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

Signed-off-by: DarkLight1337 <[email protected]>
Signed-off-by: mgoin <[email protected]>
Signed-off-by: Roger Wang <[email protected]>
Co-authored-by: DarkLight1337 <[email protected]>
Co-authored-by: mgoin <[email protected]>
Co-authored-by: Jennifer Zhao <[email protected]>

Author: 4 people

docs/source/contributing/model/multimodal.md

@@ -860,8 +860,8 @@ prompt_tokens, prompts_length = _tokenize_prompts_with_image_and_batch(
 )
 ```
 
-To accommodate this, instead of a string you can return an instance of {class}`~vllm.multimodal.processing.PromptUpdateDetails`
-with different `full` and `feature` attributes:
+To assign the vision embeddings to only the image tokens, instead of a string
+you can return an instance of {class}`~vllm.multimodal.processing.PromptUpdateDetails`:
 
 ```python
 hf_config = self.info.get_hf_config()
@@ -879,9 +879,9 @@ def get_replacement_fuyu(item_idx: int):
     image_tokens = ([_IMAGE_TOKEN_ID] * ncols +
                     [_NEWLINE_TOKEN_ID]) * nrows
 
-    return PromptUpdateDetails(
-        full=image_tokens + [bos_token_id],
-        features=image_tokens,
+    return PromptUpdateDetails.select_token_id(
+        image_tokens + [bos_token_id],
+        embed_token_id=_IMAGE_TOKEN_ID,
     )
 ```
 
@@ -914,9 +914,9 @@ def _get_prompt_updates(
         image_tokens = ([_IMAGE_TOKEN_ID] * ncols +
                         [_NEWLINE_TOKEN_ID]) * nrows
 
-        return PromptUpdateDetails(
-            full=image_tokens + [bos_token_id],
-            features=image_tokens,
+        return PromptUpdateDetails.select_token_id(
+            image_tokens + [bos_token_id],
+            embed_token_id=_IMAGE_TOKEN_ID,
         )
 
     return [

docs/source/models/supported_models.md

@@ -1006,9 +1006,6 @@ See [this page](#generative-models) for more information on how to use generativ
 <sup>+</sup> Multiple items can be inputted per text prompt for this modality.
 
 :::{important}
-To use Gemma3 series models, you have to install Hugging Face Transformers library from source via
-`pip install git+https://github.com/huggingface/transformers`.
-
 Pan-and-scan image pre-processing is currently supported on V0 (but not V1).
 You can enable it by passing `--mm-processor-kwargs '{"do_pan_and_scan": True}'`.
 :::

tests/models/decoder_only/vision_language/test_models.py

@@ -330,9 +330,8 @@
         max_num_seqs=4,
         dtype="bfloat16",
         auto_cls=AutoModelForImageTextToText,
-        tensor_parallel_size=8,
-        vllm_runner_kwargs={"gpu_memory_utilization": 0.8},
-        marks=multi_gpu_marks(num_gpus=8),
+        tensor_parallel_size=4,
+        marks=multi_gpu_marks(num_gpus=4),
     ),
     "llava_next": VLMTestInfo(
         models=["llava-hf/llava-v1.6-mistral-7b-hf"],

tests/models/decoder_only/vision_language/test_pixtral.py

@@ -200,22 +200,14 @@ def test_chat(
 
 
 @large_gpu_test(min_gb=48)
-@pytest.mark.parametrize(
-    "prompt,expected_ranges",
-    [(_create_engine_inputs_hf(IMG_URLS[:1]), [{
-        "offset": 11,
-        "length": 494
-    }]),
-     (_create_engine_inputs_hf(IMG_URLS[1:4]), [{
-         "offset": 11,
-         "length": 266
-     }, {
-         "offset": 277,
-         "length": 1056
-     }, {
-         "offset": 1333,
-         "length": 418
-     }])])
+@pytest.mark.parametrize("prompt,expected_ranges",
+                         [(_create_engine_inputs_hf(IMG_URLS[:1]),
+                           [PlaceholderRange(offset=11, length=494)]),
+                          (_create_engine_inputs_hf(IMG_URLS[1:4]), [
+                              PlaceholderRange(offset=11, length=266),
+                              PlaceholderRange(offset=277, length=1056),
+                              PlaceholderRange(offset=1333, length=418)
+                          ])])
 def test_multi_modal_placeholders(vllm_runner, prompt,
                                   expected_ranges: list[PlaceholderRange],
                                   monkeypatch) -> None:

tests/models/multimodal/processing/test_llama4.py

@@ -71,13 +71,11 @@ def test_processor_override(
     # image token offsets
     img_locs = processed_inputs["mm_placeholders"].get("image", [])
     assert len(img_locs) == num_imgs
-    assert [img_loc["offset"] for img_loc in img_locs] == \
+    assert [img_loc.offset for img_loc in img_locs] == \
         [i for i, v in enumerate(prompt_token_ids) \
         if v == config.boi_token_index]
 
     # patch sizes and masks
-    assert prompt_token_ids.count(config.image_token_index) \
-        == sum(img_patch.sum() for img_patch in mm_kwargs["embed_is_patch"])
     patch_token_id = vocab[hf_processor.img_patch_token]
     num_patches = processed_inputs["prompt_token_ids"].count(patch_token_id)
     mm_counts = {"image": num_imgs}
@@ -89,11 +87,3 @@ def test_processor_override(
         == mm_kwargs["patches_per_image"].sum() * num_patches_per_chunk
     assert mm_kwargs["pixel_values"].shape[0] \
         == mm_kwargs["patches_per_image"].sum()
-
-    for embed_is_patch, aspect_ratio in zip(mm_kwargs["embed_is_patch"],
-                                            mm_kwargs["aspect_ratios"]):
-        assert embed_is_patch.shape[0] == \
-            len(tokenizer.encode(
-                hf_processor._prompt_split_image(
-                    aspect_ratio, num_patches_per_chunk),
-                add_special_tokens=False))

tests/models/multimodal/processing/test_llava_next.py

@@ -92,8 +92,8 @@ def _validate_image_prompt_replacements_one(
