[`PEFT`] Peft integration alternative design #25077

younesbelkada · 2023-07-25T11:12:34Z

What does this PR do?

From the offline discussion + the comments from @patrickvonplaten in #24827 (comment) I propose a new design for tightly integrating PEFT into transformers.
This integration enables loading any PEFT adapter that is saved locally or on the Hub directly into PEFT without dispatching the entire model creation process to PEFT as introduced in #24827.

This would also enable an easier pipeline integration (a one-liner to load adapter weights) | EDIT: pipeline should work out of the box

Let's constraint this integration to few PEFT methods only, for simplicity and redirect users to use PEFT for advanced features (e.g. merge and unload) and advanced PEFT methods (adaptation prompt, prompt learning).

Current API:

Load a model with an adapter locally or from the Hub:

import torch
from transformers import AutoModelForCausalLM, OPTForCausalLM

model_id = "facebook/opt-350m"
adapter_model_id = "ybelkada/opt-350m-lora"

# directly on from_pretrained
model = AutoModelForCausalLM.from_pretrained(adapter_model_id)
print(model)

# directly on from_pretrained
model = OPTForCausalLM.from_pretrained(adapter_model_id)
print(model)

Load and attach adapter to an existing model

from transformers import AutoModelForCausalLM

# with load_adapter
model = AutoModelForCausalLM.from_pretrained(model_id)
model.load_adapter(adapter_model_id)

print(model)

# 8-bit + multiGPU compatiblity
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_8bit=True, device_map="balanced")
model.load_adapter(adapter_model_id)

print(model)
print(set(model.hf_device_map.values()))

_ = model(torch.LongTensor([[0, 1, 2, 3]]).to(0))

Attach an adapter, iteratively enable / disable adapters

from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig

model_id = "facebook/opt-350m"
adapter_model_id = "ybelkada/opt-350m-lora"
tokenizer = AutoTokenizer.from_pretrained(model_id)
text = "Hello"
inputs = tokenizer(text, return_tensors="pt")

model = AutoModelForCausalLM.from_pretrained(model_id)
peft_config = PeftConfig.from_pretrained(adapter_model_id)

# To get random weights
peft_config.init_lora_weights = False

model.add_adapter(peft_config)
print(model)

model.disable_adapters()
output_disabled = model.generate(**inputs)
print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
>>> Hello, I'm a newbie to this sub. I'm looking for a good place to

model.enable_adapters()
output_enabled = model.generate(**inputs)
print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
>>> Hello, MMMMMMMM

Add multiple adapters

from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig, LoraConfig

model_id = "facebook/opt-350m"

# directly on from_pretrained
model = AutoModelForCausalLM.from_pretrained(model_id)

lora_config = LoraConfig(
    target_modules=["q_proj", "k_proj"],
    init_lora_weights=False
)

model.add_adapter(lora_config, adapter_name="adapter_1")

# attach new adapter with same config
model.add_adapter(lora_config, adapter_name="adapter_2")

model.set_adapter("adapter_1")
output_disabled = model.generate(**inputs)
print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
>>> Hello, I'm a newbie to this sub. I'm looking for a good place to

model.set_adapter("adapter_2")
output_enabled = model.generate(**inputs)
print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
>>> Hello, I'm a newbie to the game. I'm looking for a good way to

Save adapters

from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig, LoraConfig

model_id = "facebook/opt-350m"

# directly on from_pretrained
model = AutoModelForCausalLM.from_pretrained(model_id)

lora_config = LoraConfig(
    target_modules=["q_proj", "k_proj"],
)

model.add_adapter(lora_config)

... # train here

model.save_pretrained(save_dir) 

# you can either load it back with transformers or PEFT

from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(save_dir)

# or

model = AutoModelForCausalLM.from_pretrained(save_dir)

Train adapters using Trainer

Check this gist: https://gist.github.com/younesbelkada/cdda6e4abcb09e58f6324d75e0d88862

This PR is on par with: huggingface/peft#749

Features to support:

loading PEFT adapters
using multiple adapters
deal with models loaded with accelerate
Loading directly from from_pretrained
Merging adapter weights - to not support
Unload adapter weights - to not support
Training with BC with expected PEFT checkpoints format (do we really want to support training? Shall we just redirect users to load a classic PeftModel if they want to train a model?)
What about save_pretrained ?

