Skip to content

GitLab

Unverified Commit bce5e224 authored by Nicolas Patry's avatar Nicolas Patry Committed by GitHub
Browse files

Adding Idefics multi modal model. (#842) 


Co-Authored-By: default avatarVictor Sanh <victorsanh@gmail.com>


# What does this PR do?

<!--
Congratulations! You've made it this far! You're not quite done yet
though.

Once merged, your PR is going to appear in the release notes with the
title you set, so make sure it's a great title that fully reflects the
extent of your awesome contribution.

Then, please replace this with a description of the change and which
issue is fixed (if applicable). Please also include relevant motivation
and context. List any dependencies (if any) that are required for this
change.

Once you're done, someone will review your PR shortly (see the section
"Who can review?" below to tag some potential reviewers). They may
suggest changes to make the code even better. If no one reviewed your PR
after a week has passed, don't hesitate to post a new comment
@-mentioning the same persons---sometimes notifications get lost.
-->

<!-- Remove if not applicable -->

Fixes # (issue)


## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Did you read the [contributor
guideline](https://github.com/huggingface/transformers/blob/main/CONTRIBUTING.md#start-contributing-pull-requests),
      Pull Request section?
- [ ] Was this discussed/approved via a Github issue or the
[forum](https://discuss.huggingface.co/)? Please add a link
      to it if that's the case.
- [ ] Did you make sure to update the documentation with your changes?
Here are the
[documentation
guidelines](https://github.com/huggingface/transformers/tree/main/docs),
and
[here are tips on formatting
docstrings](https://github.com/huggingface/transformers/tree/main/docs#writing-source-documentation

).
- [ ] Did you write any new necessary tests?


## Who can review?

Anyone in the community is free to review the PR once the tests have
passed. Feel free to tag
members/contributors who may be interested in your PR.

<!-- Your PR will be replied to more quickly if you can figure out the
right person to tag with @


