sglang 0.4.4.post3__py3-none-any.whl → 0.4.4.post4__py3-none-any.whl

sglang/srt/models/minicpmo.py

@@ -40,16 +40,19 @@ from transformers.models.whisper.modeling_whisper import (
 from sglang.srt.layers.quantization import QuantizationConfig
 from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternTokenPairs,
-    embed_mm_inputs,
-    get_multimodal_data_bounds,
+    general_mm_embed_routine,
 )
-from sglang.srt.managers.schedule_batch import MultimodalInputs
-from sglang.srt.model_executor.forward_batch_info import ForwardBatch, ForwardMode
+from sglang.srt.managers.schedule_batch import (
+    MultimodalDataItem,
+    MultimodalInputs,
+    flatten_nested_list,
+)
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.utils import set_default_torch_dtype
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.minicpmv import (
     Idefics2VisionTransformer,
-    MiniCPMVBaseModel,
+    MiniCPMBaseModel,
     Resampler2_5,
 )
 from sglang.srt.models.qwen2 import Qwen2ForCausalLM
@@ -1409,7 +1412,7 @@ class MultiModalProjector(nn.Module):
         return hidden_states
 
 
-class MiniCPMO(MiniCPMVBaseModel):
+class MiniCPMO(MiniCPMBaseModel):
     def __init__(
         self,
         config: PretrainedConfig,
@@ -1537,7 +1540,7 @@ class MiniCPMO(MiniCPMVBaseModel):
 
         return input_lengths_after_cnn, input_lengths_after_pooling
 
-    def get_audio_embedding_streaming(self, multimodal_input: MultimodalInputs):
+    def get_audio_embedding_streaming(self, items: List[MultimodalDataItem]):
         r"""
         Extract audio embeddings in a streaming manner using cached key-value pairs.
 
@@ -1545,26 +1548,15 @@ class MiniCPMO(MiniCPMVBaseModel):
         for faster inference on subsequent audio frames. It only supports batch_size=1 and is intended
         for streaming scenarios.
 
-        Args:
-            multimodal_input (dict):
-                - **"audio_features"** (`torch.FloatTensor`): Input mel-spectrograms of shape `(batch_size, 80, frames)`.
-                - **"audio_feature_lens"** (List[List[int]]): Lengths of each audio segment for each item in the batch.
-
         Returns:
             List[List[torch.Tensor]]: audio embeddings
         """
-        # print("audio embedding")
-
-        wavforms = (
-            []
-            if multimodal_input.audio_features is None
-            else multimodal_input.audio_features
+        wavforms = flatten_nested_list(
+            [item.audio_features for item in items if item.audio_features]
         )
         # list, [[x1, x2], [y1], [z1]]
-        audio_feature_lens_raw = (
-            []
-            if multimodal_input.audio_feature_lens is None
-            else multimodal_input.audio_feature_lens
+        audio_feature_lens_raw = flatten_nested_list(
+            [item.audio_feature_lens for item in items if item.audio_feature_lens]
         )
 
         # exist audio
@@ -1650,7 +1642,7 @@ class MiniCPMO(MiniCPMVBaseModel):
             ret[i, start:ending] = True
         return ret
 
-    def get_audio_embedding(self, multimodal_input: MultimodalInputs, chunk_length=-1):
+    def get_audio_embedding(self, items: List[MultimodalDataItem], chunk_length=-1):
         r"""
         Extract full audio embeddings with optional chunk-based attention.
 
@@ -1659,31 +1651,25 @@ class MiniCPMO(MiniCPMVBaseModel):
         not use key-value caching and is suitable for non-streaming inference.
 
         Args:
-            multimodal_input (dict):
-                - **"audio_features"** (`torch.FloatTensor`): Input mel-spectrograms of shape `(batch_size, 80, frames)`.
-                - **"audio_feature_lens"** (List[List[int]]): Lengths of each audio segment for each item in the batch.
             chunk_length (int, optional): Determines whether to use full attention (-1) or chunk-based
                 attention (>0) during embedding computation.
 
