diff --git a/byte_infer_perf/general_perf/backends/CPU/calculate_cpu_diff.sh b/byte_infer_perf/general_perf/backends/CPU/calculate_cpu_diff.sh old mode 100644 new mode 100755 index f786dfa0b..f7960afdf --- a/byte_infer_perf/general_perf/backends/CPU/calculate_cpu_diff.sh +++ b/byte_infer_perf/general_perf/backends/CPU/calculate_cpu_diff.sh @@ -1,12 +1,16 @@ #！bin/bash -if [ ! -d "general_perf/backends/CPU/venv" ];then - virtualenv -p python3 general_perf/backends/CPU/venv - source general_perf/backends/CPU/venv/bin/activate - general_perf/backends/CPU/venv/bin/python3 -m pip install --upgrade pip -q - general_perf/backends/CPU/venv/bin/python3 -m pip install -r general_perf/backends/CPU/requirements.txt -q +if [ "$3" != 'TPU' ]; then + if [ ! -d "general_perf/backends/CPU/venv" ];then + virtualenv -p python3 general_perf/backends/CPU/venv + source general_perf/backends/CPU/venv/bin/activate + general_perf/backends/CPU/venv/bin/python3 -m pip install --upgrade pip -q + general_perf/backends/CPU/venv/bin/python3 -m pip install -r general_perf/backends/CPU/requirements.txt -q + else + source general_perf/backends/CPU/venv/bin/activate + general_perf/backends/CPU/venv/bin/python3 -m pip install -r general_perf/backends/CPU/requirements.txt -q + fi else - source general_perf/backends/CPU/venv/bin/activate - general_perf/backends/CPU/venv/bin/python3 -m pip install -r general_perf/backends/CPU/requirements.txt -q + echo "Hardware Type is TPU, Skip venv..." fi python3 general_perf/backends/CPU/calculate_cpu_diff.py --task $1 --batch_size $2 diff --git a/byte_infer_perf/general_perf/backends/CPU/compile_backend_cpu.py b/byte_infer_perf/general_perf/backends/CPU/compile_backend_cpu.py old mode 100644 new mode 100755 index 3d88a1114..76fa45675 --- a/byte_infer_perf/general_perf/backends/CPU/compile_backend_cpu.py +++ b/byte_infer_perf/general_perf/backends/CPU/compile_backend_cpu.py @@ -2,8 +2,8 @@ import json import logging os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' -import tensorflow as tf import torch +import tensorflow as tf import onnxruntime import time import numpy as np diff --git a/byte_infer_perf/general_perf/backends/CPU/runtime_backend_cpu.py b/byte_infer_perf/general_perf/backends/CPU/runtime_backend_cpu.py old mode 100644 new mode 100755 index eec8c98b2..78fa11efc --- a/byte_infer_perf/general_perf/backends/CPU/runtime_backend_cpu.py +++ b/byte_infer_perf/general_perf/backends/CPU/runtime_backend_cpu.py @@ -2,8 +2,8 @@ import json import logging os.environ['TF_CPP_MIN_LOG_LEVEL'] = '3' -import tensorflow as tf import torch +import tensorflow as tf import onnxruntime import time import numpy as np diff --git a/byte_infer_perf/general_perf/backends/TPU/README.md b/byte_infer_perf/general_perf/backends/TPU/README.md new file mode 100644 index 000000000..9313a42ed --- /dev/null +++ b/byte_infer_perf/general_perf/backends/TPU/README.md @@ -0,0 +1,32 @@ + + +# How to run + +## 1. Create docker container + +```bash +docker pull sophgo/tpuc_dev:latest +docker run --privileged --name TPUPerf -td -v /dev/:/dev/ -v /opt/:/opt/ -v :/workspace/ --entrypoint bash sophgo/tpuc_dev:latest +docker exec -it TPUPerf bash +``` + +## 2. Environment Initialization + +```bash +pip3 install tpu_mlir +apt install unzip +pip3 install dfss +python3 -m dfss --url=open@sophgo.com:/sophon-demo/Stable_diffusion_3/BM1690/sophon-sail2.zip +unzip sophon-sail2.zip +# 依照sail2目录下的README，在当前环境编译出whl并安装 +``` + +## 3. Run ByteMLPerf for TPU backend + +```bash +python3 launch.py --task yolov5-onnx-fp32 --hardware_type TPU +python3 launch.py --task resnet50-torch-fp32 --hardware_type TPU +``` + +# Notes +> Support FP32 and INT8 quantization for resnet50-torch-fp32 now, . \ No newline at end of file diff --git a/byte_infer_perf/general_perf/backends/TPU/compile_backend_tpu.py b/byte_infer_perf/general_perf/backends/TPU/compile_backend_tpu.py new file mode 100755 index 000000000..261203068 --- /dev/null +++ b/byte_infer_perf/general_perf/backends/TPU/compile_backend_tpu.py @@ -0,0 +1,178 @@ +# Copyright 2023 Graphcore Ltd. +# +# 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 copy +import json +import logging +import os +import subprocess +from pathlib import Path +from typing import Any, Dict +import tpu_mlir +import shutil + +from general_perf.backends import compile_backend + +log = logging.getLogger("CompileBackendTPU") + +class CompileBackendTPU(compile_backend.CompileBackend): + def __init__(self): + super().__init__() + self.hardware_type = "TPU" + self.need_reload = False + self.need_quant = False + self.current_dir = os.path.split(os.path.abspath(__file__))[0] + self.model_config = None + self.precision = "fp32" + self.model_precision = "F32" + self.mean = "0.0,0.0,0.0" + self.scale = "1.0,1.0,1.0" + self.pixel_format = "rgb" + self.input_num = 200 + + def version(self) -> str: + """ + Return compile backend version details + """ + return tpu_mlir.distribution + + def pre_optimize(self, configs: Dict[str, Any]): + """Model pre-optimization interface. + + Requirements: Model pre-optimization + cannot change the model format. Torch model export to ONNX is allowed. + """ + + return configs + + def compile(self, configs: Dict[str, Any], dataloader=None) -> Dict[str, Any]: + if not self.model_config: + self.model_config = configs + + self.model_info = configs["model_info"] + self.interact_info = configs["interact_info"] + self.model_path = self.model_info["model_path"] + self.input_shapes = self.model_info["input_shape"][self.model_info["inputs"]] + self.input_shapes_str = ','.join(str(num) for num in self.input_shapes) + self.model_name = self.model_info["model"] + if("model_precision" in self.interact_info.keys()): + self.model_precision = self.interact_info["model_precision"] + self.mean = self.interact_info["mean"] + self.scale = self.interact_info["scale"] + self.pixel_format = self.interact_info["pixel_format"] + self.input_num = self.interact_info["input_num"] + + self.precision=self.model_precision.upper() + gen_mlir_commands = f'model_transform \ + --model_name {self.model_name} \ + --model_def ../../{self.model_path} \ + --mean {self.mean} \ + --scale {self.scale} \ + --pixel_format {self.pixel_format} \ + --input_shapes [[{self.input_shapes_str}]] \ + --mlir {self.model_name}.mlir' + gen_mlir_logs = './model_transform.log' + + current_dir = os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) + origin_dir = os.getcwd() + self.compile_dir_name = current_dir + '/compiled_models/' + if os.path.exists(self.compile_dir_name): + shutil.rmtree(self.compile_dir_name) + os.mkdir(self.compile_dir_name) + os.chdir(self.compile_dir_name) + with open(gen_mlir_logs, 'w') as logfile: + subprocess.call(gen_mlir_commands, stdout=logfile, stderr=subprocess.STDOUT, shell=True) + if(self.precision == "INT8"): + self.dataset_path = current_dir+"/datasets/"+self.model_info["dataset_name"]+"/"+self.interact_info["dataset_path"] + + run_calibration_commands = f'run_calibration {self.model_name}.mlir \ + --dataset {self.dataset_path} \ + --input_num {self.input_num} \ + -o {self.model_name}_cali_table' + + run_calibration_logs = './run_calibration.log' + + + with open(run_calibration_logs , 'w') as logfile: + subprocess.call(run_calibration_commands, stdout=logfile, stderr=subprocess.STDOUT, shell=True) + + deploy_commands = f'model_deploy \ + --mlir {self.model_name}.mlir \ + --quantize {self.model_precision} \ + --chip bm1690 \ + --calibration_table {self.model_name}_cali_table \ + --model {self.model_name}.bmodel' + else: + deploy_commands = f'model_deploy \ + --mlir {self.model_name}.mlir \ + --quantize {self.model_precision} \ + --chip bm1690 \ + --model {self.model_name}.bmodel' + deploy_commands_logs = './model_deploy.log' + + with open(deploy_commands_logs, 'w') as logfile: + subprocess.call(deploy_commands, stdout=logfile, stderr=subprocess.STDOUT, shell=True) + + os.chdir(origin_dir) + + result = { + "model": self.model_name, + "framework": configs["model_info"]["framework"], + "compile_precision": self.precision, + "input_type": configs["model_info"]["input_type"].split(","), + "max_batch_size": configs["model_info"]["max_batch_size"], + "compile_status": "success", + "optimizations": {}, + "instance_count": 1, + "device_count": 1, + "sg_percent": 100, + "segments": [ + { + "sg_idx": 0, + "is_fallback": False, + "input_tensor_map": configs["model_info"]["input_shape"], + "output_tensor_map": configs["model_info"]["outputs"], + "compiled_model": [ + { + "compiled_bs": 1, + "compiled_obj": configs["model_info"]["model_path"], + }, + ], + }, + ], + "interact_info": self.model_config, + } + + return result + + def get_interact_profile(self, config: Dict[str, Any]): + """Collect information for core engine to let user interactively fill in configurations.""" + # load the interact_info by model name + model_profile = [] + + interact_info_file = os.path.join( + self.current_dir, "interact_infos", config["model_info"]["model"] + ".json" + ) + file_path = os.path.join(self.current_dir, self.hardware_type + ".json") + + with open(interact_info_file, "r") as f: + interact_info = json.load(f) + + + + return interact_info + + def get_best_batch_size(self) -> compile_backend.List[int] | None: + return None \ No newline at end of file diff --git a/byte_infer_perf/general_perf/backends/TPU/interact_infos/resnet50-torch-fp32.json b/byte_infer_perf/general_perf/backends/TPU/interact_infos/resnet50-torch-fp32.json new file mode 100644 index 000000000..90dec8edb --- /dev/null +++ b/byte_infer_perf/general_perf/backends/TPU/interact_infos/resnet50-torch-fp32.json @@ -0,0 +1,52 @@ +[ + + { + "name": "model_precision", + "default": "FP32", + "depends": null, + "note" : "选择数据格式", + "type": "string", + "options":["F32", "INT8", "F16"], + "dialog_type": "Radiolist Dialog" + }, + { + "name": "dataset_path", + "default": "ILSVRC2012_img_val", + "depends": null, + "note" : "量化数据集路径", + "type": "string", + "dialog_type": "Input Dialog" + }, + { + "name": "mean", + "default": "103.53,116.28,123.67", + "depends": null, + "note" : "前处理均值", + "type": "string", + "dialog_type": "Input Dialog" + }, + { + "name": "scale", + "default": "0.01742919,0.017507,0.01712475", + "depends": null, + "note" : "前处理比例", + "type": "string", + "dialog_type": "Input Dialog" + }, + { + "name": "pixel_format", + "default": "rgb", + "depends": null, + "note" : "前处理像素格式", + "type": "string", + "dialog_type": "Input Dialog" + }, + { + "name": "input_num", + "default": "200", + "depends": null, + "note" : "量化数据集数量", + "type": "int", + "dialog_type": "Input Dialog" + } +] diff --git a/byte_infer_perf/general_perf/backends/TPU/runtime_backend_tpu.py b/byte_infer_perf/general_perf/backends/TPU/runtime_backend_tpu.py new file mode 100755 index 000000000..0b19c76c6 --- /dev/null +++ b/byte_infer_perf/general_perf/backends/TPU/runtime_backend_tpu.py @@ -0,0 +1,143 @@ +# Copyright 2023 ByteDance and/or its affiliates. +# +# 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 os +import time + +import numpy as np +import sophon.sail as sail +from general_perf.backends import runtime_backend +import multiprocessing +log = logging.getLogger("RuntimeBackendTPU") + +class RuntimeBackendTPU(runtime_backend.RuntimeBackend): + def __init__(self): + super().__init__() + self.hardware_type = "TPU" + self.need_reload = False + self.model_runtimes = [] + self.configs = None + self.pack_config = None + self.batch_size = -1 + self.pack_bs = -1 + self.packrunner = False + self.engine = None + self.runner_name = "SAIL" + self.compiled_dir = ( + os.path.dirname(os.path.dirname(os.path.dirname(os.path.abspath(__file__)))) + "/compiled_models/" + ) + self.precision = "fp32" + self.max_time=multiprocessing.Value('d', -float('inf')) + self.min_time=multiprocessing.Value('d', float('inf')) + self.lock = multiprocessing.Lock() + + def version(self) -> str: + return sail.