trace( File "/home/.venv/lib/python3.10/site-packages/torch/_compile.py", line 54, in inner return disable_fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1263, in _fn return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 1533, in dispatch_trace graph = tracer.trace(root, concrete_args) # type: ignore[arg-type] File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1263, in _fn return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py", line 890, in trace (self.create_arg(fn(*args)),), File "/home/.venv/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py", line 735, in flatten_fn tree_out = root_fn(*tree_args) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 1603, in wrapped out = f(*tensors) # type:ignore[call-arg] File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture.py", line 79, in inner_f out, out_descs = call_and_expect_output_descs(f, args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 785, in call_and_expect_output_descs outs_pair = fn(*args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 1215, in inner_fn outs, outs_descs = call_and_expect_output_descs(fn, args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 785, in call_and_expect_output_descs outs_pair = fn(*args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 1157, in joint_helper return _functionalized_f_helper(primals, tangents) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 892, in _functionalized_f_helper f_outs, f_outs_descs = call_and_expect_output_descs(fn, f_args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 785, in call_and_expect_output_descs outs_pair = fn(*args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 495, in joint_helper return inner_fn_with_anomaly(primals, tangents) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 487, in inner_fn_with_anomaly return inner_fn(primals, tangents) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 451, in inner_fn backward_out = torch.autograd.grad( File "/home/.venv/lib/python3.10/site-packages/torch/autograd/init.py", line 481, in grad return handle_torch_function( File "/home/.venv/lib/python3.10/site-packages/torch/overrides.py", line 1741, in handle_torch_function result = mode.torch_function(public_api, types, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 1654, in torch_function return func(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/autograd/init.py", line 532, in grad result = _engine_run_backward( File "/home/.venv/lib/python3.10/site-packages/torch/autograd/graph.py", line 869, in _engine_run_backward return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised: RuntimeError: derivative for aten::_scaled_dot_product_flash_attention_for_cpu_backward is not implemented

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): 384 On-line CPU(s) list: 0-383 Vendor ID: AuthenticAMD Model name: AMD EPYC 9684X 96-Core Processor CPU family: 25 Model: 17 Thread(s) per core: 2 Core(s) per socket: 96 Socket(s): 2 Stepping: 2 BogoMIPS: 5099.98 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 amd_lbr_v2 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 hw_pstate ssbd mba perfmon_v2 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 user_shstk avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc amd_ibpb_ret arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d debug_swap ibpb_exit_to_user Virtualization: AMD-V L1d cache: 6 MiB (192 instances) L1i cache: 6 MiB (192 instances) L2 cache: 192 MiB (192 instances) L3 cache: 2.3 GiB (24 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-95,192-287 NUMA node1 CPU(s): 96-191,288-383 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 Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

Code Example

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.func as _torch_func

_m_t331 = nn.LayerNorm([3])
_m_t332 = nn.RReLU().eval()

torch.manual_seed(0)
t330 = torch.randn([5, 15, 9, 3])

def _fn(t330):
    t331 = _m_t331(t330)
    t332 = _m_t332(t331)
    t338 = F.scaled_dot_product_attention(t332, t332, t332)
    return t338.mean()

# Eager mode - works fine
_fgrad_eager = _torch_func.grad(_fn)(t330)

# Compiled mode - crashes
_fgrad_compiled = torch.compile(_torch_func.grad(_fn))(t330)
assert torch.allclose(_fgrad_eager, _fgrad_compiled, atol=1e-4, rtol=1e-4), (
    f'func.grad eager/compiled mismatch: max_diff={(_fgrad_eager - _fgrad_compiled).abs().max().item():.6f}'
)
RAW_BUFFERClick to expand / collapse

🐛 Describe the bug

When composing torch.func.grad with torch.compile() over a function containing LayerNorm -> RReLU -> scaled_dot_product_attention, the inductor backend crashes during AOT autograd graph capture with RuntimeError: derivative for aten::_scaled_dot_product_flash_attention_for_cpu_backward is not implemented. The same function works correctly under eager execution.