Features to not support:

disabling adapters
prompt tuning / prompt learning methods

TODOs:

docs
tests

cc @sgugger @patrickvonplaten @BenjaminBossan @pacman100

HuggingFaceDocBuilderDev · 2023-07-25T11:36:10Z

The documentation is not available anymore as the PR was closed or merged.

sgugger · 2023-07-25T11:38:58Z

Mmmm, but know we can't do model = AutoModelForCausalLM.from_pretrained("ybelkada/opt-350m-lora") despite the checkpoint having all the correct info to load the model (compared to the alternative PR). Can we still add the necessary code in from_pretrained to have the load take one line instead of two?

Apart from that, the design looks great to me!

patrickvonplaten

Design looks very nice to me! We'll be able to fully leverage this from diffusers I believe

BenjaminBossan

This looks very nice. I can't comment on the overall design, as my knowledge about transformers auto models is not very deep, but the approach looks promising from my point of view.

Let's constraint this integration to few PEFT methods only, for simplicity and redirect users to use PEFT for advanced features and advanced PEFT methods.

Features to not support: ...

Yes, this seems to be the way to go. As mentioned by others, this should be very well documented, with an error message explaining what works and what doesn't if the user tries to do something that is not supported.

I have a few small comments on the code, please take a look.

Regarding the extension of PreTrainedModel base classes with PeftAdapterMixin, it seems that it would be possible to extend it dynamically:

https://stackoverflow.com/questions/11042424/adding-base-class-to-existing-object-in-python

So in theory, we could do a check if peft is installed and only then add the mixin to the base classes, which means that PreTrainedModel would stay the same for all users who don't use peft. I'm not sure if that's a good idea or not, just throwing it out there.

src/transformers/adapters/peft_mixin.py

BenjaminBossan · 2023-07-25T12:10:35Z

src/transformers/adapters/peft_mixin.py

+        """
+        Load adapter weights from file. Requires peft as a backend to load the adapter weights
+        """
+        requires_backends(self.load_adapter, "peft")


Technically, this also requires a specific peft version, right? But it seems that requiring a version is not supported.

Yes I can add that in the next line

Agree it would be nice to get a very good error message here

src/transformers/adapters/peft_mixin.py

…tempt-2

pacman100

I really like the design, very clever way to use class methods to inject the modules and load multiple adapters. I believe with some effort, it can support all the goodies that aren't yet present such as merge_and_unload, unload, create_weighted_adapter and delete_adapter. We can take these up later on. Overall, great work @younesbelkada 🔥.

Btw, to answer Sylvain's question, we can also use load_adapter in from_pretrained so that only 1 line of code is needed instead of 2, WDYT?

src/transformers/utils/peft_utils.py

…tempt-2

younesbelkada · 2023-08-01T10:39:48Z

src/transformers/adapters/peft_mixin.py

+            )
+
+    # TODO: change it to a property but torch.jit fails. Maybe we should return None is PEFT is not available
+    def current_active_adapter(self) -> str:


If we really want to have a property maybe return none + a warning in case PEFT is not installed

pacman100 · 2023-08-02T07:06:30Z

src/transformers/adapters/peft_mixin.py

+        create_and_replace(loaded_peft_config, self, adapter_name)
+
+        adapter_state_dict = load_peft_weights(
+            peft_model_id,
+            revision=revision,
+            use_auth_token=use_auth_token,
+        )
+
+        # We need to pre-process the state dict to remove unneeded prefixes - for backward compatibility
+        processed_adapter_state_dict = {}
+        for key, value in adapter_state_dict.items():
+            if "base_model.model" in key:
+                new_key = key.replace("base_model.model.", "")
+            else:
+                new_key = key
+            processed_adapter_state_dict[new_key] = value
+
+        # Load state dict
+        incompatible_keys = set_peft_model_state_dict(self, processed_adapter_state_dict, adapter_name)
+
+        if incompatible_keys is not None:
+            # check only for unexpected keys
+            if hasattr(incompatible_keys, "unexpected_keys") and len(incompatible_keys.unexpected_keys) > 0:
+                logger.warning(
+                    f"Loading adapter weights from {peft_model_id} led to unexpected keys not found in the model: "
+                    f" {incompatible_keys.unexpected_keys}. "
+                )
+
+        # @pacman100 why this was needed?
+        if (
+            (getattr(self, "hf_device_map", None) is not None)
+            and (len(set(self.hf_device_map.values()).intersection({"cpu", "disk"})) > 0)
+            and len(self.peft_config) == 1
+        ):
+            self._dispatch_accelerate_model(
+                device_map=device_map, max_memory=max_memory, offload_dir=offload_dir, offload_index=offload_index


I think PeftModel.from_pretrained(self, peft_model_id) address all of this, WDYT?