         Returns:
             List[List[torch.Tensor]]: audio embeddings
         """
-        # print("audio embedding")
         # (bs, 80, frames) or [], multi audios need filled in advance
-        wavforms = (
-            []
-            if multimodal_input.audio_features is None
-            else multimodal_input.audio_features
+        wavforms = flatten_nested_list(
+            [item.audio_features for item in items if item.audio_features]
         )
         # list, [[x1, x2], [y1], [z1]]
-        audio_feature_lens_raw = (
-            []
-            if multimodal_input.audio_feature_lens is None
-            else multimodal_input.audio_feature_lens
+        audio_feature_lens_raw = flatten_nested_list(
+            [item.audio_feature_lens for item in items if item.audio_feature_lens]
         )
 
         final_audio_embeds = []
 
+        assert isinstance(wavforms, list)
+        assert isinstance(wavforms[0], torch.Tensor)
         # exist audio
         for wavform in wavforms:
             if len(wavform) > 0:
@@ -1757,86 +1743,46 @@ class MiniCPMO(MiniCPMVBaseModel):
                     final_audio_embeds.append(target_audio_embeds)
             return final_audio_embeds
 
+    def get_audio_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        embedding = self.get_omni_embedding(
+            items=items,
+            chunk_length=self.config.audio_chunk_length,
+            stream_input=False,
+        )
+        return embedding
+
     def get_omni_embedding(
         self,
-        input_ids,
-        multimodal_input: MultimodalInputs,
-        input_embeds: torch.Tensor,
-        forward_mode: ForwardMode,
+        items: List[MultimodalDataItem],
         chunk_length=-1,
         stream_input=False,
     ):
         """
         Args:
-            multimodal_input:
-            input_embeds:
             chunk_length: whisper use full attention or chunk attention
             stream_input: use streaming audio embedding
         Returns:
             final embeddings with audio feature
         """
-        input_embeds = input_embeds.unsqueeze(0)
-        if not forward_mode.is_decode() and multimodal_input.contains_audio_inputs():
-            audio_bounds = get_multimodal_data_bounds(
-                input_ids=input_ids,
-                pad_values=multimodal_input.pad_values,
-                token_pairs=[
-                    (multimodal_input.audio_start_id, multimodal_input.audio_end_id)
-                ],
-            )
-            if audio_bounds.numel() == 0:
-                input_embeds = input_embeds.squeeze(0)
-                # TODO
-                logger.warn("Unimplemented logic. Please try disabling chunked prefill")
-                return input_embeds
-            audio_bounds = audio_bounds.unsqueeze(0)
-            bs = len(input_embeds)
-
-            if stream_input:
-                audio_embeddings = self.get_audio_embedding_streaming(multimodal_input)
-            else:
-                audio_embeddings = self.get_audio_embedding(
-                    multimodal_input, chunk_length
-                )
-            # batch size
-            assert len(audio_embeddings) == len(input_embeds)
-            if len(audio_embeddings) > 0:
-                if self.config.chunk_input:
-                    for i in range(bs):
-                        audio_embs = torch.cat(audio_embeddings[i], dim=0).to(
-                            device=input_embeds.device, dtype=input_embeds.dtype
-                        )
-                        audio_start_pos = 0
-                        for bound in audio_bounds[i]:
-                            audio_len = bound[1] - bound[0] + 1
-                            input_embeds[0, bound[0] : bound[1] + 1] = audio_embs[
-                                audio_start_pos : audio_start_pos + audio_len, :
-                            ]
-                            audio_start_pos += audio_len
-                else:
-                    for i in range(bs):
-                        audio_embs = audio_embeddings[i]
-                        bounds = audio_bounds[i]
-                        for embs, bound in zip(audio_embs, bounds):
-                            audio_indices = torch.arange(
-                                bound[0], bound[1], dtype=torch.long
-                            ).to(input_embeds.device)
-
-                            if embs.shape[0] != len(audio_indices):
-                                raise ValueError(
-                                    f"Shape mismatch: Trying to assign embeddings of shape {embs.shape} "
-                                    f"to input indices of length {len(audio_indices)}"
-                                )
-                            input_embeds[i, audio_indices] = embs.to(input_embeds.dtype)
-        input_embeds = input_embeds.squeeze(0)
-        return input_embeds
-
-    def get_image_features(
-        self,
-        image_inputs: MultimodalInputs,
-    ) -> torch.Tensor:
-        pixel_values = image_inputs.pixel_values
-        tgt_sizes = image_inputs.tgt_sizes
+
+        if stream_input:
+            audio_embeddings = self.get_audio_embedding_streaming(items)
+        else:
+            audio_embeddings = self.get_audio_embedding(items, chunk_length)
+        bs = len(audio_embeddings)
+        # batch size
+        audio_embs = torch.cat(flatten_nested_list(audio_embeddings), dim=0)
+
+        return audio_embs
+
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        # list of tensors
+        pixel_values = flatten_nested_list([item.pixel_values for item in items])
+        tgt_sizes = torch.stack(
+            flatten_nested_list([item.tgt_size for item in items]), dim=0
+        )
+        assert len(pixel_values) == tgt_sizes.shape[0]
+
         device = self.vpm.embeddings.position_embedding.weight.device
         dtype = self.vpm.embeddings.position_embedding.weight.dtype
         all_pixel_values_lst = [
@@ -1845,10 +1791,10 @@ class MiniCPMO(MiniCPMVBaseModel):
 
