onnx_diagnostic.torch_export_patches.patch_details

class onnx_diagnostic.torch_export_patches.patch_details.PatchDetails

This class is used to store patching information. This helps understanding which rewriting was applied to which method of functions. Page Patches Diff contains all the diff for all the implemented patches.

<<<

import torch
from onnx_diagnostic.torch_export_patches import torch_export_patches
from onnx_diagnostic.torch_export_patches.patch_inputs import use_dyn_not_str
from onnx_diagnostic.torch_export_patches.patch_details import PatchDetails
from onnx_diagnostic.torch_models.hghub import get_untrained_model_with_inputs

data = get_untrained_model_with_inputs("arnir0/Tiny-LLM", verbose=0)
model, inputs, ds = data["model"], data["inputs"], data["dynamic_shapes"]
details = PatchDetails()
with torch_export_patches(
    patch_transformers=True, patch_details=details, patch_torch=False
):
    ep = torch.export.export(
        model, (), kwargs=inputs, dynamic_shapes=use_dyn_not_str(ds)
    )
patches = details.patches_involded_in_graph(ep.graph)
report = details.make_report(patches, format="rst")
print(report)

>>>

auto/patch_transformers: LlamaRotaryEmbedding.forward -> common_RotaryEmbedding.forward

--- original
+++ rewritten
@@ -1,18 +1,26 @@
-@torch.no_grad()
-@dynamic_rope_update  # power user: used with advanced RoPE types (e.g. dynamic rope)
-def forward(self, x, position_ids):
+@patched_dynamic_rope_update
+def forward(self, x, position_ids, layer_type=None):
+    if layer_type is not None:
+        # transformers>=5.0
+        inv_freq = getattr(self, f"{layer_type}_inv_freq")
+        attention_scaling = getattr(self, f"{layer_type}_attention_scaling")
+    else:
+        # transformers<5.0
+        inv_freq = self.inv_freq
+        attention_scaling = self.attention_scaling
+
     inv_freq_expanded = (
-        self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
+        inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
     )
     position_ids_expanded = position_ids[:, None, :].float()

     device_type = (
         x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
     )
-    with maybe_autocast(device_type=device_type, enabled=False):  # Force float32
+    with torch.autocast(device_type=device_type, enabled=False):  # Force float32
         freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
         emb = torch.cat((freqs, freqs), dim=-1)
-        cos = emb.cos() * self.attention_scaling
-        sin = emb.sin() * self.attention_scaling
+        cos = emb.cos() * attention_scaling
+        sin = emb.sin() * attention_scaling

     return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)

--- original
+++ rewritten
@@ -1,99 +1,193 @@
-def dynamic_rope_update(rope_forward):
-    """
-    Decorator function to update the RoPE parameters in the forward pass, if the model is using a dynamic RoPE
-    (i.e. a RoPE implementation that may recompute its frequencies in the forward pass).
+def patched_dynamic_rope_update(rope_forward):
+    """manual patch: ``[patch:transformers.modeling_rope_utils.dynamic_rope_update]``

-    Args:
-        rope_forward (Callable):
-            The forward pass of the RoPE implementation.
+    ``rope_type`` is determined in the constructor of class
+    :class:`transformers.models.phi3.modeling_phi3.Phi3RotaryEmbedding`.

-    Returns:
-        The decorated forward pass.
+    .. code-block:: python
+
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type"))
+        else:
+            self.rope_type = "default"
+
+    The original code of the patched function:
+
+    .. code-block:: python
+
+        def dynamic_rope_update(rope_forward):
+            def longrope_frequency_update(self, position_ids, device):
+                seq_len = torch.max(position_ids) + 1
+                if hasattr(self.config, "original_max_position_embeddings"):
+                    original_max_position_embeddings =
+                        self.config.original_max_position_embeddings
+                else:
+                    original_max_position_embeddings =
+                        self.config.max_position_embeddings
+                if seq_len > original_max_position_embeddings:
+                    if not hasattr(self, "long_inv_freq"):
+                        self.long_inv_freq, _ = self.rope_init_fn(
+                            self.config, device, seq_len=original_max_position_embeddings + 1
+                        )
+                    self.register_buffer("inv_freq", self.long_inv_freq, persistent=False)
+                else:
+                    self.original_inv_freq = self.original_inv_freq.to(device)
+                    self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+
+            def dynamic_frequency_update(self, position_ids, device):
+                seq_len = torch.max(position_ids) + 1
+                if seq_len > self.max_seq_len_cached:  # growth
+                    inv_freq, self.attention_scaling = self.rope_init_fn(
+                        self.config, device, seq_len=seq_len)
+                    self.register_buffer("inv_freq", inv_freq, persistent=False)
+                    self.max_seq_len_cached = seq_len
+
+                if seq_len < self.original_max_seq_len and
+                        self.max_seq_len_cached > self.original_max_seq_len:
+                    self.original_inv_freq = self.original_inv_freq.to(device)
+                    self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+                    self.max_seq_len_cached = self.original_max_seq_len
+
+            @wraps(rope_forward)
+            def wrapper(self, x, position_ids):
+                if "dynamic" in self.rope_type:
+                    dynamic_frequency_update(self, position_ids, device=x.device)
+                elif self.rope_type == "longrope":
+                    longrope_frequency_update(self, position_ids, device=x.device)
+                return rope_forward(self, x, position_ids)
+
+            return wrapper
+
     """