__version__ + + def load(self, batch_size) -> None: + log.warning("TPU Backend only support static batch_size now.") + self.bmodel_path = self.compiled_dir + self.configs["model"] + ".bmodel" + # self.input_key = self.configs["input_shape"][self.configs["inputs"]] + self.dev_id = 1 + self.net = sail.nn.Engine(self.bmodel_path, self.dev_id) + self.stream = sail.nn.Engine(self.bmodel_path, self.dev_id) + self.net_name = self.net.get_net_names()[0] + self.input_name = self.net.get_input_names(self.net_name)[0] + self.output_names = self.net.get_output_names(self.net_name) + self.input_shape = self.net.get_input_shapes(self.net_name, 0)[0] + self.output_shapes = self.net.get_output_shapes(self.net_name, 0) + self.batch_size = self.input_shape[0] + self.net_h = self.input_shape[2] + self.net_w = self.input_shape[3] + + def get_loaded_batch_size(self) -> int: + return self.batch_size + + def predict(self, data): + if isinstance(data, dict): + input_data = {0: next(iter(data.values()))} + else: + input_data = {0: data} + + output_arrays = [np.ndarray(shape=(self.output_shapes[i]), dtype=np.float32) for i in range(len(self.output_shapes))] + outputs = {i:array for i, array in enumerate(output_arrays)} + ret = self.net.process(input_data, outputs, self.stream, self.net_name) + return outputs + + def single_chip_test(self, dev_id, iter, thread_id): + net = sail.nn.Engine(self.bmodel_path, dev_id) + stream = sail.nn.Stream(dev_id) + net_name = net.get_net_names()[0] + input_shape = net.get_input_shapes(net_name, 0)[0] + output_shapes = net.get_output_shapes(net_name, 0) + + input = np.random.rand(*input_shape).astype(np.float32) + input_tensor = sail.nn.Tensor(input, sail.DataType.TPU_FLOAT32, dev_id) + input_data = {0: input_tensor} + + output_arrays = [sail.nn.Tensor(output_shapes[i], sail.DataType.TPU_FLOAT32, dev_id) for i in range(len(output_shapes))] + outputs = {i:array for i, array in enumerate(output_arrays)} + + start_time=time.time() + for i in range(iter): + net.process_async(input_data, outputs, stream, net_name) + stream.sync() + end_time=time.time() + + with self.lock: + self.min_time.value=min(self.min_time.value,start_time) + self.max_time.value=max(self.max_time.value,end_time) + + + def _run_benchmark(self, bs, iter): + chip_num, core_num, start_chip =2, 1, 0 + thread_list = [] + for chip_id in range(chip_num): + for core_id in range(core_num): + thread_list.append(multiprocessing.Process(target=self.single_chip_test, args=(chip_id+start_chip, iter, chip_id*core_num+core_id))) + + logging.info("Predict running...") + for thread in thread_list: + thread.start() + + for thread in thread_list: + thread.join() + logging.info("Predict finished") + + total_time = self.max_time.value - self.min_time.value + + frame_num = chip_num * core_num * iter + qps = frame_num / total_time + avg_latency = total_time / frame_num + tail_latency = -1 + print(f'chip_num = {chip_num}, core_num = {core_num}, frame_num = {frame_num}, qps = {qps}') + + return qps, avg_latency, tail_latency + + def benchmark(self, dataloader): + report = {} + report["BS"] = self.batch_size + interact_info = self.configs.get("interact_info", {}) + iterations = self.workload["iterations"] + + qps, avg_latency, tail_latency = self._run_benchmark( + self.batch_size, iterations*100 + ) + + report["QPS"] = int(qps) + report["AVG Latency"] = avg_latency + report["P99 Latency"] = tail_latency + + return report \ No newline at end of file diff --git a/byte_infer_perf/general_perf/backends/TPU/tpu.json b/byte_infer_perf/general_perf/backends/TPU/tpu.json new file mode 100755 index 000000000..e69de29bb diff --git a/byte_infer_perf/general_perf/backends/runtime_backend.py b/byte_infer_perf/general_perf/backends/runtime_backend.py old mode 100644 new mode 100755 index db856b0bd..8bdd4e6db --- a/byte_infer_perf/general_perf/backends/runtime_backend.py +++ b/byte_infer_perf/general_perf/backends/runtime_backend.py @@ -29,7 +29,7 @@ def version(self) -> str: def load(self, batch_size) -> str: """ - Return runtime backend version details + Load model for current backend """ raise NotImplementedError("RuntimeBackend:load") diff --git a/byte_infer_perf/general_perf/core/perf_engine.py b/byte_infer_perf/general_perf/core/perf_engine.py old mode 100644 new mode 100755 index 231211e2c..fb1d761ef --- a/byte_infer_perf/general_perf/core/perf_engine.py +++ b/byte_infer_perf/general_perf/core/perf_engine.py @@ -285,9 +285,14 @@ def get_model_info(self, model_name: str) -> Dict[str, Any]: return model_info def get_cpu_name(self): - command = "lscpu | grep 'Model name' | awk -F: '{print $2}'" - cpu_name = subprocess.check_output(command, shell=True) - return cpu_name.decode().strip() + try: + lscpu_output = subprocess.check_output(["lscpu"], text=True) + for line in lscpu_output.split('\n'): + if 'Model name' in line: + return line.split(':')[1].strip() + except subprocess.CalledProcessError as e: + print(f"Command failed: {e}") + return None def check_interact_info( self, pre_compile_config: Dict[str, Dict]) -> Dict[str, Any]: diff --git a/byte_infer_perf/general_perf/launch.py b/byte_infer_perf/general_perf/launch.py old mode 100644 new mode 100755 index 51cf30609..6b2afd6d4 --- a/byte_infer_perf/general_perf/launch.py +++ b/byte_infer_perf/general_perf/launch.py @@ -75,12 +75,13 @@ def main(): subprocess.call([ 'bash', 'general_perf/backends/CPU/calculate_cpu_diff.sh', workload['model'], - str(workload['batch_sizes'][0]) + str(workload['batch_sizes'][0]), + str(parsed_args.hardware_type) ]) cmd = f'python3 general_perf/core/perf_engine.py --hardware_type {parsed_args.hardware_type} --task {parsed_args.task}' if parsed_args.compile_only: - cmd += '--compile_only' + cmd += ' --compile_only' exit_code = subprocess.call(cmd, shell=True) sys.exit(exit_code) diff --git a/byte_infer_perf/llm_perf/backends/TPU/model_impl/__init__.py b/byte_infer_perf/llm_perf/backends/TPU/model_impl/__init__.py new file mode 100644 index 000000000..d0f1edaca --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/model_impl/__init__.py @@ -0,0 +1,21 @@ +## __all__ is a dict: +## key is model_name in `model_zoo/chatglm-xx.json` +## value is vendor specify model impl +# __all__ = { +# "chatglm" : ChatGLMForConditionalGeneration, +# "chatglm2" : ChatGLM2ForConditionalGeneration +# } + +from typing import Dict, Tuple, Any + +import torch +import torch.nn as nn +import torch_tpu + +from .tpu_llama import TPULlama + +from llm_perf.utils.logger import logger + +__all__ = { + "llama3.1": TPULlama, +} \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/backends/TPU/model_impl/modeling_llama3.py b/byte_infer_perf/llm_perf/backends/TPU/model_impl/modeling_llama3.py new file mode 100755 index 000000000..e77f8a659 --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/model_impl/modeling_llama3.py @@ -0,0 +1,1445 @@ +# coding=utf-8 +# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. +# +# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX +# and OPT implementations in this library. It has been modified from its +# original forms to accommodate minor architectural differences compared +# to GPT-NeoX and OPT used by the Meta AI team that trained the model. +# +# 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 math +from typing import List, Optional, Tuple, Union + +import torch +import torch_tpu +import torch.utils.checkpoint +from torch import nn + +from transformers.activations import ACT2FN +from transformers.cache_utils import Cache, DynamicCache, StaticCache +from transformers.generation import GenerationMixin +from transformers.modeling_attn_mask_utils import AttentionMaskConverter +from transformers.modeling_flash_attention_utils import FlashAttentionKwargs, _flash_attention_forward +from transformers.modeling_outputs import ( + BaseModelOutputWithPast, + CausalLMOutputWithPast, + QuestionAnsweringModelOutput, + SequenceClassifierOutputWithPast, + TokenClassifierOutput, +) +from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS +from transformers.modeling_utils import PreTrainedModel +from transformers.processing_utils import Unpack +from transformers.pytorch_utils import ALL_LAYERNORM_LAYERS +from transformers.utils import ( + LossKwargs, + add_code_sample_docstrings, + add_start_docstrings, + add_start_docstrings_to_model_forward, + is_flash_attn_greater_or_equal_2_10, + logging, + replace_return_docstrings, +) +from transformers.models.llama.configuration_llama import LlamaConfig + + +logger = logging.get_logger(__name__) + +_CHECKPOINT_FOR_DOC = "meta-llama/Llama-2-7b-hf" +_CONFIG_FOR_DOC = "LlamaConfig" + + +class LlamaRMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + LlamaRMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + def extra_repr(self): + return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}" + + +ALL_LAYERNORM_LAYERS.append(LlamaRMSNorm) + + +class LlamaRotaryEmbedding(nn.Module): + def __init__( + self, + dim=None, + max_position_embeddings=2048, + base=10000, + device=None, + scaling_factor=1.0, + rope_type="default", + config: Optional[LlamaConfig] = None, + ): + super().__init__() + # TODO (joao): remove the `if` below, only used for BC + self.rope_kwargs = {} + if config is None: + logger.warning_once( + "`LlamaRotaryEmbedding` can now be fully parameterized by passing the model config through the " + "`config` argument. All other arguments will be removed in v4.46" + ) + self.rope_kwargs = { + "rope_type": rope_type, + "factor": scaling_factor, + "dim": dim, + "base": base, + "max_position_embeddings": max_position_embeddings, + } + self.rope_type = rope_type + self.max_seq_len_cached = max_position_embeddings + self.original_max_seq_len = max_position_embeddings + else: + # BC: "rope_type" was originally "type" + if 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" + self.max_seq_len_cached = config.max_position_embeddings + self.original_max_seq_len = config.max_position_embeddings + + self.config = config + self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type] + + inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs) + self.register_buffer("inv_freq", inv_freq, persistent=False) + self.original_inv_freq = self.inv_freq + + def _dynamic_frequency_update(self, position_ids, device): + """ + 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) + """ + 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.rope_kwargs + ) + self.register_buffer("inv_freq", inv_freq, persistent=False) # TODO joao: may break with compilation + 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: # reset + self.register_buffer("inv_freq", self.original_inv_freq, persistent=False) + self.max_seq_len_cached = self.original_max_seq_len + + @torch.no_grad() + def forward(self, x, position_ids): + if "dynamic" in self.rope_type: + self._dynamic_frequency_update(position_ids, device=x.device) + + # Core RoPE block + inv_freq_expanded = self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1) + position_ids_expanded = position_ids[:, None, :].float() + # Force float32 (see https://github.com/huggingface/transformers/pull/29285) + device_type = x.device.type + device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu" + with torch.autocast(device_type=device_type, enabled=False): + freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2) + emb = torch.cat((freqs, freqs), dim=-1) + cos = emb.cos() + sin = emb.sin() + + # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention + cos = cos * self.attention_scaling + sin = sin * self.attention_scaling + + return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype) + + +def rotate_half(x): + """Rotates half the hidden dims of the input.""" + x1 = x[..., : x.shape[-1] // 2] + x2 = x[..., x.shape[-1] // 2 :] + return torch.cat((-x2, x1), dim=-1) + + +def apply_rotary_pos_emb(q, k, cos, sin, position_ids=None, unsqueeze_dim=1): + """Applies Rotary Position Embedding to the query and key tensors. + + Args: + q (`torch.Tensor`): The query tensor. + k (`torch.Tensor`): The key tensor. + cos (`torch.Tensor`): The cosine part of the rotary embedding. + sin (`torch.Tensor`): The sine part of the rotary embedding. + position_ids (`torch.Tensor`, *optional*): + Deprecated and unused. + unsqueeze_dim (`int`, *optional*, defaults to 1): + The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and + sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note + that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and + k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes + cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have + the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2. + Returns: + `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding. + """ + cos = cos.unsqueeze(unsqueeze_dim) + sin = sin.unsqueeze(unsqueeze_dim) + q_embed = (q * cos) + (rotate_half(q) * sin) + k_embed = (k * cos) + (rotate_half(k) * sin) + return q_embed, k_embed + + +class LlamaMLP(nn.Module): + def __init__(self, config): + super().