         max_patches = (tgt_sizes[:, 0] * tgt_sizes[:, 1]).max().item()
         assert isinstance(max_patches, int)
-
         all_pixel_values = torch.nn.utils.rnn.pad_sequence(
             all_pixel_values_lst, batch_first=True, padding_value=0.0
         )
+
         B, L, _ = all_pixel_values.shape
         all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
         patch_attn_mask = torch.zeros(
@@ -1875,53 +1821,23 @@ class MiniCPMO(MiniCPMVBaseModel):
         forward_batch: ForwardBatch,
         **kwargs: Any,
     ) -> torch.Tensor:
-        inputs_embeds = None
-        # TODO(mick): optimize the logic here: clamp, merge and embedding should happens at most once
-        if (
-            not forward_batch.forward_mode.is_decode()
-            and forward_batch.contains_image_inputs()
-        ):
-            mm_inputs = forward_batch.merge_mm_inputs()
-            inputs_embeds = embed_mm_inputs(
-                mm_input=mm_inputs,
-                input_ids=input_ids,
-                input_embedding=self.get_input_embeddings(),
-                mm_data_embedding_func=self.get_image_features,
-                placeholder_token_ids=[mm_inputs.im_token_id] + mm_inputs.pad_values,
-            )
 
-        input_ids = input_ids.clamp(
-            min=0, max=self.get_input_embeddings().num_embeddings - 1
+        mm_input = forward_batch.merge_mm_inputs()
+        placeholder_token_ids = (
+            ([mm_input.im_token_id] + [item.pad_value for item in mm_input.mm_items])
+            if forward_batch.contains_mm_inputs()
+            else []
         )
-        if inputs_embeds is None:
-            inputs_embeds = self.llm.get_input_embeddings(input_ids)
-        if (
-            not forward_batch.forward_mode.is_decode()
-            and self.config.init_audio
-            and forward_batch.contains_audio_inputs()
-        ):
-            mm_input = forward_batch.merge_mm_inputs()
-            inputs_embeds = self.get_omni_embedding(
-                input_ids=input_ids,
-                multimodal_input=mm_input,
-                input_embeds=inputs_embeds,
-                forward_mode=forward_batch.forward_mode,
-                chunk_length=self.config.audio_chunk_length,
-                stream_input=False,
-            )
-
-        forward_batch.mm_inputs = None
-
-        hidden_states = self.llm.model(
-            input_ids=None,
-            positions=positions,
+        hidden_states = general_mm_embed_routine(
+            input_ids=input_ids,
             forward_batch=forward_batch,
-            input_embeds=inputs_embeds,
-        )
-
-        return self.logits_processor(
-            input_ids, hidden_states, self.llm.lm_head, forward_batch
+            language_model=self.llm,
+            image_data_embedding_func=self.get_image_feature,
+            audio_data_embedding_func=self.get_audio_feature,
+            placeholder_token_ids=placeholder_token_ids,
+            positions=positions,
         )
+        return hidden_states
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [

sglang/srt/models/minicpmv.py

@@ -54,12 +54,12 @@ from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternTokenPairs,
     general_mm_embed_routine,
 )
-from sglang.srt.managers.schedule_batch import MultimodalInputs
+from sglang.srt.managers.schedule_batch import MultimodalDataItem, MultimodalInputs
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.utils import set_default_torch_dtype
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.qwen2 import Qwen2Config, Qwen2ForCausalLM
-from sglang.srt.utils import add_prefix
+from sglang.srt.utils import add_prefix, flatten_nested_list
 
 RawImageType = Union[Image.Image, torch.Tensor]
 
@@ -661,7 +661,7 @@ def get_version_by_config(config: PretrainedConfig) -> Tuple[int, ...]:
     return tuple(int(x) for x in version_str.split("."))
 
