onnx-diagnostic 0.8.3__py3-none-any.whl → 0.8.4__py3-none-any.whl

onnx_diagnostic/__init__.py

@@ -3,5 +3,5 @@ Patches, Investigates onnx models.
 Functions, classes to dig into a model when this one is right, slow, wrong...
 """
 
-__version__ = "0.8.3"
+__version__ = "0.8.4"
 __author__ = "Xavier Dupré"

onnx_diagnostic/_command_lines_parser.py

@@ -1286,7 +1286,13 @@ def get_parser_sbs() -> ArgumentParser:
         "--first",
         action=BooleanOptionalAction,
         default=False,
-        help="First runs the whole model.",
+        help="First runs the whole model (default is False).",
+    )
+    parser.add_argument(
+        "--sbs",
+        action=BooleanOptionalAction,
+        default=True,
+        help="Runs the side-by-side (default is True).",
     )
     parser.add_argument(
         "-2",
@@ -1342,6 +1348,20 @@ def get_parser_sbs() -> ArgumentParser:
         default="replay",
         help="If the replay is triggered, this defines the folder where everything is dumped.",
     )
+    parser.add_argument(
+        "-p",
+        "--replay-prefix-model",
+        action=BooleanOptionalAction,
+        default=False,
+        help=textwrap.dedent(
+            """
+            There are two ways to recompute an intermediate output, the first one is to "
+            produce the minimal model between torch and onnx.
+            The second one is to dump onnx models from the inputs
+            to the considered intermediate results. This enables the second one.
+            """
+        ),
+    )
 
     return parser
 
@@ -1417,6 +1437,10 @@ def _cmd_sbs(argv: List[Any]):
         print("-- done")
         del sess
 
+    if not args.sbs:
+        print("-- done")
+        return
+
     print(f"-- load onnx {args.onnx!r}")
     begin = time.perf_counter()
     onx = onnx.load(args.onnx)
@@ -1431,6 +1455,7 @@ def _cmd_sbs(argv: List[Any]):
                 set(args.replay_op_types.split(",")) if args.replay_op_types else None
             ),
             dump_folder=args.replay_folder,
+            dump_prefix_model=args.replay_prefix_model,
         )
 
     print("-- starts side-by-side")

onnx_diagnostic/export/api.py

@@ -3,6 +3,52 @@ import torch
 from .onnx_plug import EagerDirectReplacementWithOnnx
 