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.func as _torch_func

_m_t331 = nn.LayerNorm([3])
_m_t332 = nn.RReLU().eval()

torch.manual_seed(0)
t330 = torch.randn([5, 15, 9, 3])

def _fn(t330):
    t331 = _m_t331(t330)
    t332 = _m_t332(t331)
    t338 = F.scaled_dot_product_attention(t332, t332, t332)
    return t338.mean()

# Eager mode - works fine
_fgrad_eager = _torch_func.grad(_fn)(t330)

# Compiled mode - crashes
_fgrad_compiled = torch.compile(_torch_func.grad(_fn))(t330)
assert torch.allclose(_fgrad_eager, _fgrad_compiled, atol=1e-4, rtol=1e-4), (
    f'func.grad eager/compiled mismatch: max_diff={(_fgrad_eager - _fgrad_compiled).abs().max().item():.6f}'
)

Error logs

/home/.venv/lib/python3.10/site-packages/torch/autograd/graph.py:869: UserWarning: Error detected in torch::autograd::NotImplemented. Traceback of forward call that caused the error: File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/apis.py", line 433, in wrapper return eager_transforms.grad_impl(func, argnums, has_aux, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1506, in grad_impl results = grad_and_value_impl(func, argnums, has_aux, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/vmap.py", line 56, in fn return f(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1483, in grad_and_value_impl flat_grad_input = _autograd_grad( (Triggered internally at /pytorch/torch/csrc/autograd/python_anomaly_mode.cpp:122.) return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass /home/.venv/lib/python3.10/site-packages/torch/autograd/graph.py:869: UserWarning:

Previous calculation was induced by ScaledDotProductFlashAttentionForCpuBackward0. Traceback of forward call that induced the previous calculation: File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/apis.py", line 433, in wrapper return eager_transforms.grad_impl(func, argnums, has_aux, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1506, in grad_impl results = grad_and_value_impl(func, argnums, has_aux, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/vmap.py", line 56, in fn return f(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/eager_transforms.py", line 1457, in grad_and_value_impl output = func(*args, **kwargs) File "/home/bugs/crash_29f934d5_reporting.py", line 15, in _fn t338 = F.scaled_dot_product_attention(t332, t332, t332) (Triggered internally at /pytorch/torch/csrc/autograd/python_anomaly_mode.cpp:129.) return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass Traceback (most recent call last): File "/home/bugs/crash_29f934d5_reporting.py", line 22, in <module> _fgrad_compiled = torch.compile(_torch_func.grad(_fn))(t330) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1038, in compile_wrapper raise e.remove_dynamo_frames() from None # see TORCHDYNAMO_VERBOSE=1 File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1024, in compile_wrapper return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 2316, in call result = self._torchdynamo_orig_backend( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 2052, in call result = self._inner_convert( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 729, in call result = _compile( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1827, in _compile guarded_code, tracer_output = compile_inner(code, one_graph, hooks) File "/home/.venv/lib/python3.10/site-packages/torch/_utils_internal.py", line 96, in wrapper_function return function(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1500, in compile_inner return _compile_inner(code, one_graph, hooks) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1534, in _compile_inner dynamo_output = compile_frame( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1408, in compile_frame bytecode, tracer_output = transform_code_object(code, transform) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/bytecode_transformation.py", line 1608, in transform_code_object tracer_output = transformations(instructions, code_options) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 1380, in transform tracer_output = trace_frame( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 341, in _fn return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 863, in trace_frame run_tracer() File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/convert_frame.py", line 844, in run_tracer tracer.run() File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1794, in run while self.step(): File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 1459, in step self.dispatch_table[inst.opcode](self, inst) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 5570, in RETURN_VALUE self._return(inst) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/symbolic_convert.py", line 5552, in _return all_stack_locals_metadata = self.output.compile_subgraph( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 1889, in compile_subgraph self.compile_and_call_fx_graph(tx, pass2.graph_output_vars(), root) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2460, in compile_and_call_fx_graph compiled_fn = self.call_user_compiler(gm, self.example_inputs()) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2613, in call_user_compiler return self._call_user_compiler(gm, example_inputs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2689, in _call_user_compiler raise BackendCompilerFailed( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/output_graph.py", line 2664, in call_user_compiler compiled_fn = compiler_fn(gm, example_inputs) File "/home/.venv/lib/python3.10/site-packages/torch/dynamo/repro/after_dynamo.py", line 156, in call compiled_gm = compiler_fn(gm, example_inputs) File "/home/.venv/lib/python3.10/site-packages/torch/init.py", line 2461, in call return compile_fx(model, inputs, config_patches=self.config) File "/home/.venv/lib/python3.10/site-packages/torch/_inductor/compile_fx.py", line 2578, in compile_fx return _maybe_wrap_and_compile_fx_main( File "/home/.venv/lib/python3.10/site-packages/torch/_inductor/compile_fx.py", line 2655, in _maybe_wrap_and_compile_fx_main return _compile_fx_main( File "/home/.venv/lib/python3.10/site-packages/torch/_inductor/compile_fx.py", line 2850, in _compile_fx_main return aot_autograd( File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/backends/common.py", line 124, in call cg = aot_module_simplified(gm, example_inputs, **self.kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/aot_autograd.py", line 1160, in aot_module_simplified aot_graph_capture = aot_stage1_graph_capture(aot_state, functional_call) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_compile.py", line 223, in aot_stage1_graph_capture ) = aot_dispatch_autograd_graph( File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture.py", line 482, in aot_dispatch_autograd_graph fx_g = _create_graph( File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture.py", line 96, in _create_graph fx_g = make_fx( File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 2824, in wrapped return make_fx_tracer.trace(f, *args) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 2725, in trace return self._trace_inner(f, *args) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 2686, in _trace_inner t = dispatch_trace( File "/home/.venv/lib/python3.10/site-packages/torch/_compile.py", line 54, in inner return disable_fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1263, in _fn return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 1533, in dispatch_trace graph = tracer.trace(root, concrete_args) # type: ignore[arg-type] File "/home/.venv/lib/python3.10/site-packages/torch/_dynamo/eval_frame.py", line 1263, in _fn return fn(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py", line 890, in trace (self.create_arg(fn(*args)),), File "/home/.venv/lib/python3.10/site-packages/torch/fx/_symbolic_trace.py", line 735, in flatten_fn tree_out = root_fn(*tree_args) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 1603, in wrapped out = f(*tensors) # type:ignore[call-arg] File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture.py", line 79, in inner_f out, out_descs = call_and_expect_output_descs(f, args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 785, in call_and_expect_output_descs outs_pair = fn(*args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 1215, in inner_fn outs, outs_descs = call_and_expect_output_descs(fn, args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 785, in call_and_expect_output_descs outs_pair = fn(*args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 1157, in joint_helper return _functionalized_f_helper(primals, tangents) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 892, in _functionalized_f_helper f_outs, f_outs_descs = call_and_expect_output_descs(fn, f_args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/utils.py", line 785, in call_and_expect_output_descs outs_pair = fn(*args) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 495, in joint_helper return inner_fn_with_anomaly(primals, tangents) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 487, in inner_fn_with_anomaly return inner_fn(primals, tangents) File "/home/.venv/lib/python3.10/site-packages/torch/_functorch/_aot_autograd/graph_capture_wrappers.py", line 451, in inner_fn backward_out = torch.autograd.grad( File "/home/.venv/lib/python3.10/site-packages/torch/autograd/init.py", line 481, in grad return handle_torch_function( File "/home/.venv/lib/python3.10/site-packages/torch/overrides.py", line 1741, in handle_torch_function result = mode.torch_function(public_api, types, args, kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/fx/experimental/proxy_tensor.py", line 1654, in torch_function return func(*args, **kwargs) File "/home/.venv/lib/python3.10/site-packages/torch/autograd/init.py", line 532, in grad result = _engine_run_backward( File "/home/.venv/lib/python3.10/site-packages/torch/autograd/graph.py", line 869, in _engine_run_backward return Variable._execution_engine.run_backward( # Calls into the C++ engine to run the backward pass torch._dynamo.exc.BackendCompilerFailed: backend='inductor' raised: RuntimeError: derivative for aten::_scaled_dot_product_flash_attention_for_cpu_backward is not implemented

Versions

Collecting environment information... PyTorch version: 2.11.0+cu130 Is debug build: False CUDA used to build PyTorch: 13.0 ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64) GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.3) 11.4.0 Clang version: 15.0.0 ([email protected]:llvm/llvm-project.git 4ba6a9c9f65bbc8bd06e3652cb20fd4dfc846137) CMake version: version 3.22.1 Libc version: glibc-2.35

Python version: 3.10.12 (main, Mar 3 2026, 11:56:32) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-6.8.0-94-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 Is XPU available: False HIP runtime version: N/A MIOpen runtime version: N/A Is XNNPACK available: True Caching allocator config: N/A