 
-class MiniCPMVBaseModel(nn.Module):
+class MiniCPMBaseModel(nn.Module):
     """
     The abstract class of MiniCPMV can only be inherited, but cannot be
     instantiated.
@@ -853,7 +853,7 @@ class MiniCPMVBaseModel(nn.Module):
         return vlm_embedding, vision_hidden_states
 
     def get_input_embeddings(self) -> nn.Embedding:
-        return self.llm.get_input_embedding()
+        return self.llm.get_input_embeddings()
 
     def forward(
         self,
@@ -862,23 +862,14 @@ class MiniCPMVBaseModel(nn.Module):
         forward_batch: ForwardBatch,
         **kwargs: Any,
     ) -> torch.Tensor:
-        inputs_embeds = general_mm_embed_routine(
+        hidden_states = general_mm_embed_routine(
             input_ids=input_ids,
             forward_batch=forward_batch,
-            embed_tokens=self.get_input_embeddings(),
-            mm_data_embedding_func=self.get_image_features,
-        )
-
-        hidden_states = self.llm.model(
-            input_ids=None,
+            image_data_embedding_func=self.get_image_feature,
+            language_model=self.llm,
             positions=positions,
-            forward_batch=forward_batch,
-            input_embeds=inputs_embeds,
-        )
-
-        return self.logits_processor(
-            input_ids, hidden_states, self.llm.lm_head, forward_batch
         )
+        return hidden_states
 
     def init_llm(
         self,
@@ -913,11 +904,11 @@ class MiniCPMVBaseModel(nn.Module):
     ) -> torch.Tensor:
         raise NotImplementedError
 
-    def get_image_features(self, image_inputs: MultimodalInputs) -> torch.Tensor:
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
         raise NotImplementedError
 
 
-class MiniCPMV2_6(MiniCPMVBaseModel):
+class MiniCPMV2_6(MiniCPMBaseModel):
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",
@@ -1023,14 +1014,13 @@ class MiniCPMV2_6(MiniCPMVBaseModel):
         )
         return vision_embedding
 
-    def get_image_features(
-        self,
-        image_inputs: MultimodalInputs,
-    ) -> torch.Tensor:
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
         # list of tensors
-        pixel_values = image_inputs.pixel_values
-
-        tgt_sizes = image_inputs.tgt_sizes
+        pixel_values = flatten_nested_list([item.pixel_values for item in items])
+        tgt_sizes = torch.stack(
+            flatten_nested_list([item.tgt_size for item in items]), dim=0
+        )
+        assert len(pixel_values) == tgt_sizes.shape[0]
 
         device = self.vpm.embeddings.position_embedding.weight.device
         dtype = self.vpm.embeddings.position_embedding.weight.dtype
@@ -1040,10 +1030,10 @@ class MiniCPMV2_6(MiniCPMVBaseModel):
 
         max_patches = (tgt_sizes[:, 0] * tgt_sizes[:, 1]).max().item()
         assert isinstance(max_patches, int)
-
         all_pixel_values = torch.nn.utils.rnn.pad_sequence(
             all_pixel_values_lst, batch_first=True, padding_value=0.0
         )
+
         B, L, _ = all_pixel_values.shape
         all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
         patch_attn_mask = torch.zeros(

sglang/srt/models/mllama.py

@@ -796,14 +796,16 @@ class MllamaForConditionalGeneration(nn.Module):
         self.logits_processor = LogitsProcessor(config.text_config)
         self.capture_mode = False
 