Yes correct ! However I preferred to not use PeftModel and have a granular control on each thing we do (inject adapter, get state_dict, load it, dispatch if cpu offload) so that it will be easier to maintain if in the future we add more features or some important changes in PeftModel.from_pretrained.

But, on the contrary, if we fix PeftModel.from_pretrained, we need to update all the changes here too increasing the maintenance and breaking changes because from_pretrained would be always updated to be in sync with changes of PeftModel whereas here we might get an updated version of PeftModel but the current loading logic might not handle it. WDYT?

…tempt-2

sgugger

Thanks for all your work on this! Have a couple of nits and make sure to remove use_auth_token to use token instead.

sgugger · 2023-08-18T11:17:00Z

docs/source/en/peft.md

+<!--
+TODO: (@younesbelkada @stevhliu)
+-   Link to PEFT docs for further details
+-   Trainer  
+-   8-bit / 4-bit examples ?
+-->


Would be nice in a followup PR to add docs about training a model in 4/8-bit with peft and LoRA as there are many issues around this :-)

Agreed on this!

src/transformers/lib_integrations/peft/peft_mixin.py

Co-authored-by: Sylvain Gugger <[email protected]>

pacman100 · 2023-08-21T10:55:50Z

Thank you @younesbelkada for the impressive work on adding PEFT as a utility library in Transformers 🔥🚀✨

* a draft version * v2 integration * fix * make it more generic and works for IA3 * add set adapter and multiple adapters support * fixup * adapt a bit * oops * oops * oops * adapt more * fix * add more refactor * now works with model class * change it to instance method as it causes issues with `jit`. * add CR * change method name * add `add_adapter` method * clean up * Update src/transformers/adapters/peft_mixin.py Co-authored-by: Patrick von Platen <[email protected]> * add moe utils * fixup * Update src/transformers/adapters/peft_mixin.py Co-authored-by: Patrick von Platen <[email protected]> * adapt * oops * fixup * add is_peft_available * remove `requires_backend` * trainer compatibility * fixup + docstring * more details * trigger CI * Apply suggestions from code review Co-authored-by: Sylvain Gugger <[email protected]> * Update src/transformers/modeling_utils.py * fixup + is_main_process * added `save_peft_format` in save_pretrained * up * fix nits here and there * nits here and there. * docs * revert `encoding="utf-8"` * comment * added slow tests before the PEFT release. * fixup and nits * let's be on the safe zone * added more comments * v1 docs * add remaining docs * Apply suggestions from code review Co-authored-by: Steven Liu <[email protected]> * move to `lib_integrations` * fixup * this time fixup * Apply suggestions from code review Co-authored-by: Sylvain Gugger <[email protected]> * address final comments * refactor to use `token` * add PEFT to DockerFile for slow tests. * added pipeline support. --------- Co-authored-by: Patrick von Platen <[email protected]> Co-authored-by: Sylvain Gugger <[email protected]> Co-authored-by: Steven Liu <[email protected]>

BoccheseGiacomo · 2023-10-14T11:04:00Z

What does this PR do?

From the offline discussion + the comments from @patrickvonplaten in #24827 (comment) I propose a new design for tightly integrating PEFT into transformers. This integration enables loading any PEFT adapter that is saved locally or on the Hub directly into PEFT without dispatching the entire model creation process to PEFT as introduced in #24827.

This would also enable an easier pipeline integration (a one-liner to load adapter weights) | EDIT: pipeline should work out of the box

Let's constraint this integration to few PEFT methods only, for simplicity and redirect users to use PEFT for advanced features (e.g. merge and unload) and advanced PEFT methods (adaptation prompt, prompt learning).

