Add files using upload-large-folder tool

ckpts/universal/global_step120/zero/1.word_embeddings.weight/exp_avg_sq.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:69bc159dc60ca0fa15d189e9ea5a5afe2fdb1fd1843f43d423706e3ef48e0582
+size 415237419

ckpts/universal/global_step120/zero/10.input_layernorm.weight/exp_avg.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:17eaa0cccd5fef6a55674225f15d60f3ae6ea4c4a93b3558422135b52d01ee46
+size 9372

ckpts/universal/global_step120/zero/10.input_layernorm.weight/exp_avg_sq.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1dc0c4eb90385736831009786614e41bbc14440018e0ca45547e827d894162a
+size 9387

ckpts/universal/global_step120/zero/10.mlp.dense_h_to_4h_swiglu.weight/exp_avg.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dc37920f0413b3804caac3df9bab38fd1ba77bb527e0a092ca06fa4c268c1dd9
+size 33555612

ckpts/universal/global_step120/zero/10.mlp.dense_h_to_4h_swiglu.weight/exp_avg_sq.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bed44ee6d530f1d12e0af63406c2de39ddcd545f5234db79633ad832537d5fae
+size 33555627

ckpts/universal/global_step120/zero/10.mlp.dense_h_to_4h_swiglu.weight/fp32.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a2890cb784a3591974935bb944c6b8f4ad401da0bc8fce465a4714a8c04f911
+size 33555533

ckpts/universal/global_step120/zero/19.post_attention_layernorm.weight/exp_avg_sq.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7532df6a0350673df2366907c0091363072a8a4b9f12bbf9a9793c7c8e5f64e
+size 9387

ckpts/universal/global_step120/zero/4.post_attention_layernorm.weight/exp_avg.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cbc8c554279d3831a4aaba7844cb92c0de84e792dc6406c762163ca7974c1a59
+size 9372

ckpts/universal/global_step120/zero/7.mlp.dense_4h_to_h.weight/exp_avg.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4d48d36ff4c24763c96242ad3ff413431a60d7ccb26095c441b892dfffdcc692
+size 33555612

ckpts/universal/global_step120/zero/7.mlp.dense_4h_to_h.weight/fp32.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:45ee432c24153973ef44a9001e0b6398de0380a298b6676ed86d7897b1f2c23d
+size 33555533

venv/lib/python3.10/site-packages/torch/_dynamo/backends/common.py

@@ -0,0 +1,112 @@
+# mypy: ignore-errors
+
+import contextlib
+import functools
+import logging
+from unittest.mock import patch
+
+import torch
+from torch._dynamo import disable
+from torch._dynamo.utils import counters, defake
+from torch._functorch.aot_autograd import aot_module_simplified
+from torch.utils._python_dispatch import _disable_current_modes
+
+log = logging.getLogger(__name__)
+
+
+def aot_autograd(**kwargs):
+    def compiler_fn(gm: torch.fx.GraphModule, example_inputs):
+        # Hack to get around circular import problems with aot_eager_decomp_partition
+        if callable(kwargs.get("decompositions")):
+            kwargs["decompositions"] = kwargs["decompositions"]()
+
+        # NB: dont delete counter increment
+        counters["aot_autograd"]["total"] += 1
+        use_fallback = False
+
+        if use_fallback:
+            log.debug("Unable to use AOT Autograd because graph has mutation")
+            counters["aot_autograd"]["not_ok"] += 1
+            return gm
+
+        # OK attempt to compile
+
+        def _wrapped_bw_compiler(*args, **kwargs):
+            # stop TorchDynamo from trying to compile our generated backwards pass
+            return disable(disable(bw_compiler)(*args, **kwargs))
+
+        bw_compiler = kwargs.get("bw_compiler") or kwargs["fw_compiler"]
+        kwargs["bw_compiler"] = _wrapped_bw_compiler
+        kwargs["inference_compiler"] = (
+            kwargs.get("inference_compiler") or kwargs["fw_compiler"]
+        )
+
+        from functorch.compile import nop
+
+        from torch._inductor.debug import enable_aot_logging
+
+        # debug asserts slow down compile time noticeably,
+        # So only default them on when the aot_eager backend is used.
+        if kwargs.get("fw_compiler", None) == nop:
+            patch_config = patch("functorch.compile.config.debug_assert", True)
+        else:
+            patch_config = contextlib.nullcontext()
+
+        try:
+            # NB: NOT cloned!
+            with enable_aot_logging(), patch_config:
+                cg = aot_module_simplified(gm, example_inputs, **kwargs)
+                counters["aot_autograd"]["ok"] += 1
+                return disable(cg)
+        except Exception:
+            counters["aot_autograd"]["not_ok"] += 1
+            raise
+
+    return compiler_fn
+
+
+def mem_efficient_fusion_kwargs(use_decomps):
+    from functorch.compile import (
+        default_decompositions,
+        min_cut_rematerialization_partition,
+        ts_compile,
+    )
+
+    kwargs = {
+        # these are taken from memory_efficient_fusion()
+        "fw_compiler": ts_compile,
+        "bw_compiler": ts_compile,
+        "partition_fn": min_cut_rematerialization_partition,
+    }
+
+    if use_decomps:
+        kwargs["decompositions"] = default_decompositions
+
+    return kwargs
+
+
+def fake_tensor_unsupported(fn):
+    """
+    Decorator for backends that need real inputs.  We swap out fake
+    tensors for zero tensors.
+    """
+
+    @functools.wraps(fn)
+    def wrapper(model, inputs, **kwargs):
+        with _disable_current_modes():
+            inputs = list(map(defake, inputs))
+            return fn(model, inputs, **kwargs)
+
+    return wrapper
+
+
+def device_from_inputs(example_inputs) -> torch.device:
+    for x in example_inputs:
+        if hasattr(x, "device"):
+            return x.device
+
+
+def dtype_from_inputs(example_inputs) -> torch.dtype:
+    for x in example_inputs:
+        if hasattr(x, "dtype"):
+            return x.dtype

venv/lib/python3.10/site-packages/torch/_dynamo/backends/cudagraphs.py

@@ -0,0 +1,239 @@
+# mypy: ignore-errors
+
+import functools
+import operator
+from collections import defaultdict
+from typing import Dict, List, Optional
+
+import torch
+from torch._dynamo.backends.debugging import boxed_nop
+from torch._inductor.cudagraph_trees import cudagraphify_impl
+from torch._inductor.cudagraph_utils import (
+    BoxedDeviceIndex,
+    check_multiple_devices_or_any_cpu_nodes,
+    get_mutation_stack_trace,
+)
+from torch._inductor.utils import (
+    BoxedBool,
+    count_tangents,
+    has_incompatible_cudagraph_ops,
+    num_fw_fixed_arguments,
+    output_node,
+)
+from torch.multiprocessing.reductions import StorageWeakRef
+from .common import aot_autograd
+from .registry import register_backend
+
+perf_log = torch._logging.getArtifactLogger(__name__, "perf_hints")
+
+
+def find_input_mutations(g):
+    def meta_fk(meta):
+        return meta["val"] if "val" in meta else meta["fake_result"]
+
+    inputs = defaultdict(set)
+    input_idx = 0
+    mutated_inputs = set()
+    for n in g.nodes:
+        if n.op == "placeholder":
+            if isinstance(meta_fk(n.meta), torch.Tensor):
+                inputs[StorageWeakRef(meta_fk(n.meta)._typed_storage())].add(input_idx)
+            input_idx += 1
+        elif n.op == "call_function":
+            if n.target is operator.getitem:
+                continue
+            schema = n.target._schema
+            for i, arg in enumerate(schema.arguments):
+                if i < len(n.args):
+                    argument = n.args[i]
+                else:
+                    if arg.name not in n.kwargs:
+                        continue
+                    argument = n.kwargs[arg.name]
+                mut_arg = False
+                if arg.alias_info:
+                    if arg.alias_info.is_write:
+                        mut_arg = True
+                if mut_arg:
+                    # TODO: not correct for args that contain tensors in a struct
+                    # like list
+                    mutated_inputs |= inputs[
+                        StorageWeakRef(meta_fk(argument.meta)._typed_storage())
+                    ]
+
+        # TODO: error on unrecognized nodes
+    return mutated_inputs
+
+
+def get_device_node_mapping(gm: torch.fx.GraphModule):
+    device_node_mapping: Dict[torch.device, torch.fx.Node] = {}
+    for n in gm.graph.nodes:
+        t = n.meta.get("val", None)
+        if isinstance(t, torch.Tensor) and t.device not in device_node_mapping:
+            device_node_mapping[t.device] = n
+    return device_node_mapping
+
+
+def check_for_mutation(aot_model: torch.fx.GraphModule, num_fixed) -> Optional[str]:
+    mutation_indices = find_input_mutations(aot_model.graph) - set(range(num_fixed))
+    if not mutation_indices:
+        return None
+
+    return get_mutation_stack_trace(aot_model, mutation_indices)
+
+
+def check_for_skip(aot_model: torch.fx.GraphModule, num_fixed) -> Optional[str]:
+    if mut_skip := check_for_mutation(aot_model, num_fixed):
+        return mut_skip
+
+    if skip := check_multiple_devices_or_any_cpu_nodes(
+        get_device_node_mapping(aot_model)
+    ):
+        return skip
+
+    if has_incompatible_cudagraph_ops(aot_model):
+        return "skipping cudagraphs due to incompatible op"
+
+    return None
+
+
+def get_device_index(gm) -> int:
+    device = next(iter(get_device_node_mapping(gm)))
+    assert device.type == "cuda"
+    return device.index
+
+
+def get_stack_traces(gm) -> List[Optional[str]]:
+    output = output_node(gm)
+    assert len(output.args) == 1
+    return [
+        (arg.stack_trace if isinstance(arg, torch.fx.node.Node) else None)
+        for arg in output.args[0]
+    ]
+
+
+def cudagraphs(dynamo_model, dynamo_inputs):
+    do_cudagraphs = BoxedBool(True)
+    boxed_device_index = BoxedDeviceIndex(None)
+
+    def forward_cudagraphs(aot_model, aot_inputs, is_inference=False):
+        interp = boxed_nop(aot_model, aot_inputs)
+        fixed = num_fw_fixed_arguments(len(dynamo_inputs), len(aot_inputs))
+        if skip_msg := check_for_skip(aot_model, fixed):
+            BoxedBool.disable(do_cudagraphs)
+            perf_log.warning("skipping cudagraphs due to %s", skip_msg)
+            return interp
+
+        boxed_device_index.set(get_device_index(aot_model))
+
+        out = cudagraphify_impl(
+            interp,
+            aot_inputs,
+            range(fixed),
+            device_index=boxed_device_index.value,
+            is_backward=False,
+            is_inference=False,
+            stack_traces=get_stack_traces(aot_model),
+        )
+        out._boxed_call = True
+        return out
+
+    def backward_cudagraphs(aot_model, aot_inputs):
+        interp = boxed_nop(aot_model, aot_inputs)
+        if not do_cudagraphs:
+            return aot_model
+
+        fixed = count_tangents(aot_model)
+        if skip_msg := check_for_skip(aot_model, fixed):
+            perf_log.warning("skipping cudagraphs due to %s", skip_msg)
+
+            # See [Backward Generation Handling]
+            manager = torch._inductor.cudagraph_trees.get_manager(
+                boxed_device_index.value, create_if_none_exists=False
+            )
+            assert manager is not None
+
+            def fn(inputs):
+                manager.set_to_running_backward()
+                return aot_model(inputs)
+
+            fn._boxed_call = True
+            return fn
+
+        out = cudagraphify_impl(
+            interp,
+            aot_inputs,
+            range(fixed),
+            device_index=get_device_index(aot_model),
+            is_backward=True,
+            is_inference=False,
+            stack_traces=get_stack_traces(aot_model),
+        )
+        out._boxed_call = True
+        return out
+
+    aot_cudagraphs = aot_autograd(
+        fw_compiler=forward_cudagraphs,
+        bw_compiler=backward_cudagraphs,
+        inference_compiler=functools.partial(forward_cudagraphs, is_inference=True),
+        keep_inference_input_mutations=torch._dynamo.config.cudagraph_backend_keep_input_mutation,
+    )
+    return aot_cudagraphs(dynamo_model, dynamo_inputs)
+
+
+class CudagraphsBackend:
+    compiler_name = "cudagraphs"
+
+    @staticmethod
+    def reset():
+        from torch._inductor.cudagraph_trees import reset_cudagraph_trees
+
+        reset_cudagraph_trees()
+
+    @staticmethod
+    def __call__(model, inputs):
+        return cudagraphs(model, inputs)
+
+
+# aot_cudagraphs only applies CUDA graphs to the graph.  It is also helpful
+# for debugging and can serve as a perf baseline.
+register_backend(name="cudagraphs", compiler_fn=CudagraphsBackend())
+
+
+def cudagraphs_inner(model, inputs, copy_outputs=True, copy_inputs=True):
+    """This isn't registered as a backend, but is used in some benchmarks"""
+    assert isinstance(inputs, (list, tuple))
+    if copy_inputs:
+        static_inputs = [torch.zeros_like(x) for x in inputs]
+    else:
+        static_inputs = list(inputs)
+
+    # warmup
+    torch.cuda.synchronize()
+    stream = torch.cuda.Stream()
+    stream.wait_stream(torch.cuda.current_stream())
+    with torch.cuda.stream(stream):
+        model(*inputs)
+    stream.synchronize()
+    torch.cuda.current_stream().wait_stream(stream)
+    torch.cuda.synchronize()
+
+    # record
+    graph = torch.cuda.CUDAGraph()
+    with torch.cuda.graph(graph, stream=stream):
+        static_outputs = model(*static_inputs)
+    if not isinstance(static_outputs, (list, tuple)):
+        static_outputs = (static_outputs,)
+
+    def run(*new_inputs):
+        assert len(static_inputs) == len(new_inputs)
+        if copy_inputs:
+            for dst, src in zip(static_inputs, new_inputs):
+                dst.copy_(src)
+        graph.replay()
+        if copy_outputs:
+            return [x.clone() for x in static_outputs]
+        else:
+            return static_outputs
+
+    return run

