[LLAMA] KV Cache Injection

src/sparseml/exporters/transforms/kv_cache/__init__.py

@@ -22,4 +22,5 @@
 from .transforms_base import *
 from .transforms_opt import *
 from .transforms_codegen import *
+from .transforms_llama import *
 from .configs import *

src/sparseml/exporters/transforms/kv_cache/configs.py

@@ -23,6 +23,9 @@
 from sparseml.exporters.transforms.kv_cache.transforms_codegen import (
     AdditionalTransformsCodeGen,
 )
+from sparseml.exporters.transforms.kv_cache.transforms_llama import (
+    AdditionalTransformsLLAMA,
+)
 from sparseml.exporters.transforms.kv_cache.transforms_opt import (
     AdditionalTransformsOPT,
 )
@@ -112,10 +115,25 @@ class Config:
     multiply_batch_by_num_att_heads=True,
 )
 
+LLAMA_CONFIG = KeyValueCacheConfig(
+    model_name="llama",
+    additional_transforms=AdditionalTransformsLLAMA,
+    key_num_attention_heads="num_attention_heads",
+    key_num_embedding_hidden_size="hidden_size",
+    transpose_value_input=(0, 2, 1, 3),
+    transpose_key_input=None,
+    multiply_batch_by_num_att_heads=False,
+)
+
 
 def get_kv_cache_config(
     model_path: str,
-    supported_configs: List[BaseModel] = [OPT_CONFIG, CODEGEN_CONFIG, BLOOM_CONFIG],
+    supported_configs: List[BaseModel] = [
+        OPT_CONFIG,
+        CODEGEN_CONFIG,
+        BLOOM_CONFIG,
+        LLAMA_CONFIG,
+    ],
 ) -> KeyValueCacheConfig:
     """
     Get the kv cache config for the model at the given path.

