Added error when sequence length is bigger than max_position_embeddings

[Taha1506]

Fixes # 32154

[Taha1506]

Hi, It would be nice If you give some time to reviewing this pull request. @ArthurZucker . It relates to this issue.

[amyeroberts]

• No need to create a new variable if it's only used once
• We should instead store a variable from the config to avoid expensive checks on the weight size for every forward pass

Suggested change

	max_position_embedding = self.position_embedding.weight.shape[0]
	if seq_length > max_position_embedding:
	if seq_length > self.max_position_embeddings:

And in the init

    def __init__(self, config: CLIPTextConfig):
        super().__init__()
        embed_dim = config.hidden_size
        self.max_position_embeddings = config.max_position_embeddings

[amyeroberts]

I realised I asked for this change -- coming back to it, I think that we shouldn't set self.max_position_embeddings as the user can change the embeddings with set_input_embeddings --> the value from the config will become out of date. In this case, checking the shapes as before is less brittle, could you change it back to this? Apologies for flip-flopping here.

[Taha1506]

Hi! Can you take a look at this pull request again please? Thank you. @amyeroberts

[ArthurZucker]

LGTM, I checked that all config have this attribute

[ArthurZucker]

This is gonna break remote code, as they won'strore the self.max_position embedding

[ArthurZucker]

Fine for me @amyeroberts feel free to merge

[Taha1506]

@amyeroberts
Sorry for the delay, I've made the requested changes.

[Rocketknight1]

LGTM and we have reviews already. Thanks for the helpful error message!