venv/lib/python3.10/site-packages/torch/_dynamo/backends/debugging.py

@@ -0,0 +1,289 @@
+# mypy: ignore-errors
+
+import dataclasses
+import functools
+from importlib import import_module
+from typing import Any, List, Optional
+
+from functorch.compile import min_cut_rematerialization_partition
+
+import torch
+from torch import _guards
+from torch._functorch.compilers import ts_compile
+from .common import aot_autograd
+from .registry import register_debug_backend as register_backend
+
+"""
+This file contains TorchDynamo backends intended for debugging uses.
+"""
+
+
+@register_backend
+def eager(gm, fake_tensor_inputs):
+    return gm
+
+
+@register_backend
+def pre_dispatch_eager(gm, fake_tensor_inputs):
+    from torch.fx.experimental.proxy_tensor import make_fx
+
+    def runnable_gm(*args):
+        return torch.fx.Interpreter(gm).run(*args)
+
+    pre_dispatch_gm = make_fx(runnable_gm, pre_dispatch=True)(*fake_tensor_inputs)
+    pre_dispatch_gm.print_readable()
+
+    return pre_dispatch_gm
+
+
+@register_backend
+def eager_debug(gm, fake_tensor_inputs):
+    from torch._subclasses.schema_check_mode import SchemaCheckMode
+
+    # We could add more debugging bits here.
+    # Right now, this backend can be used to check for and error on
+    # custom dispatcher ops that have incorrect schemas.
+    def inner(*args):
+        with SchemaCheckMode():
+            return torch.fx.Interpreter(gm).run(*args)
+
+    return inner
+
+
+@register_backend(name="ts")
+def torchscript(gm, fake_tensor_inputs):
+    return torch.jit.script(gm)
+
+
+# used boxed call to discard inputs when they are no longer needed
+def boxed_nop(fx_g, example_inputs):
+    def run(args):
+        return torch.fx.Interpreter(fx_g).boxed_run(args)
+
+    run._boxed_call = True
+    return run
+
+
+# Useful for debugging purpose
+# aot_eager uses AOT Autograd backend with nop compiler. It is helpful in debugging.
+aot_eager = aot_autograd(
+    fw_compiler=boxed_nop, partition_fn=min_cut_rematerialization_partition
+)
+register_backend(name="aot_eager", compiler_fn=aot_eager)
+
+aot_eager_default_partitioner = aot_autograd(fw_compiler=boxed_nop)
+register_backend(
+    name="aot_eager_default_partitioner", compiler_fn=aot_eager_default_partitioner
+)
+
+# Uses TorchInductor AOT Autograd decomps and partitioner to isolate aot vs
+# inductor problems.
+# aot_eager_decomp_partition just replaces the inductor compiler with nop to help
+# isolate inductor vs aot_eager errors
+aot_eager_decomp_partition = aot_autograd(
+    # these are taken from memory_efficient_fusion()
+    fw_compiler=boxed_nop,
+    bw_compiler=boxed_nop,
+    # NB: lambda here is to delay import of inductor
+    decompositions=lambda: import_module(
+        "torch._inductor.compile_fx"
+    ).select_decomp_table(),
+    partition_fn=functools.partial(
+        min_cut_rematerialization_partition, compiler="inductor"
+    ),
+)
+register_backend(
+    name="aot_eager_decomp_partition", compiler_fn=aot_eager_decomp_partition
+)
+
+# AOT Autograd with torchscript backend. Default partitioner.
+# aot_ts uses torchscript backend. We can use this with both nnc and nvfuser
+# by using the relevant fuser with torch.jit.fuser(...)
+aot_ts = aot_autograd(fw_compiler=ts_compile)
+register_backend(name="aot_ts", compiler_fn=aot_ts)
+
+# These buggy backends are used for inducing bugs so that we can test
+# our repro extraction / minifier scripts
+
+
+class ReluCompileError(Exception):
+    pass
+
+
+class TestingOnlyCompileError(Exception):
+    pass
+
+
+@register_backend
+def relu_compile_error_TESTING_ONLY(gm: torch.fx.GraphModule, example_inputs):
+    for node in gm.graph.nodes:
+        if node.target == torch.relu:
+            raise ReluCompileError()
+    return gm
+
+
+@register_backend
+def relu_runtime_error_TESTING_ONLY(gm: torch.fx.GraphModule, example_inputs):
+    for node in gm.graph.nodes:
+        if node.target == torch.relu:
+            node.target = torch._assert
+            node.args = (False, "ReluRuntimeError")
+    gm.recompile()
+    return gm
+
+
+@register_backend
+def relu_accuracy_error_TESTING_ONLY(gm: torch.fx.GraphModule, example_inputs):
+    for node in gm.graph.nodes:
+        if node.target == torch.relu:
+            node.target = torch.add
+            node.args = (node.args[0], 1)
+    gm.recompile()
+
+    return gm
+
+
+@register_backend
+def non_leaf_compile_error_TESTING_ONLY(gm: torch.fx.GraphModule, example_inputs):
+    # Require at least one non-trivial thing in the graph,
+    # see https://github.com/pytorch/pytorch/issues/102898
+    for node in gm.graph.nodes:
+        if node.op == "call_function":
+            break
+    else:
+        return gm
+    for t in example_inputs:
+        if not t.is_leaf:
+            raise TestingOnlyCompileError()
+    return gm
+
+
+@dataclasses.dataclass
+class ExplainOutput:
+    """
+    This is the output of :func:`torch._dynamo.explain()`
+    There is no reason to create this class directly.
+    """
+
+    graphs: List[torch.fx.GraphModule]
+    graph_count: int
+    graph_break_count: int
+    break_reasons: List[
+        Any
+    ]  # Type is GraphCompileReason but doesn't matter for this purpose
+    op_count: int
+    ops_per_graph: Optional[List[torch.fx.Node]] = None
+    out_guards: Optional[List[_guards.Guard]] = None
+    compile_times: Optional[str] = None
+
+    def __str__(self):
+        output = f"Graph Count: {self.graph_count}\n"
+        output += f"Graph Break Count: {self.graph_break_count}\n"
+        output += f"Op Count: {self.op_count}\n"
+
+        output += "Break Reasons:\n"
+        for idx, break_reason in enumerate(self.break_reasons):
+            output += f"  Break Reason {idx+1}:\n"
+            output += f"    Reason: {break_reason.reason}\n"
+            output += "    User Stack:\n"
+            for frame_summary in break_reason.user_stack:
+                output += f"      {frame_summary}\n"
+
+        if self.ops_per_graph is not None:
+            output += "Ops per Graph:\n"
+            for idx, ops in enumerate(self.ops_per_graph):
+                output += f"  Ops {idx+1}:\n"
+                for op in ops:
+                    output += f"    {op}\n"
+
+        if self.out_guards is not None:
+            output += "Out Guards:\n"
+            for i, guard in enumerate(self.out_guards):
+                output += f"  Guard {i+1}:\n"
+                output += f"    {str(guard)}"
+
+        if self.compile_times is not None:
+            output += f"Compile Times: {self.compile_times}\n"
+        return output
+
+
+def _explain_graph_detail(
+    gm: torch.fx.GraphModule, graphs, op_count, ops_per_graph, break_reasons
+):
+    """
+    This function is a utility which processes a torch.fx.GraphModule and
+    accumulates information about its ops, graph breaks, and other details. It
+    is intended to be used by the ExplainWithBackend class and
+    `torch._dynamo.explain()` to provide details from Dynamo's graph capture.
+
+    Parameters:
+        gm (torch.fx.GraphModule): The GraphModule to be processed.
+        graphs (list): A list that accumulates all the GraphModules processed.
+        op_count (int): The total count of operations in all GraphModules processed so far.
+        ops_per_graph (list): A list that accumulates the operations of each GraphModule.
+        break_reasons (list): A list that accumulates the reasons for breaks in each GraphModule.
+
+    Returns:
+        tuple: A tuple containing the processed GraphModule, the updated lists of graphs,
+               operations per graph, and break reasons, and the updated operation count.
+    """
+    graphs.append(gm)
+    ops = [node.target for node in gm.graph.nodes if node.op == "call_function"]
+    op_count += len(ops)
+    ops_per_graph.append(ops)
+    if gm.compile_subgraph_reason.graph_break:
+        break_reasons.append(gm.compile_subgraph_reason)
+
+    return gm, graphs, op_count, ops_per_graph, break_reasons
+
+
+class ExplainWithBackend:
+    """
+    This class is intended to be used as a backend for `torch.compile`. It is
+    composable with other backends. When used in this way, it accumulates
+    information about graph breaks, ops, and other info and provides a string
+    representation summarizing this information.
+
+    Attributes:
+        backend (str): The name of the backend to use for optimization.
+        graphs (list): A list of the graphs captured by TorchDynamo.
+        op_count (int): The total number of operations in all optimized graphs.
+        break_reasons (list): A list of graph break reasons with stack traces.
+
+    Example Usage:
+        def fn(x):
+            x = torch.sigmoid(x)
+            return x
+
+        torch._dynamo.reset()
+        eb = ExplainWithBackend("inductor")
+        optimized_fn = torch.compile(fn, backend=eb)
+        result = optimized_fn(torch.randn(5))
+        print(eb.output())
+    """
+
+    def __init__(self, backend):
+        from .registry import lookup_backend
+
+        self.backend = lookup_backend(backend)
+        self.graphs = []
+        self.op_count = 0
+        self.break_reasons = []
+
+    def __call__(self, gm: torch.fx.GraphModule, example_inputs):
+        gm, self.graphs, self.op_count, _, self.break_reasons = _explain_graph_detail(
+            gm, self.graphs, self.op_count, [], self.break_reasons
+        )
+        return self.backend(gm, example_inputs)
+
+    def output(self) -> ExplainOutput:
+        graph_count = len(self.graphs)
+        output = ExplainOutput(
+            self.graphs,
+            graph_count,
+            graph_count - 1,
+            self.break_reasons,
+            self.op_count,
+        )
+
+        return output