__init__() + self.config = config + self.hidden_size = config.hidden_size + self.intermediate_size = config.intermediate_size + self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) + self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=config.mlp_bias) + self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=config.mlp_bias) + self.act_fn = ACT2FN[config.hidden_act] + + def forward(self, x): + down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x)) + return down_proj + + +def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor: + """ + This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch, + num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim) + """ + batch, num_key_value_heads, slen, head_dim = hidden_states.shape # [bsz, 8, q_len, 128] + if n_rep == 1: + return hidden_states + device = hidden_states.device + if device.type == 'tpu': + hidden_states = hidden_states[:, :, None, :, :].to('cpu').expand(batch, num_key_value_heads, n_rep, slen, head_dim).to('tpu') + else: + hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim) + return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim) + +class LlamaAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__(self, config: LlamaConfig, layer_idx: Optional[int] = None): + super().__init__() + self.config = config + self.layer_idx = layer_idx + if layer_idx is None: + logger.warning_once( + f"Instantiating {self.__class__.__name__} without passing a `layer_idx` is not recommended and will " + "lead to errors during the forward call if caching is used. Please make sure to provide a `layer_idx` " + "when creating this class." + ) + + self.attention_dropout = config.attention_dropout # 0.0 + self.hidden_size = config.hidden_size # 4096 + self.num_heads = config.num_attention_heads # 32 + self.head_dim = getattr(config, "head_dim", self.hidden_size // self.num_heads) # 128 + self.num_key_value_heads = config.num_key_value_heads # 8 + self.num_key_value_groups = self.num_heads // self.num_key_value_heads # 4 + self.max_position_embeddings = config.max_position_embeddings # 131072 + self.rope_theta = config.rope_theta # 5e5 + self.is_causal = True + + self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=config.attention_bias) # [4096, 4096] + self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias) # [4096, 1024] + self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=config.attention_bias) # [4096, 1024] + self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=config.attention_bias) # [4096, 4096] + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) # [bsz, q_len, 4096] + key_states = self.k_proj(hidden_states) # [bsz, q_len, 1024] + value_states = self.v_proj(hidden_states) # [bsz, q_len, 1024] + + # Multi-Head Attention + # use -1 to infer num_heads and num_key_value_heads as they may vary if tensor parallel is used + query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) # [bsz, 32, q_len, 128] + key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) # [bsz, 8, q_len, 128] + value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) # [bsz, 8, q_len, 128] + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + # kv cache，这里是把新算出来的kv和本层之前的kv在qlen维度做concat + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + key_states = repeat_kv(key_states, self.num_key_value_groups) # 在第二个维度上重复4次 + value_states = repeat_kv(value_states, self.num_key_value_groups) + # q: [bsz, 32, q_len, 128], k.transpose: [bsz, 32, 128, q_len] + # output: [bsz, 32, q_len, q_len] + attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim) + + if attention_mask is not None: # no matter the length, we just slice it + causal_mask = attention_mask[:, :, :, : key_states.shape[-2]] + attn_weights = attn_weights + causal_mask + + # upcast attention to fp32 + attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype) + attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training) + attn_output = torch.matmul(attn_weights, value_states) + # [bsz, 32, q_len, 128] + if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim): + raise ValueError( + f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is" + f" {attn_output.size()}" + ) + # [bsz, q_len, num_heads, head_dim] + attn_output = attn_output.transpose(1, 2).contiguous() + # [bsz, q_len, hidden_size] + attn_output = attn_output.reshape(bsz, q_len, -1) + # [bsz, q_len, hidden_size] + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + +class LlamaFlashAttention2(LlamaAttention): + """ + Llama flash attention module. This module inherits from `LlamaAttention` as the weights of the module stays + untouched. The only required change would be on the forward pass where it needs to correctly call the public API of + flash attention and deal with padding tokens in case the input contains any of them. + """ + + def __init__(self, *args, **kwargs): + super().__init__(*args, **kwargs) + + # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1. + # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignement, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0. + # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left). + self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10() + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.LongTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs: Unpack[FlashAttentionKwargs], + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + if isinstance(past_key_value, StaticCache): + raise ValueError( + "`static` cache implementation is not compatible with `attn_implementation==flash_attention_2` " + "make sure to use `sdpa` in the mean time, and open an issue at https://github.com/huggingface/transformers" + ) + + output_attentions = False + + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + # Flash attention requires the input to have the shape + # batch_size x seq_length x head_dim x hidden_dim + # therefore we just need to keep the original shape + query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache + # to be able to avoid many of these transpose/reshape/view. + query_states = query_states.transpose(1, 2) + key_states = key_states.transpose(1, 2) + value_states = value_states.transpose(1, 2) + + dropout_rate = self.attention_dropout if self.training else 0.0 + + # In PEFT, usually we cast the layer norms in float32 for training stability reasons + # therefore the input hidden states gets silently casted in float32. Hence, we need + # cast them back in the correct dtype just to be sure everything works as expected. + # This might slowdown training & inference so it is recommended to not cast the LayerNorms + # in fp32. (LlamaRMSNorm handles it correctly) + + input_dtype = query_states.dtype + if input_dtype == torch.float32: + if torch.is_autocast_enabled(): + target_dtype = torch.get_autocast_gpu_dtype() + # Handle the case where the model is quantized + elif hasattr(self.config, "_pre_quantization_dtype"): + target_dtype = self.config._pre_quantization_dtype + else: + target_dtype = self.q_proj.weight.dtype + + logger.warning_once( + f"The input hidden states seems to be silently casted in float32, this might be related to" + f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in" + f" {target_dtype}." + ) + + query_states = query_states.to(target_dtype) + key_states = key_states.to(target_dtype) + value_states = value_states.to(target_dtype) + + attn_output = _flash_attention_forward( + query_states, + key_states, + value_states, + attention_mask, + q_len, + position_ids=position_ids, + dropout=dropout_rate, + sliding_window=getattr(self, "sliding_window", None), + use_top_left_mask=self._flash_attn_uses_top_left_mask, + is_causal=self.is_causal, + **kwargs, + ) + + attn_output = attn_output.reshape(bsz, q_len, -1).contiguous() + attn_output = self.o_proj(attn_output) + + if not output_attentions: + attn_weights = None + + return attn_output, attn_weights, past_key_value + + +class LlamaSdpaAttention(LlamaAttention): + """ + Llama attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from + `LlamaAttention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to + SDPA API. + """ + + # Adapted from LlamaAttention.forward + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: bool = False, + use_cache: bool = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs, + ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]: + if output_attentions: + # TODO: Improve this warning with e.g. `model.config.attn_implementation = "manual"` once this is implemented. + logger.warning_once( + "LlamaModel is using LlamaSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, " + 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.' + ) + return super().forward( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + ) + + bsz, q_len, _ = hidden_states.size() + + query_states = self.q_proj(hidden_states) + key_states = self.k_proj(hidden_states) + value_states = self.v_proj(hidden_states) + + # use -1 to infer num_heads and num_key_value_heads as they may vary if tensor parallel is used + query_states = query_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) + key_states = key_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) + value_states = value_states.view(bsz, q_len, -1, self.head_dim).transpose(1, 2) + + cos, sin = position_embeddings + query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin) + + if past_key_value is not None: + # sin and cos are specific to RoPE models; cache_position needed for the static cache + cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position} + key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs) + + key_states = repeat_kv(key_states, self.num_key_value_groups) + value_states = repeat_kv(value_states, self.num_key_value_groups) + + causal_mask = attention_mask + if attention_mask is not None: + causal_mask = causal_mask[:, :, :, : key_states.shape[-2]] + + # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask, + # Reference: https://github.com/pytorch/pytorch/issues/112577. + if query_states.device.type == "cuda" and causal_mask is not None: + query_states = query_states.contiguous() + key_states = key_states.contiguous() + value_states = value_states.contiguous() + + # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment + # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling. + is_causal = True if causal_mask is None and q_len > 1 else False + + cur_device = query_states.device + + attn_output = torch.nn.functional.scaled_dot_product_attention( + query_states.to('cpu'), + key_states.to('cpu'), + value_states.to('cpu'), + attn_mask=causal_mask.to('cpu'), + dropout_p=self.attention_dropout if self.training else 0.0, + is_causal=is_causal, + ) + + # attn_output = torch.nn.functional.scaled_dot_product_attention( + # query_states, + # key_states, + # value_states, + # attn_mask=causal_mask, + # dropout_p=self.attention_dropout if self.training else 0.0, + # is_causal=is_causal, + # ) + + attn_output = attn_output.transpose(1, 2).contiguous().to(cur_device) + attn_output = attn_output.view(bsz, q_len, -1) + + attn_output = self.o_proj(attn_output) + + return attn_output, None, past_key_value + + +LLAMA_ATTENTION_CLASSES = { + "eager": LlamaAttention, + "flash_attention_2": LlamaFlashAttention2, + "sdpa": LlamaSdpaAttention, +} + + +class LlamaDecoderLayer(nn.Module): + def __init__(self, config: LlamaConfig, layer_idx: int): + super().