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): 384 On-line CPU(s) list: 0-383 Vendor ID: AuthenticAMD Model name: AMD EPYC 9684X 96-Core Processor CPU family: 25 Model: 17 Thread(s) per core: 2 Core(s) per socket: 96 Socket(s): 2 Stepping: 2 BogoMIPS: 5099.98 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 amd_lbr_v2 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 hw_pstate ssbd mba perfmon_v2 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 user_shstk avx512_bf16 clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin cppc amd_ibpb_ret arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif x2avic v_spec_ctrl vnmi avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg avx512_vpopcntdq la57 rdpid overflow_recov succor smca fsrm flush_l1d debug_swap ibpb_exit_to_user Virtualization: AMD-V L1d cache: 6 MiB (192 instances) L1i cache: 6 MiB (192 instances) L2 cache: 192 MiB (192 instances) L3 cache: 2.3 GiB (24 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-95,192-287 NUMA node1 CPU(s): 96-191,288-383 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 Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization Vulnerability Spectre v2: Mitigation; Enhanced / Automatic IBRS; IBPB conditional; STIBP always-on; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected Vulnerability Srbds: Not affected Vulnerability Tsx async abort: Not affected Vulnerability Vmscape: Mitigation; IBPB before exit to userspace

Versions of relevant libraries: [pip3] numpy==2.2.6 [pip3] nvidia-cublas==13.1.0.3 [pip3] nvidia-cuda-cupti==13.0.85 [pip3] nvidia-cuda-nvrtc==13.0.88 [pip3] nvidia-cuda-runtime==13.0.96 [pip3] nvidia-cudnn-cu13==9.19.0.56 [pip3] nvidia-cufft==12.0.0.61 [pip3] nvidia-curand==10.4.0.35 [pip3] nvidia-cusolver==12.0.4.66 [pip3] nvidia-cusparse==12.6.3.3 [pip3] nvidia-cusparselt-cu13==0.8.0 [pip3] nvidia-nccl-cu13==2.28.9 [pip3] nvidia-nvjitlink==13.0.88 [pip3] nvidia-nvtx==13.0.85 [pip3] torch==2.11.0 [pip3] triton==3.6.0 [conda] Could not collect

cc @jgong5 @mingfeima @XiaobingSuper @sanchitintel @ashokei @jingxu10 @jerryzh168 @aditew01 @chauhang @penguinwu @Chillee @samdow @kshitij12345 @bdhirsh @bobrenjc93 @aorenste @drisspg @liangel-02 @howardzhang-cv

extent analysis

TL;DR

The issue is likely due to a missing derivative implementation for aten::_scaled_dot_product_flash_attention_for_cpu_backward in the PyTorch inductor backend, and a potential workaround is to avoid using torch.compile() with torch.func.grad() for functions containing scaled_dot_product_attention.

Guidance

  • The error message indicates that the derivative for aten::_scaled_dot_product_flash_attention_for_cpu_backward is not implemented, which suggests that the inductor backend does not support this operation.
  • To verify this, try using torch.func.grad() without torch.compile() to see if the issue persists.
  • As a potential workaround, consider using torch.func.grad() without torch.compile() or using a different backend that supports the required operations.
  • If the issue is specific to the inductor backend, consider reporting it to the PyTorch developers or searching for existing issues related to this backend.

Example

No code example is provided as the issue is related to a specific PyTorch backend and operation, and the original code snippet is already provided in the issue body.

Notes

The inductor backend is a relatively new feature in PyTorch, and it may not support all operations or have all the necessary derivatives implemented. The issue may be specific to this backend, and using a different backend or avoiding torch.compile() may resolve the issue.

Recommendation

Apply workaround: Avoid using torch.compile() with torch.func.grad() for functions containing scaled_dot_product_attention until the inductor backend supports the required operations.

Vote matrix · Quick signals

Works
Did the solution work? Tap to confirm.
Easy Fix
Was it a quick fix?
Time Saver
Did it save you time?
Blocking
Was it severely blocking?
Common Issue
Are others likely hitting this too?
Flaky / Intermittent
Is it intermittent?
Verified / Reproducible
Can you reproduce it reliably?
Loading…
#api#autograd error#model save/load#optimization#mixed precision

Still need to ship something?

×6

Another batch ranked right after the header list — different links, same matching logic.

Back to top recommendations

TRENDING