venv/lib/python3.10/site-packages/torch/_dynamo/backends/distributed.py

@@ -0,0 +1,612 @@
+# mypy: ignore-errors
+
+import logging
+import traceback
+from dataclasses import dataclass, field
+from typing import Any, List, Optional
+from unittest import mock
+
+import torch
+from torch import fx
+from torch._dynamo.output_graph import GraphCompileReason
+from torch._dynamo.utils import deepcopy_to_fake_tensor, detect_fake_mode
+from torch._logging import trace_structured
+from torch.fx.node import Node
+
+# Regular log messages should go through 'log'.
+# ddp_graph_log is a separate artifact logger reserved for dumping graphs.
+# See docs/source/logging.rst for more info.
+log = logging.getLogger(__name__)
+ddp_graph_log = torch._logging.getArtifactLogger(__name__, "ddp_graphs")
+
+
+def args_str(args):
+    # a debug helper
+    if torch.is_tensor(args):
+        return f"T[{args.shape}]"
+    elif isinstance(args, tuple):
+        return f"tuple({', '.join([args_str(x) for x in args])})"
+    elif isinstance(args, list):
+        return f"list({', '.join([args_str(x) for x in args])})"
+    else:
+        return str(args)
+
+
+@dataclass
+class Bucket:
+    size: int = 0
+    params: List[str] = field(default_factory=list)
+    nodes: List[fx.Node] = field(default_factory=list)
+
+    # param_ids is just used for unit testing
+    param_ids: List = field(default_factory=list)
+
+    # keep track of any buckets that were extended for logging purposes
+    opcount_increased_to_capture_external_output: int = 0
+    paramsize_before_opcount_increase: int = 0
+
+
+def bucket_has_external_output(bucket: Bucket) -> bool:
+    nodes_in_bucket = set()
+    # we want to iterate in reverse order, but clumsi-luckily the bucket.nodes list was already created backwards
+    # so we don't reverse it here
+    for node in bucket.nodes:
+        # assume node.op != output, since those are filtered in the original iteration
+        nodes_in_bucket.add(node)
+        for user in node.users:
+            if user not in nodes_in_bucket:
+                return True
+    return False
+
+
+def pretty_print_buckets(buckets: List[Bucket], bucket_bytes_cap: int):
+    headers = ("Index", "Size (b)", "Param Names")
+    rows = []
+    extended_buckets = []
+    for idx, bucket in enumerate(reversed(buckets)):
+        if len(bucket.params) > 0:
+            rows.append((idx, bucket.size, bucket.params[0]))
+            for param in bucket.params[1:]:
+                rows.append((None, None, param))
+        if bucket.opcount_increased_to_capture_external_output > 0:
+            extended_buckets.append(
+                (
+                    idx,
+                    bucket.opcount_increased_to_capture_external_output,
+                    bucket.size - bucket.paramsize_before_opcount_increase,
+                )
+            )
+
+    if len(rows):
+        log.info(
+            "\nDDPOptimizer used bucket cap %s and created %d buckets. Enable debug logs for detailed bucket info.",
+            bucket_bytes_cap,
+            len(buckets),
+        )
+
+        if len(extended_buckets):
+            log.warning(
+                "Some buckets were extended beyond their requested parameter capacities"
+                " in order to ensure each subgraph has an output node, required for fx graph partitioning."
+                " This can be the case when a subgraph would have only contained nodes performing inplace mutation,"
+                " and returning no logical outputs. This should not be a problem, unless it results in too few graph"
+                " partitions for optimal DDP performance."
+            )
+
+        try:
+            from tabulate import tabulate
+
+            log.debug(
+                "\nDDPOptimizer produced the following bucket assignments:\n%s",
+                tabulate(rows, headers=headers, tablefmt="simple_grid"),
+            )
+
+            if len(extended_buckets):
+                log.warning(
+                    "DDPOptimizer extended these buckets to ensure per-subgraph output nodes:\n%s",
+                    tabulate(
+                        extended_buckets,
+                        headers=("Index", "Extra Ops", "Extra Param Size (b)"),
+                        tablefmt="simple_grid",
+                    ),
+                )
+        except ImportError:
+            log.debug(
+                "Please `pip install tabulate` in order to display ddp bucket sizes and diagnostic information."
+            )
+    else:
+        log.debug("DDPOptimizer captured no parameters and did not split this graph.")
+
+
+def has_higher_order_op(gm):
+    # Check if there is a higher order op in the graph
+    for node in gm.graph.nodes:
+        if node.op == "get_attr":
+            maybe_param = getattr(gm, node.target)
+            if isinstance(maybe_param, torch.fx.GraphModule):
+                return True
+    return False
+
+
+# 3 (lazy compile): Replace submodules with lazily compiling submodule
+class SubmoduleReplacer(torch.fx.interpreter.Interpreter):
+    def __init__(self, module, compiler):
+        super().__init__(module)
+        self.compiler = compiler
+
+    def lazily_compiled_submod(self, input_mod):
+        """
+        Create a wrapper around submodules which:
+        - lazily compiles each of the partitioned submodules using the user-provided compiler
+        - unpacks singleton tuples/lists into flat arg
+        """
+
+        class LazilyCompiledModule(torch.nn.Module):
+            def __init__(self, submod, compiler, unwrap_singleton_tuple):
+                super().__init__()
+                self.submod = submod
+                self.compiler = compiler
+                self.compiled = False
+                self.unwrap_singleton_tuple = unwrap_singleton_tuple
+
+            def forward(self, *args):
+                if not self.compiled:
+                    # First compile with args as example_inputs
+                    # These args will be fakeified if using Inductor/AOTAutograd
+                    new_submod = self.compiler(self.submod, args)
+                    del self.submod
+                    self.submod = new_submod
+                    self.compiled = True
+                    self.compiler = None
+
+                x = self.submod(*args)
+                # we must let 'input_mod' return a tuple, to make AOT happy.
+                # (aot_autograd compile_fn literally requires that the output of a graph it compiles is a tuple).
+                # however, we don't acutally want this tuple to be returned, since the fx logic that calls the submod
+                # will again wrap outputs from the submod in a tuple.  So we unwrap it, and count on it being re-wrapped
+                if self.unwrap_singleton_tuple and isinstance(x, (tuple, list)):
+                    return x[0]
+                return x
+
+        unwrap_singleton_tuple = False
+        for sn in input_mod.graph.nodes:
+            if sn.op == "output":
+                if not isinstance(sn.args[0], tuple):
+                    unwrap_singleton_tuple = True
+                    sn.args = (sn.args,)
+
+        input_mod.recompile()
+        input_mod.compile_subgraph_reason = GraphCompileReason(
+            "DDPOptimizer intentional graph-break (See Note [DDPOptimizer])."
+            " Set `torch._dynamo.config.optimize_ddp = False` to disable.",
+            [
+                # it's close to useless to get a real stacktrace here, and quite verbose.
+                traceback.FrameSummary(__file__, 0, DDPOptimizer),
+            ],
+        )
+        wrapper = LazilyCompiledModule(
+            input_mod,
+            self.compiler,
+            unwrap_singleton_tuple,
+        )
+        return wrapper
+
+    # We replace the submodules with lazy submodules which compile
+    # the corresponding submodules when they are run with real values
+    # Always returns `None` - we do not need to propagate values in order
+    # to replace submodules.
+    def run_node(self, n: Node) -> Any:
+        if n.op == "call_module":
+            real_mod = self.fetch_attr(n.target)
+
+            ddp_graph_log.debug("\n---%s graph---\n%s", n.target, real_mod.graph)
+
+            assert len(n.kwargs) == 0, "We assume only args for these modules"
+            lazily_compiled_submod = self.lazily_compiled_submod(real_mod)
+
+            # We update the original (outer) graph with a call into the compiled module
+            # instead of the uncompiled one.
+            self.module.delete_submodule(n.target)
+            n.target = "compiled_" + n.target
+            self.module.add_submodule(n.target, lazily_compiled_submod)
+
+
+# 3 (no lazy compile): compile each of the partitioned submodules using the user-provided compiler
+class SubmodCompiler(torch.fx.interpreter.Interpreter):
+    def __init__(self, module, compiler, fake_mode):
+        super().__init__(module)
+        self.compiler = compiler
+        self.fake_mode = fake_mode
+
+    def compile_submod(self, input_mod, args, kwargs):
+        """
+        Compile the submodule,
+        using a wrapper to make sure its output is always a tuple,
+        which is required by AotAutograd based compilers
+        """
+        assert len(kwargs) == 0, "We assume only args for these modules"
+
+        class WrapperModule(torch.nn.Module):
+            def __init__(self, submod, unwrap_singleton_tuple):
+                super().__init__()
+                self.submod = submod
+                self.unwrap_singleton_tuple = unwrap_singleton_tuple
+
+            def forward(self, *args):
+                x = self.submod(*args)
+                # TODO(whc)
+                # for some reason the isinstance check is necessary if I split one node per submod
+                # - even though I supposedly wrapped the output in a tuple in those cases, the real
+                # compiled module was still returning a tensor
+                if self.unwrap_singleton_tuple and isinstance(x, (tuple, list)):
+                    return x[0]
+                return x
+
+        unwrap_singleton_tuple = False
+        for sn in input_mod.graph.nodes:
+            if sn.op == "output":
+                if not isinstance(sn.args[0], tuple):
+                    unwrap_singleton_tuple = True
+                    sn.args = (sn.args,)
+
+        input_mod.recompile()
+        input_mod.compile_subgraph_reason = GraphCompileReason(
+            "DDPOptimizer intentional graph-break (See Note [DDPOptimizer])."
+            " Set `torch._dynamo.config.optimize_ddp = False` to disable.",
+            [
+                # it's close to useless to get a real stacktrace here, and quite verbose.
+                traceback.FrameSummary(__file__, 0, DDPOptimizer),
+            ],
+        )
+
+        wrapper = WrapperModule(
+            self.compiler(input_mod, args),
+            unwrap_singleton_tuple,
+        )
+        return wrapper
+
+    # Note:
+    #
+    # The way distributed works today around fake tensors can be somewhat confusing.
+    # Some of these codepaths are shared in both runtime, and compile time. The presence
+    # of a fake_mode, read off of fake tensor inputs, dictates how we will operate.
+    #
+    # A few things to keep in mind:
+    #
+    # 1) We invoke `compile_submod` with a real module. The output of that gets stored
+    # on the graph via `self.module.add_submodule(n.target, compiled_submod_real)`.
+    #
+    # 2) When running a call_module targeted node, if we have a fake_mode, we fakify the
+    # module we got from self.fetch_attr(n.target). Regardless of fake_mode, we then execute it.
+    #
+    # 3) Fake tensors should always be around during compile time.
+    #
+    # 4) Fake tensors should never be around at runtime.
+    #
+    # 5) We end up with a compilation mode that takes a real submodule and fake tensors,
+    # to match what aot_autograd expects. See Note: [Fake Modules and AOTAutograd]
+    def run_node(self, n: Node) -> Any:
+        args, kwargs = self.fetch_args_kwargs_from_env(n)
+        new_args = []
+        assert self.fake_mode
+        for arg in args:
+            if isinstance(arg, torch.Tensor) and not isinstance(
+                arg, torch._subclasses.FakeTensor
+            ):
+                new_args.append(torch._dynamo.utils.to_fake_tensor(arg, self.fake_mode))
+            else:
+                new_args.append(arg)
+
+        log.debug("run_node %s, %s got args %s", n.op, n.target, args_str(args))
+        assert isinstance(args, tuple)
+        assert isinstance(kwargs, dict)
+
+        if n.op == "call_module":
+            real_mod = self.fetch_attr(n.target)
+            if self.fake_mode:
+                curr_submod = deepcopy_to_fake_tensor(real_mod, self.fake_mode)
+            else:
+                curr_submod = real_mod
+
+            ddp_graph_log.debug("\n---%s graph---\n%s", n.target, curr_submod.graph)
+
+            # When calling the compiler on the submod, inputs (new_args) are expected to
+            # be FakeTensors already since Dynamo would have made them FakeTensors in the
+            # non-DDP flow.  However, the parameters are _not_ expected to be FakeTensors,
+            # since this wrapping happens during compilation
+
+            # Note: Returning Fake Tensors on First AOT Autograd Call
+            #
+            # Inductor will optimize strides of outputs when it deems it profitable.
+            # For instance, converting to channels last. When we split the graph here
+            # into multiple inductor compilations, we need to make sure that the
+            # output strides of one compilation is appropriately passed to the subsequent
+            # compilations. However, the mapping from inductor output to dynamo output
+            # is non-trivial due to aot_autograd's deduping, de-aliasing, mutation, re-writing,
+            # subclass handling, etc. In order to replay all this logic we set a flag such that
+            # the first invocation of inductor in aot_autograd will return Fake Tensors with
+            # appropriate strides. Then, all of aot autograd's runtime logic is replayed.
+            # This gives us the appropriately strided outputs here which will reflect runtime strides.
+
+            class FakeifyFirstAOTInvocationGuard:
+                def __init__(self):
+                    self.tc = torch._guards.TracingContext.try_get()
+                    assert self.tc
+                    torch._guards.TracingContext.try_get().fakify_first_call = True
+
+                def __del__(self):
+                    self.tc.fakify_first_call = False
+
+            # For aot_eager and other backends, tracing context is not set
+            has_tracing_context = torch._guards.TracingContext.try_get() is not None
+            if has_tracing_context:
+                g = FakeifyFirstAOTInvocationGuard()
+
+            from torch._dynamo.utils import counters
+
+            init = counters["aot_autograd"]["total"]
+            compiled_submod_real = self.compile_submod(real_mod, new_args, kwargs)
+
+            # TODO - better way of doing this?
+            # Only aot autograd handles fakifying first call
+            invoked_aot_autograd = init != counters["aot_autograd"]["total"]
+
+            # We update the original (outer) graph with a call into the compiled module
+            # instead of the uncompiled one.
+            self.module.delete_submodule(n.target)
+            n.target = "compiled_" + n.target
+            self.module.add_submodule(n.target, compiled_submod_real)
+
+            # Finally, we have to produce inputs for use compiling the next submodule,
+            # and these need to be FakeTensors, so we execute the module under fake_mode
+            # Because parameters are not fake we patch fake tensor mode to allow non fake inputs
+            with self.fake_mode, mock.patch.object(
+                self.fake_mode, "allow_non_fake_inputs", True
+            ):
+                if has_tracing_context and invoked_aot_autograd:
+                    out = compiled_submod_real(*new_args, **kwargs)
+                    # output should be fake or subclass
+                    assert all(
+                        (not isinstance(t, torch.Tensor) or type(t) is not torch.Tensor)
+                        for t in (out if isinstance(out, (list, tuple)) else [out])
+                    )
+                    return out
+                else:
+                    return curr_submod(*new_args, **kwargs)
+        else:
+            # placeholder or output nodes don't need to get compiled, just executed
+            return getattr(self, n.op)(n.target, new_args, kwargs)
+
+
+class DDPOptimizer:
+
+    """Note [DDPOptimizer]
+    DDPOptimizer applies when dynamo compiles models wrapped in DistributedDataParallel (DDP),
+    breaking the dynamo graph into chunks to compile separately, with the breaks aligning to
+    the boundaries of gradient-allreduce buckets chosen by DDP.
+
+    Background/Motivation
+     - DDP uses allreduce collectives to synchronize partial gradients computed on different workers
+     - DDP groups gradient allreduces into 'buckets' to optimize communication efficiency of all-reduce
+     - Parameters grouped into buckets are assumed to be adjacent in time, so they become ready
+       at around the same time during backward and thus can share the same allreduce efficiently
+     - Allreduces must overlap with backward compute for optimal training performance
+     - DDP schedules allreduces using 'hooks' fired from the c++ autograd engine in pytorch, which
+       operates when individual grads become 'ready'
+     - Dynamo+AOTAutograd produces a single fused graph that runs 'atomically' from the perspective of the
+       autograd engine, such that all gradients become 'ready' at the same time.  Hooks fire after the whole
+       fused backward function executes, preventing any overlap of compute and communication
+
+    Algorithm
+     - DDPOptimizer starts off with an FX graph traced by dynamo which represents forward.  It can traverse
+       this graph in reverse order to determine the true order that gradients will become ready during backward.
+     - Parameter sizes are counted in reverse order, up to a bucket size limit, at which point a new bucket is started
+       and a graph break introduced
+     - Each of the subgraphs is compiled by the compiler provided to dynamo by the user, and then fused back together
+       into an outer module that is returned to the user
+
+    Notes
+     - It would be better to enforce (by adding an API to DDP) that the bucket splits chosen here are used by DDP,
+       and that DDP does not need to detect or optimize bucket order by observing execution at runtime, as it does
+       in eager.
+     - If Dynamo can't capture a whole graph for the portion of the model wrapped by DDP, this algorithm will currently
+       produce splits that do not necessarily align with the buckets used by DDP.  This should result in performance
+       degradation approaching the baseline case where graph-splits are not used, but not worse.
+     - If the backend compiler fails to compile a single subgraph, it will execute eagerly despite the rest of the
+       subgraphs being compiled
+     - DDP has a 'parameters_and_buffers_to_ignore' field, which DDPOptimizer attempts to honor by reading markers
+       left by DDP on individual parameters.  In cases where other transformations, such as reparameterization, are
+       also used, the ignore markers could be lost.  If DDPOptimizer fails to ignore a parameter ignored by DDP,
+       it is not catastrophic but could impact performance by choosing sub-optimal bucket splits.
+     - DDPOptimizer always ignores all buffers, regardless of their ignore flag, since buffers do not require gradients,
+       and therefore aren't allreduced by DDP.  (They are broadcast during forward, but this is not covered by
+       DDPOptimizer)
+
+    Debugging
+     - Generally, it is easiest to debug DDPOptimizer in a single process program, using pdb.
+     - In many cases, the log messages are helpful (they show bucket size assignments)-
+       just set TORCH_LOGS env to include any of 'dynamo', 'distributed', or 'dist_ddp'.
+     - See `benchmarks/dynamo/distributed.py` for a simple harness that will run a toy model or a torchbench model
+       in a single process (or with torchrun, in multiple processes)
+
+    Args:
+        bucket_bytes_cap (int): Controls the size of buckets, in bytes, used to determine graphbreaks.  Should be
+            set to match the equivalent parameter on the original DDP module.
+
+        backend_compile_fn (callable): A dynamo compiler function, to be invoked to compile each subgraph.
+
+        first_bucket_cap (int): Controls the size of the first bucket.  Should match DDP's first bucket cap.  DDP
+            special-cases the first bucket size since it is sometimes optimal to start a small allreduce early.
+
+    """
+
+    def __init__(
+        self,
+        bucket_bytes_cap: int,
+        backend_compile_fn,
+        first_bucket_cap: Optional[int] = None,
+    ):
+        if first_bucket_cap is not None:
+            self.first_bucket_cap = first_bucket_cap
+        elif torch.distributed.is_available():
+            # this constant comes from C10D lib which is not always built
+            self.first_bucket_cap = torch.distributed._DEFAULT_FIRST_BUCKET_BYTES
+        else:
+            self.first_bucket_cap = bucket_bytes_cap
+
+        self.bucket_bytes_cap = bucket_bytes_cap
+        assert (
+            self.first_bucket_cap <= self.bucket_bytes_cap
+        ), "First bucket should be smaller/equal to other buckets to get comms warmed up ASAP"
+
+        self.backend_compile_fn = backend_compile_fn
+
+    def _ignore_parameter(self, parameter):
+        return hasattr(parameter, "_ddp_ignored") and parameter._ddp_ignored
+
+    def compile_fn(self, gm: fx.GraphModule, example_inputs: List[torch.Tensor]):
+        """
+        Implements graph splitting, first determining a set of of buckets by counting
+        parameter sizes in reverse graph order, then invoking the user/backend compiler
+        to compile each subgraph. Finally, stiches compiled graphs into one graphmodule
+        and returns its callable.
+        """
+        if has_higher_order_op(gm):
+            # This indicates presence of a higher order op. For now, we
+            # have no way to break the higher order op into two buckets.
+            # Allowing higher order ops in the graph also requires
+            # changes in the split_module, becuase graph splitter
+            # currently assumes that all the args of all ops are
+            # tensors, but in the case of higher order ops, it could be
+            # a graph module. As a workaround, we are shortcircuiting
+            raise NotImplementedError(
+                "DDPOptimizer backend: Found a higher order op in the graph. "
+                "This is not supported. Please turn off DDP optimizer using "
+                "torch._dynamo.config.optimize_ddp=False. Note that this can "
+                "cause performance degradation because there will be one bucket "
+                "for the entire Dynamo graph. Please refer to this issue - "
+                "https://github.com/pytorch/pytorch/issues/104674."
+            )
+
+        # 1: compute the partition map according to DDP bucket logic
+        buckets = [Bucket()]  # (size, param_names)
+        for node in reversed(gm.graph.nodes):
+            if node.op in ("output", "placeholder"):
+                continue
+
+            if (
+                buckets[0].size >= self.bucket_bytes_cap
+                or len(buckets) == 1
+                and buckets[0].size >= self.first_bucket_cap
+            ):
+                if bucket_has_external_output(buckets[0]):
+                    buckets.insert(0, Bucket())
+                else:
+                    # continue building this bucket past the point of filling its parameter capacity,
+                    # to increase chances it contains at least one node that is either a global output or
+                    # passed as input to a subsequent graph
+
+                    if buckets[0].opcount_increased_to_capture_external_output == 0:
+                        buckets[0].paramsize_before_opcount_increase = buckets[0].size
+                    buckets[0].opcount_increased_to_capture_external_output += 1
+
+            if node.op == "call_module":
+                target = gm.get_submodule(node.target)
+                for name, param in target.named_parameters():
+                    if param.requires_grad and not self._ignore_parameter(param):
+                        buckets[0].size += param.untyped_storage().nbytes()
+                        buckets[0].params.append(f"{node.target}_{name}")
+                        buckets[0].param_ids.append(id(param))
+            elif node.op == "get_attr":
+                maybe_param = getattr(gm, node.target)
+                if maybe_param.requires_grad and not self._ignore_parameter(
+                    maybe_param
+                ):
+                    buckets[0].size += maybe_param.untyped_storage().nbytes()
+                    buckets[0].params.append(node.target)
+                    buckets[0].param_ids.append(id(maybe_param))
+
+            # All nodes have to be mapped to a bucket, even if they don't have their own params
+            # Ignored params still end up in buckets, we just don't count them towards the capacity
+            buckets[0].nodes.append(node)
+
+        if len(buckets) > 1 and buckets[0].size == 0:
+            # we collected a small preamble graph with ops that don't include parameters, fuse it back
+            buckets[1].nodes.extend(buckets[0].nodes)
+            assert len(buckets[0].params) == 0, "Params should be empty if size is 0"
+            del buckets[0]
+
+        # stash buckets for testing/debugging purposes
+        self.buckets = buckets
+        pretty_print_buckets(buckets, self.bucket_bytes_cap)
+
+        if len(buckets) == 1:
+            # bypass split/fuse logic if there is only one bucket
+            return self.backend_compile_fn(gm, example_inputs)
+
+        # 2: partition the graphmodule according to bucket capacity
+        partition_map = {}
+        for idx, b in enumerate(buckets):
+            for node in b.nodes:
+                partition_map[node] = idx
+
+        split_gm = fx.passes.split_module.split_module(
+            gm, None, lambda node: partition_map[node]
+        )
+
+        debug_str = (
+            f"\n---orig graph---\n{gm.graph}\n"
+            + f"\n---split graph---\n{split_gm.graph}\n"
+        )
+        for name, module in split_gm.named_modules():
+            if "." not in name and len(name):
+                # only print the submod graphs, not their children
+                debug_str += f"\n---{name} graph---\n{module.graph}\n"
+        debug_str += "\n---------------\n"
+        ddp_graph_log.debug(debug_str)
+
+        trace_structured(
+            "optimize_ddp_split_graph",
+            payload_fn=lambda: split_gm.print_readable(print_output=False),
+        )
+        for name, module in split_gm.named_modules():
+            if "." not in name and len(name):
+                trace_structured(
+                    "optimize_ddp_split_child",
+                    lambda: {"name": name},
+                    payload_fn=lambda: module.print_readable(print_output=False),
+                )
+
+        # NOTE, we want to enable `optimize_ddp_lazy_compile` by default as soon as possible,
+        # becuase it will fix stride mismatch errors (see motivation: https://github.com/pytorch/pytorch/pull/114154).
+        # However, lazy compile currently causes shape mismatch in other cases (`test_graph_split_inductor_transpose`)
+        # and we need to fix them before we can enable it by default.
+        if not torch._dynamo.config.optimize_ddp_lazy_compile:
+            # Today, optimize_ddp=True and keep_output_stride=False can lead to silent
+            # correctness issues. The problem is that ddp_optimizer works by partitioning
+            # the dynamo graph, sending each subgraph through aot autograd to inductor,
+            # and creates example inputs by eagerly interpreting each subgraph to get
+            # an output that with the same metadata that we'd get from eager mode.
+            # This is a problem though, for torch._inductor.config.keep_output_stride.
+            # The above config can cause the outputs of the first graph to have
+            # **different** strides from eager, causing the inputs that we pass
+            # to the second graph to be wrong.
+            # To really fix this, we would need to faithfully ask inductor
+            # what the outputs to each graph it expects are.
+            fake_mode = detect_fake_mode(example_inputs)
+            if fake_mode is None:
+                fake_mode = torch._subclasses.fake_tensor.FakeTensorMode()
+
+        if torch._dynamo.config.optimize_ddp_lazy_compile:
+            submod_compiler = SubmoduleReplacer(split_gm, self.backend_compile_fn)
+        else:
+            submod_compiler = SubmodCompiler(
+                split_gm, self.backend_compile_fn, fake_mode
+            )
+        submod_compiler.run(*example_inputs)
+        split_gm.recompile()
+
+        ddp_graph_log.debug(
+            "\n---final graph---\n%s\n---------------\n", split_gm.graph
+        )
+        return split_gm