__init__() + self.hidden_size = config.hidden_size + + self.self_attn = LLAMA_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx) + + self.mlp = LlamaMLP(config) + self.input_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + hidden_states: torch.Tensor, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_value: Optional[Cache] = None, + output_attentions: Optional[bool] = False, + use_cache: Optional[bool] = False, + cache_position: Optional[torch.LongTensor] = None, + position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None, # necessary, but kept here for BC + **kwargs, + ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]: + """ + Args: + hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)` + attention_mask (`torch.FloatTensor`, *optional*): + attention mask of size `(batch_size, sequence_length)` if flash attention is used or `(batch_size, 1, + query_sequence_length, key_sequence_length)` if default attention is used. + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under + returned tensors for more detail. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding + (see `past_key_values`). + past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence + position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*): + Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`, + with `head_dim` being the embedding dimension of each attention head. + kwargs (`dict`, *optional*): + Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code + into the model + """ + residual = hidden_states + + hidden_states = self.input_layernorm(hidden_states) + + # Self Attention + hidden_states, self_attn_weights, present_key_value = self.self_attn( + hidden_states=hidden_states, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_value=past_key_value, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **kwargs, + ) + hidden_states = residual + hidden_states + + # Fully Connected + residual = hidden_states + hidden_states = self.post_attention_layernorm(hidden_states) + hidden_states = self.mlp(hidden_states) + hidden_states = residual + hidden_states + + outputs = (hidden_states,) + + if output_attentions: + outputs += (self_attn_weights,) + + if use_cache: + outputs += (present_key_value,) + + return outputs + + +LLAMA_START_DOCSTRING = r""" + This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the + library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads + etc.) + + This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass. + Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage + and behavior. + + Parameters: + config ([`LlamaConfig`]): + Model configuration class with all the parameters of the model. Initializing with a config file does not + load the weights associated with the model, only the configuration. Check out the + [`~PreTrainedModel.from_pretrained`] method to load the model weights. +""" + + +@add_start_docstrings( + "The bare LLaMA Model outputting raw hidden-states without any specific head on top.", + LLAMA_START_DOCSTRING, +) +class LlamaPreTrainedModel(PreTrainedModel): + config_class = LlamaConfig + base_model_prefix = "model" + supports_gradient_checkpointing = True + _no_split_modules = ["LlamaDecoderLayer"] + _skip_keys_device_placement = ["past_key_values"] + _supports_flash_attn_2 = True + _supports_sdpa = True + _supports_cache_class = True + _supports_quantized_cache = True + _supports_static_cache = True + + def _init_weights(self, module): + std = self.config.initializer_range + if isinstance(module, nn.Linear): + module.weight.data.normal_(mean=0.0, std=std) + if module.bias is not None: + module.bias.data.zero_() + elif isinstance(module, nn.Embedding): + module.weight.data.normal_(mean=0.0, std=std) + if module.padding_idx is not None: + module.weight.data[module.padding_idx].zero_() + + +LLAMA_INPUTS_DOCSTRING = r""" + Args: + input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`): + Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide + it. + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + [What are input IDs?](../glossary#input-ids) + attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*): + Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`: + + - 1 for tokens that are **not masked**, + - 0 for tokens that are **masked**. + + [What are attention masks?](../glossary#attention-mask) + + Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and + [`PreTrainedTokenizer.__call__`] for details. + + If `past_key_values` is used, optionally only the last `input_ids` have to be input (see + `past_key_values`). + + If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`] + and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more + information on the default strategy. + + - 1 indicates the head is **not masked**, + - 0 indicates the head is **masked**. + position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0, + config.n_positions - 1]`. + + [What are position IDs?](../glossary#position-ids) + past_key_values (`Cache` or `tuple(tuple(torch.FloatTensor))`, *optional*): + Pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention + blocks) that can be used to speed up sequential decoding. This typically consists in the `past_key_values` + returned by the model at a previous stage of decoding, when `use_cache=True` or `config.use_cache=True`. + + Two formats are allowed: + - a [`~cache_utils.Cache`] instance, see our + [kv cache guide](https://huggingface.co/docs/transformers/en/kv_cache); + - Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of + shape `(batch_size, num_heads, sequence_length, embed_size_per_head)`). This is also known as the legacy + cache format. + + The model will output the same cache format that is fed as input. If no `past_key_values` are passed, the + legacy cache format will be returned. + + If `past_key_values` are used, the user can optionally input only the last `input_ids` (those that don't + have their past key value states given to this model) of shape `(batch_size, 1)` instead of all `input_ids` + of shape `(batch_size, sequence_length)`. + inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*): + Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This + is useful if you want more control over how to convert `input_ids` indices into associated vectors than the + model's internal embedding lookup matrix. + use_cache (`bool`, *optional*): + If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see + `past_key_values`). + output_attentions (`bool`, *optional*): + Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned + tensors for more detail. + output_hidden_states (`bool`, *optional*): + Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for + more detail. + return_dict (`bool`, *optional*): + Whether or not to return a [`~utils.ModelOutput`] instead of a plain tuple. + cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*): + Indices depicting the position of the input sequence tokens in the sequence. Contrarily to `position_ids`, + this tensor is not affected by padding. It is used to update the cache in the correct position and to infer + the complete sequence length. +""" + + +@add_start_docstrings( + "The bare LLaMA Model outputting raw hidden-states without any specific head on top.", + LLAMA_START_DOCSTRING, +) +class LlamaModel(LlamaPreTrainedModel): + """ + Transformer decoder consisting of *config.num_hidden_layers* layers. Each layer is a [`LlamaDecoderLayer`] + + Args: + config: LlamaConfig + """ + + def __init__(self, config: LlamaConfig): + super().__init__(config) + self.padding_idx = config.pad_token_id + self.vocab_size = config.vocab_size + + self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx) + self.layers = nn.ModuleList( + [LlamaDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)] + ) + self.norm = LlamaRMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.rotary_emb = LlamaRotaryEmbedding(config=config) + + self.gradient_checkpointing = False + if getattr(config, "pretraining_tp", 1) != 1: + logger.warn("`pretraining_tp` is deprecated, please use `model.tensor_parallel` instead.") + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.embed_tokens + + def set_input_embeddings(self, value): + self.embed_tokens = value + + @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + **flash_attn_kwargs: Unpack[FlashAttentionKwargs], + ) -> Union[Tuple, BaseModelOutputWithPast]: + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + use_cache = use_cache if use_cache is not None else self.config.use_cache + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + if (input_ids is None) ^ (inputs_embeds is not None): + raise ValueError("You must specify exactly one of input_ids or inputs_embeds") + + if self.gradient_checkpointing and self.training and use_cache: + logger.warning_once( + "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`." + ) + use_cache = False + + if inputs_embeds is None: + inputs_embeds = self.embed_tokens(input_ids) + + # kept for BC (non `Cache` `past_key_values` inputs) + return_legacy_cache = False + if use_cache and not isinstance(past_key_values, Cache): + return_legacy_cache = True + if past_key_values is None: + past_key_values = DynamicCache() + else: + past_key_values = DynamicCache.from_legacy_cache(past_key_values) + logger.warning_once( + "We detected that you are passing `past_key_values` as a tuple of tuples. This is deprecated and " + "will be removed in v4.47. Please convert your cache or use an appropriate `Cache` class " + "(https://huggingface.co/docs/transformers/kv_cache#legacy-cache-format)" + ) + + if cache_position is None: + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + cache_position = torch.arange( + past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device + ) + if position_ids is None: + position_ids = cache_position.unsqueeze(0) + + causal_mask = self._update_causal_mask( + attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions + ) + hidden_states = inputs_embeds + + # create position embeddings to be shared across the decoder layers + position_embeddings = self.rotary_emb(hidden_states, position_ids) + + # decoder layers + all_hidden_states = () if output_hidden_states else None + all_self_attns = () if output_attentions else None + next_decoder_cache = None + + for decoder_layer in self.layers[: self.config.num_hidden_layers]: + if output_hidden_states: + all_hidden_states += (hidden_states,) + + if self.gradient_checkpointing and self.training: + layer_outputs = self._gradient_checkpointing_func( + decoder_layer.