model: support mllama4 (#5144)

python/sglang/srt/configs/model_config.py

@@ -486,8 +486,8 @@ multimodal_model_archs = [
     "Gemma3ForConditionalGeneration",
     "Grok1VForCausalLM",
     "Grok1AForCausalLM",
-    # TODO: add multimodal support for "Llama4ForConditionalGeneration",
     "LlavaLlamaForCausalLM",
+    "Llama4ForConditionalGeneration",
     "LlavaMistralForCausalLM",
     "LlavaQwenForCausalLM",
     "LlavaVidForCausalLM",

python/sglang/srt/managers/mm_utils.py

@@ -148,7 +148,8 @@ def get_embedding_and_mask(
         placeholder_tensor,
     ).unsqueeze(-1)
 
-    num_mm_tokens_in_input_ids = special_multimodal_mask.sum()
+    num_mm_tokens_in_input_ids = special_multimodal_mask.sum().item()
+
     if num_mm_tokens_in_input_ids != num_mm_tokens_in_embedding:
         logger.warning(
             f"Number of tokens in multimodal embedding does not match those in the input text."
@@ -172,7 +173,7 @@ def get_embedding_and_mask(
                 embedding = embedding[-num_multimodal:, :]
         else:
             raise RuntimeError(
-                "Insufficient multimodal embedding length. This is an internal error"
+                f"Insufficient multimodal embedding length: {num_mm_tokens_in_input_ids=} vs {num_mm_tokens_in_embedding=}. This is an internal error"
             )
 
     return embedding, special_multimodal_mask

python/sglang/srt/managers/multimodal_processors/mllama4.py

@@ -1,10 +1,8 @@
-from typing import List, Mapping, Optional, Tuple, Union
+from typing import List, Union
 
 import torch
-from PIL import Image
-from transformers import Llama4Processor
 from transformers.image_utils import SizeDict
-from transformers.models.llama4.image_processing_llama4 import (
+from transformers.models.llama4.image_processing_llama4_fast import (
     find_supported_resolutions,
     get_best_fit,
 )
@@ -15,7 +13,6 @@ from sglang.srt.managers.multimodal_processors.base_processor import (
 )
 from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
 from sglang.srt.models.mllama4 import Llama4ForConditionalGeneration
-from sglang.srt.utils import load_image
 
 
 class Mllama4ImageProcessor(BaseMultimodalProcessor):
@@ -25,6 +22,9 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
         super().__init__(hf_config, server_args, _processor)
         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(
             image_token=_processor.image_token
         )
@@ -54,19 +54,16 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
         )
 
         # Process the images using the processor
-        processor = Llama4Processor.from_pretrained(
+        processor = self._processor
-            self.server_args.model_path, **kwargs
-        )
 
         # Process the prompt and images
-        image_inputs = processor(
+        processor_output = self.process_mm_data(
-            text=processed_data.input_text,
+            input_text=processed_data.input_text,
             images=processed_data.images,
-            return_tensors="pt",
         )
 
         # Handle image resolutions and aspect ratios
-        if "pixel_values" in image_inputs:
+        if "pixel_values" in processor_output:
             image_processor = processor.image_processor
             tokenizer = self._processor.tokenizer
 
@@ -100,8 +97,8 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
             ]
 
             # Add to image_inputs
-            image_inputs["aspect_ratios"] = aspect_ratios
+            processor_output["aspect_ratios"] = aspect_ratios
-            image_inputs["patches_per_image"] = torch.tensor(patches_per_image)
+            processor_output["patches_per_image"] = torch.tensor(patches_per_image)
 
             # Process embed_is_patch
             vocab = tokenizer.get_vocab()
@@ -109,7 +106,7 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
             image_end_id = vocab.get(processor.end_of_img_token, -1)
 
             if patch_id != -1 and image_end_id != -1:
-                input_ids = image_inputs["input_ids"].view(-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:
@@ -129,33 +126,21 @@ class Mllama4ImageProcessor(BaseMultimodalProcessor):
                     for per_image_input_ids in split_input_ids:
                         embed_is_patch.append(per_image_input_ids == patch_id)
 
-                    image_inputs["embed_is_patch"] = embed_is_patch
+                    processor_output["embed_is_patch"] = embed_is_patch
 
         # Convert to the format expected by SGLang
-        image_inputs["input_ids"] = image_inputs["input_ids"].tolist()[0]
+        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
 
         # Add metadata for image processing
-        image_inputs["mm_items"] = [
+        processor_output["mm_items"] = [
             MultimodalDataItem(
-                pixel_values=image_inputs["pixel_values"],
+                pixel_values=processor_output["pixel_values"],
                 modality=Modality.IMAGE,
-                # Add additional metadata needed for Llama4 vision processing
-                embed_is_patch=image_inputs.get("embed_is_patch", None),
-                aspect_ratios=image_inputs.get("aspect_ratios", None),
-                patches_per_image=image_inputs.get("patches_per_image", None),
             )
         ]
 