@OlivierDehaene OR @Narsil

 -->

---------
Co-authored-by: default avatarVictor Sanh <victorsanh@gmail.com>
parent b9e33c49
Expand all Hide whitespace changes
Inline Side-by-side
Showing with 5227 additions and 331 deletions
integration-tests/models/__snapshots__/test_idefics/test_idefics.json 0 → 100644
View file @ bce5e224
{
"details": {
"best_of_sequences": null,
"finish_reason": "length",
"generated_tokens": 10,
"prefill": [
{
"id": 1,
"logprob": null,
"text": "<s>"
},
{
"id": 4911,
"logprob": -5.3632812,
"text": "User"
},
{
"id": 29901,
"logprob": -0.00762558,
"text": ":"
},
{
"id": 32000,
"logprob": -0.7739258,
"text": "<fake_token_around_image>"
},
{
"id": 32001,
"logprob": -9.775162e-05,
"text": "<image>"
},
{
"id": 32000,
"logprob": -1.1920929e-07,
"text": "<fake_token_around_image>"
},
{
"id": 1815,
"logprob": -4.4140625,
"text": "Can"
},
{
"id": 366,
"logprob": -0.01436615,
"text": "you"
},
{
"id": 2649,
"logprob": -4.9414062,
"text": "tell"
},
{
"id": 592,
"logprob": -0.3005371,
"text": "me"
},
{
"id": 263,
"logprob": -3.5703125,
"text": "a"
},
{
"id": 1407,
"logprob": -9.4296875,
"text": "very"
},
{
"id": 3273,
"logprob": -1.9111328,
"text": "short"
},
{
"id": 5828,
"logprob": -0.28881836,
"text": "story"
},
{
"id": 2729,
"logprob": -3.4179688,
"text": "based"
},
{
"id": 373,
"logprob": -0.00056886673,
"text": "on"
},
{
"id": 278,
"logprob": -0.14123535,
"text": "the"
},
{
"id": 1967,
"logprob": -0.053985596,
"text": "image"
},
{
"id": 29973,
"logprob": -0.15771484,
"text": "?"
}
],
"seed": null,
"tokens": [
{
"id": 32002,
"logprob": -0.004295349,
"special": true,
"text": "<end_of_utterance>"
},
{
"id": 29871,
"logprob": -7.43866e-05,
"special": false,
"text": " "
},
{
"id": 13,
"logprob": -2.3126602e-05,
"special": false,
"text": "\n"
},
{
"id": 7900,
"logprob": -3.9339066e-06,
"special": false,
"text": "Ass"
},
{
"id": 22137,
"logprob": 0.0,
"special": false,
"text": "istant"
},
{
"id": 29901,
"logprob": -2.6226044e-06,
"special": false,
"text": ":"
},
{
"id": 319,
"logprob": -0.87841797,
"special": false,
"text": " A"
},
{
"id": 521,
"logprob": -1.3837891,
"special": false,
"text": " ch"
},
{
"id": 21475,
"logprob": -0.00051641464,
"special": false,
"text": "icken"
},
{
"id": 338,
"logprob": -1.1435547,
"special": false,
"text": " is"
}
]
},
"generated_text": "\nAssistant: A chicken is"
}
integration-tests/models/__snapshots__/test_idefics/test_idefics_load.json 0 → 100644
View file @ bce5e224
[
{
"details": {
"best_of_sequences": null,
"finish_reason": "length",
"generated_tokens": 10,
"prefill": [
{
"id": 1,
"logprob": null,
"text": "<s>"
},
{
"id": 4911,
"logprob": -5.3476562,
"text": "User"
},
{
"id": 29901,
"logprob": -0.0075531006,
"text": ":"
},
{
"id": 32000,
"logprob": -0.7729492,
"text": "<fake_token_around_image>"
},
{
"id": 32001,
"logprob": -9.787083e-05,
"text": "<image>"
},
{
"id": 32000,
"logprob": -1.1920929e-07,
"text": "<fake_token_around_image>"
},
{
"id": 1815,
"logprob": -4.4296875,
"text": "Can"
},
{
"id": 366,
"logprob": -0.01424408,
"text": "you"
},
{
"id": 2649,
"logprob": -4.9335938,
"text": "tell"
},
{
"id": 592,
"logprob": -0.2993164,
"text": "me"
},
{
"id": 263,
"logprob": -3.5664062,
"text": "a"
},
{
"id": 1407,
"logprob": -9.4453125,
"text": "very"
},
{
"id": 3273,
"logprob": -1.9306641,
"text": "short"
},
{
"id": 5828,
"logprob": -0.2836914,
"text": "story"
},
{
"id": 2729,
"logprob": -3.4179688,
"text": "based"
},
{
"id": 373,
"logprob": -0.00056934357,
"text": "on"
},
{
"id": 278,
"logprob": -0.13928223,
"text": "the"
},
{
"id": 1967,
"logprob": -0.05355835,
"text": "image"
},
{
"id": 29973,
"logprob": -0.15771484,
"text": "?"
}
],
"seed": null,
"tokens": [
{
"id": 32002,
"logprob": -0.004333496,
"special": true,
"text": "<end_of_utterance>"
},
{
"id": 29871,
"logprob": -7.426739e-05,
"special": false,
"text": " "
},
{
"id": 13,
"logprob": -2.348423e-05,
"special": false,
"text": "\n"
},
{
"id": 7900,
"logprob": -3.9339066e-06,
"special": false,
"text": "Ass"
},
{
"id": 22137,
"logprob": 0.0,
"special": false,
"text": "istant"
},
{
"id": 29901,
"logprob": -2.861023e-06,
"special": false,
"text": ":"
},
{
"id": 319,
"logprob": -0.8828125,
"special": false,
"text": " A"
},
{
"id": 521,
"logprob": -1.3759766,
"special": false,
"text": " ch"
},
{
"id": 21475,
"logprob": -0.0005083084,
"special": false,
"text": "icken"
},
{
"id": 338,
"logprob": -1.1367188,
"special": false,
"text": " is"
}
]
},
"generated_text": "\nAssistant: A chicken is"
},
{
"details": {
"best_of_sequences": null,
"finish_reason": "length",
"generated_tokens": 10,
"prefill": [
{
"id": 1,
"logprob": null,
"text": "<s>"
},
{
"id": 4911,
"logprob": -5.3476562,
"text": "User"
},
{
"id": 29901,
"logprob": -0.0075531006,
"text": ":"
},
{
"id": 32000,
"logprob": -0.7729492,
"text": "<fake_token_around_image>"
},
{
"id": 32001,
"logprob": -9.787083e-05,
"text": "<image>"
},
{
"id": 32000,
"logprob": -1.1920929e-07,
"text": "<fake_token_around_image>"
},
{
"id": 1815,
"logprob": -4.4296875,
"text": "Can"
},
{
"id": 366,
"logprob": -0.01423645,
"text": "you"
},
{
"id": 2649,
"logprob": -4.9335938,
"text": "tell"
},
{
"id": 592,
"logprob": -0.2993164,
"text": "me"
},
{
"id": 263,
"logprob": -3.5664062,
"text": "a"
},
{
"id": 1407,
"logprob": -9.4453125,
"text": "very"
},
{
"id": 3273,
"logprob": -1.9306641,
"text": "short"
},
{
"id": 5828,
"logprob": -0.2836914,
"text": "story"
},
{
"id": 2729,
"logprob": -3.4179688,
"text": "based"
},
{
"id": 373,
"logprob": -0.00056934357,
"text": "on"
},
{
"id": 278,
"logprob": -0.13928223,
"text": "the"
},
{
"id": 1967,
"logprob": -0.05355835,
"text": "image"
},
{
"id": 29973,
"logprob": -0.15771484,
"text": "?"
}
],
"seed": null,
"tokens": [
{
"id": 32002,
"logprob": -0.004333496,
"special": true,
"text": "<end_of_utterance>"
},
{
"id": 29871,
"logprob": -7.4505806e-05,
"special": false,
"text": " "
},
{
"id": 13,
"logprob": -2.3722649e-05,
"special": false,
"text": "\n"
},
{
"id": 7900,
"logprob": -3.9339066e-06,
"special": false,
"text": "Ass"
},
{
"id": 22137,
"logprob": 0.0,
"special": false,
"text": "istant"
},
{
"id": 29901,
"logprob": -2.861023e-06,
"special": false,
"text": ":"
},
{
"id": 319,
"logprob": -0.8828125,
"special": false,
"text": " A"
},
{
"id": 521,
"logprob": -1.3759766,
"special": false,
"text": " ch"
},
{
"id": 21475,
"logprob": -0.00050878525,
"special": false,
"text": "icken"
},
{
"id": 338,
"logprob": -1.1367188,
"special": false,
"text": " is"
}
]
},
"generated_text": "\nAssistant: A chicken is"
},
{
"details": {
"best_of_sequences": null,
"finish_reason": "length",
"generated_tokens": 10,
"prefill": [
{
"id": 1,
"logprob": null,
"text": "<s>"
},
{
"id": 4911,
"logprob": -5.3476562,
"text": "User"
},
{
"id": 29901,
"logprob": -0.0075531006,
"text": ":"
},
{
"id": 32000,
"logprob": -0.7729492,
"text": "<fake_token_around_image>"
},
{
"id": 32001,
"logprob": -9.775162e-05,
"text": "<image>"
},
{
"id": 32000,
"logprob": -1.1920929e-07,
"text": "<fake_token_around_image>"
},
{
"id": 1815,
"logprob": -4.4296875,
"text": "Can"
},
{
"id": 366,
"logprob": -0.01424408,
"text": "you"
},
{
"id": 2649,
"logprob": -4.9335938,
"text": "tell"
},
{
"id": 592,
"logprob": -0.2993164,
"text": "me"
},
{
"id": 263,
"logprob": -3.5664062,
"text": "a"
},
{
"id": 1407,
"logprob": -9.4453125,
"text": "very"
},
{
"id": 3273,
"logprob": -1.9306641,
"text": "short"
},
{
"id": 5828,
"logprob": -0.2836914,
"text": "story"
},
{
"id": 2729,
"logprob": -3.4179688,
"text": "based"
},
{
"id": 373,
"logprob": -0.00056934357,
"text": "on"
},
{
"id": 278,
"logprob": -0.13928223,
"text": "the"
},
{
"id": 1967,
"logprob": -0.05355835,
"text": "image"
},
{
"id": 29973,
"logprob": -0.15771484,
"text": "?"
}
],
"seed": null,
"tokens": [
{
"id": 32002,
"logprob": -0.004333496,
"special": true,
"text": "<end_of_utterance>"
},
{
"id": 29871,
"logprob": -7.43866e-05,
"special": false,
"text": " "
},
{
"id": 13,
"logprob": -2.360344e-05,
"special": false,
"text": "\n"
},
{
"id": 7900,
"logprob": -3.9339066e-06,
"special": false,
"text": "Ass"
},
{
"id": 22137,
"logprob": 0.0,
"special": false,
"text": "istant"
},
{
"id": 29901,
"logprob": -2.7418137e-06,
"special": false,
"text": ":"
},
{
"id": 319,
"logprob": -0.8828125,
"special": false,
"text": " A"
},
{
"id": 521,
"logprob": -1.3759766,
"special": false,
"text": " ch"
},
{
"id": 21475,
"logprob": -0.00050878525,
"special": false,
"text": "icken"
},
{
"id": 338,
"logprob": -1.1367188,
"special": false,
"text": " is"
}
]
},
"generated_text": "\nAssistant: A chicken is"
},
{
"details": {
"best_of_sequences": null,
"finish_reason": "length",
"generated_tokens": 10,
"prefill": [
{
"id": 1,
"logprob": null,
"text": "<s>"
},
{
"id": 4911,
"logprob": -5.3632812,
"text": "User"
},
{
"id": 29901,
"logprob": -0.00762558,
"text": ":"
},
{
"id": 32000,
"logprob": -0.7739258,
"text": "<fake_token_around_image>"
},
{
"id": 32001,
"logprob": -9.775162e-05,
"text": "<image>"
},
{
"id": 32000,
"logprob": -1.1920929e-07,
"text": "<fake_token_around_image>"
},
{
"id": 1815,
"logprob": -4.4140625,
"text": "Can"
},
{
"id": 366,
"logprob": -0.01436615,
"text": "you"
},
{
"id": 2649,
"logprob": -4.9414062,
"text": "tell"
},
{
"id": 592,
"logprob": -0.3005371,
"text": "me"
},
{
"id": 263,
"logprob": -3.5703125,
"text": "a"
},
{
"id": 1407,
"logprob": -9.4296875,
"text": "very"
},
{
"id": 3273,
"logprob": -1.9111328,
"text": "short"
},
{
"id": 5828,
"logprob": -0.28881836,
"text": "story"
},
{
"id": 2729,
"logprob": -3.4179688,
"text": "based"
},
{
"id": 373,
"logprob": -0.00056886673,
"text": "on"
},
{
"id": 278,
"logprob": -0.14123535,
"text": "the"
},
{
"id": 1967,
"logprob": -0.053985596,
"text": "image"
},
{
"id": 29973,
"logprob": -0.15771484,
"text": "?"
}
],
"seed": null,
"tokens": [
{
"id": 32002,
"logprob": -0.004295349,
"special": true,
"text": "<end_of_utterance>"
},
{
"id": 29871,
"logprob": -7.43866e-05,
"special": false,
"text": " "
},
{
"id": 13,
"logprob": -2.3126602e-05,
"special": false,
"text": "\n"
},
{
"id": 7900,
"logprob": -3.9339066e-06,
"special": false,
"text": "Ass"
},
{
"id": 22137,
"logprob": 0.0,
"special": false,
"text": "istant"
},
{
"id": 29901,
