Add support assisted decoding in ipex 2.4 (#823)

jiqing-feng · web-flow · commit 5db1ac7697e9 · 2024-09-09T11:39:16.000+02:00
* support assisted decoding in ipex 2.5

* Update optimum/intel/ipex/modeling_base.py

Co-authored-by: Ella Charlaix &lt;80481427+echarlaix@users.noreply.github.com&gt;

* fix tests fail

* fix style

* ipex onnx config

* patch before generate and un-patch after generate

* only patch functions in assisted decoding

* try and cache the genration result and do un-patch

* raise error

* fix style

* ipex 2.4 supports assisted decoding

* fix inputs

* fix generate

* enable assisted decoding tests

* more tests on assisted decoding

* fix config name

* unpatch target model's generation
diff --git a/optimum/exporters/ipex/model_config.py b/optimum/exporters/ipex/model_config.py
@@ -0,0 +1,98 @@
+#  Copyright 2024 The HuggingFace Team. 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.
+
+from typing import Optional, Tuple
+
+from optimum.exporters.onnx.model_configs import (
+    FalconOnnxConfig,
+    GPT2OnnxConfig,
+    LlamaOnnxConfig,
+)
+from optimum.utils import DEFAULT_DUMMY_SHAPES
+from optimum.utils.input_generators import DummyPastKeyValuesGenerator, DummyTextInputGenerator
+from optimum.utils.normalized_config import NormalizedTextConfig
+
+
+DEFAULT_DUMMY_SHAPES["batch_size"] = 1
+
+
+class IPEXDummyPastKeyValuesGenerator(DummyPastKeyValuesGenerator):
+    def __init__(
+        self,
+        task: str,
+        normalized_config: NormalizedTextConfig,
+        batch_size: int = DEFAULT_DUMMY_SHAPES["batch_size"],
+        sequence_length: int = DEFAULT_DUMMY_SHAPES["sequence_length"],
+        random_batch_size_range: Optional[Tuple[int, int]] = None,
+        random_sequence_length_range: Optional[Tuple[int, int]] = None,
+        **kwargs,
+    ):
+        super().__init__(
+            task=task,
+            normalized_config=normalized_config,
+            batch_size=batch_size,
+            sequence_length=sequence_length,
+            random_batch_size_range=random_batch_size_range,
+            random_sequence_length_range=random_sequence_length_range,
+        )
+        self.num_key_value_heads = getattr(normalized_config, "num_key_value_heads", 1)
+        self.max_position_embeddings = normalized_config.max_position_embeddings
+
+    def generate(self, input_name: str, framework: str = "pt", int_dtype: str = "int64", float_dtype: str = "fp32"):
+        shape_init = (1, self.sequence_length, self.sequence_length, 1)
+        shape_beam_idx_tmp = (self.max_position_embeddings, self.batch_size)
+        shape_kv = (
+            self.max_position_embeddings,
+            self.batch_size,
+            self.num_key_value_heads,
+            self.hidden_size // self.num_attention_heads,
+        )
+        return [
+            (
+                self.random_int_tensor(shape_init, max_value=1, framework=framework).contiguous(),
+                self.random_float_tensor(shape_kv, framework=framework, dtype=float_dtype).contiguous(),
+                self.random_float_tensor(shape_kv, framework=framework, dtype=float_dtype).contiguous(),
+                self.random_int_tensor(shape_beam_idx_tmp, max_value=1, framework=framework).contiguous(),
+            )
+            for _ in range(self.num_layers)
+        ]
+
+
+class IPEXDummyTextInputGenerator(DummyTextInputGenerator):
+    def __init__(
+        self,
+        task: str,
+        normalized_config: NormalizedTextConfig,
+        batch_size: int = DEFAULT_DUMMY_SHAPES["batch_size"],
+        **kwargs,
+    ):
+        super().__init__(task, normalized_config, batch_size, **kwargs)
+
+
+class LlamaIPEXConfig(LlamaOnnxConfig):
+    DUMMY_INPUT_GENERATOR_CLASSES = (IPEXDummyTextInputGenerator, IPEXDummyPastKeyValuesGenerator)
+    DUMMY_PKV_GENERATOR_CLASS = IPEXDummyPastKeyValuesGenerator
+
+
+class FalconIPEXConfig(FalconOnnxConfig):
+    DUMMY_INPUT_GENERATOR_CLASSES = (IPEXDummyTextInputGenerator, IPEXDummyPastKeyValuesGenerator)
+    DUMMY_PKV_GENERATOR_CLASS = IPEXDummyPastKeyValuesGenerator
+
+
+class GPT2IPEXConfig(GPT2OnnxConfig):
+    DUMMY_INPUT_GENERATOR_CLASSES = (IPEXDummyTextInputGenerator, IPEXDummyPastKeyValuesGenerator)
+    DUMMY_PKV_GENERATOR_CLASS = IPEXDummyPastKeyValuesGenerator
+
+
+ipex_onnx_config = {"llama": LlamaIPEXConfig, "falcon": FalconIPEXConfig, "gpt2": GPT2IPEXConfig}
diff --git a/optimum/intel/ipex/modeling_base.py b/optimum/intel/ipex/modeling_base.py
@@ -23,10 +23,10 @@
 