__call__, + hidden_states, + causal_mask, + position_ids, + past_key_values, + output_attentions, + use_cache, + cache_position, + position_embeddings, + ) + else: + layer_outputs = decoder_layer( + hidden_states, + attention_mask=causal_mask, + position_ids=position_ids, + past_key_value=past_key_values, + output_attentions=output_attentions, + use_cache=use_cache, + cache_position=cache_position, + position_embeddings=position_embeddings, + **flash_attn_kwargs, + ) + + hidden_states = layer_outputs[0] + + if use_cache: + next_decoder_cache = layer_outputs[2 if output_attentions else 1] + + if output_attentions: + all_self_attns += (layer_outputs[1],) + + hidden_states = self.norm(hidden_states) + + # add hidden states from the last decoder layer + if output_hidden_states: + all_hidden_states += (hidden_states,) + + next_cache = next_decoder_cache if use_cache else None + if return_legacy_cache: + next_cache = next_cache.to_legacy_cache() + + if not return_dict: + return tuple(v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None) + return BaseModelOutputWithPast( + last_hidden_state=hidden_states, + past_key_values=next_cache, + hidden_states=all_hidden_states, + attentions=all_self_attns, + ) + + def _update_causal_mask( + self, + attention_mask: torch.Tensor, + input_tensor: torch.Tensor, + cache_position: torch.Tensor, + past_key_values: Cache, + output_attentions: bool, + ): + if self.config._attn_implementation == "flash_attention_2": + if attention_mask is not None and 0.0 in attention_mask: + return attention_mask + return None + + # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in + # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail + # to infer the attention mask. + past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0 + using_static_cache = isinstance(past_key_values, StaticCache) + + # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward + # if False: + if self.config._attn_implementation == "sdpa" and not using_static_cache and not output_attentions: + if AttentionMaskConverter._ignore_causal_mask_sdpa( + attention_mask, + inputs_embeds=input_tensor, + past_key_values_length=past_seen_tokens, + is_training=self.training, + ): + return None + + + dtype, device = input_tensor.dtype, input_tensor.device + sequence_length = input_tensor.shape[1] + if using_static_cache: + target_length = past_key_values.get_max_cache_shape() + else: + target_length = ( + attention_mask.shape[-1] + if isinstance(attention_mask, torch.Tensor) + else past_seen_tokens + sequence_length + 1 + ) + + # In case the provided `attention` mask is 2D, we generate a causal mask here (4D). + causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position( + attention_mask, + sequence_length=sequence_length, + target_length=target_length, + dtype=dtype, + device=device, + cache_position=cache_position, + batch_size=input_tensor.shape[0], + ) + + if ( + self.config._attn_implementation == "sdpa" + and attention_mask is not None + and attention_mask.device.type == "cuda" + and not output_attentions + ): + # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when + # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path. + # Details: https://github.com/pytorch/pytorch/issues/110213 + min_dtype = torch.finfo(dtype).min + causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype) + + return causal_mask + + @staticmethod + def _prepare_4d_causal_attention_mask_with_cache_position( + attention_mask: torch.Tensor, + sequence_length: int, + target_length: int, + dtype: torch.dtype, + device: torch.device, + cache_position: torch.Tensor, + batch_size: int, + **kwargs, + ): + """ + Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape + `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing. + + Args: + attention_mask (`torch.Tensor`): + A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape + `(batch_size, 1, query_length, key_value_length)`. + sequence_length (`int`): + The sequence length being processed. + target_length (`int`): + The target length: when generating with static cache, the mask should be as long as the static cache, + to account for the 0 padding, the part of the cache that is not filled yet. + dtype (`torch.dtype`): + The dtype to use for the 4D attention mask. + device (`torch.device`): + The device to plcae the 4D attention mask on. + cache_position (`torch.Tensor`): + Indices depicting the position of the input sequence tokens in the sequence. + batch_size (`torch.Tensor`): + Batch size. + """ + if attention_mask is not None and attention_mask.dim() == 4: + # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing. + causal_mask = attention_mask + else: + + min_dtype = torch.finfo(dtype).min + causal_mask = torch.full( + (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device + ) + if sequence_length != 1: + causal_mask = torch.triu(causal_mask, diagonal=1) + + # origin_dtype = causal_mask.dtype + causal_mask *= torch.arange(target_length, device=device).to(torch.int32) > cache_position.reshape(-1, 1).to(torch.int32) + # causal_mask.to_(origin_dtype) + causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1) + if attention_mask is not None: + causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit + mask_length = attention_mask.shape[-1] + padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :] + padding_mask = padding_mask == 0 + causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill( + padding_mask, min_dtype + ) + + return causal_mask + + +class KwargsForCausalLM(FlashAttentionKwargs, LossKwargs): ... + + +class LlamaForCausalLM(LlamaPreTrainedModel, GenerationMixin): + _tied_weights_keys = ["lm_head.weight"] + _tp_plan = {"lm_head": "colwise_rep"} + + def __init__(self, config): + super().__init__(config) + self.model = LlamaModel(config) + self.vocab_size = config.vocab_size # 128256 + self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False) # 4096, 128256 + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + def get_output_embeddings(self): + return self.lm_head + + def set_output_embeddings(self, new_embeddings): + self.lm_head = new_embeddings + + def set_decoder(self, decoder): + self.model = decoder + + def get_decoder(self): + return self.model + + @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) + @replace_return_docstrings(output_type=CausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC) + def forward( + self, + input_ids: torch.LongTensor = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + cache_position: Optional[torch.LongTensor] = None, + num_logits_to_keep: int = 0, + **kwargs: Unpack[KwargsForCausalLM], + ) -> Union[Tuple, CausalLMOutputWithPast]: + r""" + Args: + labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*): + Labels for computing the masked language modeling loss. Indices should either be in `[0, ..., + config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored + (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`. + + num_logits_to_keep (`int`, *optional*): + Calculate logits for the last `num_logits_to_keep` tokens. If `0`, calculate logits for all + `input_ids` (special case). Only last token logits are needed for generation, and calculating them only for that + token can save memory, which becomes pretty significant for long sequences or large vocabulary size. + + Returns: + + Example: + + ```python + >>> from transformers import AutoTokenizer, LlamaForCausalLM + + >>> model = LlamaForCausalLM.from_pretrained("meta-llama/Llama-2-7b-hf") + >>> tokenizer = AutoTokenizer.from_pretrained("meta-llama/Llama-2-7b-hf") + + >>> prompt = "Hey, are you conscious? Can you talk to me?" + >>> inputs = tokenizer(prompt, return_tensors="pt") + + >>> # Generate + >>> generate_ids = model.generate(inputs.input_ids, max_length=30) + >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0] + "Hey, are you conscious? Can you talk to me?\nI'm not conscious, but I can talk to you." + ```""" + output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions + output_hidden_states = ( + output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states + ) + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn) + + + outputs = self.model( + input_ids=input_ids, + attention_mask=attention_mask, + position_ids=position_ids, # [[0, 1, 2, 3, 4]] + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, # None + use_cache=use_cache, # True + output_attentions=output_attentions, # False + output_hidden_states=output_hidden_states, # False + return_dict=return_dict, # True + cache_position=cache_position, # [0, 1, 2, 3, 4] + **kwargs, + ) + + hidden_states = outputs[0] + # Only compute necessary logits, and do not upcast them to float if we are not computing the loss + logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :]) + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, vocab_size=self.config.vocab_size, **kwargs) + + if not return_dict: + output = (logits,) + outputs[1:] + return (loss,) + output if loss is not None else output + + return CausalLMOutputWithPast( + loss=loss, + logits=logits, + past_key_values=outputs.past_key_values, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@add_start_docstrings( + """ + The LLaMa Model transformer with a sequence classification head on top (linear layer). + + [`LlamaForSequenceClassification`] uses the last token in order to do the classification, as other causal models + (e.g. GPT-2) do. + + Since it does classification on the last token, it requires to know the position of the last token. If a + `pad_token_id` is defined in the configuration, it finds the last token that is not a padding token in each row. If + no `pad_token_id` is defined, it simply takes the last value in each row of the batch. Since it cannot guess the + padding tokens when `inputs_embeds` are passed instead of `input_ids`, it does the same (take the last value in + each row of the batch). + """, + LLAMA_START_DOCSTRING, +) +class LlamaForSequenceClassification(LlamaPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = LlamaModel(config) + self.score = nn.Linear(config.hidden_size, self.num_labels, bias=False) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, SequenceClassifierOutputWithPast]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + transformer_outputs = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + hidden_states = transformer_outputs[0] + logits = self.score(hidden_states) + + if input_ids is not None: + batch_size = input_ids.shape[0] + else: + batch_size = inputs_embeds.shape[0] + + if self.config.pad_token_id is None and batch_size != 1: + raise ValueError("Cannot handle batch sizes > 1 if no padding token is defined.") + if self.config.pad_token_id is None: + sequence_lengths = -1 + else: + if input_ids is not None: + # if no pad token found, use modulo instead of reverse indexing for ONNX compatibility + sequence_lengths = torch.eq(input_ids, self.config.pad_token_id).int().argmax(-1) - 1 + sequence_lengths = sequence_lengths % input_ids.shape[-1] + sequence_lengths = sequence_lengths.to(logits.device) + else: + sequence_lengths = -1 + + pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] + + loss = None + if labels is not None: + loss = self.loss_function(logits=logits, labels=labels, pooled_logits=pooled_logits, config=self.config) + + if not return_dict: + output = (pooled_logits,) + transformer_outputs[1:] + return ((loss,) + output) if loss is not None else output + + return SequenceClassifierOutputWithPast( + loss=loss, + logits=pooled_logits, + past_key_values=transformer_outputs.past_key_values, + hidden_states=transformer_outputs.hidden_states, + attentions=transformer_outputs.attentions, + ) + + +@add_start_docstrings( + """ +The Llama Model transformer with a span classification head on top for extractive question-answering tasks like +SQuAD (a linear layer on top of the hidden-states output to compute `span start logits` and `span end logits`). + """, + LLAMA_START_DOCSTRING, +) +class LlamaForQuestionAnswering(LlamaPreTrainedModel): + base_model_prefix = "transformer" + + # Copied from transformers.models.bloom.modeling_bloom.BloomForQuestionAnswering.