src/sparseml/exporters/transforms/kv_cache/transforms_llama.py

@@ -0,0 +1,223 @@
+# Copyright (c) 2021 - present / Neuralmagic, Inc. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#    http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import logging
+
+import numpy
+import onnx
+from onnx import ModelProto, numpy_helper
+
+from sparseml.exporters.transforms.kv_cache.transforms_base import (
+    AdditionalTransformsBase,
+)
+from sparseml.onnx.utils.graph_editor import ONNXGraph
+from sparseml.onnx.utils.helpers import get_nodes_by_input_id
+
+
+__all__ = ["AdditionalTransformsLLAMA"]
+
+_LOGGER = logging.getLogger(__name__)
+
+
+class AdditionalTransformsLLAMA(AdditionalTransformsBase):
+
+    POSITION_IDS_MATCHING_PATTERN = dict(op_type="Range", children_ops=[["Unsqueeze"]])
+    CAUSAL_MASK_MATCHING_PATTERN = dict(op_type="Expand", children_ops=[["Add"]])
+    SLICE_MAX_INT_NAME = "slice_max_int"
+
+    def transform(self, model: ModelProto) -> ModelProto:
+        """
+        1  Updates the Slice nodes in the attention heads by extending the `ends`
+        operator
+        2. Adds `positions` as an input to the model
+        3. Adds `causal_mask` as an input to the model
+        4. Finds the node that initially creates the `position_ids` tensor
+        5. Updates the node to use the positions input instead of
+           computing it from the Range op
+        6. Finds the nodes that initially create the `causal_mask` tensors
+        7. Updates the nodes to use the causal_mask input instead of
+              computing it from the Expand op
+        8. Update the masks to be floats, as expected by the model
+
+        :param model: model to update
+        :return: updated model
+        """
+
+        model = self.update_slice_nodes_for_positions_input(model)
+        model = self.add_positions_input(model)
+        model = self.add_causal_mask_input(model)
+
+        position_ids_nodes = self.find_nodes_by_pattern(
+            model, pattern=self.POSITION_IDS_MATCHING_PATTERN
+        )
+
+        if len(position_ids_nodes) != 1:
+            raise ValueError(
+                "Expected to find exactly one node matching "
+                f"the pattern {self.POSITION_IDS_MATCHING_PATTERN}, "
+                f"found {len(position_ids_nodes)}"
+            )
+
+        model = self.inject_positions(model, position_ids_nodes, "Unsqueeze")
+
+        causal_mask_nodes = self.find_nodes_by_pattern(
+            model, pattern=self.CAUSAL_MASK_MATCHING_PATTERN
+        )
+        model = self.inject_causal_mask(model, causal_mask_nodes, "Add")
+        model = self.adjust_causal_mask(model)
+        return model
+
+    def update_slice_nodes_for_positions_input(self, model: ModelProto) -> ModelProto:
+        """
+        Update the Slice nodes in the attention heads such that the `ends` operator is
+        set to the max int value. This value is missing from the export and is required
+        for the position ids injection. This is because the onnx export limits access to
+        the entire sin_cached and cos_cached tables, which results in an index error
+        with the position ids:
+
+        https://github.com/huggingface/transformers/blob/
+        7a6efe1e9f756f585f2ffe5ada22cf6b15edd23b/src/transformers/models/llama/
+        modeling_llama.py#L180.
+
+        By updating the `ends` operator, access is allowed to the entire tables.
+        The Slice nodes are identified based on the `data` operator which does not have
+        a parent input (as identified using the `get_node_single_parent` function).
+
+        :param model: model to update
+        :return: updated model with Slice nodes in the attention heads updated
+        """
+        arr = numpy.array(numpy.iinfo(numpy.intp).max).reshape(
+            1,
+        )
+        max_int_tensor = numpy_helper.from_array(arr, name=self.SLICE_MAX_INT_NAME)
+
+        nodes_found = 0
+        for node in model.graph.node:
+            if node.op_type == "Slice":
+                data_parent = ONNXGraph(model).get_node_single_parent(node, 0)
+                if data_parent is not None and len(data_parent.input) == 0:
+                    nodes_found += 1
+                    node.input[2] = self.SLICE_MAX_INT_NAME
+                    self.log_match(node)
+
+        _LOGGER.info(f"Found {nodes_found} slice nodes to update")
+
+        model.graph.initializer.append(max_int_tensor)
+        ONNXGraph(model).delete_orphaned_node_branches()
+        return model
+
+    def adjust_causal_mask(self, model: ModelProto) -> ModelProto:
+        """
+        Insert a `Cast`, `Sub` and `Mul` nodes after the causal mask input to change
+        the initial int64, to a mask of floats expected by the model.
+
+        Transform:
+        ```
+        |       causal_mask
+        |            |
+        |   causal_mask_input_child
+        ```
+        to:
+        ```
+        |       causal_mask (1 and 0)
+        |            |
+        |          Cast  (output -> 1.0 and 0.0)
+        |            |
+        |           Sub (output -> 0.0 and -1.0)
+        |            |
+        |           Mul (output -> 0.0 and numpy.finfo(numpy.float32).min)
+        |            |
+        |   causal_mask_input_child
+
+        The resulting node will change the input int64 mask
+        e.g.
+        ```
+        causal_mask =
+            [[[[1, 1, 1, 0, 0, 0],
+                [1, 1, 1, 1, 0, 0],
+                [1, 1, 1, 1, 1, 0],
+                [1, 1, 1, 1, 1, 1]]]]
+        ```
+
+        to a mask of floats:
+        ```
+        x = numpy.finfo(numpy.float32).min
+        causal_mask_adjusted =
+            [[[[0.0, 0.0, 0.0, x, x, x],
+                [0.0, 0.0, 0.0, 0.0, x, x],
+                [0.0, 0.0, 0.0, 0.0, 0.0, x],
+                [0.0, 0.0, 0.0, 0.0, 0.0, 0.0]]]]
+        ```
+
+        :param model: the model to update
+        :return: the updated model
+        """
+
+        graph = ONNXGraph(model)
+
+        ones_initializer = onnx.helper.make_tensor(
+            name="ones_initializer",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=[1],
+            vals=[1.0],
+        )
+
+        floating_point_limit_initializer = onnx.helper.make_tensor(
+            name="floating_point_limit_initializer",
+            data_type=onnx.TensorProto.FLOAT,
+            dims=[1],
+            vals=[-numpy.finfo(numpy.float32).min],
+        )
+
+        cast_node = onnx.helper.make_node(
+            "Cast",
+            inputs=[self.CAUSAL_MASK_NAME],
+            outputs=[f"{self.CAUSAL_MASK_NAME}_cast"],
+            to=onnx.TensorProto.FLOAT,
+        )
+
+        sub_node = onnx.helper.make_node(
+            "Sub",
+            inputs=[f"{self.CAUSAL_MASK_NAME}_cast", ones_initializer.name],
+            outputs=[f"{self.CAUSAL_MASK_NAME}_sub"],
+        )
+
+        mul_node = onnx.helper.make_node(
+            "Mul",
+            inputs=[
+                f"{self.CAUSAL_MASK_NAME}_sub",
+                floating_point_limit_initializer.name,
+            ],
+            outputs=[f"{self.CAUSAL_MASK_NAME}_mul"],
+        )
+
+        new_nodes = [cast_node, sub_node, mul_node]
+
+        # get the node that takes the causal mask as input
+        # and replace the input with the adjusted causal mask input
+        causal_mask_input_child = get_nodes_by_input_id(model, self.CAUSAL_MASK_NAME)[0]
+
+        for idx, input_name in enumerate(causal_mask_input_child.input):
+            if input_name == self.CAUSAL_MASK_NAME:
+                causal_mask_input_child.input[idx] = f"{self.CAUSAL_MASK_NAME}_mul"
+
+        for node in new_nodes:
+            graph.add_node(node)
+            self.log_match(node)
+
+        model.graph.initializer.extend(
+            [ones_initializer, floating_point_limit_initializer]
+        )
+        _LOGGER.info(f"Successfully adjusted the {self.CAUSAL_MASK_NAME} input")
+
+        return model