"logprob": -2.6226044e-06,
"special": false,
"text": ":"
},
{
"id": 319,
"logprob": -0.87841797,
"special": false,
"text": " A"
},
{
"id": 521,
"logprob": -1.3837891,
"special": false,
"text": " ch"
},
{
"id": 21475,
"logprob": -0.00051641464,
"special": false,
"text": "icken"
},
{
"id": 338,
"logprob": -1.1435547,
"special": false,
"text": " is"
}
]
},
"generated_text": "\nAssistant: A chicken is"
}
]
integration-tests/models/test_idefics.py 0 → 100644
View file @ bce5e224
import pytest
@pytest.fixture(scope="module")
def idefics_handle(launcher):
with launcher(
"HuggingFaceM4/idefics-9b-instruct", num_shard=2
) as handle:
yield handle
@pytest.fixture(scope="module")
async def idefics(idefics_handle):
await idefics_handle.health(300)
return idefics_handle.client
@pytest.mark.asyncio
async def test_idefics(idefics, response_snapshot):
response = await idefics.generate(
"User:![](https://temp-5681.s3.us-west-2.amazonaws.com/chicken_on_money.png)Can you tell me a very short story based on the image?",
max_new_tokens=10,
decoder_input_details=True,
)
assert response.details.generated_tokens == 10
assert response == response_snapshot
@pytest.mark.asyncio
async def test_idefics_load(idefics, generate_load, response_snapshot):
responses = await generate_load(
idefics,
"User:![](https://temp-5681.s3.us-west-2.amazonaws.com/chicken_on_money.png)Can you tell me a very short story based on the image?",
max_new_tokens=10,
n=4,
)
generated_texts = [r.generated_text for r in responses]
assert len(generated_texts) == 4
assert generated_texts, all(
[text == generated_texts[0] for text in generated_texts]
)
assert responses == response_snapshot
server/poetry.lock
View file @ bce5e224
This diff is collapsed.
server/pyproject.toml
View file @ bce5e224
...@@ -31,8 +31,9 @@ einops = "^0.6.1" ...@@ -31,8 +31,9 @@ einops = "^0.6.1"
texttable = { version = "^1.6.7", optional = true } texttable = { version = "^1.6.7", optional = true }
datasets = { version = "^2.14.0", optional = true } datasets = { version = "^2.14.0", optional = true }
peft = "^0.4.0" peft = "^0.4.0"
torch = {version = "^2.0.1+cu118", source = "pytorch-gpu-src"} torch = { version = "^2.0.1" }
scipy = "^1.11.1" scipy = "^1.11.1"
pillow = "^10.0.0"
[tool.poetry.extras] [tool.poetry.extras]
accelerate = ["accelerate"] accelerate = ["accelerate"]
......
server/requirements.txt
View file @ bce5e224
--extra-index-url https://download.pytorch.org/whl/cu118
accelerate==0.20.3 ; python_version >= "3.9" and python_version < "3.13" accelerate==0.20.3 ; python_version >= "3.9" and python_version < "3.13"
backoff==2.2.1 ; python_version >= "3.9" and python_version < "3.13" backoff==2.2.1 ; python_version >= "3.9" and python_version < "3.13"
certifi==2023.7.22 ; python_version >= "3.9" and python_version < "3.13" certifi==2023.7.22 ; python_version >= "3.9" and python_version < "3.13"
charset-normalizer==3.2.0 ; python_version >= "3.9" and python_version < "3.13" charset-normalizer==3.2.0 ; python_version >= "3.9" and python_version < "3.13"
click==8.1.6 ; python_version >= "3.9" and python_version < "3.13" click==8.1.6 ; python_version >= "3.9" and python_version < "3.13"
cmake==3.27.0 ; platform_system == "Linux" and platform_machine == "x86_64" and python_version >= "3.9" and python_version < "3.13"
colorama==0.4.6 ; python_version >= "3.9" and python_version < "3.13" and (sys_platform == "win32" or platform_system == "Windows") colorama==0.4.6 ; python_version >= "3.9" and python_version < "3.13" and (sys_platform == "win32" or platform_system == "Windows")
deprecated==1.2.14 ; python_version >= "3.9" and python_version < "3.13" deprecated==1.2.14 ; python_version >= "3.9" and python_version < "3.13"
einops==0.6.1 ; python_version >= "3.9" and python_version < "3.13" einops==0.6.1 ; python_version >= "3.9" and python_version < "3.13"
...@@ -13,14 +10,13 @@ filelock==3.12.2 ; python_version >= "3.9" and python_version < "3.13" ...@@ -13,14 +10,13 @@ filelock==3.12.2 ; python_version >= "3.9" and python_version < "3.13"
fsspec==2023.6.0 ; python_version >= "3.9" and python_version < "3.13" fsspec==2023.6.0 ; python_version >= "3.9" and python_version < "3.13"
googleapis-common-protos==1.60.0 ; python_version >= "3.9" and python_version < "3.13" googleapis-common-protos==1.60.0 ; python_version >= "3.9" and python_version < "3.13"
grpc-interceptor==0.15.2 ; python_version >= "3.9" and python_version < "3.13" grpc-interceptor==0.15.2 ; python_version >= "3.9" and python_version < "3.13"
grpcio-reflection==1.56.2 ; python_version >= "3.9" and python_version < "3.13" grpcio-reflection==1.57.0 ; python_version >= "3.9" and python_version < "3.13"
grpcio-status==1.56.2 ; python_version >= "3.9" and python_version < "3.13" grpcio-status==1.57.0 ; python_version >= "3.9" and python_version < "3.13"
grpcio==1.56.2 ; python_version >= "3.9" and python_version < "3.13" grpcio==1.57.0 ; python_version >= "3.9" and python_version < "3.13"
hf-transfer==0.1.3 ; python_version >= "3.9" and python_version < "3.13" hf-transfer==0.1.3 ; python_version >= "3.9" and python_version < "3.13"
huggingface-hub==0.14.1 ; python_version >= "3.9" and python_version < "3.13" huggingface-hub==0.14.1 ; python_version >= "3.9" and python_version < "3.13"
idna==3.4 ; python_version >= "3.9" and python_version < "3.13" idna==3.4 ; python_version >= "3.9" and python_version < "3.13"
jinja2==3.1.2 ; python_version >= "3.9" and python_version < "3.13" jinja2==3.1.2 ; python_version >= "3.9" and python_version < "3.13"
lit==16.0.6 ; platform_system == "Linux" and platform_machine == "x86_64" and python_version >= "3.9" and python_version < "3.13"
loguru==0.6.0 ; python_version >= "3.9" and python_version < "3.13" loguru==0.6.0 ; python_version >= "3.9" and python_version < "3.13"
markupsafe==2.1.3 ; python_version >= "3.9" and python_version < "3.13" markupsafe==2.1.3 ; python_version >= "3.9" and python_version < "3.13"
mpmath==1.3.0 ; python_version >= "3.9" and python_version < "3.13" mpmath==1.3.0 ; python_version >= "3.9" and python_version < "3.13"
...@@ -37,10 +33,11 @@ opentelemetry-sdk==1.15.0 ; python_version >= "3.9" and python_version < "3.13" ...@@ -37,10 +33,11 @@ opentelemetry-sdk==1.15.0 ; python_version >= "3.9" and python_version < "3.13"
opentelemetry-semantic-conventions==0.36b0 ; python_version >= "3.9" and python_version < "3.13" opentelemetry-semantic-conventions==0.36b0 ; python_version >= "3.9" and python_version < "3.13"
packaging==23.1 ; python_version >= "3.9" and python_version < "3.13" packaging==23.1 ; python_version >= "3.9" and python_version < "3.13"
peft==0.4.0 ; python_version >= "3.9" and python_version < "3.13" peft==0.4.0 ; python_version >= "3.9" and python_version < "3.13"
protobuf==4.23.4 ; python_version >= "3.9" and python_version < "3.13" pillow==10.0.0 ; python_version >= "3.9" and python_version < "3.13"
protobuf==4.24.0 ; python_version >= "3.9" and python_version < "3.13"
psutil==5.9.5 ; python_version >= "3.9" and python_version < "3.13" psutil==5.9.5 ; python_version >= "3.9" and python_version < "3.13"
pyyaml==6.0.1 ; python_version >= "3.9" and python_version < "3.13" pyyaml==6.0.1 ; python_version >= "3.9" and python_version < "3.13"
regex==2023.6.3 ; python_version >= "3.9" and python_version < "3.13" regex==2023.8.8 ; python_version >= "3.9" and python_version < "3.13"
requests==2.31.0 ; python_version >= "3.9" and python_version < "3.13" requests==2.31.0 ; python_version >= "3.9" and python_version < "3.13"
safetensors==0.3.2 ; python_version >= "3.9" and python_version < "3.13" safetensors==0.3.2 ; python_version >= "3.9" and python_version < "3.13"
scipy==1.11.1 ; python_version >= "3.9" and python_version < "3.13" scipy==1.11.1 ; python_version >= "3.9" and python_version < "3.13"
...@@ -48,10 +45,9 @@ sentencepiece==0.1.99 ; python_version >= "3.9" and python_version < "3.13" ...@@ -48,10 +45,9 @@ sentencepiece==0.1.99 ; python_version >= "3.9" and python_version < "3.13"
setuptools==68.0.0 ; python_version >= "3.9" and python_version < "3.13" setuptools==68.0.0 ; python_version >= "3.9" and python_version < "3.13"
sympy==1.12 ; python_version >= "3.9" and python_version < "3.13" sympy==1.12 ; python_version >= "3.9" and python_version < "3.13"
tokenizers==0.13.3 ; python_version >= "3.9" and python_version < "3.13" tokenizers==0.13.3 ; python_version >= "3.9" and python_version < "3.13"
torch==2.0.1+cu118 ; python_version >= "3.9" and python_version < "3.13" torch==2.0.1 ; python_version >= "3.9" and python_version < "3.13"
tqdm==4.65.0 ; python_version >= "3.9" and python_version < "3.13" tqdm==4.66.1 ; python_version >= "3.9" and python_version < "3.13"
transformers==4.31.0 ; python_version >= "3.9" and python_version < "3.13" transformers==4.31.0 ; python_version >= "3.9" and python_version < "3.13"
triton==2.0.0 ; platform_system == "Linux" and platform_machine == "x86_64" and python_version >= "3.9" and python_version < "3.13"
typer==0.6.1 ; python_version >= "3.9" and python_version < "3.13" typer==0.6.1 ; python_version >= "3.9" and python_version < "3.13"
typing-extensions==4.7.1 ; python_version >= "3.9" and python_version < "3.13" typing-extensions==4.7.1 ; python_version >= "3.9" and python_version < "3.13"
urllib3==2.0.4 ; python_version >= "3.9" and python_version < "3.13" urllib3==2.0.4 ; python_version >= "3.9" and python_version < "3.13"
......
server/text_generation_server/models/__init__.py
View file @ bce5e224
...@@ -54,6 +54,7 @@ try: ...@@ -54,6 +54,7 @@ try:
from text_generation_server.models.flash_santacoder import ( from text_generation_server.models.flash_santacoder import (
FlashSantacoderSharded, FlashSantacoderSharded,
) )
from text_generation_server.models.idefics import IDEFICSSharded
except ImportError as e: except ImportError as e:
logger.warning(f"Could not import Flash Attention enabled models: {e}") logger.warning(f"Could not import Flash Attention enabled models: {e}")
...@@ -64,6 +65,7 @@ if FLASH_ATTENTION: ...@@ -64,6 +65,7 @@ if FLASH_ATTENTION:
__all__.append(FlashRWSharded) __all__.append(FlashRWSharded)
__all__.append(FlashSantacoderSharded) __all__.append(FlashSantacoderSharded)
__all__.append(FlashLlama) __all__.append(FlashLlama)
__all__.append(IDEFICSSharded)
def get_model( def get_model(
...@@ -248,6 +250,17 @@ def get_model( ...@@ -248,6 +250,17 @@ def get_model(
dtype=dtype, dtype=dtype,
trust_remote_code=trust_remote_code, trust_remote_code=trust_remote_code,
) )
elif model_type == "idefics":
if FLASH_ATTENTION:
return IDEFICSSharded(
model_id,
revision,
quantize=quantize,
dtype=dtype,
trust_remote_code=trust_remote_code,
)
else:
raise NotImplementedError(FLASH_ATT_ERROR_MESSAGE.format("Idefics"))
if sharded: if sharded:
raise ValueError("sharded is not supported for AutoModel") raise ValueError("sharded is not supported for AutoModel")
......
server/text_generation_server/models/custom_modeling/idefics_config.py 0 → 100644
View file @ bce5e224
# 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.
""" Idefics model configuration"""
import copy
from transformers import PretrainedConfig
IDEFICS_PRETRAINED_CONFIG_ARCHIVE_MAP = {
"HuggingFaceM4/idefics-9b": "https://huggingface.co/HuggingFaceM4/idefics-9b/blob/main/config.json",
"HuggingFaceM4/idefics-80b": "https://huggingface.co/HuggingFaceM4/idefics-80b/blob/main/config.json",
}
class IdeficsVisionConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`IdeficsModel`]. It is used to instantiate an
Idefics model according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of the Idefics-9B.
e.g. [HuggingFaceM4/idefics-9b](https://huggingface.co/HuggingFaceM4/idefics-9b)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
hidden_size (`int`, *optional*, defaults to 768):
Dimensionality of the encoder layers and the pooler layer. (elsewhere referred to as `hidden_size`)
image_size (`int`, *optional*, defaults to 224):
The size (resolution) of each image.
intermediate_size (`int`, *optional*, defaults to 5120):
Dimensionality of the "intermediate" (i.e., feed-forward) layer in the Transformer encoder.
patch_size (`int`, *optional*, defaults to 14):
The size (resolution) of each patch.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 16):
Number of attention heads for each attention layer in the Transformer encoder.
image_num_channels (`int`, *optional*, defaults to `3`):
Number of image channels.
hidden_act (`str` or `function`, *optional*, defaults to `"quick_gelu"`):
The non-linear activation function (function or string) in the encoder and pooler. If string, `"gelu"`,
`"relu"`, `"selu"` and `"gelu_new"` ``"quick_gelu"` are supported.
layer_norm_eps (`float`, *optional*, defaults to 1e-5):
The epsilon used by the layer normalization layers.
attention_dropout (`float`, *optional*, defaults to 0.0):
The dropout ratio for the attention probabilities.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
initializer_factor (`float`, *optional*, defaults to 1.0):
A factor for initializing all weight matrices (should be kept to 1.0, used internally for initialization
testing).
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
"""
model_type = "idefics"
attribute_map = {
"hidden_size": "embed_dim",
}
def __init__(
self,
embed_dim=768,
image_size=224,
intermediate_size=5120,
patch_size=14,
num_hidden_layers=32,
num_attention_heads=16,
num_channels=3,
hidden_act="quick_gelu",
layer_norm_eps=1e-5,
attention_dropout=0.0,
initializer_range=0.02,
initializer_factor=1.0,
**kwargs,
):
self.embed_dim = embed_dim
self.image_size = image_size
self.intermediate_size = intermediate_size
self.patch_size = patch_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.num_channels = num_channels
self.layer_norm_eps = layer_norm_eps
self.attention_dropout = attention_dropout
self.initializer_range = initializer_range
self.initializer_factor = initializer_factor
self.hidden_act = hidden_act
super().__init__(**kwargs)
class IdeficsPerceiverConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`IdeficsModel`]. It is used to instantiate an
Idefics model according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of the Idefics-9B.
e.g. [HuggingFaceM4/idefics-9b](https://huggingface.co/HuggingFaceM4/idefics-9b)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
use_resampler (`bool`, *optional*, defaults to `False`):
Whether or not to use the resampler
resampler_n_latents (`int`, *optional*, defaults to ):
Number of latent embeddings to resample ("compress") the input sequence to (usually < 128).
resampler_depth (`int`, *optional*, defaults to 6):
Depth of the Perceiver Resampler (Transformer w/ cross attention). Should be shallow (< 3).
resampler_n_heads (`int`, *optional*, defaults to 16):
Number of heads in each Transformer block (for multi-headed self-attention).
resampler_head_dim (`int`, *optional*, defaults to 96):
Dimensionality of each head projection in the Transformer block.
qk_layer_norms_perceiver (`bool`, *optional*, defaults to `False`):
Whether or not to use qk layer norms in perceiver
"""
model_type = "idefics"
def __init__(
self,
use_resampler=False,
resampler_n_latents=64,
resampler_depth=6,
resampler_n_heads=16,
resampler_head_dim=96,
qk_layer_norms_perceiver=False,
**kwargs,
):
self.use_resampler = use_resampler
self.resampler_n_latents = resampler_n_latents
self.resampler_depth = resampler_depth
self.resampler_n_heads = resampler_n_heads
self.resampler_head_dim = resampler_head_dim
self.qk_layer_norms_perceiver = qk_layer_norms_perceiver
super().__init__(**kwargs)
class IdeficsConfig(PretrainedConfig):
r"""
This is the configuration class to store the configuration of a [`IdeficsModel`]. It is used to instantiate an
Idefics model according to the specified arguments, defining the model architecture. Instantiating a configuration
with the defaults will yield a similar configuration to that of the Idefics-9B.
e.g. [HuggingFaceM4/idefics-9b](https://huggingface.co/HuggingFaceM4/idefics-9b)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:
additional_vocab_size (`int`, *optional`, defaults to 0):
Additional vocabulary size of the model, typically for the special "<img>" token. Additional vocab tokens
are always trainable whereas regular vocab tokens can be frozen or not.
vocab_size (`int`, *optional*, defaults to 32000):
Vocabulary size of the Idefics model. Defines the number of different tokens that can be represented by the
`inputs_ids` passed when calling [`~IdeficsModel`]
hidden_size (`int`, *optional*, defaults to 4096):
Dimension of the hidden representations.
intermediate_size (`int`, *optional*, defaults to 11008):
Dimension of the MLP representations.
num_hidden_layers (`int`, *optional*, defaults to 32):
Number of hidden layers in the Transformer encoder.
num_attention_heads (`int`, *optional*, defaults to 32):
Number of attention heads for each attention layer in the Transformer encoder.
dropout (`float`, *optional*, defaults to 0.0):
The dropout probability for all fully connected layers in the embeddings, encoder, and pooler.
hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
The non-linear activation function (function or string) in the decoder.
initializer_range (`float`, *optional*, defaults to 0.02):
The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
alpha_initializer (`str`, *optional*, defaults to `"zeros"`):
Initialization type for the alphas.
alphas_initializer_range (`float`, *optional*, defaults to 0.0):
The standard deviation of the truncated_normal_initializer for initializing the alphas in the Gated Cross
Attention.
alpha_type (`str`, *optional*, defaults to `"float"`):
Whether the gating alphas should be vectors or single floats.
rms_norm_eps (`float`, *optional*, defaults to 1e-6):
The epsilon used by the rms normalization layers.
use_cache (`bool`, *optional*, defaults to `True`):
Whether or not the model should return the last key/values attentions (not used by all models). Only
relevant if `config.is_decoder=True`.
pad_token_id (`int`, *optional*, defaults to 0)
Padding token id.
bos_token_id (`int`, *optional*, defaults to 1)
Beginning of stream token id.
eos_token_id (`int`, *optional*, defaults to 2)
End of stream token id.
tie_word_embeddings(`bool`, *optional*, defaults to `False`):
Whether to tie weight embeddings
cross_layer_interval (`int`, *optional*, default to 1)
Interval for cross attention (from text to image) layers.
qk_layer_norms (`bool`, *optional*, defaults to `False`): Whether to add layer norm after q and k
freeze_text_layers (`bool`, *optional*, defaults to `True`): Whether to freeze text layers
freeze_text_module_exceptions (`bool`, *optional*, defaults to `[]`):
Exceptions to freezing text layers when `freeze_text_layers` is `True`
freeze_lm_head (`bool`, *optional*, defaults to `False`): Whether to freeze lm head
freeze_vision_layers (`bool`, *optional*, defaults to `True`): Whether to freeze vision layers
freeze_vision_module_exceptions (`bool`, *optional*, defaults to `[]`):
Exceptions to freezing vision layers when `freeze_vision_layers` is `True`
use_resampler (`bool`, *optional*, defaults to `False`): Whether to use the Resampler
vision_config (`IdeficsVisionConfig`, *optional*): Custom vision config or dict
perceiver_config (`IdeficsPerceiverConfig`, *optional*): Custom perceiver config or dict
Example:
```python
>>> from transformers import IdeficsModel, IdeficsConfig
>>> # Initializing a Idefics idefics-9b style configuration
>>> configuration = IdeficsConfig()
>>> # Initializing a model from the idefics-9b style configuration
>>> model = IdeficsModel(configuration)
>>> # Accessing the model configuration
>>> configuration = model.config
```"""
model_type = "idefics"
is_composition = True
def __init__(
self,
vocab_size=32000,
additional_vocab_size=0,
hidden_size=4096,
intermediate_size=11008,
num_hidden_layers=32,
num_attention_heads=32,
dropout=0.0,
hidden_act="silu",
initializer_range=0.02,
alpha_initializer="zeros",
alphas_initializer_range=0.0,
alpha_type="float",
rms_norm_eps=1e-6,
use_cache=True,
pad_token_id=0,
bos_token_id=1,
eos_token_id=2,
tie_word_embeddings=False,
cross_layer_interval=1,
qk_layer_norms=False,
freeze_text_layers=True,
freeze_text_module_exceptions=[],
freeze_lm_head=False,
freeze_vision_layers=True,