-    def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
-        pixel_values = image_inputs.pixel_values
-        pad_values = image_inputs.pad_values
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
+        pixel_values = torch.cat(
+            [item.pixel_values for item in mm_inputs.mm_items], dim=0
+        )
+        pad_values = [item.pad_value for item in mm_inputs.mm_items]
 
         num_concurrent_media, num_tiles = pixel_values.shape[1:3]
         num_patches = self.vision_model.num_patches
         image_len = num_concurrent_media * num_tiles * num_patches
-        image_inputs.num_image_tokens = image_len
+        mm_inputs.num_image_tokens = image_len
 
         pad_ids = pad_values * ((image_len + len(pad_values)) // len(pad_values))
 
@@ -815,10 +817,16 @@ class MllamaForConditionalGeneration(nn.Module):
 
         # pixel_values: shape (bs, num_image, num_tiles, 3, image_res, image_res)
         max_num_images = max_num_tiles = bs = 0
-        for i, im in enumerate(forward_batch.mm_inputs):
-            if not forward_batch.encoder_cached[i] and im is not None:
-                max_num_images = max(max_num_images, im.pixel_values.shape[1])
-                max_num_tiles = max(max_num_tiles, im.pixel_values.shape[2])
+        for i, mm_input in enumerate(forward_batch.mm_inputs):
+
+            if not forward_batch.encoder_cached[i] and mm_input is not None:
+                pixel_values = torch.cat(
+                    [item.pixel_values for item in mm_input.mm_items], dim=0
+                )
+                # max_num_images = max(max_num_images, sum(1 if item.is_image() else 0 for item in mm_input.items))
+                max_num_images = max(max_num_images, pixel_values.shape[1])
+
+                max_num_tiles = max(max_num_tiles, pixel_values.shape[2])
                 bs += 1
 
         if max_num_images * max_num_tiles * bs == 0:
@@ -842,17 +850,24 @@ class MllamaForConditionalGeneration(nn.Module):
             )
             i = 0
             encoder_lens_need = []
-            for k, im in enumerate(forward_batch.mm_inputs):
-                if forward_batch.encoder_cached[k] or im is None:
+
+            for k, mm_input in enumerate(forward_batch.mm_inputs):
+                if forward_batch.encoder_cached[k] or mm_input is None:
                     continue
 
                 encoder_lens_need.append(forward_batch.encoder_lens[k])
-                for j in range(im.pixel_values.shape[1]):
-                    img = im.pixel_values[0, j]
+                pixel_values = torch.cat(
+                    [item.pixel_values for item in mm_input.mm_items], dim=0
+                )
+                for j in range(pixel_values.shape[1]):
+                    img = pixel_values[0, j]
                     num_tiles = img.shape[0]
                     batched_images[i, j, :num_tiles] = img
-                    batched_ar_ids[i, j] = im.aspect_ratio_ids[0, j]
-                    batched_ar_mask[i, j, :num_tiles] = im.aspect_ratio_mask[0, j]
+                    batched_ar_ids[i, j] = mm_input.mm_items[0].aspect_ratio_id[0, j]
+
+                    batched_ar_mask[i, j, :num_tiles] = mm_input.mm_items[
+                        0
+                    ].aspect_ratio_mask[0, j]
                 i += 1
 
         return batched_images, batched_ar_ids, batched_ar_mask, encoder_lens_need

sglang/srt/models/qwen2.py

@@ -261,11 +261,14 @@ class Qwen2Model(nn.Module):
         )
         self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
 
-    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
+    def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor:
         if hasattr(self.config, "scale_emb"):
-            return self.embed_tokens(input_ids) * self.config.scale_emb
+            return self.get_input_embeddings()(input_ids) * self.config.scale_emb
         else:
-            return self.embed_tokens(input_ids)
+            return self.get_input_embeddings()(input_ids)
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embed_tokens
 
     def forward(
         self,
@@ -358,10 +361,10 @@ class Qwen2ForCausalLM(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
-    def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor:
-        return self.model.get_input_embeddings(input_ids)
+    def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.model.get_input_embedding(input_ids)
 
-    def get_input_embedding(self) -> nn.Embedding:
+    def get_input_embeddings(self) -> nn.Embedding:
         return self.model.embed_tokens
 
