Adding a brand new fine-tuning method, Bone. #2148

JL-er · 2024-10-14T09:03:12Z

In the paper "https://arxiv.org/pdf/2409.15371", we introduce a brand new PEFT (Parameter-Efficient Fine-Tuning) method: BONE (BLOCK AFFINE). This is a completely new structure that differs from the LoRA series. In terms of performance, it has already surpassed PISSA.

BenjaminBossan

Thanks for this PR to add the Bone method to PEFT. It's already looking fine but I found a few issues, please check out my comments.

Apart from those, to get to the next stage, could you please:

Run make style
Clarify the following: Is this implementation using row or column blocks? Would it make sense to add both as an option?
Related to that, depending on the option and the choice of r, we probably need to add some sanity checks on the shape of the targeted weight matrix to make sure that we can evenly divide it into blocks, right?
Let's add a few tests cases to get started. For this, check these lines and implement something similar for Bone. Then you can run the tests using pytest tests/test_custom_models.py -k bone -v to help identify potential issues.

src/peft/tuners/bone/config.py

src/peft/tuners/bone/layer.py

BenjaminBossan · 2024-10-15T12:24:43Z

src/peft/tuners/bone/layer.py

+                orig_weight = self.base_layer.weight.data.clone()
+                delta_weight = self.get_delta_weight(active_adapter, orig_weight)
+                orig_weight += delta_weight
+
+            result = F.linear(input=x, weight=orig_weight, bias=self.base_layer.bias)


Just an idea from my side:

Right now, you're calculating W2 = W + dW, then result = W2@x + bias. Instead, I wonder if we could re-arrange this somehow to use result = self.base_layer(x, *args, **kwargs), then result = result + bone_result. The bone_result would ideally be calculated in a parameter-efficient way, which should hopefully alleviate the memory overhead that Bone has compared to LoRA. For this, I wonder if the Bone contribution can be calculated group-wise and added incrementally instead of instantiating a big dW, so basically trading more compute for less memory. Do you think that would be possible?

I think 'result = result + bone_result' directly is not feasible. The Bone repository has flash-bone that integrates linear and bone calculations, but I'm not very skilled at writing operators, so the current optimization speed is insufficient.

If we have Y = (W + W@dW + dW)@X + b, we could rewrite it as Y = W@X + (W@dW + dW)@X + b, right? And since W@X + b is the base result, i.e. self.base_layer(x, *args, **kwargs), this should work?

Another question would be if we can re-define delta_weight in a way that would allow to reduce the memory overhead.

https://github.com/huggingface/peft/pull/2148/files#diff-34b7ebce9fc88273adb34c46598bee8a92b72f5c1e7c420585e03938396a1448R210-R214

My thought was if we can somehow calculate the activations per block and add them incrementally to the base result. Maybe it's not possible or it would not help to reduce memory, but if we can trade off compute for less memory, it is usually a worthwhile trade. If it's simple not possible, feel free to ignore my comment :)

Is it possible to add checkpointing specifically for bone_result? The optimization here indeed requires careful consideration.

I would not know how. I think users should just use the existing checkpointing utilities if they want to make use of this.

src/peft/tuners/bone/layer.py

src/peft/tuners/bone/model.py

JL-er · 2024-10-17T08:32:58Z

Thanks for this PR to add the Bone method to PEFT. It's already looking fine but I found a few issues, please check out my comments.

Apart from those, to get to the next stage, could you please:

Run make style

Clarify the following: Is this implementation using row or column blocks? Would it make sense to add both as an option?

Related to that, depending on the option and the choice of r, we probably need to add some sanity checks on the shape of the targeted weight matrix to make sure that we can evenly divide it into blocks, right?

Let's add a few tests cases to get started. For this, check these lines and implement something similar for Bone. Then you can run the tests using pytest tests/test_custom_models.py -k bone -v to help identify potential issues.

Run make style √
Currently, column grouping is supported. This will be explained in the code (as there are still many issues to address, such as quantization training and memory saving). Other grouping methods will be added later.
I added a check [here] (https://github.com/JL-er/PEFT-Bone/blob/main/src/peft/tuners/bone/layer.py#L64)
Currently, it is possible to fine-tune models like LLaMA normally, but due to the checks mentioned in issue 3, some structures do not pass.

I believe it can be merged now.

BenjaminBossan

Thanks for the updates to the PR. We still have a few boxes to tick before we can merge though.

