From 1fff6746ab363e897fc08d742e7dfde4304d70ac Mon Sep 17 00:00:00 2001 From: Victor SANH Date: Tue, 26 Sep 2023 15:41:15 +0200 Subject: [PATCH] Fix position ids logic instantiation of idefics vision part (#1064) Problem and fix is described here: https://huggingface.co/HuggingFaceM4/idefics-9b/discussions/9 --------- Co-authored-by: Nicolas Patry --- .../models/custom_modeling/idefics_vision.py | 4 +--- 1 file changed, 1 insertion(+), 3 deletions(-) diff --git a/server/text_generation_server/models/custom_modeling/idefics_vision.py b/server/text_generation_server/models/custom_modeling/idefics_vision.py index d933d7c1..30f07095 100644 --- a/server/text_generation_server/models/custom_modeling/idefics_vision.py +++ b/server/text_generation_server/models/custom_modeling/idefics_vision.py @@ -88,12 +88,10 @@ class IdeficsVisionEmbeddings(nn.Module): self.num_patches = (self.image_size // self.patch_size) ** 2 self.num_positions = self.num_patches + 1 - # self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim) self.position_embedding = TensorParallelEmbedding( prefix="model.vision_model.embeddings.position_embedding", weights=weights ) - # self.register_buffer("position_ids", torch.arange(self.num_positions).expand((1, -1)), persistent=False) - self.position_ids = weights.get_tensor(f"{prefix}.position_ids") + self.position_ids = torch.arange(self.num_positions).expand((1, -1)).to(device=weights.device) def forward(self, pixel_values: torch.FloatTensor) -> torch.Tensor: batch_size = pixel_values.shape[0]