venv/lib/python3.10/site-packages/torch/_dynamo/backends/inductor.py

@@ -0,0 +1,16 @@
+# mypy: ignore-errors
+
+import sys
+
+from torch._dynamo import register_backend
+
+
+@register_backend
+def inductor(*args, **kwargs):
+    if sys.platform == "win32":
+        raise RuntimeError("Windows not yet supported for inductor")
+
+    # do import here to avoid loading inductor into memory when it is not used
+    from torch._inductor.compile_fx import compile_fx
+
+    return compile_fx(*args, **kwargs)

venv/lib/python3.10/site-packages/torch/_dynamo/backends/onnxrt.py

@@ -0,0 +1,37 @@
+# mypy: ignore-errors
+
+# This backend is maintained by ONNX team. To direct issues
+# to the right people, please tag related GitHub issues with `module: onnx`.
+#
+# Maintainers' Github IDs: wschin, thiagocrepaldi, BowenBao, abock
+from torch.onnx._internal.onnxruntime import (
+    is_onnxrt_backend_supported,
+    torch_compile_backend,
+)
+from .registry import register_backend
+
+
+def has_onnxruntime():
+    # FIXME(abock): update test/dynamo/test_backends.py to call is_onnxrt_backend_supported()
+    return is_onnxrt_backend_supported()
+
+
+if is_onnxrt_backend_supported():
+    register_backend(name="onnxrt", compiler_fn=torch_compile_backend)
+else:
+
+    def information_displaying_backend(*args, **kwargs):
+        raise ImportError(
+            "onnxrt is not registered as a backend. "
+            "Please make sure all dependencies such as "
+            "numpy, onnx, onnxscript, and onnxruntime-training are installed. "
+            "Suggested procedure to fix dependency problem:\n"
+            "  (1) pip or conda install numpy onnx onnxscript onnxruntime-training.\n"
+            "  (2) Open a new python terminal.\n"
+            "  (3) Call the API `torch.onnx.is_onnxrt_backend_supported()`:\n"
+            "  (4)   If it returns `True`, then you can use `onnxrt` backend.\n"
+            "  (5)   If it returns `False`, please execute the package importing section in "
+            "torch/onnx/_internal/onnxruntime.py under pdb line-by-line to see which import fails."
+        )
+
+    register_backend(name="onnxrt", compiler_fn=information_displaying_backend)

venv/lib/python3.10/site-packages/torch/_dynamo/backends/registry.py

@@ -0,0 +1,115 @@
+# mypy: ignore-errors
+
+import functools
+import sys
+from typing import Callable, Dict, List, Optional, Protocol, Sequence, Tuple
+
+import torch
+from torch import fx
+
+
+class CompiledFn(Protocol):
+    def __call__(self, *args: torch.Tensor) -> Tuple[torch.Tensor, ...]:
+        ...
+
+
+CompilerFn = Callable[[fx.GraphModule, List[torch.Tensor]], CompiledFn]
+
+_BACKENDS: Dict[str, CompilerFn] = dict()
+
+
+def register_backend(
+    compiler_fn: Optional[CompilerFn] = None,
+    name: Optional[str] = None,
+    tags: Sequence[str] = (),
+):
+    """
+    Decorator to add a given compiler to the registry to allow calling
+    `torch.compile` with string shorthand.  Note: for projects not
+    imported by default, it might be easier to pass a function directly
+    as a backend and not use a string.
+
+    Args:
+        compiler_fn: Callable taking a FX graph and fake tensor inputs
+        name: Optional name, defaults to `compiler_fn.__name__`
+        tags: Optional set of string tags to categorize backend with
+    """
+    if compiler_fn is None:
+        # @register_backend(name="") syntax
+        return functools.partial(register_backend, name=name, tags=tags)
+    assert callable(compiler_fn)
+    name = name or compiler_fn.__name__
+    assert name not in _BACKENDS, f"duplicate name: {name}"
+    _BACKENDS[name] = compiler_fn
+    compiler_fn._tags = tuple(tags)
+    return compiler_fn
+
+
+register_debug_backend = functools.partial(register_backend, tags=("debug",))
+register_experimental_backend = functools.partial(
+    register_backend, tags=("experimental",)
+)
+
+
+def lookup_backend(compiler_fn):
+    """Expand backend strings to functions"""
+    if isinstance(compiler_fn, str):
+        if compiler_fn not in _BACKENDS:
+            _lazy_import()
+        if compiler_fn not in _BACKENDS:
+            _lazy_import_entry_point(compiler_fn)
+        if compiler_fn not in _BACKENDS:
+            from ..exc import InvalidBackend
+
+            raise InvalidBackend(name=compiler_fn)
+        compiler_fn = _BACKENDS[compiler_fn]
+    return compiler_fn
+
+
+def list_backends(exclude_tags=("debug", "experimental")) -> List[str]:
+    """
+    Return valid strings that can be passed to:
+
+        torch.compile(..., backend="name")
+    """
+    _lazy_import()
+    exclude_tags = set(exclude_tags or ())
+    return sorted(
+        [
+            name
+            for name, backend in _BACKENDS.items()
+            if not exclude_tags.intersection(backend._tags)
+        ]
+    )
+
+
+@functools.lru_cache(None)
+def _lazy_import():
+    from .. import backends
+    from ..utils import import_submodule
+
+    import_submodule(backends)
+
+    from ..repro.after_dynamo import dynamo_minifier_backend
+
+    assert dynamo_minifier_backend is not None
+
+
+@functools.lru_cache(None)
+def _lazy_import_entry_point(backend_name: str):
+    from importlib.metadata import entry_points
+
+    compiler_fn = None
+    group_name = "torch_dynamo_backends"
+    if sys.version_info < (3, 10):
+        backend_eps = entry_points()
+        eps = [ep for ep in backend_eps.get(group_name, ()) if ep.name == backend_name]
+        if len(eps) > 0:
+            compiler_fn = eps[0].load()
+    else:
+        backend_eps = entry_points(group=group_name)
+        if backend_name in backend_eps.names:
+            compiler_fn = backend_eps[backend_name].load()
+
+    if compiler_fn is not None and backend_name not in list_backends(tuple()):
+        register_backend(compiler_fn=compiler_fn, name=backend_name)

venv/lib/python3.10/site-packages/torch/_dynamo/backends/tensorrt.py

@@ -0,0 +1,14 @@
+# mypy: ignore-errors
+
+# import torch  # type: ignore[import]
+# from .common import device_from_inputs, fake_tensor_unsupported  # type: ignore[import]
+# from .registry import register_backend  # type: ignore[import]
+
+"""
+Placeholder for TensorRT backend for dynamo via torch-tensorrt
+"""
+
+# @register_backend
+# def tensorrt(gm, example_inputs):
+#    import torch_tensorrt # type: ignore[import]
+#    pass

venv/lib/python3.10/site-packages/torch/_dynamo/backends/torchxla.py

@@ -0,0 +1,75 @@
+# mypy: ignore-errors
+
+import logging
+import warnings
+
+from functorch.compile import make_boxed_func
+
+from ..backends.common import aot_autograd
+from .registry import register_backend, register_experimental_backend
+
+log = logging.getLogger(__name__)
+
+
+@register_experimental_backend
+def torchxla_trivial(gm, fake_tensor_inputs):
+    return gm
+
+
+@register_experimental_backend
+def torchxla_trace_once(model, fake_tensor_inputs):
+    warnings.warn(
+        "This backend will be deprecated in 2.2, please use `openxla` backend instead"
+    )
+
+    return xla_backend_helper(model, fake_tensor_inputs)
+
+
+@register_backend
+def openxla_eval(model, fake_tensor_inputs):
+    return xla_backend_helper(model, fake_tensor_inputs, boxed=False)
+
+
+def openxla_eval_boxed(model, fake_tensor_inputs):
+    return xla_backend_helper(model, fake_tensor_inputs, boxed=True)
+
+
+def xla_backend_helper(model, fake_tensor_inputs, boxed=False):
+    try:
+        import torch_xla.core.dynamo_bridge as bridge
+    except ImportError as e:
+        raise ImportError(
+            "Please follow the instruction in https://github.com/pytorch/xla#pytorchxla to install torch_xla"
+        ) from e
+
+    compiled_graph = None
+
+    def fwd(*args):
+        nonlocal model
+        nonlocal compiled_graph
+        if compiled_graph is None:
+            compiled_graph = bridge.extract_compiled_graph(model, args)
+            del model
+        return compiled_graph(*args)
+
+    return make_boxed_func(fwd) if boxed else fwd
+
+
+aot_torchxla_trivial = aot_autograd(
+    fw_compiler=torchxla_trivial,
+)
+register_experimental_backend(
+    name="aot_torchxla_trivial", compiler_fn=aot_torchxla_trivial
+)
+
+aot_torchxla_trace_once = aot_autograd(
+    fw_compiler=torchxla_trace_once,
+)
+register_experimental_backend(
+    name="aot_torchxla_trace_once", compiler_fn=aot_torchxla_trace_once
+)
+
+openxla = aot_autograd(
+    fw_compiler=openxla_eval_boxed,
+)
+register_backend(name="openxla", compiler_fn=openxla)

