- **Description:** Add to check pad_token_id and eos_token_id of model
config. It seems that this is the same bug as the HuggingFace TGI bug.
It's same bug as #29434
- **Issue:** #29431
- **Dependencies:** none
- **Twitter handle:** tell14
Example code is followings:
```python
from langchain_huggingface.llms import HuggingFacePipeline
hf = HuggingFacePipeline.from_model_id(
model_id="meta-llama/Llama-3.2-3B-Instruct",
task="text-generation",
pipeline_kwargs={"max_new_tokens": 10},
)
from langchain_core.prompts import PromptTemplate
template = """Question: {question}
Answer: Let's think step by step."""
prompt = PromptTemplate.from_template(template)
chain = prompt | hf
question = "What is electroencephalography?"
print(chain.invoke({"question": question}))
```
## Description
I encountered an error while using the` gemma-2-2b-it model` with the
`HuggingFacePipeline` class and have implemented a fix to resolve this
issue.
### What is Problem
```python
model_id="google/gemma-2-2b-it"
gemma_2_model = AutoModelForCausalLM.from_pretrained(model_id)
gemma_2_tokenizer = AutoTokenizer.from_pretrained(model_id)
gen = pipeline(
task='text-generation',
model=gemma_2_model,
tokenizer=gemma_2_tokenizer,
max_new_tokens=1024,
device=0 if torch.cuda.is_available() else -1,
temperature=.5,
top_p=0.7,
repetition_penalty=1.1,
do_sample=True,
)
llm = HuggingFacePipeline(pipeline=gen)
for chunk in llm.stream("Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World."):
print(chunk, end="", flush=True)
```
This code outputs the following error message:
```
/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py:1258: UserWarning: Using the model-agnostic default `max_length` (=20) to control the generation length. We recommend setting `max_new_tokens` to control the maximum length of the generation.
warnings.warn(
Exception in thread Thread-19 (generate):
Traceback (most recent call last):
File "/usr/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
self.run()
File "/usr/lib/python3.10/threading.py", line 953, in run
self._target(*self._args, **self._kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py", line 1874, in generate
self._validate_generated_length(generation_config, input_ids_length, has_default_max_length)
File "/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py", line 1266, in _validate_generated_length
raise ValueError(
ValueError: Input length of input_ids is 31, but `max_length` is set to 20. This can lead to unexpected behavior. You should consider increasing `max_length` or, better yet, setting `max_new_tokens`.
```
In addition, the following error occurs when the number of tokens is
reduced.
```python
for chunk in llm.stream("Hello World"):
print(chunk, end="", flush=True)
```
```
/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py:1258: UserWarning: Using the model-agnostic default `max_length` (=20) to control the generation length. We recommend setting `max_new_tokens` to control the maximum length of the generation.
warnings.warn(
/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py:1885: UserWarning: You are calling .generate() with the `input_ids` being on a device type different than your model's device. `input_ids` is on cpu, whereas the model is on cuda. You may experience unexpected behaviors or slower generation. Please make sure that you have put `input_ids` to the correct device by calling for example input_ids = input_ids.to('cuda') before running `.generate()`.
warnings.warn(
Exception in thread Thread-20 (generate):
Traceback (most recent call last):
File "/usr/lib/python3.10/threading.py", line 1016, in _bootstrap_inner
self.run()
File "/usr/lib/python3.10/threading.py", line 953, in run
self._target(*self._args, **self._kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/utils/_contextlib.py", line 116, in decorate_context
return func(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py", line 2024, in generate
result = self._sample(
File "/usr/local/lib/python3.10/dist-packages/transformers/generation/utils.py", line 2982, in _sample
outputs = self(**model_inputs, return_dict=True)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/transformers/models/gemma2/modeling_gemma2.py", line 994, in forward
outputs = self.model(
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/transformers/models/gemma2/modeling_gemma2.py", line 803, in forward
inputs_embeds = self.embed_tokens(input_ids)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1553, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/module.py", line 1562, in _call_impl
return forward_call(*args, **kwargs)
File "/usr/local/lib/python3.10/dist-packages/torch/nn/modules/sparse.py", line 164, in forward
return F.embedding(
File "/usr/local/lib/python3.10/dist-packages/torch/nn/functional.py", line 2267, in embedding
return torch.embedding(weight, input, padding_idx, scale_grad_by_freq, sparse)
RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cpu! (when checking argument for argument index in method wrapper_CUDA__index_select)
```
On the other hand, in the case of invoke, the output is normal:
```
llm.invoke("Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World.")
```
```
'Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World. Hello World.\n\nThis is a simple program that prints the phrase "Hello World" to the console. \n\n**Here\'s how it works:**\n\n* **`print("Hello World")`**: This line of code uses the `print()` function, which is a built-in function in most programming languages (like Python). The `print()` function takes whatever you put inside its parentheses and displays it on the screen.\n* **`"Hello World"`**: The text within the double quotes (`"`) is called a string. It represents the message we want to print.\n\n\nLet me know if you\'d like to explore other programming concepts or see more examples! \n'
```
### Problem Analysis
- Apparently, I put kwargs in while generating pipelines and it applied
to `invoke()`, but it's not applied in the `stream()`.
- When using the stream, `inputs = self.pipeline.tokenizer (prompt,
return_tensors = "pt")` enters cpu.
- This can crash when the model is in gpu.
### Solution
Just use `self.pipeline` instead of `self.pipeline.model.generate`.