freeze_vision_module_exceptions=[],
use_resampler=False,
vision_config=None,
perceiver_config=None,
**kwargs,
):
self.vocab_size = vocab_size
self.additional_vocab_size = additional_vocab_size
self.hidden_size = hidden_size
self.intermediate_size = intermediate_size
self.num_hidden_layers = num_hidden_layers
self.num_attention_heads = num_attention_heads
self.dropout = dropout
self.hidden_act = hidden_act
self.initializer_range = initializer_range
self.alpha_initializer = alpha_initializer
self.alphas_initializer_range = alphas_initializer_range
self.alpha_type = alpha_type
self.rms_norm_eps = rms_norm_eps
self.use_cache = use_cache
self.cross_layer_interval = cross_layer_interval
self.qk_layer_norms = qk_layer_norms
self.freeze_vision_layers = freeze_vision_layers
self.freeze_text_layers = freeze_text_layers
self.freeze_text_module_exceptions = freeze_text_module_exceptions
self.freeze_vision_module_exceptions = freeze_vision_module_exceptions
self.freeze_lm_head = freeze_lm_head
self.use_resampler = use_resampler
if perceiver_config is None:
self.perceiver_config = IdeficsPerceiverConfig()
elif isinstance(perceiver_config, dict):
self.perceiver_config = IdeficsPerceiverConfig(**perceiver_config)
elif isinstance(perceiver_config, IdeficsPerceiverConfig):
self.perceiver_config = perceiver_config
if vision_config is None:
self.vision_config = IdeficsVisionConfig()
elif isinstance(vision_config, dict):
self.vision_config = IdeficsVisionConfig(**vision_config)
elif isinstance(vision_config, IdeficsVisionConfig):
self.vision_config = vision_config
super().__init__(
pad_token_id=pad_token_id,
bos_token_id=bos_token_id,
eos_token_id=eos_token_id,
tie_word_embeddings=tie_word_embeddings,
**kwargs,
)
# IMPORTANT: Do not do any __init__ args-based checks in the constructor, since
# PretrainedConfig.from_dict first instantiates the class with the config dict and only then
# updates the config object with `kwargs` from from_pretrained, so during the instantiation
# of this object many attributes have default values and haven't yet been overridden.
# Do any required checks inside `from_pretrained` once the superclass' `from_pretrained` was run.
def to_dict(self):
"""
Serializes this instance to a Python dictionary. Override the default [`~PretrainedConfig.to_dict`].
Returns:
`Dict[str, any]`: Dictionary of all the attributes that make up this configuration instance,
"""
output = copy.deepcopy(self.__dict__)
output["vision_config"] = self.vision_config.to_dict()
output["perceiver_config"] = self.perceiver_config.to_dict()
output["model_type"] = self.__class__.model_type
return output
server/text_generation_server/models/custom_modeling/idefics_image_processing.py 0 → 100644
View file @ bce5e224
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team. All rights reserved.
#
# 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.
"""Image processor class for Idefics."""
from typing import Callable, Dict, List, Optional, Union, Iterable
import numpy as np
from PIL import Image
from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
from transformers.image_transforms import resize, to_channel_dimension_format, rescale, normalize
from transformers.image_utils import (
ChannelDimension,
ImageInput,
PILImageResampling,
make_list_of_images,
to_numpy_array,
valid_images,
)
from io import BytesIO
import requests
from transformers import TensorType, is_torch_available
IDEFICS_STANDARD_MEAN = [0.48145466, 0.4578275, 0.40821073]
IDEFICS_STANDARD_STD = [0.26862954, 0.26130258, 0.27577711]
def convert_to_rgb(image):
# `image.convert("RGB")` would only work for .jpg images, as it creates a wrong background
# for transparent images. The call to `alpha_composite` handles this case
if image.mode == "RGB":
return image
image_rgba = image.convert("RGBA")
background = Image.new("RGBA", image_rgba.size, (255, 255, 255))
alpha_composite = Image.alpha_composite(background, image_rgba)
alpha_composite = alpha_composite.convert("RGB")
return alpha_composite
class IdeficsImageProcessor(BaseImageProcessor):
r"""
Constructs a Idefics image processor.
Args:
image_size (`int`, *optional*, defaults to `224`):
Resize to image size
image_num_channels (`int`, *optional*, defaults to `3`):
Number of image channels.
image_mean (`float` or `List[float]`, *optional*, defaults to `IDEFICS_STANDARD_MEAN`):
Mean to use if normalizing the image. This is a float or list of floats the length of the number of
channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method. Can be
overridden by the `image_mean` parameter in the `preprocess` method.
image_std (`float` or `List[float]`, *optional*, defaults to `IDEFICS_STANDARD_STD`):
Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess` method.
Can be overridden by the `image_std` parameter in the `preprocess` method.
"""
model_input_names = ["pixel_values"]
def __init__(
self,
image_size: int = 224,
image_mean: Optional[Union[float, List[float]]] = None,
image_std: Optional[Union[float, List[float]]] = None,
image_num_channels: Optional[int] = 3,
**kwargs,
) -> None:
super().__init__(**kwargs)
self.image_size = image_size
self.image_num_channels = image_num_channels
self.image_mean = image_mean
self.image_std = image_std
def preprocess(
self,
images: ImageInput,
image_num_channels: Optional[int] = 3,
image_size: Optional[Dict[str, int]] = None,
image_mean: Optional[Union[float, List[float]]] = None,
image_std: Optional[Union[float, List[float]]] = None,
transform: Callable = None,
**kwargs,
) -> TensorType.PYTORCH:
"""
Preprocess a batch of images.
Args:
images (`ImageInput`):
A list of images to preprocess.
image_size (`int`, *optional*, defaults to `self.image_size`):
Resize to image size
image_num_channels (`int`, *optional*, defaults to `self.image_num_channels`):
Number of image channels.
image_mean (`float` or `List[float]`, *optional*, defaults to `IDEFICS_STANDARD_MEAN`):
Mean to use if normalizing the image. This is a float or list of floats the length of the number of
channels in the image. Can be overridden by the `image_mean` parameter in the `preprocess` method. Can
be overridden by the `image_mean` parameter in the `preprocess` method.
image_std (`float` or `List[float]`, *optional*, defaults to `IDEFICS_STANDARD_STD`):
Standard deviation to use if normalizing the image. This is a float or list of floats the length of the
number of channels in the image. Can be overridden by the `image_std` parameter in the `preprocess`
method. Can be overridden by the `image_std` parameter in the `preprocess` method.
transform (`Callable`, *optional*, defaults to `None`):
A custom transform function that accepts a single image can be passed for training. For example,
`torchvision.Compose` can be used to compose multiple transforms. If `None` - an inference mode is
assumed - and then a preset of inference-specific transforms will be applied to the images
Returns:
a PyTorch tensor of the processed images
"""
image_size = image_size if image_size is not None else self.image_size
image_num_channels = image_num_channels if image_num_channels is not None else self.image_num_channels
image_mean = image_mean if image_mean is not None else self.image_mean
image_std = image_std if image_std is not None else self.image_std
size = (image_size, image_size)
if len(images) == 0:
return []
images = make_list_of_images(images)
if not valid_images(images):
raise ValueError(
"Invalid image type. Must be of type PIL.Image.Image, numpy.ndarray, "
"torch.Tensor, tf.Tensor or jax.ndarray."
)
# For training a user needs to pass their own set of transforms as a Callable.
# For reference this is what was used in the original IDEFICS training:
# transform = transforms.Compose([
# convert_to_rgb,
# transforms.RandomResizedCrop((size, size), scale=(0.9, 1.0), interpolation=transforms.InterpolationMode.BICUBIC),
# transforms.ToTensor(),
# transforms.Normalize(mean=image_mean, std=image_std),
# ])
if transform is not None:
if not is_torch_available():
raise ImportError("To pass in `transform` torch must be installed")
import torch
images = [transform(x) for x in images]
return torch.stack(images)
# for inference we do the exact transforms that were used to train IDEFICS
images = [convert_to_rgb(x) for x in images]
# further transforms expect numpy arrays
images = [to_numpy_array(x) for x in images]
images = [resize(x, size, resample=PILImageResampling.BICUBIC) for x in images]
images = [self.rescale(image=image, scale=1 / 255) for image in images]
images = [self.normalize(x, mean=image_mean, std=image_std) for x in images]
images = [to_channel_dimension_format(x, ChannelDimension.FIRST) for x in images]
# TODO: this converts to torch tensors - switch to convert_to_tensors once it becomes available
images = BatchFeature(data={"pixel_values": images}, tensor_type=TensorType.PYTORCH)["pixel_values"]
return images
def fetch_images(self, image_url_or_urls: Union[str, List[str]]):
"""
Convert a single or a list of urls into the corresponding `PIL.Image` objects.
If a single url is passed, the return value will be a single object. If a list is passed a list of objects is
returned.
"""
headers = {
"User-Agent": (
"Mozilla/5.0 (Macintosh; Intel Mac OS X 10_15_7) AppleWebKit/537.36 (KHTML, like Gecko) Chrome/114.0.0.0"
" Safari/537.36"
)
}
if isinstance(image_url_or_urls, list):
return [self.fetch_images(x) for x in image_url_or_urls]
elif isinstance(image_url_or_urls, str):
response = requests.get(image_url_or_urls, stream=True, headers=headers)
response.raise_for_status()
return Image.open(BytesIO(response.content))
else:
raise ValueError(f"only a single or a list of entries is supported but got type={type(image_url_or_urls)}")
def rescale(
self,
image: np.ndarray,
scale: float,
data_format: Optional[Union[str, ChannelDimension]] = None,
input_data_format: Optional[Union[str, ChannelDimension]] = None,
**kwargs,
) -> np.ndarray:
"""
Rescale an image by a scale factor. image = image * scale.
Args:
image (`np.ndarray`):
Image to rescale.
scale (`float`):
The scaling factor to rescale pixel values by.
data_format (`str` or `ChannelDimension`, *optional*):