         first_placeholder = image_placeholders[0]
 
         # NOTE: There is a BOS token
-        assert first_placeholder["offset"] == 1
-        assert first_placeholder["length"] == (
+        assert first_placeholder.offset == 1
+        assert first_placeholder.length == (
             len(processed_inputs["prompt_token_ids"]) - 1) // num_imgs
 
     except Exception as exc:

tests/models/multimodal/processing/test_llava_onevision.py

@@ -92,8 +92,8 @@ def _validate_image_prompt_replacements_one(
 
         first_placeholder = image_placeholders[0]
 
-        assert first_placeholder["offset"] == 0
-        assert first_placeholder["length"] == len(
+        assert first_placeholder.offset == 0
+        assert first_placeholder.length == len(
             processed_inputs["prompt_token_ids"]) // num_imgs
     except Exception as exc:
         failed_size_excs.append((image_size, exc))

tests/multimodal/test_processing.py

@@ -785,6 +785,7 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=6,
                         tokens=[32000, 32000],
+                        is_embed=None,
                     ),
                 ],
                 "pattern_4": [
@@ -793,6 +794,7 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=3,
                         tokens=[32000],
+                        is_embed=None,
                     ),
                 ],
             }
@@ -807,12 +809,14 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=1,
                         tokens=[32000, 32000],
+                        is_embed=None,
                     ),
                     PlaceholderFeaturesInfo(
                         modality="pattern_1",
                         item_idx=1,
                         start_idx=5,
                         tokens=[32000, 32000],
+                        is_embed=None,
                     ),
                 ],
                 "pattern_3": [
@@ -821,6 +825,7 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=7,
                         tokens=[1550, 918, 1550],
+                        is_embed=None,
                     ),
                 ],
                 # No match for pattern_4 as it has lower priority than pattern_1
@@ -835,12 +840,14 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=1,
                         tokens=[32000, 32000],
+                        is_embed=None,
                     ),
                     PlaceholderFeaturesInfo(
                         modality="pattern_1",
                         item_idx=1,
                         start_idx=3,
                         tokens=[32000, 32000],
+                        is_embed=None,
                     ),
                 ],
                 "pattern_4": [
@@ -849,6 +856,7 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=5,
                         tokens=[32000],
+                        is_embed=None,
                     ),
                 ],
                 "pattern_3": [
@@ -857,6 +865,7 @@ def test_find_update_tokens(
                         item_idx=0,
                         start_idx=6,
                         tokens=[1550, 918, 1550],
+                        is_embed=None,
                     ),
                 ],
             }

tests/v1/core/test_kv_cache_utils.py

@@ -3,7 +3,7 @@
 import pytest
 import torch
 
-from vllm.multimodal.inputs import MultiModalKwargs
+from vllm.multimodal.inputs import MultiModalKwargs, PlaceholderRange
 from vllm.sampling_params import SamplingParams
 from vllm.utils import sha256
 # disable yapf here as it formats differently than isort such that both fail
@@ -158,13 +158,10 @@ def test_generate_block_hash_extra_keys():
     request = make_request(
         request_id=0,
         prompt_token_ids=[_ for _ in range(20)],
-        mm_positions=[{
-            "offset": 0,
-            "length": 5
-        }, {
-            "offset": 10,
-            "length": 5
-        }],
+        mm_positions=[
+            PlaceholderRange(offset=0, length=5),
+            PlaceholderRange(offset=10, length=5),
+        ],
         mm_hashes=["hash1", "hash2"],
     )
 
@@ -222,13 +219,10 @@ def test_hash_request_tokens(hash_fn):
     request = make_request(
         request_id=0,
         prompt_token_ids=[_ for _ in range(6)],
-        mm_positions=[{
-            "offset": 0,
-            "length": 3
-        }, {
-            "offset": 3,
-            "length": 3
-        }],
+        mm_positions=[
+            PlaceholderRange(offset=0, length=3),
+            PlaceholderRange(offset=3, length=3),
+        ],
         mm_hashes=["hash1", "hash2"],
     )
 
@@ -253,25 +247,19 @@ def test_hash_tokens_different_mm_input(hash_fn):
     request1 = make_request(
         request_id=0,
         prompt_token_ids=[_ for _ in range(6)],
-        mm_positions=[{
-            "offset": 0,
-            "length": 3
-        }, {
-            "offset": 3,
-            "length": 3
-        }],
+        mm_positions=[
+            PlaceholderRange(offset=0, length=3),
+            PlaceholderRange(offset=3, length=3),
+        ],
         mm_hashes=["hash1", "hash2"],
     )
     request2 = make_request(
         request_id=1,
         prompt_token_ids=[_ for _ in range(6)],
-        mm_positions=[{
-            "offset": 0,
-            "length": 3
-        }, {
-            "offset": 3,
-            "length": 3
-        }],
+        mm_positions=[
+            PlaceholderRange(offset=0, length=3),
+            PlaceholderRange(offset=3, length=3),
+        ],
         mm_hashes=["hash3", "hash2"],
     )
     block_size = 3