- **Original Code**
```python
stopping_criteria = StoppingCriteriaList([StopOnTokens()])
inputs = self.pipeline.tokenizer(prompt, return_tensors="pt")
streamer = TextIteratorStreamer(
self.pipeline.tokenizer,
timeout=60.0,
skip_prompt=skip_prompt,
skip_special_tokens=True,
)
generation_kwargs = dict(
inputs,
streamer=streamer,
stopping_criteria=stopping_criteria,
**pipeline_kwargs,
)
t1 = Thread(target=self.pipeline.model.generate, kwargs=generation_kwargs)
t1.start()
```
- **Updated Code**
```python
stopping_criteria = StoppingCriteriaList([StopOnTokens()])
streamer = TextIteratorStreamer(
self.pipeline.tokenizer,
timeout=60.0,
skip_prompt=skip_prompt,
skip_special_tokens=True,
)
generation_kwargs = dict(
text_inputs= prompt,
streamer=streamer,
stopping_criteria=stopping_criteria,
**pipeline_kwargs,
)
t1 = Thread(target=self.pipeline, kwargs=generation_kwargs)
t1.start()
```
By using the `pipeline` directly, the `kwargs` of the pipeline are
applied, and there is no need to consider the `device` of the `tensor`
made with the `tokenizer`.
> According to the change to use `pipeline`, it was modified to put
`text_inputs=prompts` directly into `generation_kwargs`.
## Issue
None
## Dependencies
None
## Twitter handle
None
---------
Co-authored-by: Vadym Barda <vadym@langchain.dev>
- [ ] **Description:**
- pass the device_map into model_kwargs
- removing the unused device_map variable in the hf_pipeline function
call
- [ ] **Issue:** issue #13128
When using the from_model_id function to load a Hugging Face model for
text generation across multiple GPUs, the model defaults to loading on
the CPU despite multiple GPUs being available using the expected format
``` python
llm = HuggingFacePipeline.from_model_id(
model_id="model-id",
task="text-generation",
device_map="auto",
)
```
Currently, to enable multiple GPU , we have to pass in variable in this
format instead
``` python
llm = HuggingFacePipeline.from_model_id(
model_id="model-id",
task="text-generation",
device=None,
model_kwargs={
"device_map": "auto",
}
)
```
This issue arises due to improper handling of the device and device_map
parameters.
- [ ] **Explanation:**
1. In from_model_id, the model is created using model_kwargs and passed
as the model variable of the pipeline function. So at this moment, to
load the model with multiple GPUs, "device_map" needs to be set to
"auto" within model_kwargs. Otherwise, the model defaults to loading on
the CPU.
2. The device_map variable in from_model_id is not utilized correctly.
In the pipeline function's source code of tnansformer:
- The device_map variable is stored in the model_kwargs dictionary
(lines 867-878 of transformers/src/transformers/pipelines/\__init__.py).
```python
if device_map is not None:
......
model_kwargs["device_map"] = device_map
```
- The model is constructed with model_kwargs containing the device_map
value ONLY IF it is a string (lines 893-903 of
transformers/src/transformers/pipelines/\__init__.py).
```python
if isinstance(model, str) or framework is None:
model_classes = {"tf": targeted_task["tf"], "pt": targeted_task["pt"]}
framework, model = infer_framework_load_model( ... , **model_kwargs, )
```
- Consequently, since a model object is already passed to the pipeline
function, the device_map variable from from_model_id is never used.
3. The device_map variable in from_model_id not only appears unused but
also causes errors. Without explicitly setting device=None, attempting
to load the model on multiple GPUs may result in the following error:
```
Device has 2 GPUs available. Provide device={deviceId} to
`from_model_id` to use available GPUs for execution. deviceId is -1
(default) for CPU and can be a positive integer associated with CUDA
device id.
Traceback (most recent call last):
File "foo.py", line 15, in <module>
llm = HuggingFacePipeline.from_model_id(
File
"foo\site-packages\langchain_huggingface\llms\huggingface_pipeline.py",
line 217, in from_model_id
pipeline = hf_pipeline(
File "foo\lib\site-packages\transformers\pipelines\__init__.py", line
1108, in pipeline
return pipeline_class(model=model, framework=framework, task=task,
**kwargs)
File "foo\lib\site-packages\transformers\pipelines\text_generation.py",
line 96, in __init__
super().__init__(*args, **kwargs)
File "foo\lib\site-packages\transformers\pipelines\base.py", line 835,
in __init__
raise ValueError(
ValueError: The model has been loaded with `accelerate` and therefore
cannot be moved to a specific device. Please discard the `device`
argument when creating your pipeline object.
```
This error occurs because, in from_model_id, the default values in from_model_id for device and device_map are -1 and None, respectively. It would passes the statement (`device_map is not None and device < 0`) and keep the device as -1 so the pipeline function later raises an error when trying to move a GPU-loaded model back to the CPU.
19eb82e68b/libs/community/langchain_community/llms/huggingface_pipeline.py (L204-L213)
If no one reviews your PR within a few days, please @-mention one of baskaryan, efriis, eyurtsev, ccurme, vbarda, hwchase17.
---------
Co-authored-by: William FH <13333726+hinthornw@users.noreply.github.com>
Co-authored-by: Erick Friis <erick@langchain.dev>
Co-authored-by: vbarda <vadym@langchain.dev>
Backwards compatible change that converts pydantic extras to literals
which is consistent with pydantic 2 usage.
- fireworks
- voyage ai
- mistralai
- mistral ai
- together ai
- huggigng face
- pinecone
First Pr for the langchain_huggingface partner Package
- Moved some of the hugging face related class from `community` to the
new `partner package`
Still needed :
- Documentation
- Tests
- Support for the new apply_chat_template in `ChatHuggingFace`
- Confirm choice of class to support for embeddings witht he
sentence-transformer team.
cc : @efriis
---------
Co-authored-by: Cyril Kondratenko <kkn1993@gmail.com>
Co-authored-by: Erick Friis <erick@langchain.dev>