 
+def get_main_dispatcher(
+    use_control_flow_dispatcher: bool = False,
+    onnx_plugs: Optional[List[EagerDirectReplacementWithOnnx]] = None,
+) -> Any:  # Dispatcher
+    """Creates a custom dispatcher for the custom exporter."""
+    from experimental_experiment.torch_interpreter import Dispatcher
+
+    if use_control_flow_dispatcher:
+        from .control_flow_onnx import create_global_dispatcher
+
+        control_flow_dispatcher = create_global_dispatcher()
+    else:
+        control_flow_dispatcher = None
+
+    class MainDispatcher(Dispatcher):
+        def __init__(self, previous_dispatcher=None):
+            super().__init__({})
+            self.previous_dispatcher = previous_dispatcher
+
+        @property
+        def supported(self):
+            if self.previous_dispatcher:
+                return set(self.registered_functions) | self.previous_dispatcher.supported
+            return set(self.registered_functions)
+
+        def find_function(self, name: Any):
+            if self.previous_dispatcher:
+                find = self.previous_dispatcher.find_function(name)
+                if find:
+                    return find
+            return Dispatcher.find_function(self, name)
+
+        def find_method(self, name: Any):
+            if self.previous_dispatcher:
+                find = self.previous_dispatcher.find_method(name)
+                if find:
+                    return find
+            return Dispatcher.find_method(self, name)
+
+    main_dispatcher = MainDispatcher(control_flow_dispatcher)
+    if onnx_plugs:
+        for plug in onnx_plugs:
+            main_dispatcher.registered_functions[plug.target_name] = plug.custom_converter()
+    return main_dispatcher
+
+
 def to_onnx(
     mod: Union["torch.nn.Module", "torch.fx.GraphModule"],  # noqa: F821
     args: Optional[Sequence["torch.Tensor"]] = None,  # noqa: F821
@@ -82,51 +128,11 @@ def to_onnx(
             options = exporter_kwargs.pop("options", None)
         if options is None:
             options = OptimizationOptions(patterns="default+onnxruntime")
-        if onnx_plugs or use_control_flow_dispatcher:
-            from experimental_experiment.torch_interpreter import Dispatcher
-
-            if use_control_flow_dispatcher:
-                from .control_flow_onnx import create_global_dispatcher
-
-                control_flow_dispatcher = create_global_dispatcher()
-            else:
-                control_flow_dispatcher = None
-
-            class MainDispatcher(Dispatcher):
-                def __init__(self, previous_dispatcher=None):
-                    super().__init__({})
-                    self.previous_dispatcher = previous_dispatcher
-
-                @property
-                def supported(self):
-                    if self.previous_dispatcher:
-                        return (
-                            set(self.registered_functions) | self.previous_dispatcher.supported
-                        )
-                    return set(self.registered_functions)
-
-                def find_function(self, name: Any):
-                    if self.previous_dispatcher:
-                        find = self.previous_dispatcher.find_function(name)
-                        if find:
-                            return find
-                    return Dispatcher.find_function(self, name)
-
-                def find_method(self, name: Any):
-                    if self.previous_dispatcher:
-                        find = self.previous_dispatcher.find_method(name)
-                        if find:
-                            return find
-                    return Dispatcher.find_method(self, name)
-
-            main_dispatcher = MainDispatcher(control_flow_dispatcher)
-            if onnx_plugs:
-                for plug in onnx_plugs:
-                    main_dispatcher.registered_functions[plug.target_name] = (
-                        plug.custom_converter()
-                    )
-        else:
-            main_dispatcher = None
+        main_dispatcher = (
+            get_main_dispatcher(use_control_flow_dispatcher, onnx_plugs)
+            if onnx_plugs or use_control_flow_dispatcher
+            else None
+        )
 
         return _to_onnx(
             mod,
@@ -149,6 +155,7 @@ def to_onnx(
 
     if exporter in ("dynamo", "onnx-dynamo"):
         import os
+        from ..helpers import flatten_object
         import onnxscript.rewriter.ort_fusions as ort_fusions
 
         assert (
@@ -180,7 +187,20 @@ def to_onnx(
             import onnx_ir as ir
             import onnx_ir.passes.common as common_passes
 
-            irfunctions = [ir.from_proto(plug.function_proto) for plug in onnx_plugs]
+            opset = (
+                18
+                if target_opset is None
+                else (target_opset if isinstance(target_opset, int) else target_opset[""])
+            )
+
+            irfunctions = [
+                ir.from_proto(
+                    plug.get_function_proto(
+                        opset, *flatten_object((args, kwargs), drop_keys=True)
+                    )
+                )
+                for plug in onnx_plugs
+            ]
             for func in irfunctions:
                 epo.model.functions[func.identifier()] = func
             if inline:

onnx_diagnostic/export/control_flow_research.py

@@ -92,10 +92,11 @@ def simple_loop_for(
 
     from torch._higher_order_ops.utils import setup_compilation_env
 
-    with setup_compilation_env() as backend:
-        return torch.compile(_loop_for_op_wrapper, backend=backend, fullgraph=True)(
-            n_iter, body_fn, operands
-        )
+    with setup_compilation_env() as _backend:
+        return _loop_for_op_wrapper(n_iter, body_fn, *operands)
+        # return torch.compile(_loop_for_op_wrapper, backend=backend, fullgraph=True)(
+        #    n_iter, body_fn, operands
+        # )
 
 
 def trace_loop_for(proxy_mode, func_overload, n_iter, body_fn, operands):
@@ -127,9 +128,13 @@ def loop_for_op_dense(n_iter, body_fn, operands):
     ), f"Dense implementation operands must be a list of tensors and ints {operands}"
     mode = _get_current_dispatch_mode()
     assert mode is None, "Mode should never be enabled for CPU/CUDA key"
-    return _loop_for_onnx_fn(body_fn, n_iter, None, *operands)
+    return _loop_for_onnx_fn(body_fn, n_iter, None, operands)
 