Current API:

Load a model with an adapter locally or from the Hub:
import torch
from transformers import AutoModelForCausalLM, OPTForCausalLM

model_id = "facebook/opt-350m"
adapter_model_id = "ybelkada/opt-350m-lora"

# directly on from_pretrained
model = AutoModelForCausalLM.from_pretrained(adapter_model_id)
print(model)

# directly on from_pretrained
model = OPTForCausalLM.from_pretrained(adapter_model_id)
print(model)
Load and attach adapter to an existing model
from transformers import AutoModelForCausalLM

# with load_adapter
model = AutoModelForCausalLM.from_pretrained(model_id)
model.load_adapter(adapter_model_id)

print(model)

# 8-bit + multiGPU compatiblity
model = AutoModelForCausalLM.from_pretrained(model_id, load_in_8bit=True, device_map="balanced")
model.load_adapter(adapter_model_id)

print(model)
print(set(model.hf_device_map.values()))

_ = model(torch.LongTensor([[0, 1, 2, 3]]).to(0))
Attach an adapter, iteratively enable / disable adapters
from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig

model_id = "facebook/opt-350m"
adapter_model_id = "ybelkada/opt-350m-lora"
tokenizer = AutoTokenizer.from_pretrained(model_id)
text = "Hello"
inputs = tokenizer(text, return_tensors="pt")

model = AutoModelForCausalLM.from_pretrained(model_id)
peft_config = PeftConfig.from_pretrained(adapter_model_id)

# To get random weights
peft_config.init_lora_weights = False

model.add_adapter(peft_config)
print(model)

model.disable_adapters()
output_disabled = model.generate(**inputs)
print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
>>> Hello, I'm a newbie to this sub. I'm looking for a good place to

model.enable_adapters()
output_enabled = model.generate(**inputs)
print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
>>> Hello, MMMMMMMM
Add multiple adapters
from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig, LoraConfig

model_id = "facebook/opt-350m"

# directly on from_pretrained
model = AutoModelForCausalLM.from_pretrained(model_id)

lora_config = LoraConfig(
    target_modules=["q_proj", "k_proj"],
    init_lora_weights=False
)

model.add_adapter(lora_config, adapter_name="adapter_1")

# attach new adapter with same config
model.add_adapter(lora_config, adapter_name="adapter_2")

model.set_adapter("adapter_1")
output_disabled = model.generate(**inputs)
print(tokenizer.decode(output_disabled[0], skip_special_tokens=True))
>>> Hello, I'm a newbie to this sub. I'm looking for a good place to

model.set_adapter("adapter_2")
output_enabled = model.generate(**inputs)
print(tokenizer.decode(output_enabled[0], skip_special_tokens=True))
>>> Hello, I'm a newbie to the game. I'm looking for a good way to
Save adapters
from transformers import AutoModelForCausalLM, OPTForCausalLM, AutoTokenizer
from peft import PeftConfig, LoraConfig

model_id = "facebook/opt-350m"

# directly on from_pretrained
model = AutoModelForCausalLM.from_pretrained(model_id)

lora_config = LoraConfig(
    target_modules=["q_proj", "k_proj"],
)

model.add_adapter(lora_config)

... # train here

model.save_pretrained(save_dir) 

# you can either load it back with transformers or PEFT

from peft import AutoPeftModelForCausalLM

model = AutoPeftModelForCausalLM.from_pretrained(save_dir)

# or

model = AutoModelForCausalLM.from_pretrained(save_dir)
Train adapters using Trainer

Check this gist: https://gist.github.com/younesbelkada/cdda6e4abcb09e58f6324d75e0d88862

This PR is on par with: huggingface/peft#749

Features to support:

loading PEFT adapters

using multiple adapters

deal with models loaded with accelerate

Loading directly from from_pretrained

Merging adapter weights - to not support

Unload adapter weights - to not support

Training with BC with expected PEFT checkpoints format (do we really want to support training? Shall we just redirect users to load a classic PeftModel if they want to train a model?)

What about save_pretrained ?

Features to not support:

disabling adapters

prompt tuning / prompt learning methods

TODOs:

docs

tests

cc @sgugger @patrickvonplaten @BenjaminBossan @pacman100

What packages to install do i need to run this code?

younesbelkada · 2023-10-15T18:26:40Z

What packages to install do i need to run this code?

