[Feature] Hybrid Data Pipeline

xtuner/configs/internlm/internlm2_chat_1_8b/hybrid/agent.json

@@ -0,0 +1,62 @@
+{
+    "messages": [
+        {"role": "system", "content": "You are InternLM2-Chat, a harmless AI assistant"},
+        {
+            "role": "user",
+            "content": "Please help me process and visualize this dataset.",
+            "files": [{"path": "data.csv", "size": "10K"}]
+        },
+        {
+            "role": "assistant",
+            "content": "I have processed the data and visualized it for you.",
+            "code_interpreter_call": "```python\nimport plotly.express as px\nimport pandas as pd\n\n# Load the data into a pandas dataframe\ndf = pd.read_csv('data.csv')\n\n# Create a scatter plot of rainfall vs wind direction\nfig = px.scatter(df, x='WindDir9am', y='Rainfall', color='WindDir3pm',\n                 labels={'WindDir9am': 'Wind Direction 9am', 'Rainfall': '\n\nRainfall', 'WindDir3pm': 'Wind Direction 3pm'},\n                 title='Rainfall vs Wind Direction',\n                 template='plotly_dark',\n                 width=600, height=500)\n\n# Add a hover effect to show the date\nfig.update_traces(hovertemplate='<b>Date: %{text}</b><br>Wind Direction 9am: %{x}<br>Rainfall: %{y}<br>Wind Direction 3pm: %{marker.color}')\n\n# Show the plot\nfig.show()\n```"
+        },
+        {
+            "role": "code_interpreter",
+            "content": "![image](xxx.png)"
+        },
+        {
+            "role": "assistant",
+            "content": "Since the code output is not included here, I cannot provide specific chart content. However, if the code executed correctly, it should display a polar plot with two filled areas representing the relationship between wind direction at 9 am and rainfall, and between wind direction at 3 pm and rainfall, respectively. The values for each direction are based on the average rainfall calculated from the provided dataset. The chart should have a clear title, a legend, and be intuitive for comparing rainfall with different wind directions. Given the use of a dark theme, the overall appearance of the chart should be bright lines and filled areas on a dark background."
+        },
+        {
+            "role": "user",
+            "content": "I want to know today's weather in Shanghai"
+        },
+        {
+            "role": "assistant",
+            "content": "Sure, I will search for the weather of Shanghai.",
+            "function_call": {
+                "name": "get_current_weather",
+                "parameters": {"location": "Shanghai"}
+            }
+        },
+        {
+            "role": "function",
+            "name": "get_current_weather",
+            "content": "{'temperature': 22}"
+        },
+        {
+            "role": "assistant",
+            "content": "The weather in Shanghai is 22 celsius"
+        }
+    ],
+
+    "functions": [
+        {
+            "name": "get_current_weather",
+            "description": "Get the current weather in a given location",
+            "parameters": {
+                "type": "object",
+                "properties": {
+                    "location": {
+                        "type": "string",
+                        "description": "The city and state, e.g. San Francisco, CA",
+                        "unit": {"type": "string"}},
+                        "required": ["location"]
+                    }
+            }
+        }
+    ],
+
+    "code_interpreter": "You now have access to a Jupyter notebook environment supporting Python code execution. Just send code to python to run in this stateful environment. This feature is suitable for:\n- Data analysis or processing (such as data manipulation and graphic creation)\n- Complex calculations (such as math and physics problems)\n- Programming examples (for understanding programming concepts or language features)\n- Text processing and analysis (including text analysis and natural language processing)\n- Machine learning and data science (model training and data visualization)\n- File operations and data import (handling CSV, JSON, etc. formats)"}

