Support molmo in turbomind (#2716)

* initial moe support * dynamic grouped gemm * benchmark * moe benchmark * moe sampling * split-k * refactor tuning * simplify * n-major weight * add `num` for `MatrixLayout` * packed rows * packed cols * dispatch for packed rows * w4a16 moe * refactor model loading * fix pytorch loader * refactor * dispatch w4a16 moe * fix loader * add comment * fix msvc build * fix msvc build * fix msvc build * fix ut * fix ut * fix p-lora * add all support arches * minor * fix lint * fix lint * fix lint * fix ut * bf16 support * minor * checkin molmo conversion * add chat template * refactor * fix lint * fix ut * Just for test: hardcode vocab_size * minor * minor * minor * fix inter_size config * load with non-standard filenames * fix loader * fix missing default param * defer the loading of misc weights for safetensors * add embedding_size * update * update * tmp * tmp * update molmo template * vision embedding * fix * update * fix * fix messages2prompt in templates * fix order of out_messages * fix * add user guide * update is_supported --------- Co-authored-by: Li Zhang <[email protected]>
InternLM · Nov 14, 2024 · fd8906c · fd8906c
1 parent a21def9
commit fd8906c
Show file tree

Hide file tree

Showing 19 changed files with 653 additions and 8 deletions.
diff --git a/docs/en/multi_modal/index.rst b/docs/en/multi_modal/index.rst
@@ -12,3 +12,5 @@ Vision-Language Models
    minicpmv.md
    phi3.md
    mllama.md
+   qwen2_vl.md
+   molmo.md
diff --git a/docs/en/multi_modal/molmo.md b/docs/en/multi_modal/molmo.md
@@ -0,0 +1,92 @@
+# Molmo
+
+LMDeploy supports the following molmo series of models, which are detailed in the table below:
+
+|      Model      | Size | Supported Inference Engine |
+| :-------------: | :--: | :------------------------: |
+| Molmo-7B-D-0924 |  7B  |         TurboMind          |
+|  Molmo-72-0924  | 72B  |         TurboMind          |
+
+The next chapter demonstrates how to deploy a molmo model using LMDeploy, with [Molmo-7B-D-0924](https://huggingface.co/allenai/Molmo-7B-D-0924) as an example.
+
+## Installation
+
+Please install LMDeploy by following the [installation guide](../get_started/installation.md)
+
+## Offline inference
+
+The following sample code shows the basic usage of VLM pipeline. For detailed information, please refer to [VLM Offline Inference Pipeline](./vl_pipeline.md)
+
+```python
+from lmdeploy import pipeline
+from lmdeploy.vl import load_image
+
+pipe = pipeline('allenai/Molmo-7B-D-0924')
+
+image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
+response = pipe((f'describe this image', image))
+print(response)
+```
+
+More examples are listed below:
+
+<details>
+  <summary>
+    <b>multi-image multi-round conversation, combined images</b>
+  </summary>
+
+```python
+from lmdeploy import pipeline, GenerationConfig
+
+pipe = pipeline('allenai/Molmo-7B-D-0924', log_level='INFO')
+messages = [
+    dict(role='user', content=[
+        dict(type='text', text='Describe the two images in detail.'),
+        dict(type='image_url', image_url=dict(url='https://raw.githubusercontent.com/QwenLM/Qwen-VL/master/assets/mm_tutorial/Beijing_Small.jpeg')),
+        dict(type='image_url', image_url=dict(url='https://raw.githubusercontent.com/QwenLM/Qwen-VL/master/assets/mm_tutorial/Chongqing_Small.jpeg'))
+    ])
+]
+out = pipe(messages, gen_config=GenerationConfig(do_sample=False))
+
+messages.append(dict(role='assistant', content=out.text))
+messages.append(dict(role='user', content='What are the similarities and differences between these two images.'))
+out = pipe(messages, gen_config=GenerationConfig(do_sample=False))
+```
+
+</details>
+
+## Online serving
+
+You can launch the server by the `lmdeploy serve api_server` CLI:
+
+```shell
+lmdeploy serve api_server allenai/Molmo-7B-D-0924
+```
+
+You can also start the service using the docker image:
+
+```shell
+docker run --runtime nvidia --gpus all \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
+    -p 23333:23333 \
+    --ipc=host \
+    openmmlab/lmdeploy:latest \
+    lmdeploy serve api_server allenai/Molmo-7B-D-0924
+```
+
+If you find the following logs, it means the service launches successfully.
+
+```text
+HINT:    Please open  http://0.0.0.0:23333   in a browser for detailed api usage!!!