venv/lib/python3.10/site-packages/torch/_dynamo/backends/tvm.py

@@ -0,0 +1,172 @@
+# mypy: ignore-errors
+
+import functools
+import importlib
+import logging
+import os
+import tempfile
+
+import torch
+from .common import device_from_inputs, fake_tensor_unsupported
+
+from .registry import register_backend
+
+log = logging.getLogger(__name__)
+
+
+@register_backend
+@fake_tensor_unsupported
+def tvm(gm, example_inputs, *, scheduler=None, trials=20000):
+    import tvm  # type: ignore[import]
+    from tvm import relay  # type: ignore[import]
+    from tvm.contrib import graph_executor  # type: ignore[import]
+
+    jit_mod = torch.jit.trace(gm, example_inputs)
+    device = device_from_inputs(example_inputs)
+    shape_list = [(f"inp_{idx}", i.shape) for idx, i in enumerate(example_inputs)]
+    example_outputs = gm(*example_inputs)
+    if len(example_outputs) == 0:
+        log.warning("Explicitly fall back to eager due to zero output")
+        return gm.forward
+    mod, params = relay.frontend.from_pytorch(jit_mod, shape_list)
+    if device.type == "cuda":
+        dev = tvm.cuda(device.index)
+        target = tvm.target.cuda()
+    else:
+        dev = tvm.cpu(0)
+        target = tvm.target.Target(llvm_target())
+
+    if scheduler is None:
+        scheduler = os.environ.get("TVM_SCHEDULER", None)
+
+    if scheduler == "auto_scheduler":
+        from tvm import auto_scheduler
+
+        log_file = tempfile.NamedTemporaryFile()
+
+        if not os.path.exists(log_file):
+            tasks, task_weights = auto_scheduler.extract_tasks(
+                mod["main"], params, target
+            )
+            for task in tasks:
+                print(task.compute_dag)
+            else:
+                print("No tasks")
+            if len(tasks) != 0:
+                tuner = auto_scheduler.TaskScheduler(tasks, task_weights)
+                if not os.path.exists(log_file):
+                    assert trials > 0
+                    tune_option = auto_scheduler.TuningOptions(
+                        num_measure_trials=trials,
+                        measure_callbacks=[auto_scheduler.RecordToFile(log_file)],
+                        early_stopping=2000,
+                    )
+                    try:
+                        tuner.tune(tune_option)
+                    except Exception:
+                        if os.path.exists(log_file):
+                            os.unlink(log_file)
+                        raise
+
+        with auto_scheduler.ApplyHistoryBest(log_file):
+            with tvm.transform.PassContext(
+                opt_level=3, config={"relay.backend.use_auto_scheduler": True}
+            ):
+                lib = relay.build(mod, target=target, params=params)
+    elif scheduler == "meta_schedule":
+        from tvm import meta_schedule as ms
+
+        with tempfile.TemporaryDirectory() as work_dir:
+            if device.type != "cuda":
+                # meta_schedule needs num-cores to be specified
+                # here we use the maximum core count
+                target = tvm.target.Target(
+                    f"{llvm_target()} --num-cores {ms.utils.cpu_count(logical=False)}"
+                )
+            # TODO(shingjan): This could be replaced by tvm.contrib.torch.optimize_torch
+            # once USE_PT_TVMDSOOP is updated and turned on by default in TVM.
+            database = ms.relay_integration.tune_relay(
+                mod=mod,
+                target=target,
+                work_dir=work_dir,
+                max_trials_global=20000,
+                num_trials_per_iter=64,
+                params=params,
+                strategy="evolutionary",
+            )
+            lib = ms.relay_integration.compile_relay(
+                database=database,
+                mod=mod,
+                target=target,
+                params=params,
+            )
+    elif scheduler == "default" or not scheduler:
+        # no autotuning
+        with tvm.transform.PassContext(opt_level=10):
+            lib = relay.build(mod, target=target, params=params)
+    else:
+        raise NotImplementedError(
+            "This tuning option is invalid/not implemented for torchdynamo's TVM-related backend. "
+            "There are three available options: default, auto_scheduler and meta_schedule."
+        )
+    m = graph_executor.GraphModule(lib["default"](dev))
+
+    def to_torch_tensor(nd_tensor):
+        """A helper function to transfer a NDArray to torch.tensor."""
+        if nd_tensor.dtype == "bool":
+            # DLPack does not support boolean so it can't be handled by
+            # torch.utils.dlpack.from_pack. Workaround by going through
+            # numpy, although this brings additional data copy overhead.
+            return torch.from_numpy(nd_tensor.numpy())
+        return torch.utils.dlpack.from_dlpack(nd_tensor.to_dlpack())
+
+    def to_tvm_tensor(torch_tensor):
+        """A helper function to transfer a torch.tensor to NDArray."""
+        if torch_tensor.dtype == torch.bool:
+            # same reason as above, fallback to numpy conversion which
+            # could introduce data copy overhead
+            return tvm.nd.array(torch_tensor.cpu().numpy())
+        return tvm.nd.from_dlpack(torch_tensor)
+
+    def exec_tvm(*i_args):
+        args = [a.contiguous() for a in i_args]
+        shape_info, _ = m.get_input_info()
+        active_inputs = {name for name, _ in shape_info.items()}
+        for idx, arg in enumerate(args, 0):
+            if arg.dim() != 0:
+                if arg.requires_grad:
+                    arg = arg.detach()
+                inp_name = f"inp_{idx}"
+                if inp_name not in active_inputs:
+                    log.warning(
+                        "input %s skipped as not found in tvm's runtime library",
+                        inp_name,
+                    )
+                    continue
+                m.set_input(
+                    inp_name,
+                    to_tvm_tensor(arg),
+                )
+        m.run()
+        return [to_torch_tensor(m.get_output(i)) for i in range(m.get_num_outputs())]
+
+    return exec_tvm
+
+
+tvm_meta_schedule = functools.partial(tvm, scheduler="meta_schedule")
+tvm_auto_scheduler = functools.partial(tvm, scheduler="auto_scheduler")
+
+
+def has_tvm():
+    try:
+        importlib.import_module("tvm")
+        return True
+    except ImportError:
+        return False
+
+
+@functools.lru_cache(None)
+def llvm_target():
+    if "avx512" in open("/proc/cpuinfo").read():
+        return "llvm -mcpu=skylake-avx512"
+    return "llvm -mcpu=core-avx2"