-        return image_inputs
+        return processor_output
-
-    def get_patch_per_chunk(self):
-        """Calculate patches per chunk based on vision config"""
-        image_size = self.vision_config.image_size
-        patch_size = self.vision_config.patch_size
-
-        assert (
-            image_size % patch_size == 0
-        ), f"chunk size {image_size} should be multiple of patch_size {patch_size}"
-
-        ds_ratio = int(round(1.0 / (self.vision_config.pixel_shuffle_ratio**2)))
-        return (image_size // patch_size) ** 2 // ds_ratio

python/sglang/srt/managers/schedule_batch.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import hashlib
 from enum import Enum, auto
 
 # Copyright 2023-2024 SGLang Team
@@ -157,7 +158,7 @@ class Modality(Enum):
 @dataclasses.dataclass
 class MultimodalDataItem:
     """
-    A single multimodal data, from a single image/video/audio or other
+    A single multimodal data, from a single image/video/audio or others
     """
 
     modality: Modality
@@ -195,25 +196,54 @@ class MultimodalDataItem:
 
     def set_pad_value(self):
         """
-        Set the pad value after first hashign the data
+        Set the pad value after first hashing the data
         """
 
-        def tensor_hash(f):
+        def data_hash(data) -> int:
-            f_list = flatten_nested_list(f)
+            hash_bytes = hashlib.sha256(data).digest()[:8]
-            f_list = [x.flatten() if isinstance(x, torch.Tensor) else x for x in f_list]
+            return int.from_bytes(hash_bytes, byteorder="big", signed=False)
-            f_cat = torch.concat(f_list).contiguous().numpy().tobytes()
+
-            return hash(f_cat)
+        def tensor_hash(tensor_list) -> int:
+            """
+            hash a tensor or a tensor list
+            """
+            tensor = tensor_list
+            if isinstance(tensor_list, list):
+                tensor_list = flatten_nested_list(tensor_list)
+                tensor_list = [
+                    x.flatten() if isinstance(x, torch.Tensor) else x
+                    for x in tensor_list
+                ]
+                tensor = torch.concat(tensor_list)
+
+            tensor = tensor.detach().contiguous()
+
+            if tensor.dtype == torch.bfloat16:
+                # memoryview() doesn't support PyTorch's BFloat16 dtype
+                tensor = tensor.float()
+
+            if tensor.is_cuda:
+                tensor_cpu = torch.frombuffer(
+                    tensor.storage().untyped(), dtype=tensor.dtype, count=tensor.numel()
+                ).clone()
+            else:
+                tensor_cpu = tensor
+
+            mv = memoryview(tensor_cpu.numpy())
+            return data_hash(mv.tobytes())
 
         def hash_feature(f):
             if isinstance(f, list):
                 if isinstance(f[0], torch.Tensor):
                     return tensor_hash(f)
-                return hash(tuple(flatten_nested_list(f)))
+                return data_hash(tuple(flatten_nested_list(f)))
             elif isinstance(f, np.ndarray):
                 arr = np.ascontiguousarray(f)
                 arr_bytes = arr.tobytes()
-                return hash(arr_bytes)
+                return data_hash(arr_bytes)
-            return hash(f)
+            elif isinstance(f, torch.Tensor):
+                return tensor_hash([f])
+            return data_hash(f)
 
         if self.is_audio():
             self.hash = hash_feature(self.audio_features)
@@ -256,7 +286,7 @@ class MultimodalInputs:
     mrope_position_delta: Optional[torch.Tensor] = None
 
     # image
-    im_token_id: Optional[torch.Tensor] = None
+    im_token_id: Optional[int] = None
     im_start_id: Optional[int] = None
     im_end_id: Optional[int] = None
     slice_start_id: Optional[int] = None
@@ -330,10 +360,8 @@ class MultimodalInputs:
 
         # args needed to be merged
         optional_args = [
-            "items",
+            "mm_items",
-            "image_offsets",
             "image_pad_len",
-            # "modalities", # modalities should be ["multi-images"] (one entry) even for multiple images
         ]
         for arg in optional_args:
             self_arg = getattr(self, arg, None)

python/sglang/srt/models/llama4.py

@@ -466,6 +466,9 @@ class Llama4ForCausalLM(LlamaForCausalLM):
     ):
         super().__init__(config, quant_config, prefix)
 