__init__ with Bloom->Llama + def __init__(self, config): + super().__init__(config) + self.transformer = LlamaModel(config) + self.qa_outputs = nn.Linear(config.hidden_size, 2) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.transformer.embed_tokens + + def set_input_embeddings(self, value): + self.transformer.embed_tokens = value + + @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.FloatTensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[Union[Cache, List[torch.FloatTensor]]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + start_positions: Optional[torch.LongTensor] = None, + end_positions: Optional[torch.LongTensor] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + **kwargs, + ) -> Union[Tuple, QuestionAnsweringModelOutput]: + r""" + start_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the start of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + end_positions (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for position (index) of the end of the labelled span for computing the token classification loss. + Positions are clamped to the length of the sequence (`sequence_length`). Position outside of the sequence + are not taken into account for computing the loss. + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + outputs = self.transformer( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + + sequence_output = outputs[0] + + logits = self.qa_outputs(sequence_output) + start_logits, end_logits = logits.split(1, dim=-1) + start_logits = start_logits.squeeze(-1).contiguous() + end_logits = end_logits.squeeze(-1).contiguous() + + loss = None + if start_positions is not None and end_positions is not None: + loss = self.loss_function(start_logits, end_logits, start_positions, end_positions, **kwargs) + + if not return_dict: + output = (start_logits, end_logits) + outputs[2:] + return ((loss,) + output) if loss is not None else output + + return QuestionAnsweringModelOutput( + loss=loss, + start_logits=start_logits, + end_logits=end_logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) + + +@add_start_docstrings( + """ + The Llama Model transformer with a token classification head on top (a linear layer on top of the hidden-states + output) e.g. for Named-Entity-Recognition (NER) tasks. + """, + LLAMA_START_DOCSTRING, +) +class LlamaForTokenClassification(LlamaPreTrainedModel): + def __init__(self, config): + super().__init__(config) + self.num_labels = config.num_labels + self.model = LlamaModel(config) + if getattr(config, "classifier_dropout", None) is not None: + classifier_dropout = config.classifier_dropout + elif getattr(config, "hidden_dropout", None) is not None: + classifier_dropout = config.hidden_dropout + else: + classifier_dropout = 0.1 + self.dropout = nn.Dropout(classifier_dropout) + self.score = nn.Linear(config.hidden_size, config.num_labels) + + # Initialize weights and apply final processing + self.post_init() + + def get_input_embeddings(self): + return self.model.embed_tokens + + def set_input_embeddings(self, value): + self.model.embed_tokens = value + + @add_start_docstrings_to_model_forward(LLAMA_INPUTS_DOCSTRING) + @add_code_sample_docstrings( + checkpoint=_CHECKPOINT_FOR_DOC, + output_type=TokenClassifierOutput, + config_class=_CONFIG_FOR_DOC, + ) + def forward( + self, + input_ids: Optional[torch.LongTensor] = None, + attention_mask: Optional[torch.Tensor] = None, + position_ids: Optional[torch.LongTensor] = None, + past_key_values: Optional[List[torch.FloatTensor]] = None, + inputs_embeds: Optional[torch.FloatTensor] = None, + labels: Optional[torch.LongTensor] = None, + use_cache: Optional[bool] = None, + output_attentions: Optional[bool] = None, + output_hidden_states: Optional[bool] = None, + return_dict: Optional[bool] = None, + ) -> Union[Tuple, TokenClassifierOutput]: + r""" + labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*): + Labels for computing the sequence classification/regression loss. Indices should be in `[0, ..., + config.num_labels - 1]`. If `config.num_labels == 1` a regression loss is computed (Mean-Square loss), If + `config.num_labels > 1` a classification loss is computed (Cross-Entropy). + """ + return_dict = return_dict if return_dict is not None else self.config.use_return_dict + + outputs = self.model( + input_ids, + attention_mask=attention_mask, + position_ids=position_ids, + past_key_values=past_key_values, + inputs_embeds=inputs_embeds, + use_cache=use_cache, + output_attentions=output_attentions, + output_hidden_states=output_hidden_states, + return_dict=return_dict, + ) + sequence_output = outputs[0] + sequence_output = self.dropout(sequence_output) + logits = self.score(sequence_output) + + loss = None + if labels is not None: + loss = self.loss_function(logits, labels, self.config) + + if not return_dict: + output = (logits,) + outputs[2:] + return ((loss,) + output) if loss is not None else output + + return TokenClassifierOutput( + loss=loss, + logits=logits, + hidden_states=outputs.hidden_states, + attentions=outputs.attentions, + ) diff --git a/byte_infer_perf/llm_perf/backends/TPU/model_impl/tpu_llama.py b/byte_infer_perf/llm_perf/backends/TPU/model_impl/tpu_llama.py new file mode 100755 index 000000000..806e36601 --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/model_impl/tpu_llama.py @@ -0,0 +1,168 @@ +import os +import pathlib + +import torch +import torch.nn as nn +import torch.distributed as dist + +from typing import Dict, Any +from llm_perf.utils.logger import logger +from llm_perf.utils.ps_utils import check_memory_usage +from llm_perf.utils.dist_utils import check_dist + +from accelerate import init_empty_weights + +from llm_perf.backends.TPU.tpu_ckpt_loader import TpuCkptLoader +from llm_perf.core.ckpt_loader import Llama_ModelLoader +from transformers import LlamaConfig +from transformers.cache_utils import DynamicCache, StaticCache +from .modeling_llama3 import LlamaForCausalLM + + +class TPULlamaLoader(TpuCkptLoader): + def __init__( + self, + model : LlamaForCausalLM, + model_config : LlamaConfig, + ckpt_path : str = "" + ): + mp_size = int(os.environ.get("WORLD_SIZE", "1")) + local_rank = int(os.environ.get("LOCAL_RANK", "0")) + + super().__init__("", model, mp_size, local_rank, ckpt_path) + self.model_config = model_config + + def parallel_loader(self): + self.state_dict = {} + + model_dir = pathlib.Path(self.ckpt_path).absolute() + if not model_dir.exists() or not model_dir.is_dir(): + if self.mp_rank == 0: + print(f"{model_dir} not exists or is not a directory") + return + + split_model_dir = model_dir.joinpath(f"TP{self.mp_size}") + if not split_model_dir.exists() and self.mp_size == 1: + split_model_dir = model_dir + real_model_dir = split_model_dir + elif not split_model_dir.exists() or not split_model_dir.is_dir(): + if self.mp_rank == 0: + print(f"{split_model_dir} not exists or is not a directory, please split model first.") + return + elif split_model_dir.exists(): + real_model_dir = split_model_dir / f"device_{self.mp_rank}" + + model_loader = Llama_ModelLoader(real_model_dir) + self.state_dict = model_loader.load_weight() + + def infusion_to_model(self): + self.model.model.embed_tokens.weight = self.to_parameter(self.state_dict["model.embed_tokens.weight"]) + for i in range(self.model_config.num_hidden_layers): + self.model.model.layers[i].input_layernorm.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.input_layernorm.weight"]) + + self.model.model.layers[i].self_attn.q_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.self_attn.q_proj.weight"]) + self.model.model.layers[i].self_attn.k_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.self_attn.k_proj.weight"]) + self.model.model.layers[i].self_attn.v_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.self_attn.v_proj.weight"]) + self.model.model.layers[i].self_attn.o_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.self_attn.o_proj.weight"]) + + self.model.model.layers[i].post_attention_layernorm.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.post_attention_layernorm.weight"]) + + self.model.model.layers[i].mlp.gate_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.mlp.gate_proj.weight"]) + self.model.model.layers[i].mlp.up_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.mlp.up_proj.weight"]) + self.model.model.layers[i].mlp.down_proj.weight = self.to_parameter(self.state_dict[f"model.layers.{i}.mlp.down_proj.weight"]) + + self.model.model.norm.weight = self.to_parameter(self.state_dict["model.norm.weight"]) + self.model.lm_head.weight = self.to_parameter(self.state_dict["lm_head.weight"]) + + +class TPULlama(nn.Module): + def __init__(self, xpu_cfg: Dict[str, Any]) -> None: + super().__init__() + + self.xpu_cfg = xpu_cfg + self.model_config = xpu_cfg["model_config"] + + self.model_name = self.model_config["model_name"] + self.model_path = self.model_config["model_path"] + self.model_network = self.model_config["network"] + + self.llama_config : LlamaConfig = LlamaConfig(**self.model_network) + # print(self.llama_config) + + # dist config + self.mp_size = int(os.environ.get("WORLD_SIZE", "1")) + self.local_rank = int(os.environ.get("LOCAL_RANK", "0")) + + self.transformer_model : LlamaForCausalLM = None + + + def init_inference(self): + torch.tpu.set_device(self.local_rank) + + if self.mp_size > 1: + logger.info(f"RANK: {self.local_rank} {self.mp_size} init_process_group...") + dist.init_process_group( + backend="sccl", + world_size=self.mp_size, + rank=self.local_rank + ) + check_dist() + + # check_memory_usage("Begin") + + with init_empty_weights(): + self.transformer_model = LlamaForCausalLM(self.llama_config).to(self.llama_config.torch_dtype).eval() + + # check_memory_usage("After build model") + + self.load_weight(self.model_path) + + # check_memory_usage("After load_weight") + + self.transformer_model.tpu() + + # check_memory_usage("After model to device") + + self.kv_cache = self.init_kvcache(self.llama_config.torch_dtype) + + if self.mp_size > 1: + dist.barrier() + + def finalize_inference(self): + if self.mp_size > 1 and dist.is_initialized(): + dist.destroy_process_group() + + def load_weight(self, ckpt_path): + p_loader = TPULlamaLoader(self.transformer_model, self.llama_config, ckpt_path) + p_loader.parallel_loader() + p_loader.infusion_to_model() + + def init_kvcache(self, dtype): + max_batch_size = self.xpu_cfg["max_batch_size"] + cur_device = self.transformer_model.device + # cache = DynamicCache(num_layers) + + cache = StaticCache(self.llama_config, + max_batch_size, + 4096, + torch.device('cpu'), # torch.zeros not support bf16 with TPU now + dtype, + max_batch_size).to(cur_device) + return cache + + + def forward(self, inputs : Dict[str, torch.Tensor]): + # inputs = inputs.to(torch.int32) + model_outputs = self.transformer_model.forward( + **inputs, + past_key_values=self.kv_cache + # past_key_values=None + ) + # context: [1, seq_len] --> [1, seq_len, vocab_size] or [1, 1, vocab_size] + # decode: [max_batch_size, 1] + logits = model_outputs.logits + + output_dict = { + "logits": logits + } + return output_dict \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/backends/TPU/setup.py b/byte_infer_perf/llm_perf/backends/TPU/setup.py new file mode 100644 index 000000000..84aa5154c --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/setup.py @@ -0,0 +1,61 @@ +import torch +import importlib +from typing import Any, Dict + +from llm_perf.core.scheduler import CoreScheduler +from llm_perf.backends.TPU.tpu_inferencer import TpuInferencer +from llm_perf.backends.TPU.tpu_sampler import TpuSampler +from llm_perf.backends.TPU.tpu_scheduler import TpuScheduler +from llm_perf.backends.TPU.tpu_mp_engine import TpuMpEngine +from llm_perf.utils.logger import logger + + + +def get_engine(xpu_cfg) -> TpuMpEngine: + # get model impl + hardware_type = xpu_cfg["hardware_type"] + model_config = xpu_cfg["model_config"] + model_name = model_config["model_name"] + + vendor_model_path = f"llm_perf/backends/{hardware_type}/model_impl" + vendor_model_impl = importlib.import_module( + ".", package=vendor_model_path.replace("/", ".") + ) + vendor_model = vendor_model_impl.