Currently, it is possible to fine-tune models like LLaMA normally, but due to the checks mentioned in issue 3, some structures do not pass.

This needs to be addressed. Can we change the values so that the check passes? In the worst case, we can define a new module class where the shapes fit. The goal is that all tests are green.

Also:

Great additions to the test but we should also add tests for the other model types. For instance, check how HRA did it.
There is still some commented code here and there.
We need an entry to the docs, again, check HRA. It's important to mention to pros and cons of the method.
Let's also add an example that users can use to get started. This will greatly help with adoption. Ideally, this would be a script similar to the experiments in your paper/original repo. That way, we can verify that the PEFT implementation is correct.
And after you finished, don't forget to call make style.

BenjaminBossan · 2024-10-17T15:42:17Z

src/peft/tuners/bone/layer.py

+                orig_weight = self.base_layer.weight.data.clone()
+                delta_weight = self.get_delta_weight(active_adapter, orig_weight)
+                orig_weight += delta_weight
+
+            result = F.linear(input=x, weight=orig_weight, bias=self.base_layer.bias)


I would not know how. I think users should just use the existing checkpointing utilities if they want to make use of this.

BenjaminBossan · 2024-10-17T15:42:28Z

src/peft/tuners/bone/layer.py

+            raise TypeError(f"Bone is not implemented for base layers of type {type(base_layer).__name__}")
+
+        # Initialize weights
+    #     if init_weights:


Did you consider this?

JL-er · 2024-10-18T08:29:37Z

Thanks for the updates to the PR. We still have a few boxes to tick before we can merge though.

Currently, it is possible to fine-tune models like LLaMA normally, but due to the checks mentioned in issue 3, some structures do not pass.

This needs to be addressed. Can we change the values so that the check passes? In the worst case, we can define a new module class where the shapes fit. The goal is that all tests are green.

Also:

Great additions to the test but we should also add tests for the other model types. For instance, check how HRA did it.

There is still some commented code here and there.

We need an entry to the docs, again, check HRA. It's important to mention to pros and cons of the method.

Let's also add an example that users can use to get started. This will greatly help with adoption. Ideally, this would be a script similar to the experiments in your paper/original repo. That way, we can verify that the PEFT implementation is correct.

And after you finished, don't forget to call make style.

Most of the tests have now passed, but due to Bone being initialized to 0, an error occurs when encountering
#sanity check that the logits are different
assert not torch.allclose(logits_base, logits_peft, atol=atol, rtol=rtol)

BenjaminBossan · 2024-10-18T15:18:13Z

Most of the tests have now passed, but due to Bone being initialized to 0, an error occurs when encountering
#sanity check that the logits are different

Thanks for working on those. Regarding that, I have a suggestion. Right now, we have:

        if init_weights:
            self.reset_bone_parameters(adapter_name, r)

...

    def reset_bone_parameters(self, adapter_name: str, r):
        self.bone_block[adapter_name] = nn.Parameter(torch.zeros(self.out_features // r, r, r), requires_grad=True)

I would suggest that for init_weights=False, we create the bone_block with non-zero entries, e.g. with random uniform weights. That way, for those tests, we can just pass init_weights=False and they should pass. A lot of other methods already do that.

JL-er · 2024-10-19T16:21:44Z

Most of the tests have now passed, but due to Bone being initialized to 0, an error occurs when encountering
#sanity check that the logits are different

Thanks for working on those. Regarding that, I have a suggestion. Right now, we have:
        if init_weights:
            self.reset_bone_parameters(adapter_name, r)

...

    def reset_bone_parameters(self, adapter_name: str, r):
        self.bone_block[adapter_name] = nn.Parameter(torch.zeros(self.out_features // r, r, r), requires_grad=True)
I would suggest that for init_weights=False, we create the bone_block with non-zero entries, e.g. with random uniform weights. That way, for those tests, we can just pass init_weights=False and they should pass. A lot of other methods already do that.

The last error, I don't understand why 'different' can pass but 'same' cannot.

JL-er · 2024-10-21T07:44:14Z

Most of the tests have now passed, but due to Bone being initialized to 0, an error occurs when encountering
#sanity check that the logits are different

Thanks for working on those. Regarding that, I have a suggestion. Right now, we have:
        if init_weights:
            self.reset_bone_parameters(adapter_name, r)

...

    def reset_bone_parameters(self, adapter_name: str, r):
        self.bone_block[adapter_name] = nn.Parameter(torch.zeros(self.out_features // r, r, r), requires_grad=True)
I would suggest that for init_weights=False, we create the bone_block with non-zero entries, e.g. with random uniform weights. That way, for those tests, we can just pass init_weights=False and they should pass. A lot of other methods already do that.
The last error, I don't understand why 'different' can pass but 'same' cannot.