xtuner/configs/internlm/internlm2_chat_1_8b/hybrid/example.py

@@ -0,0 +1,33 @@
+import json
+
+from xtuner.types import HybridChatTemplate, TrainingHybridChatMessages
+
+chat_template = HybridChatTemplate(
+    system='<|im_start|>system\n{system}<|im_end|>\n',
+    user='<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n',
+    assistant='{assistant}<|im_end|>\n',
+    stop_words=['<|im_end|>'],
+    image_token='<image>',
+    files='<|im_start|>user name=file\n{files}<|im_end|>\n',
+    function_call=
+    '{assistant}<|action_start|><|plugin|>\n{function_call}<|action_end|><|im_end|>\n',  # noqa: E501, E251
+    function_result=
+    '<|im_start|>environment name=<|plugin|>\n{function_result}<|im_end|>\n<|im_start|>assistant\n',  # noqa: E501, E251
+    functions='<|im_start|>system name=<|plugin|>\n{functions}<|im_end|>\n',
+    code_interpreter_call=
+    '{assistant}<|action_start|><|interpreter|>\n{code_interpreter_call}<|action_end|><|im_end|>\n',  # noqa: E501, E251
+    code_interpreter_result=
+    '<|im_start|>environment name=<|interpreter|>\n{code_interpreter_result}<|im_end|>\n<|im_start|>assistant\n',  # noqa: E501, E251
+    code_interpreter=
+    '<|im_start|>system name=<|interpreter|>\n{code_interpreter}<|im_end|>\n')
+
+agent_data = json.load(open('agent.json'))
+
+msg = TrainingHybridChatMessages.from_dict(agent_data)
+print(msg.apply_chat_template(chat_template))
+
+from transformers import AutoTokenizer
+
+tokenizer = AutoTokenizer.from_pretrained(
+    'internlm/internlm2-chat-7b', trust_remote_code=True)
+print(msg.tokenize(tokenizer, chat_template))

xtuner/configs/internlm/internlm2_chat_1_8b/hybrid/function_call.json

@@ -0,0 +1,52 @@
+[
+    {
+        "messages": [
+            {
+                "role": "user",
+                "content": "I want to know today's weather in Shanghai"
+            },
+
+            {
+                "role": "assistant",
+                "content": "Sure, I will search for the weather of Shanghai.",
+                "function_call": {
+                    "name": "get_current_weather",
+                    "parameters": {
+                        "location": "Shanghai"
+                    }
+                }
+            },
+
+            {
+                "role": "function",
+                "name": "get_current_weather",
+                "content": "{'temperature': 22}"
+            },
+            {
+                "role": "assistant",
+                "content": "The weather in Shanghai is 22 celsius"
+            }
+
+
+        ],
+
+        "functions": [
+            {
+                "name": "get_current_weather",
+                "description": "Get the current weather in a given location",
+                "parameters": {
+                    "type": "object",
+                    "properties": {
+                        "location": {
+                            "type": "string",
+                            "description": "The city and state, e.g. San Francisco, CA",
+                            "unit": {"type": "string"}
+                        },
+                        "required": ["location"]
+                    }
+                }
+            }
+        ]
+    }
+
+]

xtuner/configs/internlm/internlm2_chat_1_8b/hybrid/internlm2_chat_1_8b_function_call.py

@@ -0,0 +1,193 @@
+# Copyright (c) OpenMMLab. All rights reserved.
+import torch
+from mmengine.dataset import DefaultSampler
+from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
+                            LoggerHook, ParamSchedulerHook)
+from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
+from torch.optim import AdamW
+from transformers import AutoTokenizer
+
+from xtuner.dataset.hybrid import HybridDataset, hybrid_collate_fn
+from xtuner.dataset.hybrid.mappings import openai_to_raw_training
+from xtuner.engine.hooks import DatasetInfoHook
+from xtuner.engine.runner import TrainLoop
+from xtuner.model import AgentFinetune, AutoModelForCausalLM
+from xtuner.types import HybridChatTemplate
+
+#######################################################################
+#                          PART 1  Settings                           #
+#######################################################################
+# Model
+llm_name_or_path = '/mnt/petrelfs/share_data/basemodel/checkpoints/llm/hf_hub/models--internlm--internlm2-chat-1_8b/snapshots/aa8a7450c2227a3b6733b3c6fe33fefbb2ca54f9/'
+
+# Data
+data_dir = './'
+data_files = ['agentlego.json']
+max_length = 2048
+
+# Chat Template
+chat_template = dict(
+    type=HybridChatTemplate,
+    system='<|im_start|>system\n{system}<|im_end|>\n',
+    user='<|im_start|>user\n{user}<|im_end|>\n<|im_start|>assistant\n',
+    assistant='{assistant}<|im_end|>\n',
+    stop_words=['<|im_end|>'],
+    image_token='<image>',
+    function_call=
+    '{assistant}<|action_start|><|plugin|>\n{function_call}<|action_end|><|im_end|>\n',  # noqa: E501, E251
+    function_result=
+    '<|im_start|>environment name=<|plugin|>\n{function_result}<|im_end|>\n<|im_start|>assistant\n',  # noqa: E501, E251
+    functions='<|im_start|>system name=<|plugin|>\n{functions}<|im_end|>\n')
+
+# Scheduler & Optimizer
+batch_size = 1  # per_device
+accumulative_counts = 1
+dataloader_num_workers = 0
+max_epochs = 1
+optim_type = AdamW
+lr = 2e-4
+betas = (0.9, 0.999)
+weight_decay = 0
+max_norm = 1  # grad clip
+warmup_ratio = 0.03
+
+# Save
+save_steps = 500
+save_total_limit = 2  # Maximum checkpoints to keep (-1 means unlimited)
+
+#######################################################################
+#            PART 2  Model & Tokenizer & Image Processor              #
+#######################################################################
+tokenizer = dict(
+    type=AutoTokenizer.from_pretrained,
+    pretrained_model_name_or_path=llm_name_or_path,
+    trust_remote_code=True,
+    padding_side='right')
+
+model = dict(
+    type=AgentFinetune,
+    llm=dict(
+        type=AutoModelForCausalLM.from_pretrained,
+        pretrained_model_name_or_path=llm_name_or_path,
+        trust_remote_code=True,
+        torch_dtype=torch.float16))
+
+#######################################################################
+#                      PART 3  Dataset & Dataloader                   #
+#######################################################################
+dataset = dict(
+    type=HybridDataset,
+    data_dir=data_dir,
+    data_files=data_files,
+    sample_ratio=1,
+    tokenizer=tokenizer,
+    chat_template=chat_template,
+    max_length=max_length,
+    pack_to_max_length=True,
+    num_workers=4,
+    mappings=[openai_to_raw_training])
+
+train_dataloader = dict(
+    batch_size=batch_size,
+    num_workers=dataloader_num_workers,
+    dataset=dataset,
+    sampler=dict(type=DefaultSampler, shuffle=True),
+    collate_fn=dict(type=hybrid_collate_fn))
+
+#######################################################################
+#                    PART 4  Scheduler & Optimizer                    #
+#######################################################################
+# optimizer
+optim_wrapper = dict(
+    type=AmpOptimWrapper,
+    optimizer=dict(
+        type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
+    clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
+    accumulative_counts=accumulative_counts,
+    loss_scale='dynamic',
+    dtype='float16')
+
+# learning policy
+# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md  # noqa: E501
+param_scheduler = [
+    dict(
+        type=LinearLR,
+        start_factor=1e-5,
+        by_epoch=True,
+        begin=0,
+        end=warmup_ratio * max_epochs,
+        convert_to_iter_based=True),
+    dict(
+        type=CosineAnnealingLR,
+        eta_min=0.0,
+        by_epoch=True,
+        begin=warmup_ratio * max_epochs,
+        end=max_epochs,
+        convert_to_iter_based=True)
+]
+
+# train, val, test setting
+train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
+
+#######################################################################
+#                           PART 5  Runtime                           #
+#######################################################################
+# Log the dialogue periodically during the training process, optional
+custom_hooks = [
+    dict(type=DatasetInfoHook, tokenizer=tokenizer),
+    # dict(
+    #     type=EvaluateChatHook,
+    #     tokenizer=tokenizer,
+    #     image_processor=image_processor,
+    #     every_n_iters=evaluation_freq,
+    #     evaluation_inputs=evaluation_inputs,
+    #     evaluation_images=evaluation_images,
+    #     system=SYSTEM,
+    #     prompt_template=prompt_template)
+]
+
+# configure default hooks
+default_hooks = dict(
+    # record the time of every iteration.
+    timer=dict(type=IterTimerHook),
+    # print log every 10 iterations.
+    logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=10),
+    # enable the parameter scheduler.
+    param_scheduler=dict(type=ParamSchedulerHook),
+    # save checkpoint per `save_steps`.
+    checkpoint=dict(
+        type=CheckpointHook,
+        by_epoch=False,
+        interval=save_steps,
+        max_keep_ckpts=save_total_limit),
+    # set sampler seed in distributed evrionment.
+    sampler_seed=dict(type=DistSamplerSeedHook),
+)
+
+# configure environment
+env_cfg = dict(
+    # whether to enable cudnn benchmark
+    cudnn_benchmark=False,
+    # set multi process parameters
+    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
+    # set distributed parameters
+    dist_cfg=dict(backend='nccl'),
+)
+
+# set visualizer
+visualizer = None
+
+# set log level
+log_level = 'INFO'
+
+# load from which checkpoint
+load_from = None
+
+# whether to resume training from the loaded checkpoint
+resume = False
+
+# Defaults to use random seed and disable `deterministic`
+randomness = dict(seed=None, deterministic=False)
+
+# set log processor
+log_processor = dict(by_epoch=False)