 
 @simple_loop_for_op.py_impl(ProxyTorchDispatchMode)
 def inner(mode, n_iter, body_fn, operands):
     return trace_loop_for(mode, simple_loop_for_op, n_iter, body_fn, operands)
+
+
+simple_loop_for_op.fallthrough(torch._C.DispatchKey.AutogradCPU)
+simple_loop_for_op.fallthrough(torch._C.DispatchKey.AutogradCUDA)

onnx_diagnostic/export/onnx_plug.py

@@ -3,8 +3,12 @@ from dataclasses import dataclass
 from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 import onnx
 import torch
-from ..helpers import max_diff
-from ..helpers.torch_helper import torch_dtype_to_onnx_dtype
+from ..helpers import max_diff, string_type
+from ..helpers.torch_helper import (
+    torch_dtype_to_onnx_dtype,
+    onnx_dtype_to_torch_dtype,
+    int_device_to_torch_device,
+)
 from ..reference import OnnxruntimeEvaluator
 
 TUPLE_TENSORS = Tuple[torch.Tensor, ...]
@@ -50,6 +54,10 @@ class EagerDirectReplacementWithOnnx:
         only tensors must be counted
     :param name: the name of the custom op, the function name if not specified
     :param kwargs: constants parameters with their default values
+    :param version_selector: selects the version based on the arguments,
+        see below for an example, this allows the user to define different
+        onnx version depending on the inputs
+    :param default_opset: opset to use by default
     :param verbose: verbose level
 
     Here is an example:
@@ -133,27 +141,87 @@ class EagerDirectReplacementWithOnnx:
         ).model_proto
 
         print(pretty_onnx(onx))
+
+    This shows how to define multiple versions depending on the device,
+    the type or the targeted onnx opset.
+
+    .. code-block:: python
+
+        def qwen_version_selector(opset: int, *args: torch.Tensor) -> Tuple[str, torch.dtype]:
+            first_tensor = next(a for a in args if a is not None)
+            dtype = first_tensor.dtype
+            itype = torch_dtype_to_onnx_dtype(dtype)
+            if dtype == torch.float32:
+                if opset >= 24:
+                    return "LOOPA24", itype
+                return "LOOPMHA", itype
+            if dtype == torch.float16:
+                if first_tensor.is_cuda:
+                    return "PACKED", itype
+                return "LOOPMHA", itype
+            raise AssertionError(
+                f"Unable to handle type {torch.dtype} (itype={itype}) "
+                f"on device {torch.device} with opset={opset}"
+            )
+
+        qwen_sdpa_attention_versatile = EagerDirectReplacementWithOnnx(
+            qwen_sdpa_attention,
+            lambda qs, *args, **kwargs: torch.empty(
+                (qs.shape[0], qs.shape[2], qs.shape[1], qs.shape[3]),
+                dtype=qs.dtype,
+                device=qs.device,
+            ),
+            {
+                ("PACKED", onnx.TensorProto.FLOAT16): _add_com_microsoft_opset(
+                    PackedAttention.to_function_proto()
+                ),
+                ("LOOPA24", onnx.TensorProto.FLOAT): LoopAttention24.to_function_proto(),
+                ("LOOPA24", onnx.TensorProto.FLOAT16): _update_sequence_type(
+                    onnx.TensorProto.FLOAT16, LoopAttention24.to_function_proto()
+                ),
+                ("LOOPMHA", onnx.TensorProto.FLOAT): _add_com_microsoft_opset(
+                    LoopMHAAttention.to_function_proto()
+                ),
+                ("LOOPMHA", onnx.TensorProto.FLOAT16): _update_sequence_type(
+                    onnx.TensorProto.FLOAT16,
+                    _add_com_microsoft_opset(LoopMHAAttention.to_function_proto()),
+                ),
+            },
+            n_inputs=4,
+            n_outputs=1,
+            kwargs=dict(scaling=0.11180339887498948, num_heads=16),
+            name="qwen_sdpa_attention_versatile",
+            version_selector=qwen_version_selector,
+        )
     """
 
     def __init__(
         self,
         eager_fn: Callable[[TUPLE_TENSORS], TUPLE_TENSORS],
         shape_fn: Callable[[TUPLE_TENSORS], TUPLE_TENSORS],
-        function_proto: onnx.FunctionProto,
+        function_proto: Union[onnx.FunctionProto, Dict[Any, onnx.FunctionProto]],
         n_inputs: Optional[int] = None,
         n_outputs: Optional[int] = None,
         name: Optional[str] = None,
         kwargs: Optional[Dict[str, Union[int, float]]] = None,
         verbose: int = 0,
+        version_selector: Optional[Callable[..., Tuple[Any, ...]]] = None,
+        default_opset: int = 22,
     ):
-        assert isinstance(
-            function_proto, onnx.FunctionProto
+        assert isinstance(function_proto, onnx.FunctionProto) or (
+            isinstance(function_proto, dict)
+            or all(isinstance(v, onnx.FunctionProto) for v in function_proto.values())
         ), f"Unexpected type {type(function_proto)} for function_proto"
         assert isinstance(n_inputs, int), f"not implemented yet when n_inputs={n_inputs}"
-        assert isinstance(n_outputs, int), f"not implemented yet when n_inputs={n_outputs}"
+        assert isinstance(n_outputs, int), f"not implemented yet when n_outputs={n_outputs}"
         self.eager_fn = eager_fn
         self.shape_fn = shape_fn
-        self.function_proto = function_proto
+        self._function_proto = (
+            function_proto if isinstance(function_proto, onnx.FunctionProto) else None
+        )
+        self._function_proto_versioned = (
+            function_proto if isinstance(function_proto, dict) else {}
+        )
         self.n_inputs = n_inputs
         self.n_outputs = n_outputs
         self.name = name or (
@@ -170,24 +238,73 @@ class EagerDirectReplacementWithOnnx:
         )
         sig = inspect.signature(self.eager_fn)
         params = list(sig.parameters)
-        assert (
-            len(params) >= n_inputs
-        ), f"{self.eager_fn} accepts {params} as parameters < n_inputs={n_inputs}"
-        assert n_inputs == len(function_proto.input), (
-            f"Input mismatch n_inputs={n_inputs} but "
-            f"function_proto.input={function_proto.input}"
-        )
-        assert n_outputs == len(function_proto.output), (
-            f"Output mismatch n_outputs={n_outputs} but "
-            f"function_proto.output={function_proto.output}"
-        )
-        assert (
-            function_proto.domain == self.domain
-        ), f"Function domain must be {self.domain!r} but it is {function_proto.domain!r}"
         self.args_name = [p for p in params if p not in self.kwargs]
         self.kwargs_name = [p for p in params if p in self.kwargs]
         self.verbose = verbose
         self.custom_op = self._register()
+        self.version_selector = version_selector
+        self.default_opset = default_opset
+        self._check_protos(params)
+
+    def _check_protos(self, params):
+        assert (
+            len(params) >= self.n_inputs
+        ), f"{self.eager_fn} accepts {params} as parameters < n_inputs={self.n_inputs}"
+
+        # one proto
+        assert self._function_proto is None or self.n_inputs == len(
+            self._function_proto.input
+        ), (
+            f"Input mismatch n_inputs={self.n_inputs} but "
+            f"function_proto.input={self._function_proto.input}"
+        )
+        assert self._function_proto is None or self.n_outputs == len(
+            self._function_proto.output
+        ), (
+            f"Output mismatch n_outputs={self.n_outputs} but "
+            f"function_proto.output={self._function_proto.output}"
+        )
+        assert self._function_proto is None or (
+            self._function_proto.domain == self.domain
+        ), f"Function domain must be {self.domain!r} but it is {self._function_proto.domain!r}"
+
+        # multiple protos
+        assert all(
+            self.n_inputs == len(v.input) for v in self._function_proto_versioned.values()
+        ), f"Output mismatch n_inputs={self.n_inputs} but one version is wrong"
+        assert all(
+            self.n_outputs == len(v.output) for v in self._function_proto_versioned.values()
+        ), f"Output mismatch n_outputs={self.n_outputs} but one version is wrong"
+        assert all(
+            v.domain == self.domain for v in self._function_proto_versioned.values()
+        ), f"Function domain must be {self.domain!r} but it is different in one version"
+        assert (
+            not self._function_proto_versioned or self.version_selector
+        ), "version_selector is needed when multiple protos are given."
+
+    def get_function_proto(self, opset: int, *args) -> onnx.FunctionProto:
+        """Returns the correct version based on the inputs."""
+        if self._function_proto:
+            return self._function_proto
+        assert isinstance(
+            opset, int
+        ), f"The first argument must be an integer for the onnx opset but it is {type(opset)}"
+        assert any(
+            a is not None for a in args
+        ), f"Unexpected args={string_type(args, with_shape=True)}"
+        try:
+            key = self.version_selector(opset, *args)  # type: ignore[misc]
+        except (ValueError, AttributeError) as e:
+            raise AssertionError(
+                f"Unable to select a version, fails to get a key, available="
+                f"{set(self._function_proto_versioned)}, "
+                f"args={string_type(args,with_shape=True)}"
+            ) from e
+        assert key in self._function_proto_versioned, (
+            f"Unable to select a version, key={key}, available="
+            f"{set(self._function_proto_versioned)}, args={string_type(args,with_shape=True)}"
+        )
+        return self._function_proto_versioned[key]
 
     @property
     def domain(self) -> str:
@@ -219,6 +336,8 @@ class EagerDirectReplacementWithOnnx:
                 input_args.append(f"int {p}={val}")
             elif isinstance(val, float):
                 input_args.append(f"float {p}={val}")
+            elif isinstance(val, str):
+                input_args.append(f"str {p}={val}")
             else:
                 raise NotImplementedError(
                     f"kwargs {p!r} has a default value of unsupported type {type(val)}"
@@ -243,6 +362,7 @@ class EagerDirectReplacementWithOnnx:
         *args,
         engine: Optional[Callable] = None,
         dump_onnx_model: Optional[str] = None,
+        opset: int = 22,
         **kwargs,
     ) -> VerifyResult:
         """
@@ -257,6 +377,7 @@ class EagerDirectReplacementWithOnnx:
             :class:`onnx_diagnostic.reference.OnnxruntimeEvaluator`.
         :param dump_onnx_model: to dump the onnx model used to verify
             eager and onnx produce the same results
+        :param opset: onnx opset to use
         :param kwargs: additional arguments to the function
         :return: outputs of :func:`onnx_diagnostic.helpers.max_diff`
         """
@@ -291,7 +412,7 @@ class EagerDirectReplacementWithOnnx:
         assert engine is None, f"Not implemented yet with engine={engine!r}"
         ags, kws = self._make_args_kwargs(*args, **kwargs)
         sess = OnnxruntimeEvaluator(
-            self.function_proto,
+            self.get_function_proto(opset, *args),
             whole=True,
             dump_onnx_model=dump_onnx_model,
             function_kwargs=kws,
@@ -324,16 +445,25 @@ class EagerDirectReplacementWithOnnx:
             *args,
             **kwargs,
         ) -> Any:
-            if not g.has_local_function(
-                self.function_proto.name, domain=self.function_proto.domain
-            ):
-                g.add_function(self.function_proto)
+            has_devices = [a for a in args if isinstance(a, str) and g.has_device(a)]
+            assert (
+                has_devices
+            ), f"Missing device for any of the inputs {args}{g.get_debug_msg()}"
+            arg_device = has_devices[0]
+            fake_tensor = torch.empty(
+                tuple([(_ if isinstance(_, int) else 2) for _ in g.get_shape(args[0])]),
+                dtype=onnx_dtype_to_torch_dtype(g.get_type(args[0])),
+                device=int_device_to_torch_device(g.get_device(arg_device)),
+            )
+            function_proto = self.get_function_proto(g.main_opset, fake_tensor)
+            if not g.has_local_function(function_proto.name, domain=function_proto.domain):
+                g.add_function(function_proto)
             ags, kws = self._make_args_kwargs(*args, **kwargs)
             res = g.make_node(
-                self.function_proto.name,
+                function_proto.name,
                 ags,
                 outputs,
-                domain=self.function_proto.domain,
+                domain=function_proto.domain,
                 name=self.target_name,
                 **kws,
             )
@@ -356,41 +486,46 @@ class EagerDirectReplacementWithOnnx:
         """
         import onnxscript
 
-        onnx_plug_op = onnxscript.values.Opset(domain=self.function_proto.domain, version=1)
-        schema = onnx_plug_op[self.function_proto.name]
-        if schema is None:
-            all_types = [
-                "tensor(float)",
-                "tensor(float16)",
-                "tensor(bfloat16)",
-                "tensor(double)",
-                "tensor(int64)",
-                "tensor(int32)",
-            ]
-            type_constraints = []
-            for i in range(self.n_inputs):
-                type_constraints.append((f"T{i}", all_types, ""))
-            for i in range(self.n_outputs):
-                type_constraints.append((f"U{i}", all_types, ""))
-            schema = onnx.defs.OpSchema(
-                self.function_proto.name,
-                self.function_proto.domain,
-                1,
-                inputs=[
-                    onnx.defs.OpSchema.FormalParameter(f"arg_{i}", f"T{i}")
-                    for i in range(self.n_inputs)
-                ],
-                outputs=[
-                    onnx.defs.OpSchema.FormalParameter(f"res_{i}", f"U{i}")
-                    for i in range(self.n_outputs)
-                ],
-                type_constraints=type_constraints,
-            )
-            onnx.defs.register_schema(schema)
-        op = onnxscript.values.Op(onnx_plug_op, self.function_proto.name, schema)
+        onnx_plug_op = onnxscript.values.Opset(domain=self.domain, version=1)
+
+        def get_proto(*args):
+            function_proto = self.get_function_proto(self.default_opset, *args)
+            schema = onnx_plug_op[function_proto.name]
+            if schema is None:
+                all_types = [
+                    "tensor(float)",
+                    "tensor(float16)",
+                    "tensor(bfloat16)",
+                    "tensor(double)",
+                    "tensor(int64)",
+                    "tensor(int32)",
+                ]
+                type_constraints = []
+                for i in range(self.n_inputs):
+                    type_constraints.append((f"T{i}", all_types, ""))
+                for i in range(self.n_outputs):
+                    type_constraints.append((f"U{i}", all_types, ""))
+                schema = onnx.defs.OpSchema(
+                    function_proto.name,
+                    function_proto.domain,
+                    1,
+                    inputs=[
+                        onnx.defs.OpSchema.FormalParameter(f"arg_{i}", f"T{i}")
+                        for i in range(self.n_inputs)
+                    ],
+                    outputs=[
+                        onnx.defs.OpSchema.FormalParameter(f"res_{i}", f"U{i}")
+                        for i in range(self.n_outputs)
+                    ],
+                    type_constraints=type_constraints,
+                )
+                onnx.defs.register_schema(schema)
+            op = onnxscript.values.Op(onnx_plug_op, function_proto.name, schema)
+            return op
 
         def converter(*cargs, **ckwargs):
             ags, kws = self._make_args_kwargs(*cargs, **ckwargs)
+            op = get_proto(*cargs)
             return op(*ags, n_outputs=self.n_outputs, **kws)
 
         return onnxscript.values.TracedOnnxFunction(onnx_plug_op, converter)