Support precomputed_embeddings for Llama 4 (#8156)

Signed-off-by: Xinyuan Tong <xinyuantong.cs@gmail.com>
Co-authored-by: Xiang (Kevin) Li <lik@nvidia.com>
Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

.pre-commit-config.yaml

@@ -39,7 +39,11 @@ repos:
       - id: codespell
         additional_dependencies: ['tomli']
         args: ['--toml', 'python/pyproject.toml', '-L', 'cann']
-        exclude: test/srt/test_reasoning_parser.py # Exclude the test file that is expected to fail
+        exclude: |
+          (?x)^(
+            test/srt/test_reasoning_parser\.py|
+            docs/backend/vlm_query\.ipynb
+          )$
   - repo: https://github.com/pre-commit/mirrors-clang-format
     rev: v18.1.8
     hooks:

docs/index.rst

@@ -62,6 +62,7 @@ The core features include:
    backend/quantization.md
    backend/lora.ipynb
    backend/pd_disaggregation.md
+   backend/vlm_query.ipynb
 
 .. toctree::
    :maxdepth: 1

python/sglang/srt/layers/multimodal.py

@@ -55,6 +55,9 @@ def gpu_tensor_hash(tensor: torch.Tensor) -> int:
 
     intermediate_hashes = torch.empty(n, dtype=torch.int64, device=tensor.device)
 
+    # Set cuda device to prevent ValueError: Pointer argument (at 0) cannot be accessed from Triton (cpu tensor?)
+    # Solution from Tri: https://github.com/Dao-AILab/flash-attention/issues/523#issuecomment-1707611579
+    with torch.cuda.device(tensor.device):
         hash_kernel[grid](
             tensor,
             intermediate_hashes,

python/sglang/srt/multimodal/processors/mllama4.py

@@ -22,12 +22,12 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
         super().__init__(hf_config, server_args, _processor, *args, **kwargs)
         self.vision_config = hf_config.vision_config
         self.text_config = hf_config.text_config
-        self.boi_token_index = hf_config.boi_token_index
-        self.eoi_token_index = hf_config.eoi_token_index
-        self.image_token_index = hf_config.image_token_index
-        self.multimodal_tokens = MultimodalSpecialTokens(
+        self.IM_START_TOKEN_ID = hf_config.boi_token_index
+        self.IM_END_TOKEN_ID = hf_config.eoi_token_index
+        self.IM_TOKEN_ID = hf_config.image_token_index
+        self.mm_tokens = MultimodalSpecialTokens(
             image_token=_processor.image_token,
-            image_token_id=self.image_token_index,
+            image_token_id=self.IM_TOKEN_ID,
         ).build(_processor)
 
     async def process_mm_data_async(
@@ -37,114 +37,21 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
         *args,
         **kwargs,
     ):
-        if isinstance(input_text, list):
-            assert len(input_text) and isinstance(input_text[0], int)
-            input_text = self._processor.tokenizer.decode(input_text)
-
-        # Process images and text using the base processor's load_mm_data method
-        processed_data = self.load_mm_data(
+        base_output = self.load_mm_data(
             prompt=input_text,
-            multimodal_tokens=self.multimodal_tokens,
             image_data=image_data,
-            return_text=True,
+            multimodal_tokens=self.mm_tokens,
         )
 