     def longrope_frequency_update(self, position_ids, device, layer_type=None):
-        """Longrope uses long factor if sequence is larger than original pretraining length, short otherwise."""
+        # It is no use to patch the function after the model is created
+        # as rope_init_fn is an attribute set to one function when the model
+        # is created and when no patch is applied yet.
+        # So we select the patched version here.
+        rope_init_fn = _get_rope_init_fn(self, layer_type=layer_type)
         seq_len = torch.max(position_ids) + 1
-        original_max_position_embeddings = getattr(
-            self.config, "original_max_position_embeddings", self.config.max_position_embeddings
-        )
+        if hasattr(self.config, "original_max_position_embeddings"):
+            original_max_position_embeddings = self.config.original_max_position_embeddings
+        else:
+            original_max_position_embeddings = self.config.max_position_embeddings
+
         if layer_type is None:
-            rope_type = self.rope_type
+            # rope_type = self.rope_type
             original_inv_freq = self.original_inv_freq
             prefix = ""
         else:
-            rope_type = self.rope_type[layer_type]
+            # rope_type = self.rope_type[layer_type]
             original_inv_freq = getattr(self, f"{layer_type}_original_inv_freq")
             prefix = f"{layer_type}_"

-        if seq_len > original_max_position_embeddings:
-            if not hasattr(self, f"{layer_type}_long_inv_freq"):
-                rope_init_fn = ROPE_INIT_FUNCTIONS[rope_type]
-                long_inv_freq, _ = rope_init_fn(
-                    self.config,
-                    device,
-                    seq_len=original_max_position_embeddings + 1,
-                    layer_type=layer_type,
-                )
-            self.register_buffer(f"{prefix}inv_freq", long_inv_freq, persistent=False)
-            setattr(self, f"{prefix}long_inv_freq", long_inv_freq)
-        else:
-            # This .to() is needed if the model has been moved to a device after being initialized (because
-            # the buffer is automatically moved, but not the original copy)
-            original_inv_freq = original_inv_freq.to(device)
-            self.register_buffer(f"{prefix}inv_freq", original_inv_freq, persistent=False)
-            setattr(self, f"{prefix}original_inv_freq", original_inv_freq)
+        # At export time, seq_len is unknown.
+        long_inv_freq, _ = rope_init_fn(
+            self.config, device, seq_len=original_max_position_embeddings + 1
+        )
+        original_inv_freq = self.original_inv_freq.to(device)
+
+        # PATCHED: uses torch.cond instead of a test
+        cond = (seq_len > original_max_position_embeddings).item()
+        inv_freq = torch.cond(
+            cond,
+            (lambda x, y: x.clone()),
+            (lambda x, y: y.clone()),
+            [long_inv_freq, original_inv_freq],
+        )
+        setattr(self, f"{prefix}inv_freq", inv_freq)
+        # if seq_len > original_max_position_embeddings:
+        #    self.inv_freq = self.long_inv_freq
+        # else:
+        #    self.inv_freq = self.original_inv_freq