venv/lib/python3.10/site-packages/torch/_dynamo/repro/after_aot.py

@@ -0,0 +1,932 @@
+import argparse
+import copy
+import functools
+import io
+import logging
+import os
+import shutil
+import subprocess
+import sys
+import textwrap
+import uuid
+from importlib import import_module
+from tempfile import TemporaryFile
+from typing import Any, Callable, Dict, Union
+
+import torch
+import torch.fx as fx
+import torch.nn as nn
+from torch._dynamo.debug_utils import (
+    _cuda_system_info_comment,
+    AccuracyError,
+    backend_accuracy_fails,
+    BuckTargetWriter,
+    cast_to_fp64,
+    extra_imports,
+    generate_config_string,
+    helper_for_dump_minify,
+    InputReader,
+    InputWriter,
+    MAX_CONSTANT_NUMEL_INLINE,
+    minifier_dir,
+    NNModuleToString,
+    NopInputReader,
+    same_two_models,
+)
+from torch._dynamo.utils import clone_inputs, counters, same
+from torch.fx.experimental.proxy_tensor import make_fx
+from torch.fx.experimental.symbolic_shapes import (
+    fx_placeholder_targets,
+    has_free_symbols,
+)
+from torch.hub import tqdm
+
+from .. import config
+
+log = logging.getLogger(__name__)
+
+
+inductor_config = import_module("torch._inductor.config")
+use_buck = inductor_config.is_fbcode()
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           MAIN ENTRY POINT
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def wrap_compiler_debug(unconfigured_compiler_fn, compiler_name: str):
+    """
+    Minifier for Fx Graph modules after Aot Autograd has finished. We wrap both
+    forward and backward call separately with the backend compiler_fn - like
+    inductor or nvfuser. Intercepting after Aot Autograd presents neat
+    abstraction, where all the params are lifted as graph inputs, making it easy
+    to save the graph as a string.
+    """
+
+    @functools.wraps(unconfigured_compiler_fn)
+    def debug_wrapper(gm, example_inputs, **kwargs):
+        from torch._subclasses import FakeTensorMode
+
+        compiler_fn = functools.partial(unconfigured_compiler_fn, **kwargs)
+
+        from torch._functorch.aot_autograd import get_aot_graph_name
+
+        graph_name = get_aot_graph_name()
+
+        # TODO: why do we need to deepcopy the original graph?
+        orig_graph = copy.deepcopy(gm.graph)
+        assert config.repro_after in ("dynamo", "aot", None)
+
+        try:
+            # Call the compiler_fn - which is either aot_autograd or inductor
+            # with fake inputs
+            inner_compiled_fn = compiler_fn(gm, example_inputs)
+        except Exception as e:
+            # TODO: Failures here are troublesome because no real inputs,
+            # need a different serialization strategy
+            if config.repro_after == "aot":
+                if config.repro_level == 1:
+                    dump_compiler_graph_state(
+                        fx.GraphModule(gm, orig_graph),
+                        example_inputs,
+                        compiler_name,
+                    )
+                elif config.repro_level == 2:
+                    dump_to_minify(
+                        fx.GraphModule(gm, orig_graph),
+                        example_inputs,
+                        compiler_name,
+                    )
+                log.error("CompilerError")
+            raise
+
+        # We may run regular PyTorch compute that may trigger Dynamo, do NOT
+        # recursively attempt to accuracy minify in that case!
+        def deferred_for_real_inputs(real_inputs):
+            # This is a bit obscure: if we recursively try to accuracy minify
+            # the SAME function, this would trigger.  But most of the time
+            # we should never hit this branch
+            if config.repro_after != "aot":
+                return inner_compiled_fn(real_inputs)
+            with config.patch(repro_after=None):
+                return inner_debug_fn(real_inputs)
+
+        def inner_debug_fn(real_inputs):
+            """
+            Aot Autograd fw_compiler and bw_compiler can have fake tensors. So,
+            example_inputs can be fake tensors. We can call compiler_fn (which is
+            inductor or nvfuser) with fake tensors but the actually compiled_fn
+            should be called with real tensors. Therefore, the actual invocation
+            is deferred.
+            """
+            # Copy the tensor attrs like shape, stride etc by converting to Fake Tensor
+            # because inductor clears the tensor list in its codegen. And example_inputs
+            # are available only for the first invocation.
+            fake_mode = FakeTensorMode()
+            copy_tensor_attrs = [
+                fake_mode.from_tensor(x) if isinstance(x, torch.Tensor) else x
+                for x in real_inputs
+            ]
+            if config.repro_level == 3:
+                # Always dump the original module in case we have segfaults
+                dump_to_minify(
+                    fx.GraphModule(gm, orig_graph), real_inputs, compiler_name
+                )
+
+            if config.repro_level == 4:
+                if compiler_name != "inductor":
+                    raise NotImplementedError(
+                        "Accuracy minification is supported for inductor only"
+                    )
+                if backend_aot_accuracy_fails(gm, real_inputs, compiler_fn):
+                    log.warning(
+                        "Accuracy failed for the AOT Autograd graph %s", graph_name
+                    )
+                    dump_compiler_graph_state(
+                        fx.GraphModule(gm, orig_graph),
+                        real_inputs,
+                        f"{compiler_name}_accuracy",
+                    )
+                    dump_to_minify(
+                        fx.GraphModule(gm, orig_graph),
+                        real_inputs,
+                        f"{compiler_name}_accuracy",
+                    )
+                    raise AccuracyError("Bad accuracy detected")
+                else:
+                    # Call the compiled function with real inputs
+                    return inner_compiled_fn(real_inputs)
+            else:
+                try:
+                    # Call the compiled function with real inputs
+                    out = inner_compiled_fn(real_inputs)
+                    # sync cuda kernels to ensure IMA detection
+                    for arg in example_inputs:
+                        if isinstance(arg, torch.Tensor) and arg.is_cuda:
+                            torch.cuda.synchronize()
+                            break
+                    return out
+                except Exception as e:
+                    if config.repro_level == 1:
+                        dump_compiler_graph_state(
+                            fx.GraphModule(gm, orig_graph),
+                            copy_tensor_attrs,
+                            compiler_name,
+                        )
+                    elif config.repro_level == 2:
+                        dump_to_minify(
+                            fx.GraphModule(gm, orig_graph),
+                            copy_tensor_attrs,
+                            compiler_name,
+                        )
+                    raise
+
+        if config.repro_after == "aot":
+            compiled_fn = deferred_for_real_inputs
+            compiled_fn._boxed_call = True  # type: ignore[attr-defined]
+            return compiled_fn
+        else:
+            return inner_compiled_fn
+
+    return debug_wrapper
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           DUMP REPROS
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def generate_compiler_repro_string(gm, args, *, stable_output=False, save_dir=None):
+    model_str = textwrap.dedent(
+        f"""
+import torch
+from torch import tensor, device
+import torch.fx as fx
+from torch._dynamo.testing import rand_strided
+from math import inf
+import torch._inductor.inductor_prims
+
+{generate_config_string(stable_output=stable_output)}
+
+isolate_fails_code_str = None
+
+{extra_imports}
+
+        """
+    )
+    if not stable_output:
+        model_str += f"# torch version: {torch.version.__version__}\n"
+        if hasattr(torch.version, "cuda"):
+            model_str += f"# torch cuda version: {torch.version.cuda}\n"
+        if hasattr(torch.version, "git_version"):
+            model_str += f"# torch git version: {torch.version.git_version}\n\n\n"
+        model_str += _cuda_system_info_comment()
+
+    model_str += NNModuleToString.convert(gm)
+
+    # get hint shape/stride when dynamic shape enabled
+    def hint_if_symint(x):
+        return tuple(i.node.hint if isinstance(i, torch.SymInt) else i for i in x)
+
+    writer = InputWriter(save_dir)
+    for placeholder, arg in zip(fx_placeholder_targets(gm), args):
+        if isinstance(arg, (int, torch.SymInt)):
+            writer.symint(placeholder, arg)
+        elif isinstance(arg, torch.Tensor):
+            # TODO: improve these names with FQN
+            writer.tensor(placeholder, arg)
+        else:
+            raise TypeError(f"arg is neither SymInt/int nor torch.Tensor, {arg}")
+
+    model_str += "\n".join(writer.lines()) + "\n"
+
+    model_str += "mod = Repro()\n"
+    return model_str
+
+
+def save_graph_repro(
+    fd,
+    gm,
+    args,
+    compiler_name,
+    *,
+    stable_output=False,
+    save_dir=None,
+    command="run",
+    accuracy=None,
+    tracing_mode=None,
+    check_str=None,
+):
+    fd.write(
+        generate_compiler_repro_string(
+            gm,
+            args,
+            stable_output=stable_output,
+            save_dir=save_dir,
+        )
+    )
+    if accuracy is None:
+        accuracy = "_accuracy" in compiler_name
+    if tracing_mode is None:
+        tracing_mode = "real"
+        if any(has_free_symbols(a) for a in args):
+            tracing_mode = "symbolic"
+    fd.write("if __name__ == '__main__':\n")
+    fd.write("    from torch._dynamo.repro.after_aot import run_repro\n")
+    fd.write(
+        f"    with torch.no_grad():\n"
+        f"        run_repro(mod, load_args, accuracy={accuracy!r}, command={command!r}, "
+        f"save_dir={save_dir!r}, tracing_mode={tracing_mode!r}, check_str={check_str!r}"
+        ")\n"
+    )
+
+
+def dump_compiler_graph_state(gm, args, compiler_name, *, accuracy=None):
+    subdir = os.path.join(minifier_dir(), "checkpoints")
+    if not os.path.exists(subdir):
+        os.makedirs(subdir, exist_ok=True)
+    file_name = os.path.join(subdir, f"{len(gm.graph.nodes)}.py")
+    log.warning(
+        "Writing checkpoint with %s nodes to %s", len(gm.graph.nodes), file_name
+    )
+    with open(file_name, "w") as fd:
+        save_graph_repro(
+            fd, gm, args, compiler_name, save_dir=subdir, accuracy=accuracy
+        )
+    curdir = os.getcwd()
+    repro_path = os.path.join(curdir, "repro.py")
+    try:
+        shutil.copyfile(file_name, repro_path)
+        log.warning("Copying repro file for convenience to %s", repro_path)
+        if use_buck:
+            BuckTargetWriter(file_name).write()
+    except OSError:
+        log.warning("No write permissions for %s", repro_path)
+        pass
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           DUMP MINIFIER
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def dump_to_minify(gm, args, compiler_name: str):
+    out = io.StringIO()
+    # TODO: factor this out
+    subdir = os.path.join(minifier_dir(), "checkpoints")
+    if not os.path.exists(subdir):
+        os.makedirs(subdir, exist_ok=True)
+    save_graph_repro(out, gm, args, compiler_name, save_dir=subdir, command="minify")
+    return helper_for_dump_minify(out.getvalue())
+
+
+def isolate_fails(
+    fx_g,
+    args,
+    compiler_name: str,
+    env=None,
+    save_dir=None,
+    accuracy=None,
+    tracing_mode=None,
+    check_str=None,
+):
+    if env is None:
+        env = {}
+    subdir = os.path.join(os.getcwd(), "isolate")
+    if not os.path.exists(subdir):
+        os.makedirs(subdir, exist_ok=True)
+    file_name = os.path.join(subdir, f"{str(uuid.uuid4())[:5]}.py")
+    with open(file_name, "w") as fd:
+        save_graph_repro(
+            fd,
+            fx_g,
+            args,
+            compiler_name,
+            save_dir=save_dir,
+            command="minifier-query",
+            accuracy=accuracy,
+            tracing_mode=tracing_mode,
+            check_str=check_str,
+        )
+    # with open(file_name, "r") as fd:
+    #     print(fd.read())
+    new_env = os.environ.copy()
+    new_env = {**new_env, **env}
+    stdout, stderr = TemporaryFile(), TemporaryFile()
+
+    if use_buck:
+        cmd = BuckTargetWriter(file_name).write(print_msg=False)
+    else:
+        cmd = ["python", file_name]
+
+    p = subprocess.Popen(
+        cmd,
+        cwd=subdir,
+        stdout=stdout,
+        stderr=stderr,
+        env=new_env,
+    )
+    p.wait()
+
+    stdout.seek(0)
+    stderr.seek(0)
+    print(
+        textwrap.indent(stdout.read().decode("utf-8"), prefix=">>  "), file=sys.stdout
+    )
+    print(
+        textwrap.indent(stderr.read().decode("utf-8"), prefix=">>  "), file=sys.stderr
+    )
+    # print(f"Isolated test failed - {file_name}")
+    return p.returncode != 0
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                       MINIFIER TOOLS
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def inductor_fails(fx_g, args, check_str=None):
+    has_cuda = False
+    for arg in args:
+        if isinstance(arg, torch.Tensor) and arg.is_cuda:
+            has_cuda = True
+            break
+
+    def sync():
+        if has_cuda:
+            # Ensures that segfaults are surfaced
+            torch.cuda.synchronize()
+
+    from torch._inductor.compile_fx import compile_fx_inner
+
+    try:
+        result = fx_g(*args)
+        assert isinstance(result, (tuple, list))
+        assert not any(isinstance(x, (tuple, list)) for x in result)
+    except Exception:
+        return False
+
+    sync()
+
+    try:
+        compile_mod = compile_fx_inner(fx_g, args)
+        compile_mod(args)
+        sync()
+    except Exception as e:
+        if check_str is not None and check_str not in repr(e):
+            return False
+        print(repr(e))
+        return True
+    return False
+
+
+def inductor_accuracy_fails(
+    fx_g, args, check_str=None, *, require_fp64=False, ignore_non_fp=False
+):
+    from torch._inductor.compile_fx import compile_fx_inner
+
+    return backend_aot_accuracy_fails(
+        fx_g,
+        args,
+        compile_fx_inner,
+        require_fp64=require_fp64,
+        ignore_non_fp=ignore_non_fp,
+    )
+
+
+backend_aot_accuracy_fails = functools.partial(backend_accuracy_fails, only_fwd=True)
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           REPRO MAIN
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def repro_common(options, mod, load_args):
+    # Invariant for graphs we generate with the repro script
+    assert not any(mod.named_parameters())
+    for n, b in mod.named_buffers():
+        if b.numel() > MAX_CONSTANT_NUMEL_INLINE:
+            log.warning(
+                "Constant %s was not serialized, generated random data instead. "
+                "If you think this is affecting you, please comment on "
+                "https://github.com/pytorch/pytorch/issues/100468",
+                n,
+            )
+
+    if not hasattr(load_args, "_version"):
+        log.warning(
+            "load_args does not have a _version attribute, please file a bug to PyTorch "
+            "and describe how you generate this repro script"
+        )
+    else:
+        if load_args._version > 0:
+            log.warning(
+                "load_args is version %s, but this version of PyTorch only supports "
+                "version 0.  We will try to run it anyway but there may be an incompatibility; "
+                "if so, try upgrading your version of PyTorch.",
+                load_args._version,
+            )
+
+    nop_reader = NopInputReader()
+    load_args(nop_reader)
+
+    with tqdm(desc="Loading inputs", total=nop_reader.total) as pbar:
+        input_reader = InputReader(save_dir=options.save_dir, pbar=pbar)
+        load_args(input_reader)
+        args = input_reader.args
+
+    # Turn mod into a GraphModule the slow way
