feat: integrate image tokens into inputs embeds

router/src/config.rs

@@ -112,9 +112,20 @@ pub struct ClipVisionModel {
 
 #[derive(Clone, Debug, Serialize, Deserialize)]
 #[serde(rename_all = "snake_case")]
-pub struct Idefics3 {
+pub struct Idefics3 {}
-    pub(crate) vision_encoder_max_image_size: usize,
+
-    pub(crate) image_seq_len: usize,
+impl Idefics3 {
+    pub fn get_max_longest_edge(&self) -> usize {
+        364
+    }
+
+    pub fn get_number_of_features(&self) -> usize {
+        169
+    }
+
+    pub fn get_max_longest_edge_for_image_resize(&self) -> usize {
+        1456
+    }
 }
 
 #[derive(Clone, Debug, Serialize, Deserialize)]

server/text_generation_server/models/custom_modeling/idefics2.py

@@ -753,6 +753,19 @@ class Idefics3ForConditionalGeneration(nn.Module):
             config.pad_token_id if config.pad_token_id is not None else -1
         )
 
+    def _merge_input_ids_with_image_features(
+        self,
+        input_ids: torch.Tensor,
+        inputs_embeds: torch.Tensor,
+        image_features: torch.Tensor,
+    ):
+        """In place merges in vision_embeddings with inputs_embeds."""
+        # mask = input_ids == self.config.image_token_index
+        mask = input_ids == self.config.image_token_id
+        # Let's pray we have enabled enough slots !
+        inputs_embeds[mask] = image_features.view(-1, image_features.shape[-1])
+        return inputs_embeds
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -835,25 +848,22 @@ class Idefics3ForConditionalGeneration(nn.Module):
 
                 all_states.append(image_hidden_states)
             image_hidden_states = torch.stack(all_states, dim=0)
-            # When we generate, we don't want to replace the potential image_token_id that we generated by images
+            # TODO: remove when prefill image tokens are handled correctly
-            # that simply don't exist
+            # * for now dummy tokens are added instead of the image tokens output byt the vision model
-            # TODO: finish implementing the image token replacement
+            mask_size = (input_ids == self.config.image_token_id).sum().item()
+            unrolled_image_size = (
+                image_hidden_states.shape[1] * image_hidden_states.shape[2]
+            )
+            diff = mask_size - unrolled_image_size
+            if diff > 0:
+                print(
+                    f"Mask size {mask_size} is greater than the number of images {unrolled_image_size}."
+                )
 
-            # inputs_embeds = self.inputs_merger(
+            if mask_size == unrolled_image_size:
-            #     input_ids=input_ids,
+                inputs_embeds = self._merge_input_ids_with_image_features(
-            #     inputs_embeds=inputs_embeds,
+                    input_ids, inputs_embeds, image_hidden_states
-            #     image_hidden_states=image_hidden_states,
+                )
-            # )
-
-            # import ipdb; ipdb.set_trace()
-            # num_images, _, vision_hidden_size = image_hidden_states.shape
-            # special_image_token_mask = input_ids == self.image_token_id
-            # new_inputs_embeds = inputs_embeds.clone()
-            # reshaped_image_hidden_states = image_hidden_states.view(-1, vision_hidden_size).to(
-            #     inputs_embeds.dtype
-            # )  # cast to the dtype of the input_embeds to support quantized models
-            # new_inputs_embeds[special_image_token_mask] = reshaped_image_hidden_states
-            # inputs_embeds = new_inputs_embeds
 
         hidden_states = self.text_model.model(
             inputs_embeds=inputs_embeds,

server/text_generation_server/models/vlm_causal_lm.py

@@ -23,6 +23,75 @@ tracer = trace.get_tracer(__name__)
 IDEFICS2_FAKE_TOKEN = "<fake_token_around_image>"
 IDEFICS2_IMAGE_TOKEN = "<image>"
 