     def dynamic_frequency_update(self, position_ids, device, layer_type=None):
-        """
-        dynamic RoPE layers should recompute `inv_freq` in the following situations:
-        1 - growing beyond the cached sequence length (allow scaling)
-        2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
-        """
+        # constructor:
+        # - self.max_seq_len_cached = config.max_position_embeddings
+        # - self.original_max_seq_len = config.max_position_embeddings
+        # - inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device)
+
+        # It is no use to patch the function after the model is created
+        # as rope_init_fn is an attribute set to one function when the model
+        # is created and when no patch is applied yet.
+        # So we select the patched version here.
+        rope_init_fn = _get_rope_init_fn(self, layer_type=layer_type)
+
+        # This behaviour is difficult to translate.
+        # The sequence always grows.
+        # The test should always True.
+        # So:  self.max_seq_len_cached = max(self.max_seq_len_cached, seq_len) --> seq_len
+        #
+        # if seq_len > self.max_seq_len_cached:  # growth
+        #    inv_freq, self.attention_scaling = self.rope_init_fn(
+        #        self.config, device, seq_len=seq_len
+        #    )
+        #    self.register_buffer("inv_freq", inv_freq, persistent=False)
+        #    self.max_seq_len_cached = seq_len
+        #
+        # So we should not need what follows.
+        #
+        # cond = (seq_len > self.max_seq_len_cached).item()
+        # self.attention_scaling = torch.cond(
+        #    cond,
+        #    (lambda x, y: x.clone()),
+        #    (lambda x, y: y.clone()),
+        #    [attention_scaling, self.attention_scaling],
+        # )
+
         seq_len = torch.max(position_ids) + 1
+        long_inv_freq, self.attention_scaling = rope_init_fn(self.config, device, seq_len=seq_len)
+
         if layer_type is None:
-            rope_type = self.rope_type
-            max_seq_len_cached = self.max_seq_len_cached
+            # rope_type = self.rope_type
+            # max_seq_len_cached = self.max_seq_len_cached
             original_inv_freq = self.original_inv_freq
             prefix = ""
         else:
-            rope_type = self.rope_type[layer_type]
-            max_seq_len_cached = getattr(
-                self, f"{layer_type}_max_seq_len_cached", self.max_seq_len_cached
-            )
+            # rope_type = self.rope_type[layer_type]
+            # max_seq_len_cached = getattr(
+            #     self, f"{layer_type}_max_seq_len_cached", self.max_seq_len_cached
+            # )
             original_inv_freq = getattr(self, f"{layer_type}_original_inv_freq")
             prefix = f"{layer_type}_"

-        if seq_len > max_seq_len_cached:  # growth
-            rope_init_fn = ROPE_INIT_FUNCTIONS[rope_type]
-            inv_freq, self.attention_scaling = rope_init_fn(
-                self.config,
-                device,
-                seq_len=seq_len,
-                layer_type=layer_type,
-            )
-            # TODO joao: may break with compilation
-            self.register_buffer(f"{prefix}inv_freq", inv_freq, persistent=False)
-            setattr(self, f"{layer_type}_max_seq_len_cached", seq_len)
+        # Second test to translate.
+        # Let's keep in mind, self.max_seq_len_cached = seq_len is likely to be True.
+        # But in that case the following condition is a way to restore the original cache.

-        if (
-            seq_len < self.original_max_seq_len and max_seq_len_cached > self.original_max_seq_len
-        ):  # reset
-            # This .to() is needed if the model has been moved to a device after being initialized (because
-            # the buffer is automatically moved, but not the original copy)
-            original_inv_freq = original_inv_freq.to(device)
-            self.register_buffer(f"{prefix}inv_freq", original_inv_freq, persistent=False)
-            setattr(self, f"{prefix}original_inv_freq", original_inv_freq)
-            setattr(self, f"{layer_type}_max_seq_len_cached", self.original_max_seq_len)
+        # if (
+        #    seq_len < self.original_max_seq_len
+        #    and self.max_seq_len_cached > self.original_max_seq_len
+        # ):
+        #    self.original_inv_freq = self.original_inv_freq.to(device)
+        #    self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+        #    self.max_seq_len_cached = self.original_max_seq_len
+
+        original_inv_freq = self.original_inv_freq.to(device)
+        cond = (seq_len >= self.original_max_seq_len).item()
+        # PATCHED: uses torch.cond instead of a test
+        inv_freq = torch.cond(
+            cond,
+            (lambda x, y: x.clone()),
+            (lambda x, y: y.clone()),
+            [long_inv_freq, original_inv_freq],
+        )
+        setattr(self, f"{prefix}inv_freq", inv_freq)