+    # TODO: speed this up
+    mod = make_fx(mod, tracing_mode=options.tracing_mode)(*args)
+
+    torch._inductor.config.generate_intermediate_hooks = True
+
+    return mod, args
+
+
+ACCURACY_FAILS: Dict[str, Callable[[nn.Module, Any], bool]] = {
+    "": inductor_fails,
+    # This might look inverted but it's not.  strict_accuracy means "we will
+    # minify any time we see anything that diverges", whereas accuracy is more
+    # conservative, and will only minify if there is a meaningful fp64
+    # divergence
+    "accuracy": functools.partial(
+        inductor_accuracy_fails, require_fp64=True, ignore_non_fp=True
+    ),
+    "strict_accuracy": inductor_accuracy_fails,
+}
+
+
+def repro_minifier_query(options, mod, load_args):
+    mod, args = repro_common(options, mod, load_args)
+    fail_fn = functools.partial(
+        ACCURACY_FAILS[options.accuracy], check_str=options.check_str
+    )
+    if fail_fn(mod, args):
+        sys.exit(1)
+    else:
+        sys.exit(0)
+
+
+def repro_minify(options, mod, load_args):
+    from functorch.compile import minifier
+
+    mod, args = repro_common(options, mod, load_args)
+    compiler_name = "inductor_accuracy" if options.accuracy != "" else "inductor"
+
+    favored_device = 1 if torch.cuda.device_count() >= 2 else 0
+    env_variables = {"CUDA_VISIBLE_DEVICES": str(favored_device)}
+
+    module_fails: Any
+    if options.isolate:
+        module_fails = functools.partial(
+            isolate_fails,
+            env=env_variables,
+            compiler_name=compiler_name,
+            save_dir=options.save_dir,
+            accuracy=options.accuracy,
+            tracing_mode=options.tracing_mode,
+        )
+    else:
+        module_fails = ACCURACY_FAILS[options.accuracy]
+
+    minifier(
+        mod,
+        args,
+        module_fails=functools.partial(module_fails, check_str=options.check_str),
+        dump_state=functools.partial(
+            dump_compiler_graph_state, compiler_name=compiler_name
+        ),
+        save_dir=options.save_dir,
+        offload_to_disk=options.offload_to_disk,
+        skip_offload=options.skip_saving_eager_intermediates,
+        skip_sanity=options.skip_sanity,
+        max_granularity=options.max_granularity,
+    )
+
+
+def repro_analyze(options, mod, load_args):
+    from torch._inductor.compile_fx import compile_fx_inner
+    from torch._inductor.hooks import intermediate_hook
+
+    mod, args = repro_common(options, mod, load_args)
+
+    # TODO: The logic for cloning inputs/models here is intentionally
+    # modeled off of run_fwd_maybe_bwd, but arguably it is better not to
+    # clone inputs (as you are doubling your effective GPU memory usage).
+    # It is certainly faster though!  It probably makes sense to let the
+    # user specify the offload strategy.
+
+    with tqdm(desc="Compiling"):
+        compiled = compile_fx_inner(mod, args)
+    total = counters["inductor"]["intermediate_hooks"]
+
+    known_names = set()
+
+    def save_hook(name, val):
+        known_names.add(name)
+        if not options.skip_saving_inductor_intermediates:
+            writer.write_tensor(os.path.join("inductor", name), val)
+        pbar.update(1)  # type: ignore[has-type]
+
+    writer = torch.utils._content_store.ContentStoreWriter(
+        options.save_dir, stable_hash=options.stable_hash
+    )
+    reader = torch.utils._content_store.ContentStoreReader(options.save_dir)
+
+    new_args = clone_inputs(args)
+    with intermediate_hook(save_hook), tqdm(
+        desc="Saving inductor intermediates", total=total
+    ) as pbar:
+        compiled(new_args)
+        assert not new_args
+
+    def compare_tuples(tuple1, tuple2):
+        diff_indices = [i for i in range(len(tuple1)) if tuple1[i] != tuple2[i]]
+        diff_values = [(tuple1[i], tuple2[i]) for i in diff_indices]
+
+        if not diff_values:
+            return None
+        else:
+            return " and ".join(f"{a} != {b}" for a, b in diff_values)
+
+    def check_hook(name, val):
+        meta = writer.compute_tensor_metadata(val)
+        meta2 = reader.read_tensor_metadata(os.path.join("inductor", name))
+        reason = compare_tuples(meta, meta2)
+        if reason is not None:
+            pbar.write(f"NONDETERMINISTIC INDUCTOR at {name} ({reason})")
+        pbar.update(1)
+
+    if not options.skip_check_deterministic:
+        new_args = clone_inputs(args)
+        with intermediate_hook(check_hook), tqdm(
+            desc="Checking inductor determinism", total=total
+        ) as pbar:
+            compiled(new_args)
+            assert not new_args
+
+    class WriterInterp(fx.Interpreter):
+        def __init__(self, mod, subdir):
+            super().__init__(mod)
+            self.subdir = subdir
+
+        def run_node(self, n):
+            r = super().run_node(n)
+            name = n.name
+            if name in known_names:
+                pbar.update(1)
+                writer.write_tensor(os.path.join(self.subdir, name), r)
+            return r
+
+    # NB: the module cast doesn't actually do anything, since there are no
+    # parameters/buffers on the module
+    if not options.skip_saving_float64_intermediates:
+        new_mod, new_args = cast_to_fp64(copy.deepcopy(mod), clone_inputs(args))
+        with tqdm(desc="Saving float64 intermediates", total=total) as pbar:
+            WriterInterp(new_mod, "float64").boxed_run(new_args)
+        assert not new_args
+
+    class ExactReaderInterp(fx.Interpreter):
+        def run_node(self, n):
+            r = super().run_node(n)
+            name = n.name
+            if name in known_names:
+                meta = writer.compute_tensor_metadata(r)
+                meta2 = reader.read_tensor_metadata(os.path.join("float64", name))
+                reason = compare_tuples(meta, meta2)
+                if reason is not None:
+                    pbar.write(f"NONDETERMINISTIC FLOAT64 at {name} ({reason})")
+                pbar.update(1)
+            return r
+
+    # TODO: check eager determinism
+
+    if not options.skip_check_deterministic:
+        new_mod, new_args = cast_to_fp64(copy.deepcopy(mod), clone_inputs(args))
+        with tqdm(desc="Checking float64 determinism", total=total) as pbar:
+            ExactReaderInterp(new_mod).boxed_run(new_args)
+            assert not new_args
+
+    # Now that we've saved everything, interp through the eager graph
+    # and do comparisons
+    class ReaderInterp(fx.Interpreter):
+        def run_node(self, n):
+            r = super().run_node(n)
+            name = n.name
+            if name in known_names:
+                inductor = reader.read_tensor(os.path.join("inductor", name))
+                float64 = reader.read_tensor(os.path.join("float64", name))
+                logged = False
+
+                def log_error(msg, *args):
+                    nonlocal logged
+                    logged = True
+                    pbar.write(f"DIVERGED at {name}: {msg % args}")
+
+                if not same(
+                    r,
+                    inductor,
+                    float64,
+                    tol=torch._dynamo.config.repro_tolerance,
+                    equal_nan=True,
+                    log_error=log_error,
+                ):
+                    assert logged
+                pbar.update(1)
+            return r
+
+    with tqdm(desc="Checking divergence", total=total) as pbar:
+        ReaderInterp(mod).boxed_run(args)
+    assert not args
+
+
+def repro_run(options, mod, load_args):
+    from torch._inductor.compile_fx import compile_fx_inner
+
+    mod, args = repro_common(options, mod, load_args)
+
+    from torch.cuda import synchronize
+
+    compiled = compile_fx_inner(mod, args)
+
+    if options.accuracy != "":
+        # We don't really respect --accuracy vs --strict-accuracy here, it
+        # seems counterintuitive
+        if not same_two_models(mod, compiled, args, only_fwd=True):
+            raise AccuracyError("Bad accuracy detected")
+    else:
+        need_sync = False
+        for arg in args:
+            if isinstance(arg, torch.Tensor) and arg.is_cuda:
+                need_sync = True
+                break
+        ref = compiled(list(args))
+        if need_sync:
+            synchronize()  # ensure segfaults are surfaced
+    return lambda: compiled(list(args))
+
+
+# TODO: lazily load the inputs or something, rather than cloning them
+def run_repro(
+    mod,
+    load_args,
+    *,
+    command="run",
+    accuracy: Union[bool, str] = "",
+    save_dir=None,
+    tracing_mode=None,
+    patch_code=None,
+    check_str=None,
+    **kwargs,
+):
+    for k in kwargs:
+        log.warning(
+            "Unrecognized kwarg %s; perhaps this repro was made on a newer version of PyTorch",
+            k,
+        )
+
+    if accuracy is True:
+        accuracy = "accuracy"
+    elif accuracy is False:
+        accuracy = ""
+
+    if patch_code is not None:
+        log.warning(
+            "patch_code no longer works on this version of PyTorch, silently ignoring"
+        )
+
+    parser = argparse.ArgumentParser(
+        description=f"""\
+An after_aot repro script, typically triggering a bug in PyTorch Inductor.
+When run with no arguments, this script defaults to running '{command}'.
+Extra flags may be available; to find out more, try '{command} --help'.
+There are also alternate subcommands available, see below.
+
+default settings on this script:
+  {accuracy=}
+  {tracing_mode=}
+  {save_dir=}
+  {check_str=}
+""",
+        formatter_class=argparse.RawTextHelpFormatter,
+    )
+
+    def common_flags(parser):
+        accuracy_group = parser.add_mutually_exclusive_group()
+        accuracy_group.add_argument(
+            "--no-accuracy",
+            dest="accuracy",
+            action="store_const",
+            const="",
+            default=accuracy,
+            help="do not test accuracy, just run the module and see if it errors",
+        )
+        accuracy_group.add_argument(
+            "--accuracy",
+            action="store_const",
+            const="accuracy",
+            default=accuracy,
+            help="""\
+test if the RMSE between the compiled module and the fp64 reference is greater
+than eager and the fp64 reference. This is usually more reliable than the
+standard allclose test, as we expect numeric differences from compiling, often
+improving accuracy over eager.  RMSE test allows for compiled module to
+diverge greatly from eager, as long as this divergence moves it closer to the
+'true' mathematical value of the network.  Caveats: (1) double precision can
+still suffer from rounding error, so it is not a perfect reference (see for
+example 'Herbie: Automatically Improving Floating Point Accuracy') for
+approaches that detect the necessary working precision and compute it in
+arbitrary precision floating point; unfortunately, this is not practical for
+tensor computation; (2) if there are not enough samples in the output being
+compared, we may get unlucky and have an unlucky greater RMSE than eager; this
+could be overcome by applying a more rigorous statistical test at some
+p-value, which we leave for future work.
+""",
+        )
+        accuracy_group.add_argument(
+            "--strict-accuracy",
+            dest="accuracy",
+            action="store_const",
+            const="strict_accuracy",
+            default=accuracy,
+            help="""\
+by default, when doing accuracy minification we will reject reductions which
+change the divergence from a floating point divergence to a integral/boolean
+divergence.  This is because some operations like ReLU involve temporarily
+sharp boundaries that smooth out again afterwards; without requiring
+divergence on floating point, the minifier will often fixate on divergent
+boolean tensor even though this is not the true source of the divergence.
+However, rejecting these reductions makes it more difficult for the minifier
+to make process.  Using this option will let the minifier progress for ALL
+divergences--you just might not end up with a useful repro in the end.""",
+        )
+
+        parser.add_argument(
+            "--save-dir",
+            type=str,
+            default=save_dir,
+            metavar="DIR",
+            help="directory where saved inputs live",
+        )
+        parser.add_argument(
+            "--no-save-dir",
+            dest="save_dir",
+            action="store_const",
+            const=None,
+            help="don't use any directory for saved inputs",
+        )
+        parser.add_argument(
+            "--tracing-mode",
+            type=str,
+            metavar="{real,fake,symbolic}",
+            default=tracing_mode,
+            help="how to trace the repro module into a GraphModule with metadata",
+        )
+
+    subparsers = parser.add_subparsers(
+        dest="command", metavar="{run,minify,analyze}", required=True
+    )
+
+    parser_run = subparsers.add_parser(
+        "run",
+        help="just run the repro",
+    )
+    common_flags(parser_run)
+
+    parser_minify = subparsers.add_parser(
+        "minify", help="run the minifier on the repro"
+    )
+    common_flags(parser_minify)
+    parser_minify_isolate = parser_minify.add_mutually_exclusive_group()
+    parser_minify_isolate.add_argument(
+        "--isolate",
+        action="store_true",
+        default=True,
+        help="run in separate processes to avoid interference (default)",
+    )
+    parser_minify_isolate.add_argument(
+        "--no-isolate",
+        dest="isolate",
+        action="store_false",
+        help="speed up by running all compilation in same process",
+    )
+    parser_minify.add_argument(
+        "--skip-saving-eager-intermediates",
+        action="store_true",
+        help="skip saving eager intermediates on --minify",
+    )
+    # TODO: make this an option for --analyze too
+    parser_minify.add_argument(
+        "--offload-to-disk",
+        action="store_true",
+        help="during minification, offload delta debugging intermediates to disk.  Use if you're OOMing",
+    )
+    parser_minify.add_argument(
+        "--skip-sanity",
+        action="store_true",
+        help="skip sanity check at beginning of minification on original graph",
+    )
+    parser_minify.add_argument(
+        "--max-granularity",
+        type=int,
+        default=None,
+        help="start at this granularity and work down; must be power of 2",
+    )
+    parser_minify.add_argument(
+        "--check-str",
+        type=str,
+        default=check_str,
+        help="require minified program to fail with error containing this string",
+    )
+
+    parser_analyze = subparsers.add_parser(
+        "analyze", help="run the accuracy analyzer on the repro"
+    )
+    common_flags(parser_analyze)
+    parser_analyze.add_argument(
+        "--skip-saving-inductor-intermediates",
+        action="store_true",
+        help="skip saving inductor intermediates on --analyze",
+    )
+    parser_analyze.add_argument(
+        "--skip-saving-float64-intermediates",
+        action="store_true",
+        help="skip saving float64 intermediates",
+    )
+    parser_analyze.add_argument(
+        "--skip-check-deterministic",
+        action="store_true",
+        help="skip checking that the network is deterministic",
+    )
+    parser_analyze.add_argument(
+        "--stable-hash",
+        action="store_true",
+        help="use SHA-1 checksum instead of fast (but possibly unsound) hash",
+    )
+
+    # Run the repro in the context of minification, inverting exit code meaning
+    parser_minifier_query = subparsers.add_parser(
+        "minifier-query",
+    )
+    common_flags(parser_minifier_query)
+    parser_minifier_query.add_argument(
+        "--check-str",
+        type=str,
+        default=check_str,
+        help="require minified program to fail with error containing this string",
+    )
+
+    args = None
+    if len(sys.argv) <= 1:
+        args = [command, *sys.argv[1:]]
+
+    options = parser.parse_args(args)
+    COMMAND_FNS = {
+        "minify": repro_minify,
+        "analyze": repro_analyze,
+        "minifier-query": repro_minifier_query,
+        "run": repro_run,
+    }
+    return COMMAND_FNS[options.command](options, mod, load_args)