Just install the latest transformers & peft

pip install -U peft transformers

… method (huggingface#749) Refactors a bit the internals of some PEFT models and introduces a new method inject_adapter_in_model for users that want to pass a bare model and a peft config to inject adapters in-place into the model. These changes are totally BC with the previous PEFT versions. This PR makes things easier for the PEFT integration in transformers huggingface/transformers#25077 The main goal of the PR is to expose a new API for advanced users that want to integrate PEFT method without the need to use the PeftModel wrapper. A simple use case could be someone that wants to inject adapters into a model and wants to keep the original class of the model without having to offload that to peft that will create a PeftModel. I have faced this issue in huggingface/transformers#25077 Among other things, this PR refactors some internals of PEFT library, while keeping it fully backward compatible. To tackle the main motivation I propose to differentiate things between two type of adapters 1- adapters that are injectable (LoRA, AdaLoRA, IA3) 2- adapters that are not injectable (the rest) As a first iteration this API would be supported only for the scenario 1- / therefore I decided to create 2 abstract classes to make things easy to be able to determine if the adapter layer (e.g. LoraLayer) / adapter module (e.g. LoraModel) does follow the minimal requirement (i.e. needed attributes, etc.) Other related changes: 1- Creates a new property method is_prompt_learning to avoid importing PromptLearningConfig all the way down 2- Introduces a new object TUNERS_MAPPING, which is a mapping of supported pluggable adapters 3- Creates two abstract classes 3.1- BaseTunerLayer: a mixin to check for minimal required attributes that a tuner layer should have active_adapter / _is_plugable 3.2- BaseTuner: a higher level module mixin that should be used for any injectable adapters in the future. --------- Co-authored-by: Benjamin Bossan <[email protected]> Co-authored-by: Patrick von Platen <[email protected]>

… method (#749) Refactors a bit the internals of some PEFT models and introduces a new method inject_adapter_in_model for users that want to pass a bare model and a peft config to inject adapters in-place into the model. These changes are totally BC with the previous PEFT versions. This PR makes things easier for the PEFT integration in transformers huggingface/transformers#25077 The main goal of the PR is to expose a new API for advanced users that want to integrate PEFT method without the need to use the PeftModel wrapper. A simple use case could be someone that wants to inject adapters into a model and wants to keep the original class of the model without having to offload that to peft that will create a PeftModel. I have faced this issue in huggingface/transformers#25077 Among other things, this PR refactors some internals of PEFT library, while keeping it fully backward compatible. To tackle the main motivation I propose to differentiate things between two type of adapters 1- adapters that are injectable (LoRA, AdaLoRA, IA3) 2- adapters that are not injectable (the rest) As a first iteration this API would be supported only for the scenario 1- / therefore I decided to create 2 abstract classes to make things easy to be able to determine if the adapter layer (e.g. LoraLayer) / adapter module (e.g. LoraModel) does follow the minimal requirement (i.e. needed attributes, etc.) Other related changes: 1- Creates a new property method is_prompt_learning to avoid importing PromptLearningConfig all the way down 2- Introduces a new object TUNERS_MAPPING, which is a mapping of supported pluggable adapters 3- Creates two abstract classes 3.1- BaseTunerLayer: a mixin to check for minimal required attributes that a tuner layer should have active_adapter / _is_plugable 3.2- BaseTuner: a higher level module mixin that should be used for any injectable adapters in the future. --------- Co-authored-by: Benjamin Bossan <[email protected]> Co-authored-by: Patrick von Platen <[email protected]>

younesbelkada added 2 commits July 19, 2023 12:05

a draft version

297e311

v2 integration

cedeeda

younesbelkada mentioned this pull request Jul 25, 2023

[core] PEFT refactor + introducing inject_adapter_in_model public method huggingface/peft#749