-        # Process the images using the processor
-        processor = self._processor
-
         # Process the prompt and images
-        processor_output = self.process_mm_data(
-            input_text=processed_data.input_text,
-            images=processed_data.images,
-        )
-
-        # Handle image resolutions and aspect ratios
-        if "pixel_values" not in processor_output:  # no image processed
-            return None
-
-        image_processor = processor.image_processor
-        tokenizer = self._processor.tokenizer
-
-        # Calculate tile size and find supported resolutions
-        tile_size = self.vision_config.image_size
-        max_num_tiles = getattr(self.vision_config, "max_patches", 1)
-
-        possible_resolutions = find_supported_resolutions(
-            max_num_chunks=max_num_tiles,
-            patch_size=SizeDict(height=tile_size, width=tile_size),
-        )
-
-        # Find best fit for each image
-        best_fit_sizes = [
-            get_best_fit(
-                (image.size[1], image.size[0]),  # (height, width)
-                torch.tensor(possible_resolutions),
-                resize_to_max_canvas=image_processor.resize_to_max_canvas,
-            )
-            for image in processed_data.images
-        ]
-
-        # Calculate aspect ratios and patches per image
-        aspect_ratios = [
-            (image_size[0] // tile_size, image_size[1] // tile_size)
-            for image_size in best_fit_sizes
-        ]
-
-        patches_per_image = [
-            1 if r_h * r_w == 1 else 1 + r_h * r_w for (r_h, r_w) in aspect_ratios
-        ]
-
-        # Add to image_inputs
-        processor_output["aspect_ratios"] = aspect_ratios
-        processor_output["patches_per_image"] = torch.tensor(patches_per_image)
-
-        # Process embed_is_patch
-        vocab = tokenizer.get_vocab()
-        patch_id = vocab.get(processor.img_patch_token, -1)
-        image_end_id = vocab.get(processor.end_of_img_token, -1)
-
-        if patch_id != -1 and image_end_id != -1:
-            input_ids = processor_output["input_ids"].view(-1)
-
-            # Remove BOS token if present
-            if input_ids.size(0) > 0 and input_ids[0] == tokenizer.bos_token_id:
-                input_ids = input_ids[1:]
-
-            # Find image end indices and split input_ids
-            image_end_indices = (input_ids == image_end_id).nonzero().view(-1)
-
-            if image_end_indices.size(0) > 0:
-                # Split at image boundaries
-                split_indices = (image_end_indices + 1)[:-1]
-                split_input_ids = torch.tensor_split(input_ids, split_indices)
-                split_input_ids = [x for x in split_input_ids if x.numel() > 0]
-
-                # Create embed_is_patch for each image
-                embed_is_patch = []
-                for per_image_input_ids in split_input_ids:
-                    embed_is_patch.append(per_image_input_ids == patch_id)
-
-                processor_output["embed_is_patch"] = embed_is_patch
-
-        # Convert to the format expected by SGLang
-        processor_output["input_ids"] = processor_output["input_ids"].tolist()[0]
-
-        processor_output["im_start_id"] = self.boi_token_index
-        processor_output["im_end_id"] = self.eoi_token_index
-        processor_output["im_token_id"] = self.image_token_index
-
-        image_offsets = self.get_mm_items_offset(
-            input_ids=torch.tensor(processor_output["input_ids"]),
-            mm_token_id=self.image_token_index,
-        )
-
-        # Add metadata for image processing
-        processor_output["mm_items"] = [
-            MultimodalDataItem(
-                feature=processor_output["pixel_values"],
-                modality=Modality.IMAGE,
-                offsets=image_offsets,
+        mm_items, input_ids, _ = self.process_and_combine_mm_data(
+            base_output, self.mm_tokens
         )
-        ]
 
-        return processor_output
+        return {
+            "input_ids": input_ids.tolist(),
+            "mm_items": mm_items,
+            "im_start_id": self.IM_START_TOKEN_ID,
+            "im_end_id": self.IM_END_TOKEN_ID,
+            "im_token_id": self.IM_TOKEN_ID,
+        }

test/srt/test_vlm_input_format.py

@@ -216,5 +216,43 @@ class TestKimiVLImageUnderstandsImage(
         )
 
 
+# not for CI: too large
+# class TestLlama4ImageUnderstandsImage(
+#     VLMInputTestBase, unittest.IsolatedAsyncioTestCase
+# ):
+#     model_path = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+#     chat_template = "llama_4_vision"
+
+#     def setUp(self):
+#         self.engine = Engine(
+#             model_path=self.model_path,
+#             trust_remote_code=True,
+#             chat_template=self.chat_template,
+#             enable_multimodal=True,
+#             mem_fraction_static=0.8,
+#             tp_size=4,
+#             attention_backend="fa3",
+#             context_length=65536,
+#         )
+
+#     @classmethod
+#     def _init_visual(cls):
+#         model = AutoModel.from_pretrained(cls.model_path, trust_remote_code=True, torch_dtype="auto")
+#         cls.vision_tower = model.vision_model.eval().to(cls.device)
+#         cls.mm_projector = model.multi_modal_projector.eval().to(cls.device)
+
+#         cls.visual = lambda tokenizer_output: cls.mm_projector(
+#             cls.vision_tower(
+#                 pixel_values=tokenizer_output["pixel_values"],
+#             ).last_hidden_state.flatten(0, -2)
+#         )
+
+#     def _pixel_values_image_data(self, processor_output):
+#         return dict(
+#             modality="IMAGE",
+#             pixel_values=processor_output["pixel_values"],
+#         )
+
+
 if __name__ == "__main__":
     unittest.main()