[V1] Scatter and gather placeholders in the model runner (#16076)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: mgoin <mgoin64@gmail.com>
Signed-off-by: Roger Wang <ywang@roblox.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>
Co-authored-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Jennifer Zhao <ai.jenniferzhao@gmail.com>

vllm/v1/worker/utils.py

 # SPDX-License-Identifier: Apache-2.0
+from typing import Optional
+
 import torch
 
 
@@ -27,3 +29,46 @@ def sanity_check_mm_encoder_outputs(
         f"but got tensors with shapes {[e.shape for e in mm_embeddings]} "
         "instead. This is most likely due to incorrect implementation "
         "of the model's `get_multimodal_embeddings` method.")
+
+
+def scatter_mm_placeholders(
+    embeds: torch.Tensor,
+    is_embed: Optional[torch.Tensor],
+) -> torch.Tensor:
+    """
+    Scatter the multimodal embeddings into a contiguous tensor that represents
+    the placeholder tokens.
+
+    :class:`vllm.multimodal.processing.PromptUpdateDetails.is_embed`.
+
+    Args:
+        embeds: The multimodal embeddings.
+          Shape: `(num_embeds, embed_dim)`
+        is_embed: A boolean mask indicating which positions in the placeholder
+          tokens need to be filled with multimodal embeddings.
+          Shape: `(num_placeholders, num_embeds)`
+    """
+    if is_embed is None:
+        return embeds
+
+    placeholders = embeds.new_full(
+        (is_embed.shape[0], embeds.shape[-1]),
+        fill_value=torch.nan,
+    )
+    placeholders[is_embed] = embeds
+    return placeholders
+
+
+def gather_mm_placeholders(
+    placeholders: torch.Tensor,
+    is_embed: Optional[torch.Tensor],
+) -> torch.Tensor:
+    """
+    Reconstructs the embeddings from the placeholder tokens.
+
+    This is the operation of :func:`scatter_mm_placeholders`.
+    """
+    if is_embed is None:
+        return placeholders
+
+    return placeholders[is_embed]