venv/lib/python3.10/site-packages/torch/_dynamo/repro/after_dynamo.py

@@ -0,0 +1,566 @@
+import argparse
+import copy
+import functools
+import logging
+import os
+import shutil
+import sys
+import textwrap
+from importlib import import_module
+from typing import Union
+
+import torch
+import torch.fx as fx
+
+from torch._dynamo.debug_utils import (
+    AccuracyError,
+    backend_accuracy_fails,
+    BUCK_CMD_PREFIX,
+    BuckTargetWriter,
+    extra_imports,
+    generate_config_string,
+    helper_for_dump_minify,
+    InputReader,
+    InputWriter,
+    minifier_dir,
+    NNModuleToString,
+    NopInputReader,
+    run_fwd_maybe_bwd,
+    same_two_models,
+)
+from torch.fx.experimental.symbolic_shapes import fx_placeholder_targets
+from torch.hub import tqdm
+
+from .. import config
+from ..backends.registry import lookup_backend, register_debug_backend
+from ..debug_utils import clone_inputs_retaining_gradness
+
+log = logging.getLogger(__name__)
+
+
+inductor_config = import_module("torch._inductor.config")
+use_buck = inductor_config.is_fbcode()
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           MAIN ENTRY POINT
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def wrap_backend_debug(unconfigured_compiler_fn, compiler_name: str):
+    """
+    A minifier decorator that wraps the TorchDynamo produced Fx graph modules.
+    As opposed to wrap_compiler_debug, this wrapper intercepts at the
+    TorchDynamo produced Fx Graph Module. This makes it backend-agnostic to some
+    level, e.g., it is useful for minifying issues related to Aot Autograd
+    tracing.  If an error is found, we minify and save the minified repro in
+    repro.tar.gz.
+    """
+
+    @functools.wraps(unconfigured_compiler_fn)
+    def debug_wrapper(gm, example_inputs, **kwargs):
+        compiler_fn = functools.partial(unconfigured_compiler_fn, **kwargs)
+        assert config.repro_after in ("dynamo", "aot", None)
+
+        if config.repro_after == "dynamo":
+
+            def add_paths(exc):
+                exc.minifier_path = os.path.join(minifier_dir(), "minifier_launcher.py")
+                if use_buck:
+                    exc.buck_command = " ".join(
+                        BUCK_CMD_PREFIX
+                        + [BuckTargetWriter(exc.minifier_path).cmd_line_path]
+                    )
+
+            if config.repro_level == 3:
+                dump_to_minify_after_dynamo(gm, example_inputs, compiler_name)
+
+            # Check for either accuracy (level 4) or other type of failures.
+            if config.repro_level == 4:
+                # Check Accuracy
+                compiled_gm = compiler_fn(copy.deepcopy(gm), example_inputs)
+                if backend_accuracy_fails(gm, example_inputs, compiler_fn):
+                    log.warning(
+                        "Accuracy failed for the TorchDynamo produced graph. Creating script to minify the error."
+                    )
+                    dump_to_minify_after_dynamo(
+                        fx.GraphModule(gm, copy.deepcopy(gm.graph)),
+                        example_inputs,
+                        compiler_name,
+                    )
+                    exc = AccuracyError("Bad accuracy detected.")
+                    add_paths(exc)
+                    raise exc
+            else:
+                try:
+                    compiled_gm = compiler_fn(copy.deepcopy(gm), example_inputs)
+                    run_fwd_maybe_bwd(compiled_gm, example_inputs)
+                except Exception as exc:
+                    log.warning(
+                        "Compiled Fx GraphModule failed. Creating script to minify the error."
+                    )
+                    if config.repro_level == 1:
+                        dump_state_fn = functools.partial(
+                            dump_backend_state, compiler_name=compiler_name
+                        )
+                        dump_state_fn(
+                            fx.GraphModule(gm, copy.deepcopy(gm.graph)), example_inputs
+                        )
+                    elif config.repro_level == 2:
+                        dump_to_minify_after_dynamo(
+                            fx.GraphModule(gm, copy.deepcopy(gm.graph)),
+                            example_inputs,
+                            compiler_name,
+                        )
+                    add_paths(exc)
+                    raise
+        else:
+            compiled_gm = compiler_fn(gm, example_inputs)
+
+        return compiled_gm
+
+    debug_wrapper._torchdynamo_orig_callable = unconfigured_compiler_fn  # type: ignore[attr-defined]
+    if hasattr(unconfigured_compiler_fn, "compiler_name"):
+        debug_wrapper.__name__ = unconfigured_compiler_fn.compiler_name
+    if hasattr(unconfigured_compiler_fn, "get_compiler_config"):
+        debug_wrapper.get_compiler_config = unconfigured_compiler_fn.get_compiler_config  # type: ignore[attr-defined]
+    return debug_wrapper
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           REPRO DUMPERS
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def generate_dynamo_fx_repro_string(
+    gm,
+    args,
+    compiler_name,
+    check_accuracy=False,
+    *,
+    stable_output=False,
+    save_dir=None,
+    command="run",
+):
+    """
+    Generate a repro string for backend-agnostic minified version.
+    """
+
+    model_str = NNModuleToString.convert(gm)
+
+    # TODO: Figure out why torch.compile'd hash isn't work on this codepath
+    writer = InputWriter(save_dir, stable_hash=True)
+    for placeholder, arg in zip(fx_placeholder_targets(gm), args):
+        if isinstance(arg, (int, torch.SymInt)):
+            writer.symint(placeholder, arg)
+        elif isinstance(arg, torch.Tensor):
+            # TODO: improve these names with FQN
+            writer.tensor(placeholder, arg)
+        else:
+            raise TypeError(f"arg is neither SymInt/int nor torch.Tensor, {arg}")
+    load_args = "\n".join(writer.lines())
+
+    return textwrap.dedent(
+        f"""
+from math import inf
+import torch
+from torch import tensor, device
+import torch.fx as fx
+import torch._dynamo
+from torch._dynamo.testing import rand_strided
+from torch._dynamo.debug_utils import run_fwd_maybe_bwd
+
+{generate_config_string(stable_output=stable_output)}
+
+{extra_imports}
+
+{model_str}
+mod = Repro()
+
+{load_args}
+
+if __name__ == '__main__':
+    from torch._dynamo.repro.after_dynamo import run_repro
+    run_repro(mod, load_args, accuracy={check_accuracy!r}, command={command!r},
+        save_dir={save_dir!r}, autocast={torch.is_autocast_enabled()!r}, backend={compiler_name!r})
+"""
+    )
+
+
+def dump_backend_repro_as_file(gm, args, compiler_name, check_accuracy=False):
+    """
+    Saves the repro to a repro.py file
+    """
+    curdir = os.getcwd()
+    subdir = os.path.join(os.getcwd(), "checkpoints")
+    if not os.path.exists(subdir):
+        os.makedirs(subdir, exist_ok=True)
+    file_name = os.path.join(subdir, f"minified_{len(gm.graph.nodes)}_nodes.py")
+    log.warning(
+        "Writing checkpoint with %s nodes to %s", len(gm.graph.nodes), file_name
+    )
+
+    with open(file_name, "w") as fd:
+        fd.write(
+            generate_dynamo_fx_repro_string(
+                gm, args, compiler_name, check_accuracy, save_dir=subdir
+            )
+        )
+    latest_repro = os.path.join(curdir, "repro.py")
+    log.warning("Copying %s to %s for convenience", file_name, latest_repro)
+
+    if use_buck:
+        BuckTargetWriter(latest_repro).write()
+
+    shutil.copyfile(file_name, latest_repro)
+
+
+def dump_backend_state(gm, args, compiler_name, check_accuracy=False):
+    """
+    Dumps the dynamo graph to repro the issue.
+    1) It tries to convert Fx GraphModule to a string. If we can, it writes to a
+    repro.py file.
+    2) If we can't convert Fx GraphModule to a string, we use to_folder to save
+    the module and save a tar file.
+    """
+    assert NNModuleToString.can_convert_to_string(gm)
+    return dump_backend_repro_as_file(gm, args, compiler_name, check_accuracy)
+    # return dump_backend_repro_as_tarfile(gm, args, compiler_name)
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                       MINIFIER DUMPER
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def dump_to_minify_after_dynamo(gm, args, compiler_name):
+    # TODO: factor this out
+    subdir = os.path.join(minifier_dir(), "checkpoints")
+    if not os.path.exists(subdir):
+        os.makedirs(subdir, exist_ok=True)
+    helper_for_dump_minify(
+        generate_dynamo_fx_repro_string(
+            gm,
+            args,
+            compiler_name,
+            check_accuracy=config.repro_level == 4,
+            save_dir=subdir,
+            command="minify",
+        )
+    )
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                       MINIFIER BACKENDS
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+@register_debug_backend
+def dynamo_minifier_backend(gm, example_inputs, compiler_name):
+    from functorch.compile import minifier
+
+    compiler_fn = lookup_backend(compiler_name)
+
+    # TODO: It's inconsistent to pass SymInt inputs but REAL tensors.
+    # We should pass ints and look at the GraphModule placeholders
+    # to resolve them to SymInt (if necessary)
+    example_inputs = [
+        i.node.hint if isinstance(i, torch.SymInt) else i for i in example_inputs
+    ]
+
+    try:
+        compiled_gm = compiler_fn(gm, example_inputs)
+        run_fwd_maybe_bwd(compiled_gm, example_inputs)
+        raise ValueError("No issue was detected")
+    except Exception as exc:
+        orig_failure = str(exc)
+        log.warning(
+            "Compiled Fx GraphModule failed. Creating script to minify the error."
+        )
+        dump_state_fn = functools.partial(
+            dump_backend_state, compiler_name=compiler_name
+        )
+        dump_state_fn(fx.GraphModule(gm, copy.deepcopy(gm.graph)), example_inputs)
+        fails_fn = functools.partial(
+            backend_fails,
+            compiler_fn=compiler_fn,
+            orig_failure=orig_failure,
+        )
+        minifier(
+            gm,
+            example_inputs,
+            module_fails=fails_fn,
+            dump_state=dump_state_fn,
+        )
+    return gm
+
+
+@register_debug_backend
+def dynamo_accuracy_minifier_backend(gm, example_inputs, compiler_name):
+    from functorch.compile import minifier
+
+    compiler_fn = lookup_backend(compiler_name)
+
+    # Set the eval mode to remove randomness.
+    gm.eval()
+
+    # Check Accuracy
+    if backend_accuracy_fails(
+        gm, example_inputs, compiler_fn, only_fwd=config.repro_forward_only
+    ):
+        log.warning("Accuracy failed for the TorchDynamo produced graph")
+        dump_state_fn = functools.partial(
+            dump_backend_state, compiler_name=compiler_name, check_accuracy=True
+        )
+        fails_fn = functools.partial(
+            backend_accuracy_fails,
+            compiler_fn=compiler_fn,
+            only_fwd=config.repro_forward_only,
+        )
+        dump_state_fn(fx.GraphModule(gm, copy.deepcopy(gm.graph)), example_inputs)
+        minifier(
+            gm,
+            example_inputs,
+            module_fails=fails_fn,
+            dump_state=dump_state_fn,
+        )
+    else:
+        log.error("Input graph does not fail accuracy testing")
+    return gm
+
+
+def backend_fails(gm, example_inputs, compiler_fn, orig_failure):
+    """
+    Minifier uses this function to identify if the minified graph module fails
+    with the same error.
+
+    One caveat is that minifier can potentially go into a wrong direction when
+    the resulting graph module fails for a different reason. To avoid this, we
+    save the string for the original exception and check similarity between new
+    and old exception. They can be somewhat different in some cases, when the
+    exception string depends on the failing node information. So, we have a
+    loose similarity metric to guide the minifier path.
+    """
+    from difflib import SequenceMatcher
+
+    try:
+        # Run the original gm to check eager validity
+        run_fwd_maybe_bwd(gm, clone_inputs_retaining_gradness(example_inputs))
+        compiled_gm = compiler_fn(gm, example_inputs)
+        run_fwd_maybe_bwd(compiled_gm, clone_inputs_retaining_gradness(example_inputs))
+        return False
+    except Exception as e:
+        new_failure = str(e)
+        if SequenceMatcher(None, orig_failure, new_failure).ratio() > 0.5:
+            return True
+        return False
+
+
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+#                           REPRO MAIN
+# ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ #
+
+
+def run_load_args(options, mod, load_args):
+    if not hasattr(load_args, "_version"):
+        log.warning(
+            "load_args does not have a _version attribute, please file a bug to PyTorch "
+            "and describe how you generate this repro script"
+        )
+    else:
+        if load_args._version > 0:
+            log.warning(
+                "load_args is version %s, but this version of PyTorch only supports "
+                "version 0.  We will try to run it anyway but there may be an incompatibility; "
+                "if so, try upgrading your version of PyTorch.",
+                load_args._version,
+            )
+
+    nop_reader = NopInputReader()
+    load_args(nop_reader)
+
+    with tqdm(desc="Loading inputs", total=nop_reader.total) as pbar:
+        input_reader = InputReader(save_dir=options.save_dir, pbar=pbar)
+        load_args(input_reader)
+        args = input_reader.args
+
+    return args
+
+
+def repro_minify(options, mod, load_args):
+    args = run_load_args(options, mod, load_args)
+
+    # Setup debug minifier compiler
+    if not options.accuracy:
+        compiler_fn = lookup_backend("dynamo_minifier_backend")
+    else:
+        compiler_fn = lookup_backend("dynamo_accuracy_minifier_backend")
+
+    if options.backend is None:
+        raise RuntimeError(
+            "Compiler name is None - this likely means that a custom compiler "
+            "was called by torchdynamo. Please remove this error, import your "
+            "custom compiler function, and replace the backend=None "
+            "line in run_repro to backend=<my_imported_custom_function>"
+        )
+
+    dynamo_minifier_backend = functools.partial(
+        compiler_fn,
+        compiler_name=options.backend,
+    )
+    opt_mod = torch._dynamo.optimize(dynamo_minifier_backend)(mod)
+
+    with torch.cuda.amp.autocast(enabled=options.autocast):
+        opt_mod(*args)
+
+
+def repro_run(options, mod, load_args):
+    opt_mod = torch._dynamo.optimize(options.backend)(mod)
+
+    if options.accuracy != "":
+        mod.eval()
+        opt_mod.eval()
+
+        with torch.cuda.amp.autocast(enabled=options.autocast):
+            # TODO: disable clone
+            args = run_load_args(options, mod, load_args)
+            assert same_two_models(mod, mod, args), "Eager itself failed"
+            if not same_two_models(mod, opt_mod, args):
+                raise AccuracyError("Dynamo failed")
+    else:
+        with torch.cuda.amp.autocast(enabled=options.autocast):
+            args = run_load_args(options, mod, load_args)
+            ref = run_fwd_maybe_bwd(
+                mod, args, only_fwd=options.only_fwd, disable_clone=True
+            )
+            del args
+
+            args = run_load_args(options, mod, load_args)
+            res = run_fwd_maybe_bwd(
+                opt_mod, args, only_fwd=options.only_fwd, disable_clone=True
+            )
+
+
+def run_repro(
+    mod,
+    load_args,
+    *,
+    command="run",
+    accuracy: Union[bool, str] = "",
+    save_dir=None,
+    autocast=False,
+    backend="inductor",
+    **kwargs,
+):
+    for k in kwargs:
+        log.warning(
+            "Unrecognized kwarg %s; perhaps this repro was made on a newer version of PyTorch",
+            k,
+        )
+
+    if accuracy is True:
+        accuracy = "accuracy"
+    elif accuracy is False:
+        accuracy = ""
+
+    parser = argparse.ArgumentParser(
+        description=f"""\
+An after_dynamo repro script, typically triggering a bug in Dynamo or
+AOTAutograd.  When run with no arguments, this script defaults to running
+'{command}'.  Extra flags may be available; to find out more, try '{command}
+--help'.  There are also alternate subcommands available, see below.
+
+default settings on this script:
+  {accuracy=}
+  {save_dir=}
+""",
+        formatter_class=argparse.RawTextHelpFormatter,
+    )
+
+    def common_flags(parser):
+        accuracy_group = parser.add_mutually_exclusive_group()
+        accuracy_group.add_argument(
+            "--no-accuracy",
+            dest="accuracy",
+            action="store_const",
+            const="",
+            default=accuracy,
+            help="do not test accuracy, just run the module and see if it errors",
+        )
+        accuracy_group.add_argument(
+            "--accuracy",
+            action="store_const",
+            const="accuracy",
+            default=accuracy,
+            help="test accuracy",
+        )
+        parser.add_argument(
+            "--save-dir",
+            type=str,
+            default=save_dir,
+            metavar="DIR",
+            help="directory where saved inputs live",
+        )
+        parser.add_argument(
+            "--no-save-dir",
+            dest="save_dir",
+            action="store_const",
+            const=None,
+            help="don't use any directory for saved inputs",
+        )
+        parser.add_argument(
+            "--no-isolate",
+            dest="isolate",
+            action="store_false",
+            default=False,
+            help="no isolate (doesn't do anything for after_dynamo)",
+        )
+        parser.add_argument(
+            "--autocast",
+            default=autocast,
+            action="store_true",
+            help="use torch.cuda.amp.autocast",
+        )
+        parser.add_argument(
+            "--no-autocast",
+            dest="autocast",
+            action="store_false",
+            help="don't use torch.cuda.amp.autocast",
+        )
+        parser.add_argument(
+            "--backend",
+            type=str,
+            default=backend,
+            metavar="BACKEND",
+            help="torch.compile backend to use",
+        )
+
+    subparsers = parser.add_subparsers(
+        dest="command", metavar="{run,minify}", required=True
+    )
+
+    parser_run = subparsers.add_parser(
+        "run",
+        help="just run the repro",
+    )
+    common_flags(parser_run)
+    parser_run.add_argument(
+        "--only-fwd",
+        action="store_true",
+        help="don't run backwards compilation for testing",
+    )
+
+    parser_minify = subparsers.add_parser(
+        "minify", help="run the minifier on the repro"
+    )
+    common_flags(parser_minify)
+
+    args = None
+    if len(sys.argv) <= 1:
+        args = [command, *sys.argv[1:]]
+
+    options = parser.parse_args(args)
+    COMMAND_FNS = {
+        "minify": repro_minify,
+        "run": repro_run,
+    }
+    COMMAND_FNS[options.command](options, mod, load_args)

venv/lib/python3.10/site-packages/torch/_dynamo/variables/__init__.py

@@ -0,0 +1,151 @@
+# mypy: ignore-errors
+
+from .base import VariableTracker
+from .builtin import BuiltinVariable
+from .constant import ConstantVariable, EnumVariable
+from .ctx_manager import (
+    ContextWrappingVariable,
+    DeterministicAlgorithmsVariable,
+    DisabledSavedTensorsHooksVariable,
+    GradIncrementNestingCtxManagerVariable,
+    GradInplaceRequiresGradCtxManagerVariable,
+    GradModeVariable,
+    InferenceModeVariable,
+    StreamContextVariable,
+    StreamVariable,
+    VmapIncrementNestingCtxManagerVariable,
+    WithExitFunctionVariable,
+)
+from .dicts import (
+    ConstDictVariable,
+    CustomizedDictVariable,
+    DataClassVariable,
+    DefaultDictVariable,
+    SetVariable,
+)
+from .distributed import BackwardHookVariable
+from .functions import (
+    FunctoolsPartialVariable,
+    NestedUserFunctionVariable,
+    SkipFunctionVariable,
+    UserFunctionVariable,
+    UserMethodVariable,
+)
+from .higher_order_ops import (
+    FunctorchHigherOrderVariable,
+    TorchHigherOrderOperatorVariable,
+)
+from .iter import (
+    CountIteratorVariable,
+    CycleIteratorVariable,
+    IteratorVariable,
+    ItertoolsVariable,
+    RepeatIteratorVariable,
+)
+from .lazy import LazyVariableTracker
+from .lists import (
+    BaseListVariable,
+    ListIteratorVariable,
+    ListVariable,
+    NamedTupleVariable,
+    RangeVariable,
+    RestrictedListSubclassVariable,
+    SliceVariable,
+    TupleIteratorVariable,
+    TupleVariable,
+)
+from .misc import (
+    AutogradFunctionContextVariable,
+    AutogradFunctionVariable,
+    ClosureVariable,
+    DeletedVariable,
+    GetAttrVariable,
+    InspectSignatureVariable,
+    LambdaVariable,
+    MethodWrapperVariable,
+    NewCellVariable,
+    NewGlobalVariable,
+    NumpyVariable,
+    PythonModuleVariable,
+    StringFormatVariable,
+    SuperVariable,
+    TypingVariable,
+    UnknownVariable,
+)
+from .nn_module import NNModuleVariable, UnspecializedNNModuleVariable
+from .sdpa import SDPAParamsVariable
+from .tensor import (
+    FakeItemVariable,
+    NumpyNdarrayVariable,
+    SymNodeVariable,
+    TensorVariable,
+    UnspecializedPythonVariable,
+    UntypedStorageVariable,
+)
+from .torch import TorchCtxManagerClassVariable, TorchInGraphFunctionVariable
+from .user_defined import (
+    RemovableHandleVariable,
+    UserDefinedClassVariable,
+    UserDefinedObjectVariable,
+)
+
+__all__ = [
+    "AutogradFunctionContextVariable",
+    "AutogradFunctionVariable",
+    "BackwardHookVariable",
+    "BaseListVariable",
+    "BuiltinVariable",
+    "ClosureVariable",
+    "ConstantVariable",
+    "ConstDictVariable",
+    "ContextWrappingVariable",
+    "CountIteratorVariable",
+    "CustomizedDictVariable",
+    "CycleIteratorVariable",
+    "DataClassVariable",
+    "DefaultDictVariable",
+    "DeletedVariable",
+    "DeterministicAlgorithmsVariable",
+    "EnumVariable",
+    "FakeItemVariable",
+    "GetAttrVariable",
+    "GradModeVariable",
+    "InspectSignatureVariable",
+    "IteratorVariable",
+    "ItertoolsVariable",
+    "LambdaVariable",
+    "LazyVariableTracker",
+    "ListIteratorVariable",
+    "ListVariable",
+    "NamedTupleVariable",
+    "NestedUserFunctionVariable",
+    "NewCellVariable",
+    "NewGlobalVariable",
+    "NNModuleVariable",
+    "NumpyNdarrayVariable",
+    "NumpyVariable",
+    "PythonModuleVariable",
+    "RangeVariable",
+    "RemovableHandleVariable",
+    "RepeatIteratorVariable",
+    "RestrictedListSubclassVariable",
+    "SDPAParamsVariable",
+    "SkipFunctionVariable",
+    "SliceVariable",
+    "StringFormatVariable",
+    "SuperVariable",
+    "TensorVariable",
+    "TorchCtxManagerClassVariable",
+    "TorchInGraphFunctionVariable",
+    "TupleVariable",
+    "UnknownVariable",
+    "UnspecializedNNModuleVariable",
+    "UnspecializedPythonVariable",
+    "UntypedStorageVariable",
+    "UserDefinedClassVariable",
+    "UserDefinedObjectVariable",
+    "UserFunctionVariable",
+    "UserMethodVariable",
+    "VariableTracker",
+    "WithExitFunctionVariable",
+]