+IDEFICS3_IMAGE_TOKEN = "<image>"
+IDEFICS3_FAKE_IMAGE_TOKEN = "<fake_token_around_image>"
+IDEFICS3_GLOBAL_IMG_TOKEN = "<global-img>"
+
+
+def _prompt_split_image(
+    image_seq_len,
+    image_rows,
+    image_cols,
+    fake_token_around_image,
+    image_token,
+    global_img_token,
+):
+    """Prompt with expanded image tokens for when the image is split into patches."""
+    text_split_images = ""
+    for n_h in range(image_rows):
+        for n_w in range(image_cols):
+            text_split_images += (
+                f"{fake_token_around_image}"
+                + f"<row_{n_h + 1}_col_{n_w + 1}>"
+                + f"{image_token}" * image_seq_len
+            )
+        text_split_images += "\n"
+
+    text_split_images += (
+        f"\n{fake_token_around_image}"
+        + f"{global_img_token}"
+        + f"{image_token}" * image_seq_len
+        + f"{fake_token_around_image}"
+    )
+    return text_split_images
+
+
+def _prompt_single_image(
+    image_seq_len, fake_token_around_image, image_token, global_img_token
+):
+    """Prompt with expanded image tokens for a single image."""
+    return (
+        f"{fake_token_around_image}"
+        + f"{global_img_token}"
+        + f"{image_token}" * image_seq_len
+        + f"{fake_token_around_image}"
+    )
+
+
+def get_image_prompt_string(
+    image_rows,
+    image_cols,
+    image_seq_len,
+    fake_token_around_image,
+    image_token,
+    global_img_token,
+):
+    if image_rows == 0 and image_cols == 0:
+        return _prompt_single_image(
+            image_seq_len,
+            fake_token_around_image=fake_token_around_image,
+            image_token=image_token,
+            global_img_token=global_img_token,
+        )
+    return _prompt_split_image(
+        image_seq_len,
+        image_rows,
+        image_cols,
+        fake_token_around_image,
+        image_token,
+        global_img_token,
+    )
+
 
 def get_anyres_image_grid_shape(image_size, grid_pinpoints, patch_size):
     """
@@ -55,8 +124,22 @@ def image_text_replacement(processor, image_input, config, image_id: int) -> str
             image_str *= 5
         return image_str
     if config.model_type == "idefics3":
-        image_str = f"{IDEFICS2_FAKE_TOKEN}{IDEFICS2_IMAGE_TOKEN}{IDEFICS2_FAKE_TOKEN}"
+        # TODO: implement this in a more general way
-        image_str = ""
+        n_rows = image_input["rows"][0][image_id]
+        n_cols = image_input["cols"][0][image_id]
+
+        # TODO: avoid using hardcoded values
+        image_seq_len = 169  # default value
+        # image_seq_len = int(((image_size // patch_size) ** 2) / (scale_factor**2))
+
+        image_str = get_image_prompt_string(
+            n_rows,
+            n_cols,
+            image_seq_len,
+            image_token=IDEFICS3_IMAGE_TOKEN,
+            fake_token_around_image=IDEFICS3_FAKE_IMAGE_TOKEN,
+            global_img_token=IDEFICS3_GLOBAL_IMG_TOKEN,
+        )
         return image_str
     elif config.model_type == "llava_next":
         height, width = image_input["image_sizes"][image_id]
@@ -85,6 +168,10 @@ def image_text_replacement_fixup(config, text: str) -> str:
         return text.replace(
             f"{IDEFICS2_FAKE_TOKEN}{IDEFICS2_FAKE_TOKEN}", IDEFICS2_FAKE_TOKEN
         )
+    if config.model_type == "idefics3":
+        return text.replace(
+            f"{IDEFICS2_FAKE_TOKEN}{IDEFICS2_FAKE_TOKEN}", IDEFICS2_FAKE_TOKEN
+        )
     return text
 
 
@@ -198,7 +285,9 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
                     raise RuntimeError(f"Invalid chunk type {chunk_type}")
 
         if images:
-            image_inputs = processor.image_processor(images, return_tensors="pt")
+            image_inputs = processor.image_processor(
+                images, return_tensors="pt", return_row_col_info=True
+            )
         else:
             image_inputs = None
 
@@ -212,9 +301,10 @@ class VlmCausalLMBatch(FlashCausalLMBatch):
                 if chunk_type == "text":
                     full_text += chunk.text
                 elif chunk_type == "image":
-                    full_text += image_text_replacement(
+                    replacement_text = image_text_replacement(
                         processor, image_inputs, config, image_id
                     )
+                    full_text += replacement_text
                     image_id += 1
 
             full_text = image_text_replacement_fixup(config, full_text)
@@ -289,7 +379,7 @@ class VlmCausalLM(FlashCausalLM):
             model_id,
             revision=revision,
             trust_remote_code=trust_remote_code,
-            **processor_kwargs,
+            # **processor_kwargs,
         )
         self.batch_class = batch_class
         # import ipdb; ipdb.set_trace()