The channel dimension format for the output image. If unset, the channel dimension format of the input
image is used. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
input_data_format (`ChannelDimension` or `str`, *optional*):
The channel dimension format for the input image. If unset, the channel dimension format is inferred
from the input image. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
Returns:
`np.ndarray`: The rescaled image.
"""
# return rescale(image, scale=scale, data_format=data_format, input_data_format=input_data_format, **kwargs)
# requires 4.32
return rescale(image, scale=scale, data_format=data_format, **kwargs)
def normalize(
self,
image: np.ndarray,
mean: Union[float, Iterable[float]],
std: Union[float, Iterable[float]],
data_format: Optional[Union[str, ChannelDimension]] = None,
input_data_format: Optional[Union[str, ChannelDimension]] = None,
**kwargs,
) -> np.ndarray:
"""
Normalize an image. image = (image - image_mean) / image_std.
Args:
image (`np.ndarray`):
Image to normalize.
mean (`float` or `Iterable[float]`):
Image mean to use for normalization.
std (`float` or `Iterable[float]`):
Image standard deviation to use for normalization.
data_format (`str` or `ChannelDimension`, *optional*):
The channel dimension format for the output image. If unset, the channel dimension format of the input
image is used. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
input_data_format (`ChannelDimension` or `str`, *optional*):
The channel dimension format for the input image. If unset, the channel dimension format is inferred
from the input image. Can be one of:
- `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
- `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
Returns:
`np.ndarray`: The normalized image.
"""
# TODO 4.32
return normalize(
image, mean=mean, std=std, data_format=data_format, **kwargs
)
import transformers
transformers.IdeficsImageProcessor = IdeficsImageProcessor
server/text_generation_server/models/custom_modeling/idefics_perceiver.py 0 → 100644
View file @ bce5e224
# This code was adapted from https://github.com/lucidrains/flamingo-pytorch licensed under the MIT License.
#
# MIT License
#
# Copyright (c) 2020 The Google AI Language Team Authors, The HuggingFace Inc. team and github/lonePatient
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
"""
Generic interface to various configurations of the Perceiver Resampler, that simply takes in a series of (potentially
time-indexed) contextual embeddings, and "resamples" (compresses) them down to a pre-specified number of latents! Note
that the Perceiver in general resamples based solely off the *long-range* context; there's a nice opportunity here to
prime the Perceiver Resampler with say a single layer's worth of language embeddings (the target domain), and use that
to softly "retrieve & compress" what we need --> this would be a novel contribution we should explore.
References:
- DeepMind's Flamingo: https://www.deepmind.com/blog/tackling-multiple-tasks-with-a-single-visual-language-model
- Code borrowed w/ love from: https://github.com/lucidrains/flamingo-pytorch
"""
from typing import Optional, Tuple
import torch
import torch.nn as nn
from text_generation_server.utils.layers import (
TensorParallelColumnLinear,
TensorParallelRowLinear,
)
EPS=1e-5
class IdeficsPerceiverResampler(nn.Module):
def __init__(
self,
prefix,
config,
embed_dim: int,
depth: int,
n_heads: int,
head_dim: int,
n_latents: int,
weights,
) -> None:
"""
Instantiates a Perceiver Resampler that operates over a sequence of embeddings (say from a ResNet or ViT or
MAE) of a given dimension, performs `depth` blocks of cross-attention with a fixed `n_latents` inputs, then
returns a Tensor of shape [bsz, n_latents, embed_dim]. :param embed_dim: Dimensionality of embeddings being fed
to the Perceiver Resampler (also dimensionality of latent embeddings *returned* by the Perceiver Resampler.
Could be e.g., VIT embed_dim, ResNet pool dim, and so on.
Args:
config (`IdeficsConfig`): config object
embed_dim (`int`): The size of each embedding vector
depth (`int`): Depth of the Perceiver Resampler (Transformer w/ cross attention). Should be shallow (< 3).
n_heads (`int`): Number of heads in each Transformer block (for multi-headed self-attention).
head_dim (`int`): Dimensionality of each head projection in the Transformer block.
n_latents (`int`):
Number of latent embeddings to resample ("compress") the input sequence to (usually < 128).
"""
super().__init__()
self.embed_dim, self.n_heads, self.head_dim, self.n_latents = embed_dim, n_heads, head_dim, n_latents
self.qk_layer_norms = config.perceiver_config.qk_layer_norms_perceiver
# Create Latents for Perceiver
self.latents = nn.Parameter(weights.get_tensor(f"{prefix}.latents"))
self.intermediate_dim = (
self.embed_dim * 4
if not hasattr(config.vision_config, "embed_dim")
else config.vision_config.embed_dim * 4
)
# Create Transformer Blocks
self.blocks = nn.ModuleList(
[
nn.ModuleList(
[
IdeficsPerceiverAttention(
prefix=f"{prefix}.blocks.{layer_id}.0",
config=config,
embed_dim=self.embed_dim,
n_heads=self.n_heads,
head_dim=self.head_dim,
qk_layer_norms=self.qk_layer_norms,
weights=weights,
),
IdeficsMLP(
prefix=f"{prefix}.blocks.{layer_id}.1",
intermediate_size=self.intermediate_dim,
config=config,
weights=weights
),
]
)
for layer_id in range(depth)
]
)
self.layer_norm = nn.LayerNorm.load(prefix=f"{prefix}.layer_norm", weights=weights, eps=EPS)
def forward(self, context: torch.Tensor) -> torch.Tensor:
"""Resample arbitrary length context & *compress* down to self.n_latents latent embeddings"""
# einsum.repeat(self.latents, "seq embed -> bsz seq embed", bsz=context.shape[0])
latents = self.latents.repeat(context.shape[0], 1, 1)
# Feed through Perceiver Attention blocks...
for attn, ff in self.blocks:
latents = attn(context, latents) + latents
latents = ff(latents) + latents
return self.layer_norm(latents)
class IdeficsPerceiverAttention(nn.Module):
def __init__(self,
prefix,
config,
embed_dim: int,
n_heads: int,
head_dim: int,
qk_layer_norms: bool,
weights
) -> None:
"""Perceiver Cross-Attention Module --> let long-form inputs be `context`, resampled embeddings be `latents`"""
super().__init__()
self.embed_dim, self.n_heads, self.head_dim = embed_dim, n_heads, head_dim
self.qk_layer_norms = qk_layer_norms
# Normalization & Scaling
self.context_layer_norm = nn.LayerNorm.load(prefix=f"{prefix}.context_layer_norm", weights=weights, eps=EPS)
self.latents_layer_norm = nn.LayerNorm.load(prefix=f"{prefix}.latents_layer_norm", weights=weights, eps=EPS)
if self.qk_layer_norms:
self.q_layer_norm = nn.LayerNorm.load(prefix=f"{prefix}.q_layer_norm", weights=weights, eps=EPS)
self.k_layer_norm = nn.LayerNorm.load(prefix=f"{prefix}.k_layer_norm", weights=weights, eps=EPS)
self.qk_scale = self.head_dim**-0.5
process_group = weights.process_group
if n_heads % weights.process_group.size() != 0:
raise ValueError(
f"`num_heads` must be divisible by `num_shards` (got `num_heads`: {n_heads} "
f"and `num_shards`: {weights.process_group.size()}"
)
self.n_heads //= weights.process_group.size()
# Q, K, V Projection (no bias -- detail from Perceiver/Flamingo Papers).
self.q_proj = TensorParallelColumnLinear.load(
config=config, prefix=f"{prefix}.q_proj", weights=weights, bias=False
)
self.k_proj = TensorParallelColumnLinear.load(
config=config, prefix=f"{prefix}.k_proj", weights=weights, bias=False
)
self.v_proj = TensorParallelColumnLinear.load(
config=config, prefix=f"{prefix}.v_proj", weights=weights, bias=False
)
self.output_proj = TensorParallelRowLinear.load(
config=config, prefix=f"{prefix}.output_proj", weights=weights, bias=False
)
def forward(self, context: torch.Tensor, latents: torch.Tensor) -> torch.Tensor:
"""
Runs Perceiver Self-Attention, with special (context, latents) appended along the `seq` dimension!
Args:
context (`torch.Tensor`):
Tensor of shape `[bsz, seq, embed_dim]` representing long-form context to resample.
latents (`torch.Tensor`):
Tensor of shape `[bsz, n_latents, embed_dim]` representing fixed length latents to compress to.
Returns:
`torch.Tensor`: Tensor of shape `[bsz, n_latents, embed_dim]` representing attention over latents w/ cross
from context.
"""
context = self.context_layer_norm(context)
latents = self.latents_layer_norm(latents)
batch_size, seq_length, embed_dim = context.shape[:3]
# Query, Key, Value Projections --> Note that in Flamingo, latents are *concatenated* with context prior to attn!
# Note: This results in queries w/ `seq = n_latents`, and keys, values with `seq = len(context) + n_latents`
q = self.q_proj(latents)
k = self.k_proj(torch.cat([context, latents], dim=-2))
v = self.v_proj(torch.cat([context, latents], dim=-2))
# Multiheaded Self-Attention w/ stable softmax (subtract per-row max -- `amax` -- before softmax call)
# =>> `attn` should be a 2D matrix of shape [n_latents x (context + n_latents)]
# einsum.rearrange(x, "bsz seq (heads embed) -> bsz heads seq embed", heads=self.n_heads)
q, k, v = [x.reshape(batch_size, x.shape[1], self.n_heads, self.head_dim).transpose(1, 2) for x in (q, k, v)]
if self.qk_layer_norms:
q = self.q_layer_norm(q)
k = self.k_layer_norm(k)
scores = torch.einsum("... i d, ... j d -> ... i j", q * self.qk_scale, k)
stabilized_scores = scores - (scores.amax(dim=-1, keepdim=True).detach())
attn = stabilized_scores.softmax(dim=-1)
# Attend & project back to output...
resampled = torch.einsum("... i j, ... j d -> ... i d", attn, v)
# einsum.rearrange(resampled, "bsz heads seq embed -> bsz seq (heads embed)", heads=self.n_heads)
return self.output_proj(resampled.transpose(1, 2).flatten(-2))
class IdeficsMLP(nn.Module):
def __init__(self,
prefix,
intermediate_size,
config,
weights,
):
"""Simple MLP block with intermediate_size and embedding size"""
super().__init__()
self.embed_dim = config.vision_config.embed_dim
self.ln = nn.LayerNorm.load(prefix=f"{prefix}.ln", weights=weights, eps=EPS)
self.fc = TensorParallelColumnLinear.load(
config=config, prefix=f"{prefix}.fc", weights=weights, bias=False,
)
self.act = nn.ReLU()
self.c_proj = TensorParallelRowLinear.load(
config=config, prefix=f"{prefix}.c_proj", weights=weights, bias=False,