 import intel_extension_for_pytorch as ipex
 import torch
+import transformers
 from huggingface_hub import hf_hub_download
 from huggingface_hub.constants import HUGGINGFACE_HUB_CACHE
 from intel_extension_for_pytorch.cpu._auto_kernel_selection import _enable_tpp
-from intel_extension_for_pytorch.transformers.optimize import get_dummy_input
 from transformers import (
     AutoConfig,
     AutoModel,
@@ -43,20 +43,24 @@
     is_torch_xpu_available,
 )
 from transformers.dynamic_module_utils import get_class_from_dynamic_module
+from transformers.generation.candidate_generator import _crop_past_key_values
 from transformers.modeling_outputs import CausalLMOutputWithPast, ModelOutput
 from transformers.models.auto.auto_factory import _get_model_class as get_model_class
 from transformers.utils import WEIGHTS_NAME
 
 from optimum.exporters import TasksManager
+from optimum.exporters.tasks import make_backend_config_constructor_for_task
 from optimum.modeling_base import OptimizedModel
 from optimum.utils import NormalizedConfigManager
 
+from ...exporters.ipex.model_config import ipex_onnx_config
 from ...exporters.ipex.model_patcher import (
     _IPEX_EXPORTED_GENERATION_TASKS,
     _IPEX_MINIMUM_VERSION_FOR_PATCHING,
     _patch_model,
 )
-from ..generation.modeling import prepare_jit_inputs
+from ..generation.modeling import get_float_type
+from ..utils.constant import _TASK_ALIASES
 from ..utils.import_utils import is_ipex_version, is_torch_version, is_transformers_version
 from ..utils.modeling_utils import MULTI_QUERY_ATTN_MODELS, recursive_to_device
 
@@ -86,10 +90,35 @@ def _is_patched_with_ipex(model, task):
 
 
 def _prepare_inputs_for_ipex_model(model, task, use_cache):
-    if task in _IPEX_EXPORTED_GENERATION_TASKS and _is_patched_with_ipex(model, task):
-        return get_dummy_input(model, return_dict=True)
+    task = _TASK_ALIASES.get(task, task)
+    signature = inspect.signature(model.forward) if hasattr(model, "forward") else inspect.signature(model.__call__)
+    if _is_patched_with_ipex(model, task) and model.config.model_type in ipex_onnx_config:
+        onnx_config_class = make_backend_config_constructor_for_task(
+            ipex_onnx_config[model.config.model_type], task=task
+        )
+    else:
+        onnx_config_class = TasksManager.get_exporter_config_constructor(model=model, exporter="onnx", task=task)
+    float_dtype = get_float_type(model.dtype)
+    if "text-generation" in task:
+        onnx_config = onnx_config_class(
+            model.config, use_past=use_cache, use_past_in_inputs=use_cache, float_dtype=float_dtype
+        )
     else:
-        return prepare_jit_inputs(model, task, use_cache)
+        onnx_config = onnx_config_class(model.config)
+
+    dummy_inputs = onnx_config.generate_dummy_inputs(framework="pt")
+
+    # Check attention_mask shape
+    if _is_patched_with_ipex(model, task) and model.config.model_type in ipex_onnx_config and use_cache:
+        past_len = dummy_inputs["past_key_values"][0][0].shape[-2]
+        input_len = dummy_inputs["input_ids"].shape[-1]
+        attention_len = dummy_inputs["attention_mask"].shape[-1]
+        if attention_len != input_len + past_len:
+            dummy_inputs["attention_mask"] = torch.ones([dummy_inputs["input_ids"].shape[0], input_len + past_len]).to(
+                dummy_inputs["input_ids"].dtype
+            )
+
+    return {key: dummy_inputs[key] for key in signature.parameters if dummy_inputs.get(key, None) is not None}
 
 
 def ipex_jit_trace(model, task, use_cache):
@@ -103,11 +132,7 @@ def ipex_jit_trace(model, task, use_cache):
     sample_inputs = _prepare_inputs_for_ipex_model(model, task, use_cache)
 
     model.config.return_dict = False
-
-    if "past_key_values" in sample_inputs:
-        model.config.use_cache = use_cache
-        if not use_cache:
-            sample_inputs.pop("past_key_values")
+    model.config.use_cache = use_cache
 