This issue has been resolved because 'Bone same' reuses 'base_model.weight'. However, since the Bone structure requires computation through 'base_model.weight', I modified the forward part of the calculation. The calculation method in the commented section might not be compatible with the Bone structure.

BenjaminBossan

Thanks a lot for the updates. We're making good progress. I was checking the tests and still found quite a few failures. Btw you can run them locally using pytest tests/ -k "bone".

Currently, tests involving GPT2 are failing because Bone does not support Conv1D layers. You can skip these tests, check how other methods achieve that:

peft/tests/test_decoder_models.py

Lines 70 to 83 in e8259ff

    
           def skip_adalora_or_oft_or_hra_and_gpt2(test_list): 
        
               return [ 
        
                   test 
        
                   for test in test_list 
        
                   if not ( 
        
                       ("GPT2LMHeadModel" in test[1]) 
        
                       and ( 
        
                           (test[2] == AdaLoraConfig) 
        
                           or (test[2] == BOFTConfig) 
        
                           or (test[2] == HRAConfig) 
        
                           or (test[2] == OFTConfig) 
        
                       ) 
        
                   ) 
        
               ]

We also have multiple tests failing that require Bone to be initialized as non-identity transform. For those tests to pass, we only need to ensure that init_weights=False is passed to the config. Please check how this is achieved for other methods:

peft/tests/test_decoder_models.py

Lines 212 to 263 in e8259ff

    
               @parameterized.expand( 
        
                   PeftTestConfigManager.get_grid_parameters( 
        
                       { 
        
                           "model_ids": PEFT_DECODER_MODELS_TO_TEST, 
        
                           "lora_kwargs": {"init_lora_weights": [False]}, 
        
                           "adalora_kwargs": {"init_lora_weights": [False]}, 
        
                           "ia3_kwargs": {"init_ia3_weights": [False]}, 
        
                           "boft_kwargs": {"init_weights": [False]}, 
        
                           "oft_kwargs": {"init_weights": [False]}, 
        
                           "vera_kwargs": {"init_weights": [False]}, 
        
                           "fourierft_kwargs": {"init_weights": [False]}, 
        
                           "hra_kwargs": {"init_weights": [False]}, 
        
                           "task_type": "CAUSAL_LM", 
        
                       }, 
        
                   ) 
        
               ) 
        
               def test_merge_layers(self, test_name, model_id, config_cls, config_kwargs): 
        
                   self._test_merge_layers(model_id, config_cls, config_kwargs) 
        
               @parameterized.expand( 
        
                   PeftTestConfigManager.get_grid_parameters( 
        
                       { 
        
                           "model_ids": PEFT_DECODER_MODELS_TO_TEST, 
        
                           "lora_kwargs": {"init_lora_weights": [False]}, 
        
                           "ia3_kwargs": {"init_ia3_weights": [False]}, 
        
                           "boft_kwargs": {"init_weights": [False]}, 
        
                           "oft_kwargs": {"init_weights": [False]}, 
        
                           "vera_kwargs": {"init_weights": [False]}, 
        
                           "fourierft_kwargs": {"init_weights": [False]}, 
        
                           "hra_kwargs": {"init_weights": [False]}, 
        
                           "task_type": "CAUSAL_LM", 
        
                       }, 
        
                       filter_params_func=skip_oft_or_hra_and_gpt2, 
        
                   ) 
        
               ) 
        
               def test_merge_layers_multi(self, test_name, model_id, config_cls, config_kwargs): 
        
                   self._test_merge_layers_multi(model_id, config_cls, config_kwargs) 
        
               @parameterized.expand( 
        
                   PeftTestConfigManager.get_grid_parameters( 
        
                       { 
        
                           "model_ids": PEFT_DECODER_MODELS_TO_TEST, 
        
                           "lora_kwargs": {"init_lora_weights": [False]}, 
        
                           "ia3_kwargs": {"init_ia3_weights": [False]}, 
        
                           "boft_kwargs": {"init_weights": [False]}, 
        
                           "oft_kwargs": {"init_weights": [False]}, 
        
                           "task_type": "CAUSAL_LM", 
        
                       }, 
        
                   ) 
        
