# torch/_inductor/pattern_matcher.py
# NB: This assertion might not be true in general, but it is true for
# the two use cases we have
# (triton_kernel_wrapper_functional, auto_functionalized)
assert len(graph_with_eager_vals.graph.nodes) == len(
    replacement.graph.nodes
)

---

compute_required_storage_length(shape, strides, offset)
# = 1 + offset + sum((dim - 1) * stride for dim, stride in zip(shape, strides))

---

import torch
import triton
import triton.language as tl


@triton.jit
def _fill_kernel(
    buf_ptr,
    num_tokens,
    stride_t,
    stride_i,
    INNER: tl.constexpr,
):
    pid = tl.program_id(0)
    if pid >= num_tokens:
        return
    for i in range(INNER):
        tl.store(buf_ptr + pid * stride_t + i * stride_i, 1.0)


def fn(x: torch.Tensor) -> torch.Tensor:
    num_tokens = x.shape[0]
    n_groups = 1            # ← (1) size-1 leading dim
    inner = 8
    aligned_T = ((num_tokens + 3) // 4) * 4   # padded length, SymInt-derived

    # (2) as_strided view; stride[0] = inner * aligned_T is a SymInt expression.
    buf = torch.empty(
        n_groups * inner * aligned_T,
        dtype=torch.float32,
        device=x.device,
    ).as_strided(
        (n_groups, num_tokens, inner),
        (inner * aligned_T, 1, aligned_T),
    )

    # (3) triton kernel mutates `buf` -> goes into tensors_to_clone.
    _fill_kernel[(num_tokens,)](
        buf,
        num_tokens,
        stride_t=buf.stride(1),
        stride_i=buf.stride(2),
        INNER=inner,
    )
    return buf + x.sum()


def main() -> None:
    torch.set_default_device("cuda")
    compiled = torch.compile(fn, dynamic=True, fullgraph=True)
    for n in (64, 128, 256):
        x = torch.randn(n, 8)
        out = compiled(x)
        torch.cuda.synchronize()
        print(f"n={n} ok, sum={out.sum().item():.2f}")


if __name__ == "__main__":
    main()

---

Traceback (most recent call last):
  File "/tmp/repro_inductor_bug.py", line 94, in <module>
    main()
  File "/tmp/repro_inductor_bug.py", line 88, in main
    out = compiled(x)
          ^^^^^^^^^^^
  File "/home/zjy/vllm/.venv/lib/python3.12/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/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1053, in _compile_fx_inner
    raise InductorError(e, currentframe()).with_traceback(
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1037, in _compile_fx_inner
    mb_compiled_graph = fx_codegen_and_compile(
                        ^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1798, in fx_codegen_and_compile
    return scheme.codegen_and_compile(gm, example_inputs, inputs_to_check, graph_kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 1344, in codegen_and_compile
    _recursive_post_grad_passes(gm, is_inference=is_inference)
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/compile_fx.py", line 583, in _recursive_post_grad_passes
    post_grad_passes(gm, is_inference)
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/fx_passes/post_grad.py", line 357, in post_grad_passes
    ).apply_graph_pass(decompose_triton_kernel_wrapper_functional)
      ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/fx/passes/graph_transform_observer.py", line 103, in apply_graph_pass
    return pass_fn(self.gm.graph)
           ^^^^^^^^^^^^^^^^^^^^^^
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/fx_passes/post_grad.py", line 1255, in decompose_triton_kernel_wrapper_functional
    graph_pass.apply(graph)
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/pattern_matcher.py", line 2063, in apply
    entry.apply(m, graph, node)
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/pattern_matcher.py", line 1132, in apply
    self.handler(match, *match.args, **match.kwargs)
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/fx_passes/post_grad.py", line 1253, in _
    match.replace_by_example(decomp, flat_args, run_functional_passes=False)
  File "/home/zjy/vllm/.venv/lib/python3.12/site-packages/torch/_inductor/pattern_matcher.py", line 316, in replace_by_example
    assert len(graph_with_eager_vals.graph.nodes) == len(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
torch._inductor.exc.InductorError: AssertionError: 

Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"