Merged

patrickvonplaten reviewed Jul 25, 2023

View reviewed changes

fix

cf3c325

BenjaminBossan reviewed Jul 25, 2023

View reviewed changes

younesbelkada added 4 commits July 26, 2023 13:26

make it more generic and works for IA3

c6d0848

Merge remote-tracking branch 'upstream/main' into peft-integration-at…

0343763

…tempt-2

add set adapter and multiple adapters support

867ae2c

fixup

72762de

pacman100 reviewed Jul 26, 2023

View reviewed changes

younesbelkada added 6 commits July 27, 2023 14:17

adapt a bit

02b5802

oops

067fef4

oops

e8cc945

oops

d371173

adapt more

08043d5

fix

619a5d6

younesbelkada mentioned this pull request Jul 31, 2023

config.json file not available #25215

Closed

4 tasks

younesbelkada added 2 commits July 31, 2023 19:02

add more refactor

e61de3b

now works with model class

59b3cb3

sgugger reviewed Aug 1, 2023

View reviewed changes

src/transformers/utils/peft_utils.py Show resolved Hide resolved

younesbelkada mentioned this pull request Aug 1, 2023

[Detr] Fix detr BatchNorm replacement issue #25230

Merged

younesbelkada added 2 commits August 1, 2023 10:22

Merge remote-tracking branch 'upstream/main' into peft-integration-at…

da8dfc5

…tempt-2

change it to instance method as it causes issues with jit.

e67c3c3

younesbelkada commented Aug 1, 2023

View reviewed changes

add CR

eb57382

pacman100 reviewed Aug 2, 2023

View reviewed changes

change method name

babb278

This was referenced Aug 17, 2023

Add PEFT support directly in transformers pipeline #24750

Closed

Add support for Pix2Struct huggingface/peft#796

Closed

younesbelkada added 3 commits August 18, 2023 08:06

Merge remote-tracking branch 'upstream/main' into peft-integration-at…

180545f

…tempt-2

move to lib_integrations

83d0f15

fixup

f739aee

younesbelkada marked this pull request as ready for review August 18, 2023 08:24

younesbelkada requested a review from sgugger August 18, 2023 08:24

this time fixup

fccf419

younesbelkada requested a review from BenjaminBossan August 18, 2023 08:36

sgugger approved these changes Aug 18, 2023

View reviewed changes

younesbelkada and others added 5 commits August 18, 2023 13:21

Apply suggestions from code review

fb6af42

Co-authored-by: Sylvain Gugger <[email protected]>

address final comments

616cfec

refactor to use token

70b1570

add PEFT to DockerFile for slow tests.

2934e69

added pipeline support.

3dd9211

Narsil approved these changes Aug 18, 2023

View reviewed changes

younesbelkada merged commit faed2ca into huggingface:main Aug 18, 2023

younesbelkada deleted the peft-integration-attempt-2 branch August 18, 2023 17:13

This was referenced Aug 18, 2023

🚨🚨🚨 [Refactor] Move third-party related utility files into integrations/ folder 🚨🚨🚨 #25599

Merged

Fix PEFT integration failures on nightly CI #25624

Merged

younesbelkada mentioned this pull request Aug 22, 2023

[core] PEFT integration #24827

Closed

3 tasks

muellerzr mentioned this pull request Aug 30, 2023

Skip pypi transformers until release huggingface/accelerate#1911

Merged

5 tasks

pbarker mentioned this pull request Jan 6, 2024

Support setting multiple adapters #28372

Open

[PEFT] Peft integration alternative design #25077

[PEFT] Peft integration alternative design #25077

Uh oh!

Conversation

younesbelkada commented Jul 25, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What does this PR do?

Load a model with an adapter locally or from the Hub:

Load and attach adapter to an existing model

Attach an adapter, iteratively enable / disable adapters

Add multiple adapters

Save adapters

Train adapters using Trainer

Uh oh!

HuggingFaceDocBuilderDev commented Jul 25, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

sgugger commented Jul 25, 2023 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

patrickvonplaten left a comment

Choose a reason for hiding this comment

Uh oh!

BenjaminBossan left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

pacman100 left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

sgugger left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

pacman100 commented Aug 21, 2023

Uh oh!

BoccheseGiacomo commented Oct 14, 2023

What does this PR do?

Load a model with an adapter locally or from the Hub:

Load and attach adapter to an existing model

Attach an adapter, iteratively enable / disable adapters

Add multiple adapters

Save adapters

Train adapters using Trainer

Uh oh!

younesbelkada commented Oct 15, 2023

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

[`PEFT`] Peft integration alternative design #25077

[`PEFT`] Peft integration alternative design #25077

younesbelkada commented Jul 25, 2023 •

edited

Loading

HuggingFaceDocBuilderDev commented Jul 25, 2023 •

edited

Loading

sgugger commented Jul 25, 2023 •

edited

Loading