)
def forward(self, hidden_states: Optional[Tuple[torch.FloatTensor]]) -> torch.FloatTensor:
hidden_states = self.ln(hidden_states)
hidden_states = self.fc(hidden_states)
hidden_states = self.act(hidden_states)
hidden_states = self.c_proj(hidden_states)
return hidden_states
server/text_generation_server/models/custom_modeling/idefics_processing.py 0 → 100644
View file @ bce5e224
# coding=utf-8
# Copyright 2022 The HuggingFace Inc. team.
#
# 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.
"""
Processor class for IDEFICS.
"""
from typing import Callable, List, Optional, Union
from urllib.parse import urlparse
from transformers.feature_extraction_utils import BatchFeature
from transformers.processing_utils import ProcessorMixin
from transformers.tokenization_utils_base import BatchEncoding, PaddingStrategy, TextInput, TruncationStrategy
from transformers.utils import TensorType, is_torch_available
from text_generation_server.models.custom_modeling.idefics_image_processing import IdeficsImageProcessor
if is_torch_available():
import torch
IMAGE_TOKEN = "<image>"
# copied from m4.training.packing
def incremental_to_binary_attention_mask(incremental_mask, num_classes=-1):
# This function converts: [-1, 0, 1] => [[0, 0], [1, 0], [0, 1]]
# If any of images index are more than num_classes, set them to -1.
# Words after the max number of images allowed have been seen don't attend on anything
if num_classes != -1:
incremental_mask[incremental_mask >= num_classes] = -1
negatives = incremental_mask == -1
incremental_mask[negatives] = 0
attn_mask = torch.nn.functional.one_hot(incremental_mask, num_classes=num_classes)
attn_mask[negatives, :] = 0
return attn_mask
# copied from m4.training.packing
def image_attention_mask_for_packed_input_ids(input_ids, tokenizer):
image_attention_mask = torch.full_like(input_ids, fill_value=-1)
next_image_attention_mask = torch.full_like(input_ids, fill_value=-1)
image_token_id = tokenizer.convert_tokens_to_ids(IMAGE_TOKEN)
eod_token_id = tokenizer.eos_token_id
for batch_idx in range(input_ids.size(0)):
count = -1
seen_eod = False
for idx, token_id in enumerate(input_ids[batch_idx]):
if token_id == image_token_id:
count += 1
image_attention_mask[batch_idx][idx] = count
seen_eod = False
else:
image_attention_mask[batch_idx][idx] = count
if seen_eod:
image_attention_mask[batch_idx][idx] = -1
if token_id == eod_token_id:
seen_eod = True
for batch_idx in range(input_ids.size(0)):
count = -1
seen_eod = False
for idx in range(input_ids[batch_idx].size(0) - 1, -1, -1):
token_id = input_ids[batch_idx][idx]
if token_id == image_token_id:
count += 1
next_image_attention_mask[batch_idx][idx] = count
seen_eod = False
else:
next_image_attention_mask[batch_idx][idx] = count
if token_id == eod_token_id:
seen_eod = True
if seen_eod:
next_image_attention_mask[batch_idx][idx] = -1
non_negative_indices = next_image_attention_mask[batch_idx] != -1
next_image_attention_mask[batch_idx][non_negative_indices] -= count
next_image_attention_mask[batch_idx][non_negative_indices] *= -1
return image_attention_mask, next_image_attention_mask
def is_url(string):
"""Checks if the passed string contains a valid url and nothing else. e.g. if space is included it's immediately
invalidated the url"""
if " " in string:
return False
result = urlparse(string)
return all([result.scheme, result.netloc])
class IdeficsProcessor(ProcessorMixin):
r"""
Constructs a IDEFICS processor which wraps a LLama tokenizer and IDEFICS image processor into a single processor.
[`IdeficsProcessor`] offers all the functionalities of [`IdeficsImageProcessor`] and [`LlamaTokenizerFast`]. See
the docstring of [`~IdeficsProcessor.__call__`] and [`~IdeficsProcessor.decode`] for more information.
Args:
image_processor (`IdeficsImageProcessor`):
An instance of [`IdeficsImageProcessor`]. The image processor is a required input.
tokenizer (`LlamaTokenizerFast`):
An instance of [`LlamaTokenizerFast`]. The tokenizer is a required input.
image_size (`int`, *optional*, defaults to 224): Image size (assuming a square image)
"""
attributes = ["image_processor", "tokenizer"]
image_processor_class = "IdeficsImageProcessor"
tokenizer_class = "LlamaTokenizerFast"
def __init__(self, image_processor, tokenizer=None, image_size=224, add_end_of_utterance_token=None, **kwargs):
if image_processor is None:
raise ValueError("You need to specify an `image_processor`.")
if tokenizer is None:
raise ValueError("You need to specify a `tokenizer`.")
super().__init__(image_processor, tokenizer)
self.current_processor = self.image_processor
self.image_token_id = tokenizer.convert_tokens_to_ids(IMAGE_TOKEN)
self.default_image_dims = (
self.image_processor.image_num_channels,
self.image_processor.image_size,
self.image_processor.image_size,
)
self.tokenizer_was_trained_with_end_of_utterance_token = (
True
if "<end_of_utterance>" in self.tokenizer.special_tokens_map.get("additional_special_tokens", [])
else False
)
def __call__(
self,
prompts: Union[List[TextInput], List[List[TextInput]]],
padding: Union[bool, str, PaddingStrategy] = False,
truncation: Union[bool, str, TruncationStrategy] = None,
max_length: Optional[int] = None,
transform: Callable = None,
add_eos_token=False,
add_end_of_utterance_token=None,
debug=False,
return_tensors: Optional[Union[str, TensorType]] = TensorType.PYTORCH,
) -> BatchEncoding:
"""This method takes batched or non-batched prompts made of text and images and converts them into prompts that
the model was trained on and prepares the image pixel values for the model to process.
Args:
prompts (`Union[List[TextInput], [List[List[TextInput]]]]`):
either a single prompt or a batched list of prompts - see the detailed description immediately after
the end of the arguments doc section.
padding (`bool`, `str` or [`~utils.PaddingStrategy`], *optional*, defaults to `False`):
Select a strategy to pad the returned sequences (according to the model's padding side and padding
index) among:
- `True` or `'longest'`: Pad to the longest sequence in the batch (or no padding if only a single
sequence if provided).
- `'max_length'`: Pad to a maximum length specified with the argument `max_length` or to the maximum
acceptable input length for the model if that argument is not provided.
- `False` or `'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different
lengths).
max_length (`int`, *optional*):
Maximum length of the returned list and optionally padding length (see above).
truncation (`bool`, *optional*):
Activates truncation to cut input sequences longer than `max_length` to `max_length`.
transform (`Callable`, *optional*):
A custom transform function that accepts a single image can be passed for training. For example,
`torchvision.Compose` can be used to compose multiple functions. If `None` a preset inference-specific
set of transforms will be applied to the images
add_eos_token (`bool`, *optional*, defaults to `False`):
Adds `eos_token` at the end of the final prompt if True`
add_end_of_utterance_token (`bool`, *optional*)
Whether to automatically add `<end_of_utterance>` after each prompt's text input (unless followed by an
image). If `None` the tokenizer will be checked instead and if this token is found in
`additional_special_tokens` then the value will be `True`.
debug (`bool`, *optional*, defaults to `False`):
`True` value will help debug prompt generation by dumping useful information
return_tensors (`str` or `TensorType`, *optional*, defaults to `TensorType.PYTORCH`):
The type of tensors to return. Can be one of:
- `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
Returns:
a dict with entries: `input_ids`, `attention_mask`, `pixel_values`, `image_attention_mask` which can be
directly passed to `model.generate`
Detailed explanation:
Each entry in `prompts` is either a text to be passed as is or an image that will be processed.
An image can be either an image object (`PIL.Image`) or a url from which the image can be retrieved.
When the processor encounters an image it'll inject `<fake_token_around_image><image><fake_token_around_image>`
entry into the prompt.
Example:
```python
checkpoint = "HuggingFaceM4/idefics-9b"
processor = AutoProcessor.from_pretrained(checkpoint)
url = "https://hips.hearstapps.com/hmg-prod/images/cute-photos-of-cats-in-grass-1593184777.jpg"
img = processor.image_processor.fetch_images([url])[0]
prompts = [
"User:",
img,
"Describe this image.\nAssistant: An image of two kittens in grass.\n",
"User:",
"https://hips.hearstapps.com/hmg-prod/images/dog-puns-1581708208.jpg",
"Describe this image.\nAssistant:",
]
inputs = processor(prompts, return_tensors="pt")
generated_ids = model.generate(**inputs, max_length=100)
generated_text = processor.batch_decode(generated_ids, skip_special_tokens=True)[0]
```
In this example the `prompts` will be converted into:
```
<s>User:<fake_token_around_image><image><fake_token_around_image>Describe this image.
Assistant: An image of two kittens in grass.
User:<fake_token_around_image><image><fake_token_around_image>Describe this image.
Assistant:'
```
and the two images will be massaged using [`IdeficsImageProcessor.__call__`] method and placed inside the
`pixel_values` dict entry of the return value.
This example also examplifies that images can be passed as objects or as text urls. It can be seen that the
first image is passed as object and the second one as a url.
To do training do:
```python
image_transform = transforms.Compose(
[
transforms.RandomResizedCrop(
(w, h), scale=(0.9, 1.0), interpolation=transforms.InterpolationMode.BICUBIC
),
transforms.ToTensor(),
transforms.Normalize(mean=self.image_mean, std=self.image_std),
]
)
inputs = processor(prompts, transform=image_transform, return_tensors="pt")
```
In order to help debug prompt generation enable `debug=True` which will show you what's happening.
"""
# if the value isn't overriden by the user, check if the tokenizer was trained with this token and then use it
if add_end_of_utterance_token is None:
add_end_of_utterance_token = self.tokenizer_was_trained_with_end_of_utterance_token
# turn non-batched prompts into batched
if not any(isinstance(i, list) for i in prompts):