+    def get_input_embeddings(self):
+        return self.model.embed_tokens
+
     def _init_model(
         self,
         config: Llama4TextConfig,

python/sglang/srt/models/mllama4.py

@@ -1,14 +1,19 @@
-# TODO: add Aapted from vllm/mllama4.py
 from collections.abc import Iterable
-from typing import Optional, Set, Tuple
+from typing import List, Optional, Set, Tuple
 
 import torch
 from torch import nn
-from transformers import Llama4Config
+from transformers import Llama4Config, Llama4VisionModel
+from transformers.models.llama4.modeling_llama4 import Llama4MultiModalProjector
 
 from sglang.srt.layers.logits_processor import LogitsProcessor
 from sglang.srt.layers.moe.fused_moe_triton import FusedMoE
 from sglang.srt.layers.quantization import QuantizationConfig
+from sglang.srt.managers.mm_utils import (
+    MultiModalityDataPaddingPatternImageTokens,
+    general_mm_embed_routine,
+)
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
@@ -30,6 +35,9 @@ class Llama4ForConditionalGeneration(nn.Module):
         self.config = config
         self.quant_config = quant_config
 
+        self.vision_model = Llama4VisionModel(config.vision_config)
+        self.multi_modal_projector = Llama4MultiModalProjector(config)
+
         # Initialize the language model
         from sglang.srt.models.llama4 import Llama4ForCausalLM
 
@@ -41,6 +49,29 @@ class Llama4ForConditionalGeneration(nn.Module):
 
         self.logits_processor = LogitsProcessor(config.text_config)
 
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        # Get all special token IDs
+        im_token_id: int = mm_inputs.im_token_id
+
+        pattern = MultiModalityDataPaddingPatternImageTokens(torch.tensor(im_token_id))
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
+
+    def get_image_feature(
+        self,
+        items: List[MultimodalDataItem],
+    ) -> torch.Tensor:
+        pixel_values = (
+            torch.concat([item.pixel_values for item in items])
+            .to(next(self.vision_model.parameters()).device)
+            .type(next(self.vision_model.parameters()).dtype)
+        )
+
+        image_outputs = self.vision_model(pixel_values, output_hidden_states=False)
+        image_features = image_outputs.last_hidden_state
+        vision_flat = image_features.view(-1, image_features.size(-1))
+        projected_vision_flat = self.multi_modal_projector(vision_flat)
+        return projected_vision_flat
+
     def forward(
         self,
         input_ids: torch.Tensor,
@@ -49,7 +80,15 @@ class Llama4ForConditionalGeneration(nn.Module):
         **kwargs: object,
     ) -> torch.Tensor:
 
-        return self.language_model(input_ids, positions, forward_batch)
+        hs = general_mm_embed_routine(
+            input_ids=input_ids,
+            forward_batch=forward_batch,
+            language_model=self.language_model,
+            image_data_embedding_func=self.get_image_feature,
+            positions=positions,
+        )
+
+        return hs
 
     def permute_qk_weight_for_rotary(
         self,
@@ -108,17 +147,17 @@ class Llama4ForConditionalGeneration(nn.Module):
         )
 
         for name, loaded_weight in weights:
-
+            if not "vision" in name:
-            if name.startswith("vision_model") or name.startswith(
+                name, loaded_weight = self.permute_qk_weight_for_rotary(
-                "multi_modal_projector"
+                    name, loaded_weight
-            ):
+                )
-                continue
-
-            name, loaded_weight = self.permute_qk_weight_for_rotary(name, loaded_weight)
 
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 if weight_name not in name:
                     continue
+
+                if "vision" in name:
+                    continue
                 name = name.replace(weight_name, param_name)
                 param = params_dict[name]
                 weight_loader = param.weight_loader

test/srt/test_vision_openai_server.py

@@ -307,7 +307,6 @@ class TestOpenAIVisionServer(CustomTestCase):
         self.assertGreater(len(video_response), 0)
 
     def test_regex(self):
-        return
         client = openai.Client(api_key=self.api_key, base_url=self.base_url)
 
         regex = (
@@ -683,6 +682,31 @@ class TestJanusProServer(TestOpenAIVisionServer):
         pass
 
 
+class TestLlama4Server(TestOpenAIVisionServer):
+    @classmethod
+    def setUpClass(cls):
+        cls.model = "meta-llama/Llama-4-Scout-17B-16E-Instruct"
+        cls.base_url = DEFAULT_URL_FOR_TEST
+        cls.api_key = "sk-123456"
+        cls.process = popen_launch_server(
+            cls.model,
+            cls.base_url,
+            timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
+            other_args=[
+                "--chat-template",
+                "llama-4",
+                "--mem-fraction-static",
+                "0.8",
+                "--tp-size=8",
+                "--context-length=8192",
+            ],
+        )
+        cls.base_url += "/v1"
+
+    def test_video_chat_completion(self):
+        pass
+
+
 class TestGemma3itServer(TestOpenAIVisionServer):
     @classmethod
     def setUpClass(cls):