     @wraps(rope_forward)
     def wrapper(self, x, position_ids, layer_type=None):
-        rope_type = self.rope_type if layer_type is None else self.rope_type[layer_type]
-        kwargs = {"layer_type": layer_type} if layer_type is not None else {}
-        if "dynamic" in rope_type:
-            dynamic_frequency_update(self, position_ids, device=x.device, **kwargs)
-        elif rope_type == "longrope":
-            longrope_frequency_update(self, position_ids, device=x.device, **kwargs)
-        return rope_forward(self, x, position_ids, **kwargs)
+        if layer_type is None:
+            if "dynamic" in self.rope_type:
+                dynamic_frequency_update(self, position_ids, device=x.device)
+            elif self.rope_type == "longrope":
+                longrope_frequency_update(self, position_ids, device=x.device)
+            return rope_forward(self, x, position_ids)
+
+        if "dynamic" in self.rope_type:
+            dynamic_frequency_update(self, position_ids, device=x.device, layer_type=layer_type)
+        elif self.rope_type == "longrope":
+            longrope_frequency_update(self, position_ids, device=x.device, layer_type=layer_type)
+        return rope_forward(self, x, position_ids, layer_type=layer_type)

     return wrapper

impacted nodes

aten.unsqueeze.default(b_model_rotary_emb_inv_freq, 0) -> unsqueeze_12
aten.unsqueeze.default(unsqueeze_12, 2) -> unsqueeze_13
aten._assert_tensor_metadata.default(unsqueeze_13) -> _assert_tensor_metadata_default_3
aten.to.dtype(unsqueeze_13, torch.float32) -> to_3
aten.expand.default(to_3, [sym_size_int_19, -1, 1]) -> expand_1
aten._assert_tensor_metadata.default(expand_1) -> _assert_tensor_metadata_default_4
aten.to.dtype_layout(expand_1) -> to_4
aten.slice.Tensor(position_ids, 0, 0, 9223372036854775807) -> slice_4
aten.unsqueeze.default(slice_4, 1) -> unsqueeze_14
aten.slice.Tensor(unsqueeze_14, 2, 0, 9223372036854775807) -> slice_5
aten._assert_tensor_metadata.default(slice_5) -> _assert_tensor_metadata_default_5
aten.to.dtype(slice_5, torch.float32) -> to_5
<built-in function getitem>(wrap_with_autocast, 0) -> mul
<built-in function getitem>(wrap_with_autocast, 1) -> mul_1
aten._assert_tensor_metadata.default(mul) -> _assert_tensor_metadata_default_8
aten.to.dtype(mul, torch.float32) -> to_8
aten._assert_tensor_metadata.default(mul_1) -> _assert_tensor_metadata_default_9
aten.to.dtype(mul_1, torch.float32) -> to_9

append(family: str, function_to_patch: str | Callable, patch: Callable) → PatchInfo

Stores a patch.

Parameters:

• family – a category, anything to classify the patch

• function_to_patch – function to patch

• patch – function patched

Returns:

instance of PatchInfo

data() → List[Dict[str, Any]]

Returns the data for a dataframe.

find(name: str) → PatchInfo | None

Finds a patch by name.

make_report(patches: List[Tuple[PatchInfo, List[torch.fx.Node]]], format: str = 'raw') → str

Creates a report based on the involved patches.

Parameters:

• patches – from method patches_involded_in_graph()

• format – format of the report

Returns:

report

matching_pair(patch: PatchInfo, node: torch.fx.Node) → bool

Last validation for a pair. RotaryEmbedding has many rewriting and they all end up in the same code line.

property n_patches: int

Returns the number of stored patches.

patches_involded_in_graph(graph: torch.fx.Graph) → List[Tuple[PatchInfo, List[torch.fx.Node]]]

Enumerates all patches impacting a graph. The function goes through the graph node (only the main graph) and looks into the metadata to determine if a listed patch was involved.

Parameters:

graph – fx graph

Returns:

list of nodes impacted by a patch

class onnx_diagnostic.torch_export_patches.patch_details.PatchInfo(function_to_patch: str | Callable, patch: Callable, family: str = '')

Stores information about patches.

Parameters:

• function_to_patch – function to patch

• patch – function patched

• family – a category, anything to classify the patch

format_diff(format: str = 'raw') → str

Format a diff between two function as a string.

Parameters:

format – 'raw' or 'rst'

Returns:

diff

<<<

import transformers
import onnx_diagnostic.torch_export_patches.patches.patch_transformers as ptr
from onnx_diagnostic.torch_export_patches.patch_details import PatchInfo
from onnx_diagnostic.torch_export_patches.patches.patch_transformers import (
    patched_eager_mask,
)

eager_mask = transformers.masking_utils.eager_mask
diff = PatchInfo(eager_mask, patched_eager_mask).format_diff(format="rst")
print(diff)

>>>

eager_mask -> patched_eager_mask

--- original
+++ rewritten
@@ -1,4 +1,4 @@
-def eager_mask(
+def patched_eager_mask(
     batch_size: int,
     cache_position: torch.Tensor,
     kv_length: int,
@@ -6,41 +6,14 @@
     mask_function: Callable = causal_mask_function,
     attention_mask: Optional[torch.Tensor] = None,
     dtype: torch.dtype = torch.float32,
-    allow_is_bidirectional_skip: bool = False,
-    use_vmap: bool = False,
     **kwargs,
 ) -> torch.Tensor:
-    """
-    Create a 4D float mask of shape `(batch_size, 1, query_length, kv_length)` where a value of 0 indicates that
-    the element should take part in the attention computation, and -inf (minimum value for the given `dtype`) that
-    it should not.
-
-    Args:
-        batch_size (`int`):
-            The batch size of the input sequence.
-        cache_position (`torch.Tensor`):
-            A tensor of shape (query_length,) indicating the current indices of the input sequence elements.
-        kv_length (`int`):
-            The size that the key and value states will have during the attention computation.
-        kv_offset (`int`, optional):
-            An optional offset to indicate at which first position the key and values states will refer to.
-        mask_function (`Callable`):
-            The mask factory function describing the mask pattern.
-        attention_mask (`torch.Tensor`, optional):
-            The 2D attention mask corresponding to padded tokens of shape (batch_size, number_of_seen_tokens+q_length)
-        dtype (`torch.dtype`, optional):
-            The dtype to use for the mask. By default, `torch.float32`.
-        allow_is_bidirectional_skip (`bool`, optional):
-            Whether to allow to return `None` for the mask under conditions where we do not have to add any bias,
-            i.e. full attention without any padding. Default to `False`.
-        use_vmap (`bool`, optional):
-            Whether to use `vmap` during the mask construction or not. Allows powerful custom patterns that may not be
-            index-based (for the cost of speed performance). By default `False`.
-    """
+    """manual patch for function ``transformers.masking_utils.eager_mask``."""
     # The masks for eager attention are simply boolean mask from sdpa, casted to 0 and -inf
     _ = kwargs.pop("allow_is_causal_skip", None)
-    _ = kwargs.pop("allow_torch_fix", None)
-    mask = sdpa_mask(
+    _ = kwargs.pop("allow_is_bidirectional_skip", None)
+    # PATCHED: this line called the patched version of sdpa_mask
+    mask = patched_sdpa_mask_recent_torch(
         batch_size=batch_size,
         cache_position=cache_position,
         kv_length=kv_length,
@@ -48,14 +21,15 @@
         mask_function=mask_function,
         attention_mask=attention_mask,
         allow_is_causal_skip=False,
-        allow_is_bidirectional_skip=allow_is_bidirectional_skip,
+        allow_is_bidirectional_skip=False,
         allow_torch_fix=False,
-        use_vmap=use_vmap,
         **kwargs,
     )
-    # only bidirectional masks can be skipped, otherwise we convert bool -> float
-    if mask is not None:
-        min_dtype = torch.finfo(dtype).min
-        # we need 0s where the tokens should be taken into account, and -inf otherwise (mask is already of boolean type)
-        mask = torch.where(mask, torch.tensor(0.0, device=mask.device, dtype=dtype), min_dtype)
+    min_dtype = torch.finfo(dtype).min
+    # PATCHED: the following line
+    # we need 0s where the tokens should be taken into account,
+    # and -inf otherwise (mask is already of boolean type)
+    # mask =
+    #   torch.where(mask, torch.tensor(0.0, device=mask.device, dtype=dtype), min_dtype)
+    mask = (~mask).to(dtype) * min_dtype
     return mask

make_diff() → str

Returns a diff as a string.

to_dict() → Dict[str, Any]

usual

to_tuple() → Tuple[str, Callable, Callable]

usual

onnx_diagnostic.torch_export_patches.patch_details.clean_code_with_black(code: str) → str

Changes the code style with black if available.

onnx_diagnostic.torch_export_patches.patch_details.make_diff_code(code1: str, code2: str, output: str | None = None) → str

Creates a diff between two codes.

Parameters:

• code1 – first code

• code2 – second code

• output – if not empty, stores the output in this file

Returns:

diff