     # Use Tensor Processing Primitives to accelerate linear, see https://arxiv.org/abs/2104.05755.
     # Only ipex >= 2.3.0 supports tpp. The tpp is only verified for llm in generation tasks.
@@ -372,7 +397,7 @@ def _init_warmup(self):
         # TODO : add warmup for IPEX exported model
         if not self._is_ipex_exported:
             use_cache = "past_key_values" in self.input_names
-            dummy_inputs = prepare_jit_inputs(self, self.export_feature, use_cache)
+            dummy_inputs = _prepare_inputs_for_ipex_model(self, self.export_feature, use_cache)
             if self._device.type != "cpu":
                 dummy_inputs = recursive_to_device(value=dummy_inputs, device=self._device)
             for _ in range(2):
@@ -652,11 +677,28 @@ def _prepare_generation_config(
         return generation_config, model_kwargs
 
     def generate(self, *args, **kwargs):
-        if self._is_ipex_exported and kwargs.get("assistant_model", None):
+        if is_ipex_version("<", "2.4.0") and self._is_ipex_exported and kwargs.get("assistant_model", None):
             raise ValueError(
-                f"Assisted decoding is not supported for patched models for now, support methods are {_IPEX_EXPORTED_GENERATION_METHODS}"
+                f"Assisted decoding is not supported for patched models if ipex < 2.4, support methods are {_IPEX_EXPORTED_GENERATION_METHODS}"
             )
-        return super().generate(*args, **kwargs)
+        # Patch functions to support IAKV cache
+        if self._is_ipex_exported and kwargs.get("assistant_model", None):
+            transformers.generation.utils._crop_past_key_values = _ipex_crop_past_key_values
+        elif self._is_ipex_exported:
+            transformers.generation.candidate_generator._crop_past_key_values = _ipex_crop_past_key_values
+
+        try:
+            result = super().generate(*args, **kwargs)
+        except Exception as e:
+            transformers.generation.utils._crop_past_key_values = _crop_past_key_values
+            transformers.generation.candidate_generator._crop_past_key_values = _crop_past_key_values
+            raise e
+
+        if self._is_ipex_exported and kwargs.get("assistant_model", None):
+            transformers.generation.utils._crop_past_key_values = _crop_past_key_values
+            transformers.generation.candidate_generator._crop_past_key_values = _crop_past_key_values
+
+        return result
 
 
 def _ipex_prepare_inputs_for_generation(
@@ -736,3 +778,16 @@ def _ipex_reorder_cache(
             tuple(past_state.index_select(0, beam_idx.to(past_state.device)) for past_state in layer_past)
             for layer_past in past_key_values
         )
+
+
+def _ipex_crop_past_key_values(model, past_key_values, max_length):
+    if isinstance(model, IPEXModel) and _is_patched_with_ipex(model, "text-generation"):
+        new_past_key_values = []
+        for i in range(len(past_key_values)):
+            pkv = []
+            pkv.append(past_key_values[i][0][:, :max_length, :max_length, :])
+            pkv += [past_key_values[i][_] for _ in range(1, 4)]
+            new_past_key_values.append(tuple(pkv))
+        new_past_key_values = tuple(new_past_key_values)
+        return new_past_key_values
+    return _crop_past_key_values(model, past_key_values, max_length)
diff --git a/tests/ipex/test_modeling.py b/tests/ipex/test_modeling.py
@@ -281,11 +281,10 @@ def test_pipeline(self, model_arch):
         self.assertEqual(pipe.device, model.device)
         self.assertTrue(all("This is a sample" in item["generated_text"] for item in outputs))
 
-    # High optimized model llama is not supported assisted decoding for now.
     @parameterized.expand(SUPPORTED_ARCHITECTURES)
     def test_assisted_decoding(self, model_arch):
-        # Patched models are not support assisted decoding for now.
-        if model_arch in self.IPEX_PATCHED_SUPPORTED_ARCHITECTURES:
+        # Patched models are not support assisted decoding if ipex < 2.5.
+        if model_arch in self.IPEX_PATCHED_SUPPORTED_ARCHITECTURES and is_ipex_version("<", "2.4.0"):
             return
         model_id = MODEL_NAMES[model_arch]
         tokenizer = AutoTokenizer.from_pretrained(model_id)
@@ -296,11 +295,15 @@ def test_assisted_decoding(self, model_arch):
         ipex_output_assisted = ipex_model.generate(
             **tokens, do_sample=False, assistant_model=transformers_model, max_new_tokens=4
         )
+        ipex_output_assisted_2 = ipex_model.generate(
+            **tokens, do_sample=False, assistant_model=ipex_model, max_new_tokens=4
+        )
         transformers_output = transformers_model.generate(**tokens, do_sample=False, max_new_tokens=4)
         transformers_output_assisted = transformers_model.generate(
             **tokens, do_sample=False, assistant_model=ipex_model, max_new_tokens=4
         )
         self.assertTrue(torch.equal(ipex_output, ipex_output_assisted))
+        self.assertTrue(torch.equal(ipex_output, ipex_output_assisted_2))
         self.assertTrue(torch.equal(transformers_output, transformers_output_assisted))
 
     @parameterized.expand(