               ) 
        
               def test_merge_layers_nan(self, test_name, model_id, config_cls, config_kwargs): 
        
                   self._test_merge_layers_nan(model_id, config_cls, config_kwargs)

peft/tests/test_encoder_decoder_models.py

Lines 92 to 104 in e8259ff

    
               @parameterized.expand( 
        
                   PeftTestConfigManager.get_grid_parameters( 
        
                       { 
        
                           "model_ids": PEFT_ENCODER_DECODER_MODELS_TO_TEST, 
        
                           "lora_kwargs": {"init_lora_weights": [False]}, 
        
                           "adalora_kwargs": {"init_lora_weights": [False]}, 
        
                           "ia3_kwargs": {"init_ia3_weights": [False]}, 
        
                           "vera_kwargs": {"init_weights": [False]}, 
        
                           "hra_kwargs": {"init_weights": [False]}, 
        
                           "task_type": "SEQ_2_SEQ_LM", 
        
                       }, 
        
                   ) 
        
               )

peft/tests/test_feature_extraction_models.py

Lines 109 to 123 in e8259ff

    
               @parameterized.expand( 
        
                   PeftTestConfigManager.get_grid_parameters( 
        
                       { 
        
                           "model_ids": PEFT_FEATURE_EXTRACTION_MODELS_TO_TEST, 
        
                           "lora_kwargs": {"init_lora_weights": [False]}, 
        
                           "adalora_kwargs": {"init_lora_weights": [False]}, 
        
                           "ia3_kwargs": {"init_ia3_weights": [False]}, 
        
                           "boft_kwargs": {"init_weights": [False]}, 
        
                           "oft_kwargs": {"init_weights": [False]}, 
        
                           "vera_kwargs": {"init_weights": [False]}, 
        
                           "hra_kwargs": {"init_weights": [False]}, 
        
                           "task_type": "FEATURE_EXTRACTION", 
        
                       }, 
        
                   ) 
        
               )

(there are more instances in the same files, please ensure that all are covered)

Let's fix those issues and then hopefully all the tests should pass.

tests/testing_common.py

HuggingFaceDocBuilderDev · 2024-10-21T14:19:07Z

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

JL-er · 2024-10-21T14:32:00Z

Is there anything else I need to do about this error?

JL-er · 2024-10-21T14:43:53Z

Is there anything else I need to do about this error?

I rerun make style, and it seems to be fine now. Thank you！！！

JL-er · 2024-10-21T16:32:11Z

I didn't encounter this error when running locally, but I tried to fix it this way. I'm not sure if it will work (or I could set r=2 during initialization).

BenjaminBossan

Thanks for further working on fixing the tests. One part is not working, but I added a suggestion that will fix the problem.

Also, since this PR is now getting close to the finish line, did you have time to work on these point I mentioned earlier?

We need an entry to the docs, again, check HRA. It's important to mention to pros and cons of the method.

Let's also add an example that users can use to get started. This will greatly help with adoption. Ideally, this would be a script similar to the experiments in your paper/original repo. That way, we can verify that the PEFT implementation is correct.

tests/test_custom_models.py

JL-er · 2024-10-22T14:02:23Z

We need an entry to the docs, again, check HRA. It's important to mention to pros and cons of the method.

Yes, I need to explain the pros and cons of Bone, but where should I write it?
I plan to move the finetune files from the Bone repository into the example directory for reproducibility. Regarding the correctness of PEFT adaptation, I have tested it locally 3 times and all results were normal. After merging with PEFT, I will test it once more.

BenjaminBossan · 2024-10-22T15:29:37Z

Yes, I need to explain the pros and cons of Bone, but where should I write it?

You could check this PR and how the docs are updated there (ignore the rest).

I plan to move the finetune files from the Bone repository into the example directory for reproducibility.

That would be a good addition. Let's make it easy for users to pick this method up and adapt it to their own needs.

JL-er · 2024-10-23T09:28:41Z

Yes, I need to explain the pros and cons of Bone, but where should I write it?

You could check this PR and how the docs are updated there (ignore the rest).

I plan to move the finetune files from the Bone repository into the example directory for reproducibility.

That would be a good addition. Let's make it easy for users to pick this method up and adapt it to their own needs.

The aforementioned issues have all been resolved.

