SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Add Idefics2 (#30253)

Browse Source 
* Initial add model additions

* Test

* All weights loading

* Can perform full forward pass

* Local and remote the same

* Matching local and remote

* Fixup

* Idefics2Model importable; fixup docstrings

* Don't skip by default

* Remove deprecated use_resampler arg

* Remove self.config

* DecoupledLinear takes config

* Tidy up

* Enable eager attention and tidy up

* Most tests passing

* Update for batch of processed images

* Add image processor

* Update doc pages

* Update conversion script

* Remove erroneous breakpoint

* Remove accidendtal spelling change

* Update to reflect changes on hub - make generate work

* Fix up

* Image processor tests

* Update tests

* Add a processor

* Add a processor

* Update convert script

* Update modeling file - remove fixmes

* Bug fix

* Add processing test

* Use processor

* Fix up

* Update src/transformers/models/idefics2/modeling_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update src/transformers/models/idefics2/modeling_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Fix test

* Update config - PR comments and defaults align with checkpoint

* Reviewer comments

* Add copied froms for flahs attention

* Update src/transformers/models/idefics2/modeling_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Apply suggestions from code review

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Remove qk_layer_norm and freeze_layers functionality

* Fix

* Remove freeze_layer options from config

* Sync with upstream main

* Fix attention shapes siglip

* Remove Llava-next refs - TO REBASE

* Use AutoModel for text model

* Add comment to explain vision embeddings

* Fix issue with tie_word_embeddings

* Address review comments

* Fix and fix up

* Chat templates for idefics

* Fix copies

* Fix

* Add layer norms to FA2

* Fix tests

* Apply suggestions from code review

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Fix

* Review comments

* Update src/transformers/models/idefics2/modeling_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update inputs merger

* Merge weights in correct order

* Update convert script

* Update src/transformers/models/idefics2/processing_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update template

* Model code examples (fix idefics too)

* More review comments

* Tidy up

* Update processing

* Fix attention mask preparation

* Update inputs_merger inputs

* Vectorize inputs_merger

* Update src/transformers/models/idefics2/__init__.py

Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>

* Update src/transformers/models/idefics2/modeling_idefics2.py

* Review comments

* saying bye to the `qk_layer_norms`

* Simplify

* Update latents

* Remove erroneuous readme changes

* Return images when applying chat template

* Fix bug - prompt images are for a single sample

* Update src/transformers/models/idefics2/modeling_idefics2.py

* image splitting

* fix test

* some more comment

* some comment

* Apply suggestions from code review

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update src/transformers/models/idefics2/image_processing_idefics2.py

Co-authored-by: amyeroberts <22614925+amyeroberts@users.noreply.github.com>

* Update processor

* Update model tests

* Update src/transformers/models/idefics2/processing_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update src/transformers/models/idefics2/processing_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Don't add BOS in template

* Update src/transformers/models/idefics2/processing_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Remove index in examples

* Update tests to reflect #13

* Update src/transformers/models/idefics2/processing_idefics2.py

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* PR comment - consistent typing

* Update readme and model doc

* Update docs

* Update checkpoint references

* Update examples

* Fix and update tests

* Small addition

* Update tests - remove copied from as no ignore placement copy could be found

* Update example

* small fixes

* Update docs/source/en/model_doc/idefics2.md

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update docs/source/en/model_doc/idefics2.md

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Update README.md

Co-authored-by: Victor SANH <victorsanh@gmail.com>

* Connector model as bridge

* Fix up

* Fix up

* Don't pass model inputs for generation kwargs update

* IDEFICS-2 -> Idefics2

* Remove config archive name

* IDEFICS-2 -> Idefics2

* Add back llava-next

* Update readmes

* Add requirements for processor tester

* Use custom convert_to_rgb to avoid possible BC

* Fix doc example

* Fix doc example

* Skip model doc tests - as model to large

* More doc example - account for image splitting

* Update src/transformers/image_transforms.py

* Fix config doctest

---------

Co-authored-by: Pablo Montalvo <39954772+molbap@users.noreply.github.com>
Co-authored-by: ArthurZucker <arthur.zucker@gmail.com>
Co-authored-by: Victor SANH <victorsanh@gmail.com>
Co-authored-by: Arthur <48595927+ArthurZucker@users.noreply.github.com>
This commit is contained in:
amyeroberts
2024-04-15 17:03:03 +01:00
committed by GitHub
parent 667939a2d3
commit 6b78360e6d
41 changed files with 4692 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
docs/source/en/perf_infer_gpu_one.md
View File

@@ -47,6 +47,7 @@ FlashAttention-2 is currently supported for the following architectures:
* [GPTNeo](https://huggingface.co/docs/transformers/model_doc/gpt_neo#transformers.GPTNeoModel)
* [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)
* [GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj#transformers.GPTJModel)
* [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
* [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
* [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
* [Llava](https://huggingface.co/docs/transformers/model_doc/llava)
@@ -96,8 +97,8 @@ model_id = "tiiuae/falcon-7b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
model_id,
torch_dtype=torch.bfloat16,
model_id,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
)
```
@@ -109,7 +110,7 @@ FlashAttention-2 can only be used when the model's dtype is `fp16` or `bf16`. Ma
<br>
You can also set `use_flash_attention_2=True` to enable FlashAttention-2 but it is deprecated in favor of `attn_implementation="flash_attention_2"`.
</Tip>
FlashAttention-2 can be combined with other optimization techniques like quantization to further speedup inference. For example, you can combine FlashAttention-2 with 8-bit or 4-bit quantization:
@@ -123,14 +124,14 @@ tokenizer = AutoTokenizer.from_pretrained(model_id)
# load in 8bit
model = AutoModelForCausalLM.from_pretrained(
model_id,
model_id,
load_in_8bit=True,
attn_implementation="flash_attention_2",
)
# load in 4bit
model = AutoModelForCausalLM.from_pretrained(
model_id,
model_id,
load_in_4bit=True,
attn_implementation="flash_attention_2",
)