+HINT:    Please open  http://0.0.0.0:23333   in a browser for detailed api usage!!!
+HINT:    Please open  http://0.0.0.0:23333   in a browser for detailed api usage!!!
+INFO:     Started server process [2439]
+INFO:     Waiting for application startup.
+INFO:     Application startup complete.
+INFO:     Uvicorn running on  http://0.0.0.0:23333  (Press CTRL+C to quit)
+```
+
+The arguments of `lmdeploy serve api_server` can be reviewed in detail by `lmdeploy serve api_server -h`.
+
+More information about `api_server` as well as how to access the service can be found from [here](api_server_vl.md)
diff --git a/docs/zh_cn/multi_modal/index.rst b/docs/zh_cn/multi_modal/index.rst
@@ -12,3 +12,5 @@
    minicpmv.md
    phi3.md
    mllama.md
+   qwen2_vl.md
+   molmo.md
diff --git a/docs/zh_cn/multi_modal/molmo.md b/docs/zh_cn/multi_modal/molmo.md
@@ -0,0 +1,92 @@
+# Qwen2-VL
+
+LMDeploy 支持 Molmo 系列模型，具体如下：
+
+|      Model      | Size | Supported Inference Engine |
+| :-------------: | :--: | :------------------------: |
+| Molmo-7B-D-0924 |  7B  |         TurboMind          |
+|  Molmo-72-0924  | 72B  |         TurboMind          |
+
+本文将以[Molmo-7B-D-0924](https://huggingface.co/allenai/Molmo-7B-D-0924) 为例，演示使用 LMDeploy 部署 Molmo 系列模型的方法
+
+## 安装
+
+请参考[安装文档](../get_started/installation.md)安装 LMDeploy。
+
+## 离线推理
+
+以下是使用 pipeline 进行离线推理的示例，更多用法参考[VLM离线推理 pipeline](./vl_pipeline.md)
+
+```python
+from lmdeploy import pipeline
+from lmdeploy.vl import load_image
+
+pipe = pipeline('allenai/Molmo-7B-D-0924')
+
+image = load_image('https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg')
+response = pipe((f'describe this image', image))
+print(response)
+```
+
+更多例子如下：
+
+<details>
+  <summary>
+    <b>多图多轮对话</b>
+  </summary>
+
+```python
+from lmdeploy import pipeline, GenerationConfig
+
+pipe = pipeline('Qwen/Qwen2-VL-2B-Instruct', log_level='INFO')
+messages = [
+    dict(role='user', content=[
+        dict(type='text', text='Describe the two images in detail.'),
+        dict(type='image_url', image_url=dict(url='https://raw.githubusercontent.com/QwenLM/Qwen-VL/master/assets/mm_tutorial/Beijing_Small.jpeg')),
+        dict(type='image_url', image_url=dict(url='https://raw.githubusercontent.com/QwenLM/Qwen-VL/master/assets/mm_tutorial/Chongqing_Small.jpeg'))
+    ])
+]
+out = pipe(messages, gen_config=GenerationConfig(top_k=1))
+
+messages.append(dict(role='assistant', content=out.text))
+messages.append(dict(role='user', content='What are the similarities and differences between these two images.'))
+out = pipe(messages, gen_config=GenerationConfig(top_k=1))
+```
+
+</details>
+
+## 在线服务
+
+你可以通过 `lmdeploy serve api_server` CLI 工具启动服务：
+
+```shell
+lmdeploy serve api_server Qwen/Qwen2-VL-2B-Instruct
+```
+
+也可以基于 docker image 启动服务：
+
+```shell
+docker run --runtime nvidia --gpus all \
+    -v ~/.cache/huggingface:/root/.cache/huggingface \
+    --env "HUGGING_FACE_HUB_TOKEN=<secret>" \
+    -p 23333:23333 \
+    --ipc=host \
+    openmmlab/lmdeploy:qwen2vl \
+    lmdeploy serve api_server Qwen/Qwen2-VL-2B-Instruct
+```
+
+如果日志中有如下信息，就表明服务启动成功了。
+
+```text
+HINT:    Please open  http://0.0.0.0:23333   in a browser for detailed api usage!!!
+HINT:    Please open  http://0.0.0.0:23333   in a browser for detailed api usage!!!
+HINT:    Please open  http://0.0.0.0:23333   in a browser for detailed api usage!!!
+INFO:     Started server process [2439]
+INFO:     Waiting for application startup.
+INFO:     Application startup complete.
+INFO:     Uvicorn running on  http://0.0.0.0:23333  (Press CTRL+C to quit)
+```
+
+有关 `lmdeploy serve api_server` 的详细参数可以通过`lmdeploy serve api_server -h`查阅。
+
+关于 `api_server` 更多的介绍，以及访问 `api_server` 的方法，请阅读[此处](api_server_vl.md)
diff --git a/lmdeploy/archs.py b/lmdeploy/archs.py
@@ -121,7 +121,8 @@ def check_vl_llm(config: dict) -> bool:
         'InternVLChatModel', 'MiniGeminiLlamaForCausalLM',
         'MGMLlamaForCausalLM', 'MiniCPMV', 'LlavaForConditionalGeneration',
         'LlavaNextForConditionalGeneration', 'Phi3VForCausalLM',
-        'Qwen2VLForConditionalGeneration', 'MllamaForConditionalGeneration'
+        'Qwen2VLForConditionalGeneration', 'MllamaForConditionalGeneration',
+        'MolmoForCausalLM'
     ])
     if arch == 'QWenLMHeadModel' and 'visual' in config:
         return True