JL-er · 2024-10-23T10:09:45Z

I'm sorry, due to a conflict with the official PEFT branch, I mistakenly merged it, which caused my previous repository to be deleted and the PR to be forcibly closed. I have now fixed the issue. Should I resubmit or can you reopen my old PR?

BenjaminBossan · 2024-10-23T11:32:39Z

If you can somehow salvage this branch and re-open this PR, it would be better, since that allows us to have the whole discussion in one place. But don't worry if this doesn't work, we can also continue the discussion on the new PR. Just let me know your decision.

JL-er · 2024-10-23T11:57:41Z

If you can somehow salvage this branch and re-open this PR, it would be better, since that allows us to have the whole discussion in one place. But don't worry if this doesn't work, we can also continue the discussion on the new PR. Just let me know your decision.

I am really sorry that I could not recover the old branch. Let’s discuss further question under the new PR, which is ready to be tested.

BenjaminBossan requested changes Oct 15, 2024

View reviewed changes

BenjaminBossan requested changes Oct 17, 2024

View reviewed changes

BenjaminBossan requested changes Oct 21, 2024

View reviewed changes

tests/testing_common.py Outdated Show resolved Hide resolved

BenjaminBossan requested changes Oct 22, 2024

View reviewed changes

tests/test_custom_models.py Outdated Show resolved Hide resolved

tests/test_custom_models.py Outdated Show resolved Hide resolved

JL-er closed this Oct 23, 2024

JL-er force-pushed the main branch from b031a40 to 0d58942 Compare October 23, 2024 09:32

JL-er mentioned this pull request Oct 23, 2024

Support Bone #2172

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Adding a brand new fine-tuning method, Bone. #2148

Adding a brand new fine-tuning method, Bone. #2148

JL-er commented Oct 14, 2024

BenjaminBossan left a comment

BenjaminBossan Oct 15, 2024

JL-er Oct 16, 2024

BenjaminBossan Oct 16, 2024

JL-er Oct 17, 2024

BenjaminBossan Oct 17, 2024

JL-er commented Oct 17, 2024

BenjaminBossan left a comment

BenjaminBossan Oct 17, 2024

BenjaminBossan Oct 17, 2024

JL-er commented Oct 18, 2024 •

edited

Loading

BenjaminBossan commented Oct 18, 2024

JL-er commented Oct 19, 2024

JL-er commented Oct 21, 2024

BenjaminBossan left a comment

HuggingFaceDocBuilderDev commented Oct 21, 2024

JL-er commented Oct 21, 2024

JL-er commented Oct 21, 2024 •

edited

Loading

JL-er commented Oct 21, 2024 •

edited

Loading

BenjaminBossan left a comment

JL-er commented Oct 22, 2024

BenjaminBossan commented Oct 22, 2024

JL-er commented Oct 23, 2024

JL-er commented Oct 23, 2024

BenjaminBossan commented Oct 23, 2024

JL-er commented Oct 23, 2024 •

edited

Loading

	def skip_adalora_or_oft_or_hra_and_gpt2(test_list):
	return [
	test
	for test in test_list
	if not (
	("GPT2LMHeadModel" in test[1])
	and (
	(test[2] == AdaLoraConfig)
	or (test[2] == BOFTConfig)
	or (test[2] == HRAConfig)
	or (test[2] == OFTConfig)
	)
	)
	]

	@parameterized.expand(
	PeftTestConfigManager.get_grid_parameters(
	{
	"model_ids": PEFT_DECODER_MODELS_TO_TEST,
	"lora_kwargs": {"init_lora_weights": [False]},
	"adalora_kwargs": {"init_lora_weights": [False]},
	"ia3_kwargs": {"init_ia3_weights": [False]},
	"boft_kwargs": {"init_weights": [False]},
	"oft_kwargs": {"init_weights": [False]},
	"vera_kwargs": {"init_weights": [False]},
	"fourierft_kwargs": {"init_weights": [False]},
	"hra_kwargs": {"init_weights": [False]},
	"task_type": "CAUSAL_LM",
	},
	)
	)
	def test_merge_layers(self, test_name, model_id, config_cls, config_kwargs):
	self._test_merge_layers(model_id, config_cls, config_kwargs)