     @torch.no_grad()

sglang/srt/models/qwen2_5_vl.py

@@ -30,22 +30,13 @@ import torch
 import torch.nn as nn
 import torch.nn.functional as F
 from einops import rearrange
-from transformers import AutoModel, Qwen2VLConfig
+from transformers import Qwen2VLConfig
 from transformers.activations import ACT2FN
 from transformers.models.qwen2.modeling_qwen2 import Qwen2RMSNorm
-from transformers.models.qwen2_5_vl import Qwen2_5_VLProcessor
 from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
-    Qwen2_5_VLConfig,
     Qwen2_5_VLVisionConfig,
 )
-from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
-    Qwen2_5_VLForConditionalGeneration,
-)
 
-from sglang.srt.distributed import (
-    get_tensor_model_parallel_rank,
-    get_tensor_model_parallel_world_size,
-)
 from sglang.srt.hf_transformers_utils import get_processor
 from sglang.srt.layers.attention.vision import VisionAttention
 from sglang.srt.layers.linear import ColumnParallelLinear, RowParallelLinear
@@ -57,7 +48,7 @@ from sglang.srt.managers.mm_utils import (
     MultiModalityDataPaddingPatternTokenPairs,
     general_mm_embed_routine,
 )
-from sglang.srt.managers.schedule_batch import MultimodalInputs
+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.models.qwen2 import Qwen2Model
@@ -513,19 +504,24 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
         self.logits_processor = LogitsProcessor(config)
         self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True)
 
-    def pad_input_ids(self, input_ids: List[int], image_inputs: MultimodalInputs):
+    def pad_input_ids(self, input_ids: List[int], mm_inputs: MultimodalInputs):
         # Get all special token IDs
-        im_start_id: int = image_inputs.im_start_id
-        im_end_id: int = image_inputs.im_end_id
+        im_start_id: int = mm_inputs.im_start_id
+        im_end_id: int = mm_inputs.im_end_id
 
         media_token_pairs = [(im_start_id, im_end_id)]
         pattern = MultiModalityDataPaddingPatternTokenPairs(media_token_pairs)
+        return pattern.pad_input_tokens(input_ids, mm_inputs)
 
-        return pattern.pad_input_tokens(input_ids, image_inputs)
-
-    def get_image_feature(self, image_input: MultimodalInputs) -> torch.Tensor:
-        pixel_values = image_input.pixel_values.type(self.visual.dtype)
-        image_embeds = self.visual(pixel_values, grid_thw=image_input.image_grid_thws)
+    def get_image_feature(self, items: List[MultimodalDataItem]) -> torch.Tensor:
+        # in qwen-vl, last dim is the same
+        pixel_values = torch.cat([item.pixel_values for item in items], dim=0).type(
+            self.visual.dtype
+        )
+        image_grid_thws = torch.concat([item.image_grid_thws for item in items], dim=0)
+        assert pixel_values.dim() == 2, pixel_values.dim()
+        assert image_grid_thws.dim() == 2, image_grid_thws.dim()
+        image_embeds = self.visual(pixel_values, grid_thw=image_grid_thws)
         return image_embeds
 
     def _process_video_input(self, video_input: Qwen2VLVideoInputs) -> torch.Tensor:
@@ -570,18 +566,12 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
                     f"(3, seq_len) positions, but got {positions.size()}"
                 )
 
-        inputs_embeds = general_mm_embed_routine(
+        hidden_states = general_mm_embed_routine(
             input_ids=input_ids,
             forward_batch=forward_batch,
-            embed_tokens=self.get_input_embeddings(),
-            mm_data_embedding_func=self.get_image_feature,
-        )
-
-        hidden_states = self.model(
-            input_ids=None,
+            language_model=self.model,
+            image_data_embedding_func=self.get_image_feature,
             positions=positions,
-            forward_batch=forward_batch,
-            input_embeds=inputs_embeds,
         )
 
         if not get_embedding:
@@ -594,9 +584,9 @@ class Qwen2_5_VLForConditionalGeneration(nn.Module):
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
         stacked_params_mapping = [
             # (param_name, shard_name, shard_id)
-            ("qkv_proj", "q_proj", "q"),
-            ("qkv_proj", "k_proj", "k"),
-            ("qkv_proj", "v_proj", "v"),
+            (".qkv_proj", ".q_proj", "q"),
+            (".qkv_proj", ".k_proj", "k"),
+            (".qkv_proj", ".v_proj", "v"),
             ("gate_up_proj", "up_proj", 1),
             ("gate_up_proj", "gate_proj", 0),
         ]