__all__[model_name] + + mp_engine = TpuMpEngine( + world_size=xpu_cfg["tp_size"], + model_impl=vendor_model, + xpu_cfg=xpu_cfg + ) + + return mp_engine + + + +def setup_scheduler(xpu_cfg) -> CoreScheduler: + # get model impl + hardware_type = xpu_cfg["hardware_type"] + model_config = xpu_cfg["model_config"] + model_name = model_config["model_name"] + + vendor_model_path = f"llm_perf/backends/{hardware_type}/model_impl" + vendor_model_impl = importlib.import_module( + ".", package=vendor_model_path.replace("/", ".") + ) + vendor_model = vendor_model_impl.__all__[model_name] + + # create inferencer + inferencer = TpuInferencer(vendor_model, xpu_cfg) + + # create sampler + sampler = TpuSampler() + + # create scheduler + scheduler = TpuScheduler( + inferencer=inferencer, + sampler=sampler, + xpu_cfg=xpu_cfg + ) + + return scheduler \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/backends/TPU/tpu_ckpt_loader.py b/byte_infer_perf/llm_perf/backends/TPU/tpu_ckpt_loader.py new file mode 100644 index 000000000..dbdf8a8bc --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/tpu_ckpt_loader.py @@ -0,0 +1,50 @@ +import torch +import torch_tpu +import torch.distributed as dist + +from llm_perf.core.ckpt_loader import CoreCkptLoader + +class TpuCkptLoader(CoreCkptLoader): + def __init__( + self, + prefix, model, + mp_size=1, mp_rank=0, + ckpt_path: str="" + ): + super().__init__(prefix, model, mp_size, mp_rank, ckpt_path) + + def weight_to_device(self, weight: torch.Tensor, non_blocking=False): + if self.mp_rank == 0: + weight = weight.tpu(non_blocking=non_blocking) + else: + cur_device = torch.tpu.current_device() + weight = torch.emtpy_like(weight, device=f"tpu:{cur_device}") + return weight + + + def broadcast_weight(self, key, device='cpu', non_blocking=False): + if self.mp_rank != 0: + tensor_shape = self.state_dict[key]["shape"] + tensor_dtype = self.state_dict[key]["dtype"] + tensor = torch.empty(tensor_shape, dtype=tensor_dtype) + else: + tensor = self.state_dict[key].cpu() + tensor_tpu = self.weight_to_device(tensor, non_blocking=non_blocking) + dist.broadcast(tensor_tpu, src=0) + self.state_dict[key] = tensor_tpu + + def scatter_weight(self, key, dim, split_mode='default', outter=1, device='cpu', non_blocking=False): + self.broadcast_weight(key, non_blocking=non_blocking) + weight = self.state_dict[key] + + if split_mode == 'default': + weight_split = self.split(weight, dim) + elif split_mode == 'with_outter': + weight_split = self.with_outter_split(weight, dim, outter) + elif split_mode == 'split_outter': + weight_split = self.split(weight, dim, outter) + else: + assert False, f"unknown split mode {split_mode}" + + weight_split = [x.contiguous() for x in weight_split] + self.state_dict[key] = weight_split[self.mp_rank] \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/backends/TPU/tpu_inferencer.py b/byte_infer_perf/llm_perf/backends/TPU/tpu_inferencer.py new file mode 100644 index 000000000..7b30a71f4 --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/tpu_inferencer.py @@ -0,0 +1,129 @@ +from dataclasses import dataclass +from typing import Dict, Iterable, List + +from llm_perf.core.generation import GenerateRequest +from llm_perf.core.inferencer import CoreInferencer +from llm_perf.backends.TPU.tpu_mp_engine import TpuMpEngine +from llm_perf.utils.logger import logger + +class TpuInferencer(CoreInferencer): + def __init__(self, model_impl, xpu_cfg) -> None: + super().__init__() + + self.tp_size = xpu_cfg["tp_size"] + self.pad_token_id = xpu_cfg["pad_token_id"] + self.max_batch_size = xpu_cfg["max_batch_size"] + self.mp_engine = TpuMpEngine(self.tp_size, model_impl, xpu_cfg) + + def prepare_inputs( + self, + tasks: List[CoreInferencer.Task], + **kwargs + ): + input_dict = { + "input_ids": None, + "position_ids": None, + "attention_mask": None, + "all_q_len": None, + "all_kv_len": None, + "is_context": None, + "valid_slot_ids": None + } + + is_context = kwargs.get("is_context") if "is_context" in kwargs.keys() else False + valid_slot_ids = kwargs.get("valid_slot_ids") if "valid_slot_ids" in kwargs.keys() else [i for i in range(self.max_batch_size)] + + + get_input_logits = False + for task in tasks: + if task.request.generate_config.get_input_logits: + get_input_logits = True + break + + input_dict["is_context"] = is_context + input_dict["valid_slot_ids"] = valid_slot_ids + input_dict["get_input_logits"] = get_input_logits + + if is_context: + q_len = len(tasks[0].request.input_ids) + kv_len = len(tasks[0].request.input_ids) + + input_dict["input_ids"] = [ + tasks[0].request.input_ids + ] + input_dict["position_ids"] = [ + [i for i in range(q_len)] + ] + input_dict["attention_mask"] = [ + [1 for _ in range(q_len)] + ] + input_dict["all_q_len"] = [ + q_len + ] + input_dict["all_kv_len"] = [ + kv_len + ] + else: + all_input_ids = [] + all_position_ids = [] + all_attention_mask = [] + all_q_len = [] + all_kv_len = [] + + for task in tasks: + q_len = 1 + kv_len = 0 + + if task is None: + kv_len = 1 + + input_ids = [ + self.pad_token_id + ] + position_ids = [ + 0 + ] + attention_mask = [ + 0 + ] + else: + kv_len = len(task.request.input_ids) + len(task.generate_ids) - 1 + + input_ids = [ + task.generate_ids[-1] + ] + position_ids = [ + kv_len + ] + attention_mask = [ + 1 + ] + all_input_ids.append(input_ids) + all_position_ids.append(position_ids) + all_attention_mask.append(attention_mask) + all_q_len.append(q_len) + all_kv_len.append(kv_len) + + input_dict["input_ids"] = all_input_ids + input_dict["position_ids"] = all_position_ids + input_dict["attention_mask"] = all_attention_mask + input_dict["all_q_len"] = all_q_len + input_dict["all_kv_len"] = all_kv_len + + return input_dict + + def infer( + self, + tasks: List[CoreInferencer.Task], + **kwargs + ): + input_dict = self.prepare_inputs(tasks, **kwargs) + output_dict = self.mp_engine.mp_forward(input_dict) + + logits = output_dict["logits"] + next_token_logits = logits[:, -1, :].contiguous() + infer_outputs = { + "logits": logits, + "last_logits": next_token_logits + } + return infer_outputs \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/backends/TPU/tpu_mp_engine.py b/byte_infer_perf/llm_perf/backends/TPU/tpu_mp_engine.py new file mode 100644 index 000000000..55c0ca52f --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/tpu_mp_engine.py @@ -0,0 +1,187 @@ +import os +import sys +import time +import signal +import pathlib +from multiprocessing import Queue +from typing import List +from abc import ABC + +import torch +import torch.nn as nn +import torch.distributed as dist +import torch_tpu + +from llm_perf.core.mp_engine import CoreMpEngine +from llm_perf.utils.logger import logger + + +# context: +# input_ids: [1, s_q] +# attention_mask = [1, s_q] +# full_attention_mask = [1, 1, s_q, s_kv] (sq == s_kv) +def get_context_masks( + input_ids : torch.Tensor, + padding_mask : torch.Tensor +): + # input_ids: [1, q_len] + # padding_mask = [1, q_len] + _, q_len = input_ids.shape + + # [1, q_len, q_len] + full_attention_mask = torch.ones( + 1, q_len, q_len, + device=input_ids.device + ) + # full_attention_mask.tril_() + full_attention_mask = full_attention_mask * padding_mask.unsqueeze(1) + full_attention_mask -= padding_mask.unsqueeze(-1) - 1 + full_attention_mask = (full_attention_mask < 0.5).bool() + full_attention_mask.unsqueeze_(1) + return full_attention_mask + +# decode +# input_ids: [bs, 1] +# attention_mask = [bs, 1] +# full_attention_mask = [bs, 1, 1, s_kv] +def get_decode_masks( + input_ids : torch.Tensor, + all_kv_len: List[int] +): + # input_ids: [batch_size, 1] + # padding_mask: [batch_size, 1 + max_kv_len] + batch_size, q_len = input_ids.shape + max_qkv_len = q_len + max(all_kv_len) + + # [batch_size, 1, max_qkv_len] + padding_mask = [] + for i in range(batch_size): + cur_qkv_len = q_len + all_kv_len[i] + mask_per_batch = [1] * cur_qkv_len + [0] * (max_qkv_len - cur_qkv_len) + padding_mask.append(mask_per_batch) + full_attention_mask = torch.tensor( + padding_mask, + device=input_ids.device + ).unsqueeze_(1) + full_attention_mask = (full_attention_mask < 0.5).bool() + full_attention_mask.unsqueeze_(1) + return full_attention_mask + + +class TpuMpEngine(CoreMpEngine): + def __init__(self, world_size: int, model_impl: nn.Module, xpu_cfg) -> None: + super().__init__(world_size, model_impl, xpu_cfg) + def build_inputs(self, forward_inputs): + # list --> torch.Tensor --> tpu + forward_inputs["input_ids"] = torch.tensor( + forward_inputs["input_ids"] + ).to(torch.int32).tpu() + forward_inputs["position_ids"] = torch.tensor( + forward_inputs["position_ids"] + ).to(torch.int32).tpu() + forward_inputs["attention_mask"] = torch.tensor( + forward_inputs["attention_mask"] + ).to(torch.int32).tpu() + + is_context = forward_inputs["is_context"] + if is_context: + forward_inputs["full_attention_mask"] = get_context_masks( + forward_inputs["input_ids"], + forward_inputs["attention_mask"] + ) + else: + forward_inputs["full_attention_mask"] = get_decode_masks( + forward_inputs["input_ids"], + forward_inputs["all_kv_len"] + ) + return forward_inputs + + @torch.no_grad() + def mp_loop_worker( + self, + local_rank: int, + world_size: int, + input_queue: Queue, + output_queue: Queue, + model_impl, + xpu_config + ): + try: + torch.manual_seed(1) + + # set rank and world_size + os.environ["RANK"] = str(local_rank) + os.environ["LOCAL_RANK"] = str(local_rank) + os.environ["WORLD_SIZE"] = str(world_size) + os.environ["LOCAL_WORLD_SIZE"] = str(world_size) + + # create and init model based on model_impl and xpu_config + model = model_impl(xpu_config) + if hasattr(model, 'init_inference'): + model.init_inference() + + def signal_handler(signum, frame): + logger.info(f"rank {local_rank} received signal {signum}, exiting...") + if hasattr(model, 'finalize_inference'): + model.finalize_inference() + os._exit(0) + + signal.signal(signal.SIGINT, signal_handler) + signal.signal(signal.SIGTERM, signal_handler) + + # current rank is ready + output_queue.put("ready", block=True) + logger.info(f"{local_rank}/{world_size} rank is ready") + + # model process loop + while True: + ( + forward_inputs, + ) = input_queue.get(block=True) + + log = forward_inputs.get("log", False) + workspace = forward_inputs.get("workspace", None) + + forward_inputs["log_file"] = None + if log and workspace is not None: + workspace_dir = workspace / f"rank_{local_rank}" + workspace_dir.mkdir(exist_ok=True, parents=True) + forward_inputs["log_file"] = open(workspace_dir / "run.log", "w") + + inputs_dict = self.build_inputs(forward_inputs) + start_time = time.perf_counter_ns() + + output_dict = model.forward(inputs_dict) + + torch.tpu.synchronize() + end_time = time.perf_counter_ns() + duration_ms = round((end_time - start_time) / 1e6, 3) + output_dict["duration_ms"] = duration_ms + + output_dict["logits"] = output_dict["logits"].to('cpu') + # TP realization: rank0 send result back to main process + if local_rank == 0: + output_queue.put(output_dict) + + if log and workspace is not None: + forward_inputs["log_file"].close() + + except Exception as e: + logger.exception(f"[BUG] engine _load_and_listen failed, no more requests will be handled. {e}") + output_queue.put(RuntimeError("[BUG] fatal exception in model subprocess")) + + def mp_forward(self, *args): + # extra args + # workspace: pathlib.Path, where to save files for each rank + # log: bool, whether to save logs to file + # override_hidden_states: bool, whether to override hidden_states + # random_seed: int, random seed for torch.manual_seed + + # send inputs to all subprocesses + for _ in range(self.world_size): + self._input_queues.put(args, block=True) + + # wait for one subprocess send result back to main process + output_dict = self._output_queues.get(block=True) + + return output_dict \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/backends/TPU/tpu_sampler.py b/byte_infer_perf/llm_perf/backends/TPU/tpu_sampler.py new file mode 100644 index 000000000..9502664bf --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/tpu_sampler.py @@ -0,0 +1,155 @@ +from typing import Any, Dict, List, Tuple, Union + +import torch +import torch_tpu + +from llm_perf.core.generation import GenerateResult +from llm_perf.core.inferencer import CoreInferencer +from llm_perf.core.sampler import CoreSampler + +from llm_perf.utils.logger import logger + + +class TpuSampler(CoreSampler): + def __init__(self) -> None: + super().__init__() + + def sample( + self, + tasks: List[CoreInferencer.Task], + logits: torch.FloatTensor + ) -> List[int]: + top_p = [p.request.generate_config.top_p for p in tasks] + if all(p == 1.0 for p in top_p): + top_p = None + + top_k = [p.request.generate_config.top_k for p in tasks] + if all(k == 0 for k in top_k): + top_k = None + + temperature = [p.request.generate_config.temperature for p in tasks] + if all(t == 1.0 for t in temperature): + temperature = None + + ( + sp_input_ids, + sp_cu_seqlens, + sp_max_seqlens, + repetition_penalty, + mask_eos_token, + ) = (None, None, 0, None, None) + eos_token_id = [p.request.generate_config.eos_token_id or -1 for p in tasks] + + next_tokens, softmax_out = self._sample( + logits.float(), + temperature=temperature, + top_k=top_k, + top_p=top_p, + input_ids=sp_input_ids, + cu_seqlens=sp_cu_seqlens, + max_seqlens=sp_max_seqlens, + repetition_penalty=repetition_penalty, + mask_eos_token=mask_eos_token, + min_tokens_to_keep=1, + eos_token_id=eos_token_id, + ) + + next_tokens = next_tokens.tolist() + + # The aux_data is softmax_out here + return next_tokens, softmax_out + + + def _sample( + self, + logits: torch.FloatTensor, + temperature: Union[List[float], torch.FloatTensor] = None, + top_k: Union[List[int], torch.IntTensor] = None, + top_p: Union[List[float], torch.FloatTensor] = None, + input_ids: Union[List[int], torch.IntTensor] = None, + cu_seqlens: Union[List[int], torch.IntTensor] = None, + repetition_penalty: Union[List[float], torch.FloatTensor] = None, + mask_eos_token: Union[List[int], torch.IntTensor] = None, + min_tokens_to_keep: int = 1, + eos_token_id: int = 0, + max_seqlens: int = 0, + ) -> Tuple[List[int], torch.FloatTensor]: + _is_greedy = False + _is_random = False + _is_fastpath = False + + if top_k: + assert all( + k == top_k[0] for k in top_k + ), f"expect the same batch top_k, but got {top_k}" + if all(k == 1 for k in top_k): + _is_greedy = True + elif top_p: + _is_random = True + if all(p == top_p[0] for p in top_p): + _is_fastpath = True + _top_p = top_p[0] + else: + raise RuntimeError( + f"Unsupported sample strategy, parameter top_k: {top_k} top_p: {top_p}" + ) + + if _is_greedy: + return torch.argmax(logits, dim=-1), torch.nn.functional.softmax(logits, dim=-1) + else: + raise NotImplementedError + + def postprocess( + self, + tasks: List[CoreInferencer.Task], + infer_outputs: Dict[str, torch.FloatTensor], + next_tokens: List[int], + ) -> List[GenerateResult]: + generate_result = [] + for i in range(len(tasks)): + token_id = next_tokens[i] + task = tasks[i] + + # take current generated token into account + generate_tokens_len = len(task.generate_ids) + 1 + + if token_id == task.request.generate_config.eos_token_id: + if generate_tokens_len < task.request.generate_config.min_new_tokens: + finish_reason = "" + token_id = task.request.generate_config.eos_token_id + else: + finish_reason = "stop" + elif generate_tokens_len >= task.request.generate_config.max_new_tokens: + finish_reason = "max_length" + else: + finish_reason = "" + + + if task.request.generate_config.get_input_logits: + gen_res = GenerateResult( + token_id=token_id, + finish_reason=finish_reason, + + wait_time=task.wait_time[-1], + model_time=task.model_time[-1], + post_process_time=task.post_process_time[-1], + + logits=infer_outputs["logits"][i].float().cpu(), + last_logits=infer_outputs["last_logits"][i].float().cpu(), + ) + else: + gen_res = GenerateResult( + token_id=token_id, + finish_reason=finish_reason, + + wait_time=task.wait_time[-1], + model_time=task.model_time[-1], + post_process_time=task.post_process_time[-1], + + logits=None, + last_logits=None, + ) + + generate_result.append(gen_res) + + return generate_result diff --git a/byte_infer_perf/llm_perf/backends/TPU/tpu_scheduler.py b/byte_infer_perf/llm_perf/backends/TPU/tpu_scheduler.py new file mode 100644 index 000000000..2bc97e9e3 --- /dev/null +++ b/byte_infer_perf/llm_perf/backends/TPU/tpu_scheduler.py @@ -0,0 +1,139 @@ +import sys +import time +from concurrent.futures import ThreadPoolExecutor, as_completed +from typing import List, Set + +import torch +import torch_tpu + +from llm_perf.core.scheduler import CoreScheduler +from llm_perf.core.inferencer import CoreInferencer +from llm_perf.core.sampler import CoreSampler +from llm_perf.backends.TPU.tpu_inferencer import TpuInferencer +from llm_perf.utils.logger import logger + +class TpuScheduler(CoreScheduler): + def __init__( + self, + inferencer: CoreInferencer, + sampler: CoreSampler, + xpu_cfg + ) -> None: + super().__init__(inferencer, sampler) + self.max_batch_size = xpu_cfg["max_batch_size"] + + + @torch.inference_mode() + def scheduler_loop(self): + task_slots: List[CoreInferencer.Task] = [None] * self.max_batch_size + avail_slots: List[int] = [self.max_batch_size - 1 - i for i in range(self.max_batch_size)] + context_slots: List[int] = [] + + while self.started: + while not self.task_queue.empty(): + if len(avail_slots) == 0: + break + slot = avail_slots.pop() + task_slots[slot] = self.task_queue.get() + context_slots.append(slot) + + if len(avail_slots) == self.max_batch_size: + with self.task_queue.not_empty: + self.task_queue.not_empty.wait(0.1) + continue + + + # context phase + if len(context_slots) != 0: + # do inference --> logits + select_slot = context_slots.pop(0) + select_slots= [ + select_slot + ] + + cur_task = task_slots[select_slot] + cur_tasks = [ + cur_task + ] + + cur_task.update_st("model_start") + + outputs = self.inferencer.infer( + cur_tasks, + is_context=True, + valid_slot_ids=select_slots + ) + + cur_task.update_st("model_end") + + # sample logits --> tokens + next_tokens, _ = self.sampler.sample( + tasks=cur_tasks, + logits=outputs["last_logits"] + ) + + cur_task.update_st("process_end") + + # postprocess -> gen result + generation_results = self.sampler.postprocess( + tasks=cur_tasks, + infer_outputs=outputs, + next_tokens=next_tokens, + ) + + # add result to task + cur_task.add_result(generation_results[0]) + if generation_results[0].finish_reason: + cur_task.finish() + + + # decode phase + else: + select_slots = [] + valid_tasks = [] + for i, task in enumerate(task_slots): + if task is not None: + select_slots.append(i) + valid_tasks.append(task) + + for task in valid_tasks: + task.update_st("model_start") + + outputs = self.inferencer.infer( + valid_tasks, + is_context=False, + valid_slot_ids=select_slots + ) + + for task in valid_tasks: + task.update_st("model_end") + + + # sample logits --> tokens + next_tokens, _ = self.sampler.sample( + tasks=valid_tasks, + logits=outputs["last_logits"] + ) + + for task in valid_tasks: + task.update_st("process_end") + + # postprocess -> gen result + generation_results = self.sampler.postprocess( + tasks=valid_tasks, + infer_outputs=outputs, + next_tokens=next_tokens, + ) + + # add result to task + for i, gen_res in enumerate(generation_results): + valid_tasks[i].add_result(gen_res) + if gen_res.finish_reason: + valid_tasks[i].finish() + + for i, task in enumerate(task_slots): + if task is not None and task.is_finished(): + avail_slots.append(i) + task_slots[i] = None + + avail_slots.sort(reverse=True) \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/model_zoo/llama3.1-8b-torch-bf16.json b/byte_infer_perf/llm_perf/model_zoo/llama3.1-8b-torch-bf16.json new file mode 100755 index 000000000..e0e4327cb --- /dev/null +++ b/byte_infer_perf/llm_perf/model_zoo/llama3.1-8b-torch-bf16.json @@ -0,0 +1,48 @@ +{ + "model_name": "llama3.1", + "model_path": "llm_perf/model_zoo/sota/llama3.1-8b", + "model_interface": "LlamaForCausalLM", + "tokenizer": { + "path": "llm_perf/model_zoo/sota/llama3.1-8b", + "support_chn": true + }, + "network": { + "_name_or_path": "meta-llama/Llama-3.1-8B-Instruct", + "architectures": [ + "LlamaForCausalLM" + ], + "attention_bias": false, + "attention_dropout": 0.0, + "bos_token_id": 128000, + "eos_token_id": [ + 128001, + 128008, + 128009 + ], + "hidden_act": "silu", + "hidden_size": 4096, + "initializer_range": 0.02, + "intermediate_size": 14336, + "max_position_embeddings": 131072, + "mlp_bias": false, + "model_type": "llama", + "num_attention_heads": 32, + "num_hidden_layers": 32, + "num_key_value_heads": 8, + "pretraining_tp": 1, + "rms_norm_eps": 1e-05, + "rope_scaling": { + "factor": 8.0, + "low_freq_factor": 1.0, + "high_freq_factor": 4.0, + "original_max_position_embeddings": 8192, + "rope_type": "llama3" + }, + "rope_theta": 500000.0, + "tie_word_embeddings": false, + "torch_dtype": "bfloat16", + "transformers_version": "4.42.3", + "use_cache": true, + "vocab_size": 128256 + } +} \ No newline at end of file diff --git a/byte_infer_perf/llm_perf/prepare_model.py b/byte_infer_perf/llm_perf/prepare_model.py index 431c6e05b..fde3df514 100644 --- a/byte_infer_perf/llm_perf/prepare_model.py +++ b/byte_infer_perf/llm_perf/prepare_model.py @@ -13,6 +13,7 @@ "llama3-torch-bf16-70b": ("llama3-70b", "shenzhi-wang/Llama3-70B-Chinese-Chat"), "falcon-torch-bf16-180b": ("falcon-180b", "tiiuae/falcon-180B"), "mixtral-torch-bf16-8x22b": ("mixtral-8x22b-instruct", "mistralai/Mixtral-8x22B-Instruct-v0.1"), + "llama3.1-8b-torch-bf16": ("llama3.1-8b", "meta-llama/Llama-3.1-8B-Instruct") } if __name__ == "__main__": diff --git a/byte_infer_perf/llm_perf/workloads/llama3.1-8b-torch-bf16.json b/byte_infer_perf/llm_perf/workloads/llama3.1-8b-torch-bf16.json new file mode 100644 index 000000000..39f61087d --- /dev/null +++ b/byte_infer_perf/llm_perf/workloads/llama3.1-8b-torch-bf16.json @@ -0,0 +1,18 @@ +{ + "model": "llama3.1-8b", + "test_accuracy": false, + "min_tp_size": 1, + "accuracy_config": { + "dataset": "llm_perf/datasets/merged_52_test.csv", + "min_new_tokens": 1, + "max_new_tokens": 512 + }, + "test_perf": true, + "perf_config": { + "tp_sizes": [1], + "batch_sizes": [1, 4, 8], + "input_tokens": [1024], + "output_tokens": 200, + "perf_time": 100 + } +} \ No newline at end of file