	@parameterized.expand(
	PeftTestConfigManager.get_grid_parameters(
	{
	"model_ids": PEFT_DECODER_MODELS_TO_TEST,
	"lora_kwargs": {"init_lora_weights": [False]},
	"ia3_kwargs": {"init_ia3_weights": [False]},
	"boft_kwargs": {"init_weights": [False]},
	"oft_kwargs": {"init_weights": [False]},
	"vera_kwargs": {"init_weights": [False]},
	"fourierft_kwargs": {"init_weights": [False]},
	"hra_kwargs": {"init_weights": [False]},
	"task_type": "CAUSAL_LM",
	},
	filter_params_func=skip_oft_or_hra_and_gpt2,
	)
	)
	def test_merge_layers_multi(self, test_name, model_id, config_cls, config_kwargs):
	self._test_merge_layers_multi(model_id, config_cls, config_kwargs)

	@parameterized.expand(
	PeftTestConfigManager.get_grid_parameters(
	{
	"model_ids": PEFT_DECODER_MODELS_TO_TEST,
	"lora_kwargs": {"init_lora_weights": [False]},
	"ia3_kwargs": {"init_ia3_weights": [False]},
	"boft_kwargs": {"init_weights": [False]},
	"oft_kwargs": {"init_weights": [False]},
	"task_type": "CAUSAL_LM",
	},
	)
	)
	def test_merge_layers_nan(self, test_name, model_id, config_cls, config_kwargs):
	self._test_merge_layers_nan(model_id, config_cls, config_kwargs)

	@parameterized.expand(
	PeftTestConfigManager.get_grid_parameters(
	{
	"model_ids": PEFT_ENCODER_DECODER_MODELS_TO_TEST,
	"lora_kwargs": {"init_lora_weights": [False]},
	"adalora_kwargs": {"init_lora_weights": [False]},
	"ia3_kwargs": {"init_ia3_weights": [False]},
	"vera_kwargs": {"init_weights": [False]},
	"hra_kwargs": {"init_weights": [False]},
	"task_type": "SEQ_2_SEQ_LM",
	},
	)
	)

	@parameterized.expand(
	PeftTestConfigManager.get_grid_parameters(
	{
	"model_ids": PEFT_FEATURE_EXTRACTION_MODELS_TO_TEST,
	"lora_kwargs": {"init_lora_weights": [False]},
	"adalora_kwargs": {"init_lora_weights": [False]},
	"ia3_kwargs": {"init_ia3_weights": [False]},
	"boft_kwargs": {"init_weights": [False]},
	"oft_kwargs": {"init_weights": [False]},
	"vera_kwargs": {"init_weights": [False]},
	"hra_kwargs": {"init_weights": [False]},
	"task_type": "FEATURE_EXTRACTION",
	},
	)
	)

Adding a brand new fine-tuning method, Bone. #2148

Adding a brand new fine-tuning method, Bone. #2148

Conversation

JL-er commented Oct 14, 2024

BenjaminBossan left a comment

Choose a reason for hiding this comment

BenjaminBossan Oct 15, 2024

Choose a reason for hiding this comment

JL-er Oct 16, 2024

Choose a reason for hiding this comment

BenjaminBossan Oct 16, 2024

Choose a reason for hiding this comment

JL-er Oct 17, 2024

Choose a reason for hiding this comment

BenjaminBossan Oct 17, 2024

Choose a reason for hiding this comment

JL-er commented Oct 17, 2024

BenjaminBossan left a comment

Choose a reason for hiding this comment

BenjaminBossan Oct 17, 2024

Choose a reason for hiding this comment

BenjaminBossan Oct 17, 2024

Choose a reason for hiding this comment

JL-er commented Oct 18, 2024 • edited Loading

BenjaminBossan commented Oct 18, 2024

JL-er commented Oct 19, 2024

JL-er commented Oct 21, 2024

BenjaminBossan left a comment

Choose a reason for hiding this comment

HuggingFaceDocBuilderDev commented Oct 21, 2024

JL-er commented Oct 21, 2024

JL-er commented Oct 21, 2024 • edited Loading

JL-er commented Oct 21, 2024 • edited Loading

BenjaminBossan left a comment

Choose a reason for hiding this comment

JL-er commented Oct 22, 2024

BenjaminBossan commented Oct 22, 2024

JL-er commented Oct 23, 2024

JL-er commented Oct 23, 2024

BenjaminBossan commented Oct 23, 2024

JL-er commented Oct 23, 2024 • edited Loading

JL-er commented Oct 18, 2024 •

edited

Loading

JL-er commented Oct 21, 2024 •

edited

Loading

JL-er commented Oct 21, 2024 •

edited

Loading

JL-er commented Oct 23, 2024 •

edited

Loading