diff --git a/lmdeploy/model.py b/lmdeploy/model.py
@@ -1747,6 +1747,37 @@ def match(cls, model_path: str) -> Optional[str]:
             return 'internvl-phi3'
 
 
+@MODELS.register_module(name='molmo')
+class Molmo(BaseChatTemplate):
+
+    def __init__(self,
+                 user=' User: ',
+                 eoh='',
+                 assistant=' Assistant:',
+                 eoa='',
+                 separator=' ',
+                 stop_words=['<|endoftext|>'],
+                 **kwargs):
+        super().__init__(user=user,
+                         eoh=eoh,
+                         assistant=assistant,
+                         eoa=eoa,
+                         separator=separator,
+                         stop_words=stop_words,
+                         **kwargs)
+
+    @classmethod
+    def match(cls, model_path: str) -> Optional[str]:
+        """Return the model_name that was registered to MODELS.
+
+        Args:
+            model_path (str): the model path used for matching.
+        """
+        path = model_path.lower()
+        if 'molmo' in path:
+            return 'molmo'
+
+
 def best_match_model(query: str) -> Optional[str]:
     """Get the model that matches the query.
 

diff --git a/lmdeploy/serve/vl_async_engine.py b/lmdeploy/serve/vl_async_engine.py
@@ -64,6 +64,7 @@ async def _get_prompt_input(self,
         results = {}
         input_ids = []
         from lmdeploy.vl.templates import (MllamaTempateWrapper,
+                                           MolmoChatTemplateWrapper,
                                            Qwen2VLChatTemplateWrapper)
         ranges = None
         grid_thws = None
@@ -99,6 +100,10 @@ async def _get_prompt_input(self,
                     results['cross_attention_states'] = features[0]
                     return results
 
+                if isinstance(self.vl_prompt_template,
+                              MolmoChatTemplateWrapper):
+                    return features[0]
+
             features = [x.cpu().numpy() for x in features]
             input_ids = []
             begins = []

diff --git a/lmdeploy/turbomind/deploy/config.py b/lmdeploy/turbomind/deploy/config.py
@@ -35,6 +35,13 @@ class ModelConfig:
     kv_head_num: int = None
     hidden_units: int = None
     vocab_size: int = None
+    # Turbomind used to assume token_embedding and lm_head has the same size
+    # at vocab dim, i.e. `vocab_size`
+    # But in molmo, embedding.shape is [vocab_size + 128, hidden_units]
+    # while lm_head shape is [hidden_units, vocab_size].
+    # Therefore, we add a new attr "embedding_size" to represent the vocab dim
+    # of token_embedding
+    embedding_size: int = 0
     num_layer: int = None
     inter_size: int = None
     norm_eps: float = None

diff --git a/lmdeploy/turbomind/deploy/source_model/__init__.py b/lmdeploy/turbomind/deploy/source_model/__init__.py
@@ -9,5 +9,6 @@
 from .meta_llama import MetaLlamaModel  # noqa: F401
 from .minicpmv import MiniCPMVModel  # noqa: F401
 from .mixtral import MixtralModel  # noqa: F401
+from .molmo import MolmoModel  # noqa: F401
 from .qwen import QwenModel  # noqa: F401
 from .xcomposer2 import Xcomposer2Model  # noqa: F401