prompts = [prompts]
fake_token = "<fake_token_around_image>"
image_token = "<image>"
end_of_utterance_token = "<end_of_utterance>"
def image_tokens(last_was_image):
if last_was_image:
return image_token + fake_token
else:
return fake_token + image_token + fake_token
all_texts = []
all_images = []
for sample in prompts:
# the model was trained on samples starting with <s>
full_text = f"{self.tokenizer.bos_token}"
# an image can either be an image object in the item or the url, everything else is a verbatim prompt text
image_objects = []
last_was_image = False
last_was_text = False
for i, item in enumerate(sample):
if i > 0:
last_was_text = True if not last_was_image else False
if isinstance(item, str):
item = item.strip(" ")
if is_url(item):
image = self.image_processor.fetch_images(item)
full_text += image_tokens(last_was_image)
image_objects.append(image)
last_was_image = True
else:
# we add end_of_utterance_token between each subsequent text prompts (but not at the last one!)
if add_end_of_utterance_token and last_was_text:
full_text += end_of_utterance_token
full_text += item
last_was_image = False
else:
# must be an image obj
full_text += image_tokens(last_was_image)
image_objects.append(item)
last_was_image = True
if add_eos_token:
full_text += self.tokenizer.eos_token
if debug is True:
print(f"{full_text=}")
image_objects = self.image_processor(image_objects, transform=transform)
text_encoding = self.tokenizer(
text=full_text,
add_special_tokens=False,
padding=padding,
truncation=truncation,
max_length=max_length,
)
all_texts.append(text_encoding["input_ids"])
all_images.append(image_objects)
max_seq_len = max(len(x) for x in all_texts)
# max_num_images has to be at least 1 even when there are no images
max_num_images = max(len(x) for x in all_images)
max_num_images = max(1, max_num_images)
at_least_one_image = sum(len(x) for x in all_images) > 0
output_input_ids = []
output_images = []
output_attention_masks = []
for text, images in zip(all_texts, all_images):
padded_input_ids = [self.tokenizer.pad_token_id] * max_seq_len
unpadded_seq_len = len(text)
start = max_seq_len - unpadded_seq_len
padded_input_ids[start:] = text[:max_seq_len]
attention_mask = torch.zeros((max_seq_len,), dtype=torch.long)
attention_mask[start:] = 1
image_count = padded_input_ids.count(self.image_token_id)
local_max_num_images = min(image_count, max_num_images)
current_images = images[:local_max_num_images]
if len(current_images) > 0:
padded_image_tensor = torch.zeros(max_num_images, *current_images.size()[1:])
padded_image_tensor[: current_images.size(0)] = current_images
else:
padded_image_tensor = torch.zeros(max_num_images, *self.default_image_dims)
output_images.append(padded_image_tensor)
output_input_ids.append(torch.tensor(padded_input_ids))
output_attention_masks.append(attention_mask)
output_input_ids = torch.stack(output_input_ids)
output_images = torch.stack(output_images)
output_attention_masks = torch.stack(output_attention_masks)
if at_least_one_image:
image_attention_mask, _ = image_attention_mask_for_packed_input_ids(output_input_ids, self.tokenizer)
image_attention_mask = incremental_to_binary_attention_mask(
image_attention_mask, num_classes=max_num_images
)
else:
# in full language mode we set the image mask to all-0s
image_attention_mask = torch.zeros(
output_input_ids.shape[0], output_input_ids.shape[1], 1, dtype=torch.bool
)
return BatchFeature(
data={
"input_ids": output_input_ids,
"attention_mask": output_attention_masks,
"pixel_values": output_images,
"image_attention_mask": image_attention_mask,
}
)
def batch_decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
refer to the docstring of this method for more information.
"""
return self.tokenizer.batch_decode(*args, **kwargs)
def decode(self, *args, **kwargs):
"""
This method forwards all its arguments to LlamaTokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
the docstring of this method for more information.
"""
return self.tokenizer.decode(*args, **kwargs)
@property
def model_input_names(self):
tokenizer_input_names = self.tokenizer.model_input_names
image_processor_input_names = self.image_processor.model_input_names
return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
server/text_generation_server/models/idefics.py 0 → 100644
View file @ bce5e224
import torch
import torch.distributed
from typing import List, Optional, Tuple
from transformers import (
AutoTokenizer,
AutoConfig,
AutoProcessor,
)
from text_generation_server.models.custom_modeling.idefics_config import IdeficsConfig
from text_generation_server.models.custom_modeling.idefics_processing import (
IdeficsProcessor,
)
from transformers import LlamaTokenizerFast
from text_generation_server.models.custom_modeling.idefics_modeling import (
IdeficsForVisionText2Text,
)
from text_generation_server.models.idefics_causal_lm import IdeficsCausalLM
from text_generation_server.utils import (
initialize_torch_distributed,
weight_files,
Weights,
)
class IDEFICSSharded(IdeficsCausalLM):
def __init__(
self,
model_id: str,
revision: Optional[str] = None,
quantize: Optional[str] = None,
dtype: Optional[torch.dtype] = None,
trust_remote_code: bool = False,
):
self.process_group, rank, world_size = initialize_torch_distributed()
if torch.cuda.is_available():
device = torch.device(f"cuda:{rank}")
# 9b seems to work correctly enough in float16, but 80b seems
# to be really saturating for f16.
dtype = torch.bfloat16 if dtype is None else dtype
else:
device = torch.device("cpu")
dtype = torch.float32
self.device, self.dtype = device, dtype
config = IdeficsConfig.from_pretrained(
model_id,
revision=revision,
trust_remote_code=trust_remote_code,
)
config.quantize = quantize
config.vision_config.quantize = quantize
tokenizer = LlamaTokenizerFast.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
self.processor = IdeficsProcessor.from_pretrained(
model_id,
revision=revision,
padding_side="left",
truncation_side="left",
trust_remote_code=trust_remote_code,
)
torch.distributed.barrier(group=self.process_group)
filenames = weight_files(model_id, revision=revision, extension=".safetensors")
weights = Weights(
filenames,
device=device,
dtype=dtype,
process_group=self.process_group,
)
model = IdeficsForVisionText2Text(config, weights)
torch.distributed.barrier(group=self.process_group)
super(IdeficsCausalLM, self).__init__(
model=model,
tokenizer=tokenizer,
requires_padding=True,
dtype=dtype,
device=device,
rank=rank,
world_size=world_size,
)
server/text_generation_server/models/idefics_causal_lm.py 0 → 100644
View file @ bce5e224
This diff is collapsed.
server/text_generation_server/server.py
View file @ bce5e224
...@@ -14,7 +14,7 @@ from text_generation_server.interceptor import ExceptionInterceptor ...@@ -14,7 +14,7 @@ from text_generation_server.interceptor import ExceptionInterceptor
from text_generation_server.models import Model, get_model from text_generation_server.models import Model, get_model
from text_generation_server.pb import generate_pb2_grpc, generate_pb2 from text_generation_server.pb import generate_pb2_grpc, generate_pb2
from text_generation_server.tracing import UDSOpenTelemetryAioServerInterceptor from text_generation_server.tracing import UDSOpenTelemetryAioServerInterceptor
from text_generation_server.models.idefics_causal_lm import IdeficsCausalLMBatch
class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer): class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
def __init__(self, model: Model, cache: Cache, server_urls: List[str]): def __init__(self, model: Model, cache: Cache, server_urls: List[str]):
...@@ -26,6 +26,7 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer): ...@@ -26,6 +26,7 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
# Force inference mode for the lifetime of TextGenerationService # Force inference mode for the lifetime of TextGenerationService
self._inference_mode_raii_guard = torch._C._InferenceMode(True) self._inference_mode_raii_guard = torch._C._InferenceMode(True)
async def Info(self, request, context): async def Info(self, request, context):
return self.model.info return self.model.info
...@@ -54,9 +55,14 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer): ...@@ -54,9 +55,14 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
return generate_pb2.FilterBatchResponse(batch=filtered_batch.to_pb()) return generate_pb2.FilterBatchResponse(batch=filtered_batch.to_pb())
async def Warmup(self, request, context): async def Warmup(self, request, context):
batch = self.model.batch_type.from_pb( if self.model.batch_type == IdeficsCausalLMBatch: #Hack, i would rather use kwargs in the `from_pb` call
request.batch, self.model.tokenizer, self.model.dtype, self.model.device batch = self.model.batch_type.from_pb(
) request.batch, self.model.tokenizer, self.model.processor, self.model.dtype, self.model.device
)
else:
batch = self.model.batch_type.from_pb(
request.batch, self.model.tokenizer, self.model.dtype, self.model.device
)
max_supported_total_tokens = self.model.warmup(batch) max_supported_total_tokens = self.model.warmup(batch)
return generate_pb2.WarmupResponse( return generate_pb2.WarmupResponse(
...@@ -64,9 +70,14 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer): ...@@ -64,9 +70,14 @@ class TextGenerationService(generate_pb2_grpc.TextGenerationServiceServicer):
) )
async def Prefill(self, request, context): async def Prefill(self, request, context):
batch = self.model.batch_type.from_pb( if self.model.batch_type == IdeficsCausalLMBatch: #Hack, i would rather use kwargs in the `from_pb` call
request.batch, self.model.tokenizer, self.model.dtype, self.model.device batch = self.model.batch_type.from_pb(
) request.batch, self.model.tokenizer, self.model.processor, self.model.dtype, self.model.device
)
else:
batch = self.model.batch_type.from_pb(
request.batch, self.model.tokenizer, self.model.dtype, self.model.device
)
generations, next_batch = self.model.generate_token(batch) generations, next_batch = self.model.generate_token(batch)
self.cache.set(next_batch) self.cache.set(next_batch)
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment