[MODEL] Qwen Multimodal Support (Qwen-VL / Qwen-VL-Chat) (#8029)

Signed-off-by: Alex-Brooks <Alex.Brooks@ibm.com>
Co-authored-by: DarkLight1337 <tlleungac@connect.ust.hk>

docs/source/models/supported_models.rst

@@ -242,6 +242,11 @@ Multimodal Language Models
     - Image\ :sup:`+`
     - :code:`openbmb/MiniCPM-V-2` (see note), :code:`openbmb/MiniCPM-Llama3-V-2_5`, :code:`openbmb/MiniCPM-V-2_6`, etc.
     -
+  * - :code:`QWenLMHeadModel`
+    - Qwen
+    - Image
+    - :code:`Qwen/Qwen-VL`, :code:`Qwen/Qwen-VL-Chat`, etc.
+    -
   * - :code:`UltravoxModel`
     - Ultravox
     - Audio\ :sup:`E+`

examples/offline_inference_vision_language.py

@@ -159,6 +159,20 @@ def run_blip2(question):
     return llm, prompt, stop_token_ids
 
 
+# Qwen
+def run_qwen_vl(question):
+
+    llm = LLM(
+        model="Qwen/Qwen-VL",
+        trust_remote_code=True,
+        max_num_seqs=5,
+    )
+
+    prompt = f"{question}Picture 1: <img></img>\n"
+    stop_token_ids = None
+    return llm, prompt, stop_token_ids
+
+
 model_example_map = {
     "llava": run_llava,
     "llava-next": run_llava_next,
@@ -169,6 +183,7 @@ model_example_map = {
     "minicpmv": run_minicpmv,
     "blip-2": run_blip2,
     "internvl_chat": run_internvl,
+    "qwen_vl": run_qwen_vl,
 }
 
 

tests/models/test_qwen.py

-from typing import Type
+import pathlib
+from typing import List, Optional, Type
 
 import pytest
 
-from ..conftest import HfRunner, VllmRunner
+from vllm.multimodal.utils import rescale_image_size
+
+from ..conftest import IMAGE_ASSETS, HfRunner, VllmRunner, _ImageAssets
 from .utils import check_logprobs_close
 
-models = ["qwen/qwen-vl"]
+pytestmark = pytest.mark.vlm
 
+text_only_models = [
+    "Qwen/Qwen-7B-Chat"  # Has no visual component
+]
 
-@pytest.mark.parametrize("dtype", ["half"])
-@pytest.mark.parametrize("max_tokens", [32])
-@pytest.mark.parametrize("num_logprobs", [5])
-@pytest.mark.parametrize("model", models)
-def test_text_only_qwen_model(
+multimodal_models = ["Qwen/Qwen-VL"]
+
+HF_IMAGE_PROMPTS = IMAGE_ASSETS.prompts({
+    "stop_sign":
+    "Picture 1: <img></img>\nWhat's the content of the image?: ",
+    "cherry_blossom":
+    "Picture 1: <img></img>\nWhat is the season?: ",
+})
+
+
+### Tests for multimodal Qwen models
+def run_test(
+    tmp_path: pathlib.PosixPath,
     hf_runner: Type[HfRunner],
     vllm_runner: Type[VllmRunner],
-    example_prompts,
+    image_assets: _ImageAssets,
     model: str,
     *,
+    size_factors: List[float],
     dtype: str,
     max_tokens: int,
     num_logprobs: int,
+    tensor_parallel_size: int,
+    distributed_executor_backend: Optional[str] = None,
 ):
-    # This test checks language inputs only, since the visual component
-    # for qwen-vl is still unsupported in VLLM. In the near-future, the
-    # implementation and this test will be extended to consider
-    # visual inputs as well.
-    with hf_runner(model, dtype=dtype) as hf_model:
-        hf_outputs = hf_model.generate_greedy_logprobs_limit(
-            example_prompts,
+    """Inference result should be the same between hf and vllm.
+
+    All the image fixtures for the test is under tests/images.
+    For huggingface runner, we provide the PIL images as input.
+    For vllm runner, we provide MultiModalDataDict objects
+    and corresponding MultiModalConfig as input.
+    Note, the text input is also adjusted to abide by vllm contract.
+    The text output is sanitized to be able to compare with hf.
+    """
+    images = [asset.pil_image for asset in image_assets]
+
+    # Export the images to a tempdir and substitute it into the hf prompt;
+    # the contents between <img>/</img> will be ignored by VLLM, but the
+    # transformers implementation for the visual transformer parses this to
+    # reload it in the forward call; the contents are treated as a URL or a
+    # local path.
+    for idx, asset in enumerate(image_assets):
+        image_tmp_path = tmp_path / f"{asset.name}.jpg"
+        asset.pil_image.save(image_tmp_path)
+        HF_IMAGE_PROMPTS[idx] = HF_IMAGE_PROMPTS[idx].replace(
+            "<img></img>", f"<img>{image_tmp_path}</img>")
+
+    inputs_per_image = [(
+        [prompt for _ in size_factors],
+        [rescale_image_size(image, factor) for factor in size_factors],
+    ) for image, prompt in zip(images, HF_IMAGE_PROMPTS)]
+
+    # NOTE: take care of the order. run vLLM first, and then run HF.
+    # vLLM needs a fresh new process without cuda initialization.
+    # if we run HF first, the cuda initialization will be done and it
+    # will hurt multiprocessing backend with fork method (the default method).
+
+    # max_model_len should be greater than image_feature_size
+    # Qwen encodes images into a fixed content size of 256
+    with vllm_runner(model,
+                     max_model_len=300,
+                     max_num_seqs=1,
+                     dtype=dtype,
+                     tensor_parallel_size=tensor_parallel_size,
+                     distributed_executor_backend=distributed_executor_backend,
+                     enforce_eager=True) as vllm_model:
+        vllm_outputs_per_image = [
+            vllm_model.generate_greedy_logprobs(prompts,
                                                 max_tokens,
                                                 num_logprobs=num_logprobs,
-        )
+                                                images=images)
+            for prompts, images in inputs_per_image
+        ]
 
-    with vllm_runner(model, dtype=dtype) as vllm_model:
-        vllm_outputs = vllm_model.generate_greedy_logprobs(
-            example_prompts,
+    with hf_runner(model, dtype=dtype) as hf_model:
+        hf_outputs_per_image = [
+            hf_model.generate_greedy_logprobs_limit(prompts,
                                                     max_tokens,
                                                     num_logprobs=num_logprobs,
-        )
+                                                    images=images)
+            for prompts, images in inputs_per_image
+        ]
+
+    for hf_outputs, vllm_outputs in zip(hf_outputs_per_image,
+                                        vllm_outputs_per_image):
 
         check_logprobs_close(
             outputs_0_lst=hf_outputs,
@@ -46,3 +106,60 @@ def test_text_only_qwen_model(
             name_0="hf",
             name_1="vllm",
         )
+
+
+@pytest.mark.parametrize("model", multimodal_models)
+@pytest.mark.parametrize(
+    "size_factors",
+    [
+        # No image
+        [],
+        # Single-scale
+        [1.0],
+        # Single-scale, batched
+        [1.0, 1.0, 1.0],
+        # Multi-scale
+        [0.25, 0.5, 1.0],
+    ],
+)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [8])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_multimodal_models(tmp_path, hf_runner, vllm_runner, image_assets,
+                           model, size_factors, dtype, max_tokens,
+                           num_logprobs) -> None:
+    run_test(
+        tmp_path,
+        hf_runner,
+        vllm_runner,
+        image_assets,
+        model,
+        size_factors=size_factors,
+        dtype=dtype,
+        max_tokens=max_tokens,
+        num_logprobs=num_logprobs,
+        tensor_parallel_size=1,
+    )
+
+
+# Ensure that a text-only Qwen model can still be loaded and
+# used for inference in VLLM without throwing.
+@pytest.mark.parametrize("model", text_only_models)
+@pytest.mark.parametrize("dtype", ["bfloat16"])
+@pytest.mark.parametrize("max_tokens", [32])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_text_only_qwen_model_can_be_loaded_and_run(
+    vllm_runner: Type[VllmRunner],
+    example_prompts,
+    model: str,
+    *,
+    dtype: str,
+    max_tokens: int,
+    num_logprobs: int,
+):
+    with vllm_runner(model, dtype=dtype) as vllm_model:
+        vllm_model.generate_greedy_logprobs(
+            example_prompts,
+            max_tokens,
+            num_logprobs=num_logprobs,
+        )

vllm/entrypoints/chat_utils.py

@@ -150,6 +150,8 @@ class BaseMultiModalItemTracker(ABC, Generic[_T]):
             if model_type in ("blip-2", "chatglm", "fuyu", "paligemma"):
                 # These models do not use image tokens in the prompt
                 return None
+            if model_type == "qwen":
+                return f"Picture {current_count}: <img></img>"
             if model_type.startswith("llava"):
                 return self._cached_token_str(self._tokenizer,
                                               hf_config.image_token_index)

vllm/model_executor/layers/resampler.py

+# coding=utf-8
+# Adapted from
+# https://github.com/huggingface/transformers/blob/v4.28.0/src/transformers/models/llama/modeling_llama.py
+# https://huggingface.co/Qwen/Qwen-7B/blob/main/modeling_qwen.py
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+#
+# Copyright 2023 The Qwen team.
+# Copyright 2023 The vLLM team.
+# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Shared resampler perceiver network used in multimodal models and
+related helpers for sincos positional embeddings.
+
+Example models: Qwen (Qwen-VL), Minicpmv2.0
+"""
+import math
+from functools import partial
+from typing import Callable, Optional, Tuple, Union
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from torch import nn
+from torch.nn.init import trunc_normal_
+
+from vllm.model_executor.layers.linear import ReplicatedLinear
+
+DEFAULT_LN = partial(nn.LayerNorm, eps=1e-6)
+
+
+def get_abs_pos(abs_pos: torch.Tensor, tgt_size: Union[torch.Tensor,
+                                                       int]) -> torch.Tensor:
+    # abs_pos: L, C
+    # tgt_size: (H, W)
+    # return: M, C
+    src_size = int(math.sqrt(abs_pos.size(0)))
+    dtype = abs_pos.dtype
+    if isinstance(tgt_size, int):
+        tgt_size = (tgt_size, tgt_size)
+    if (src_size == tgt_size[0] and src_size == tgt_size[1]):
+        return abs_pos
+    return (F.interpolate(
+        abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
+        size=(tgt_size[0], tgt_size[1]),
+        mode="bicubic",
+        align_corners=False,
+    ).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype))
+
+
+# sin/cos positional embedding helpers are adapted from:
+# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
+def get_1d_sincos_pos_embed_from_grid(
+    embed_dim: int, pos: np.ndarray,
+    version: Tuple[int, int] = (2, 0)) -> torch.Tensor:
+    """
+    embed_dim: output dimension for each position
+    pos: a list of positions to be encoded: size (M,) / (H, W)
+    out: (M, D) / (H, W, D)
+    """
+    assert embed_dim % 2 == 0
+    omega = np.arange(embed_dim // 2, dtype=np.float32)
+    omega /= embed_dim / 2.0
+    omega = 1.0 / 10000**omega  # (D/2,)
+
+    if version == (2, 0):
+        pos = pos.reshape(-1)  # (M,)
+        out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
+        emb_sin = np.sin(out)  # (M, D/2)
+        emb_cos = np.cos(out)  # (M, D/2)
+        emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
+    else:
+        out = np.einsum("hw,d->hwd", pos, omega)  # (H, W, D/2), outer product
+        emb_sin = np.sin(out)  # (H, W, D/2)
+        emb_cos = np.cos(out)  # (H, W, D/2)
+        emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (H, W, D)
+    return emb
+
+
+def get_2d_sincos_pos_embed_from_grid(
+    embed_dim: int, grid: np.ndarray,
+    version: Tuple[int, int] = (2, 0)) -> torch.Tensor:
+    assert embed_dim % 2 == 0
+
+    # use half of dimensions to encode grid_h
+    emb_h = get_1d_sincos_pos_embed_from_grid(
+        embed_dim // 2, grid[0], version)  # (H*W, D/2) or (H, W, D/2)
+    emb_w = get_1d_sincos_pos_embed_from_grid(
+        embed_dim // 2, grid[1], version)  # (H*W, D/2) or (H, W, D/2)
+
+    if version == (2, 0):
+        emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
+    else:
+        emb = np.concatenate([emb_h, emb_w], axis=-1)  # (H, W, D)
+    return emb
+
+
+def get_2d_sincos_pos_embed(
+        embed_dim: int,
+        grid_size: Union[int, Tuple[int, int]],
+        cls_token: bool = False,
+        version: Tuple[int, int] = (2, 0),
+) -> torch.Tensor:
+    """
+    grid_size: int of the grid height and width
+    return:
+    pos_embed: [grid_size*grid_size, embed_dim] or
+                [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+    """
+    if isinstance(grid_size, int):
+        grid_h_size, grid_w_size = grid_size, grid_size
+    else:
+        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
+
+    grid_h = np.arange(grid_h_size, dtype=np.float32)
+    grid_w = np.arange(grid_w_size, dtype=np.float32)
+    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
+    grid = np.stack(grid, axis=0)
+    assert isinstance(grid, np.ndarray) and \
+        grid.shape == (2, grid_h_size, grid_w_size)
+
+    if version == (2, 0):
+        grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
+        pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid, version)
+        if cls_token:
+            pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed],
+                                       axis=0)
+    else:
+        pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid, version)
+    return pos_embed
+
+
+class BaseResampler(nn.Module):
+    """
+    A 2D perceiver-resampler network with one cross attention layers by
+        (grid_size**2) learnable queries and 2d sincos pos_emb.
+    Outputs:
+        A tensor with the shape of (grid_size**2, embed_dim)
+    """
+
+    def __init__(
+        self,
+        num_queries: int,
+        embed_dim: int,
+        num_heads: int,
+        kv_dim: Optional[int] = None,
+        norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
+        do_post_projection: bool = True,
+    ) -> None:
+        super().__init__()
+
+        self.num_queries = num_queries
+        self.embed_dim = embed_dim
+        self.num_heads = num_heads
+
+        self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
+        trunc_normal_(self.query, std=0.02)
+        if kv_dim is not None and kv_dim != embed_dim:
+            self.kv_proj = ReplicatedLinear(kv_dim, embed_dim, bias=False)
+        else:
+            # Maintain the same return value with ReplicatedLinear.forward
+            self.kv_proj = lambda *args, **kwargs: (  # type: ignore # noqa 
+                nn.Identity()(*args, **kwargs),
+                None,
+            )
+        self.attn = nn.MultiheadAttention(embed_dim, num_heads)
+        self.ln_q = norm_layer(embed_dim)
+        self.ln_kv = norm_layer(embed_dim)
+        self.do_post_projection = do_post_projection
+        self.ln_post = norm_layer(embed_dim) if do_post_projection else None
+        self.proj = nn.Parameter(
+            (embed_dim**-0.5) *
+            torch.randn(embed_dim, embed_dim)) if do_post_projection else None
+
+    def _init_weights(self, m: nn.Module) -> None:
+        if isinstance(m, nn.Linear):
+            trunc_normal_(m.weight, std=0.02)
+            if isinstance(m, nn.Linear) and m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, nn.LayerNorm):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def _repeat(self, query, N: int):
+        return query.unsqueeze(1).repeat(1, N, 1)
+
+
+class Resampler2(BaseResampler):
+    """Resampler-perceiver network to be used for a variety of model types,
+    e.g., Qwen-vl / Minicpmv 2.0. The main difference is the addition of the
+    do_post_projection arg, which indicates whether or not there should be
+    a post layer normalization and projector after the attention. This is
+    present in minicpmv2.0, but not qwen-vl.
+    """
+
+    def __init__(
+        self,
+        grid_size: int,
+        embed_dim: int,
+        num_heads: int,
+        kv_dim: Optional[int] = None,
+        norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
+        adaptive: bool = False,
+        do_post_projection: bool = True,
+    ) -> None:
+        super().__init__(grid_size**2,
+                         embed_dim,
+                         num_heads,
+                         kv_dim,
+                         norm_layer,
+                         do_post_projection=do_post_projection)
+
+        self.adaptive = adaptive
+        pos_embed_arr = get_2d_sincos_pos_embed(embed_dim,
+                                                grid_size,
+                                                version=(2, 0))
+
+        self.pos_embed = nn.Parameter(
+            torch.from_numpy(pos_embed_arr).requires_grad_(False))
+
+        self.apply(self._init_weights)
+
+    def forward(
+        self,
+        x: torch.Tensor,
+        tgt_sizes: Optional[torch.Tensor] = None,
+        attn_mask: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if tgt_sizes is None:
+            tgt_sizes = int(math.sqrt(x.size(1)))
+        if self.adaptive:
+            pos_embed_arr = get_2d_sincos_pos_embed(self.embed_dim,
+                                                    tgt_sizes,
+                                                    version=(2, 0))
+            pos_embed = torch.from_numpy(pos_embed_arr).to(device=x.device,
+                                                           dtype=x.dtype)
+        else:
+            pos_embed = get_abs_pos(self.pos_embed,
+                                    tgt_sizes).to(device=x.device,
+                                                  dtype=x.dtype)
+
+        x, _ = self.kv_proj(x)
+        x = self.ln_kv(x).permute(1, 0, 2)
+
+        N = x.shape[1]
+        q = self.ln_q(self.query)
+        out = self.attn(
+            self._repeat(q, N) + self.pos_embed.unsqueeze(1),
+            x + pos_embed.unsqueeze(1),
+            x,
+            attn_mask=attn_mask,
+        )[0]
+        x = out.permute(1, 0, 2)
+        if self.do_post_projection:
+            x = self.ln_post(x)
+            x = x @ self.proj
+        return x

vllm/model_executor/models/__init__.py

@@ -51,7 +51,6 @@ _GENERATION_MODELS = {
     "PhiForCausalLM": ("phi", "PhiForCausalLM"),
     "Phi3ForCausalLM": ("llama", "LlamaForCausalLM"),
     "PhiMoEForCausalLM": ("phimoe", "PhiMoEForCausalLM"),
-    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
     "Qwen2ForCausalLM": ("qwen2", "Qwen2ForCausalLM"),
     "Qwen2MoeForCausalLM": ("qwen2_moe", "Qwen2MoeForCausalLM"),
     "RWForCausalLM": ("falcon", "FalconForCausalLM"),
@@ -88,6 +87,7 @@ _MULTIMODAL_MODELS = {
                                           "PaliGemmaForConditionalGeneration"),
     "Phi3VForCausalLM": ("phi3v", "Phi3VForCausalLM"),
     "UltravoxModel": ("ultravox", "UltravoxModel"),
+    "QWenLMHeadModel": ("qwen", "QWenLMHeadModel"),
 }
 _CONDITIONAL_GENERATION_MODELS = {
     "BartModel": ("bart", "BartForConditionalGeneration"),

vllm/model_executor/models/minicpmv.py

@@ -26,11 +26,9 @@ import re
 from array import array
 from functools import partial
 from typing import (Any, Callable, Iterable, List, Mapping, Optional, Tuple,
-                    TypedDict, Union)
+                    TypedDict)
 
-import numpy as np
 import torch
-import torch.nn.functional as F
 import torch.types
 from PIL import Image
 from torch import nn
@@ -44,6 +42,8 @@ from vllm.logger import init_logger
 from vllm.model_executor.layers.linear import ReplicatedLinear
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.quantization import QuantizationConfig
+from vllm.model_executor.layers.resampler import (Resampler2,
+                                                  get_2d_sincos_pos_embed)
 from vllm.model_executor.layers.sampler import Sampler, SamplerOutput
 from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
 from vllm.model_executor.model_loader.utils import set_default_torch_dtype
@@ -98,101 +98,6 @@ MiniCPMVImageInputs = MiniCPMVImagePixelInputs
 DEFAULT_LN = partial(nn.LayerNorm, eps=1e-6)
 
 
-def get_abs_pos(abs_pos: torch.Tensor, tgt_size: torch.Tensor):
-    # abs_pos: L, C
-    # tgt_size: (H, W)
-    # return: M, C
-    src_size = int(math.sqrt(abs_pos.size(0)))
-    # tgt_size = int(math.sqrt(tgt_size))
-    dtype = abs_pos.dtype
-
-    return (F.interpolate(
-        abs_pos.float().reshape(1, src_size, src_size, -1).permute(0, 3, 1, 2),
-        size=(tgt_size[0], tgt_size[1]),
-        mode="bicubic",
-        align_corners=False,
-    ).permute(0, 2, 3, 1).flatten(0, 2).to(dtype=dtype))
-
-
-# https://github.com/facebookresearch/mae/blob/efb2a8062c206524e35e47d04501ed4f544c0ae8/util/pos_embed.py#L20
-def get_2d_sincos_pos_embed(
-        embed_dim: int,
-        grid_size: Union[int, Tuple[int, int]],
-        cls_token: bool = False,
-        version: Tuple[int, int] = (2, 0),
-):
-    """
-    grid_size: int of the grid height and width
-    return:
-    pos_embed: [grid_size*grid_size, embed_dim] or
-                [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
-    """
-    if isinstance(grid_size, int):
-        grid_h_size, grid_w_size = grid_size, grid_size
-    else:
-        grid_h_size, grid_w_size = grid_size[0], grid_size[1]
-
-    grid_h = np.arange(grid_h_size, dtype=np.float32)
-    grid_w = np.arange(grid_w_size, dtype=np.float32)
-    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
-    grid = np.stack(grid, axis=0)
-
-    if version == (2, 0):
-        grid = grid.reshape([2, 1, grid_h_size, grid_w_size])
-        pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid, version)
-        if cls_token:
-            pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed],
-                                       axis=0)
-    else:
-        pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid, version)
-    return pos_embed
-
-
-def get_2d_sincos_pos_embed_from_grid(embed_dim: int,
-                                      grid: np.ndarray,
-                                      version: Tuple[int, int] = (2, 0)):
-    assert embed_dim % 2 == 0
-
-    # use half of dimensions to encode grid_h
-    emb_h = get_1d_sincos_pos_embed_from_grid(
-        embed_dim // 2, grid[0], version)  # (H*W, D/2) or (H, W, D/2)
-    emb_w = get_1d_sincos_pos_embed_from_grid(
-        embed_dim // 2, grid[1], version)  # (H*W, D/2) or (H, W, D/2)
-
-    if version == (2, 0):
-        emb = np.concatenate([emb_h, emb_w], axis=1)  # (H*W, D)
-    else:
-        emb = np.concatenate([emb_h, emb_w], axis=-1)  # (H, W, D)
-    return emb
-
-
-def get_1d_sincos_pos_embed_from_grid(embed_dim: int,
-                                      pos: np.ndarray,
-                                      version: Tuple[int, int] = (2, 0)):
-    """
-    embed_dim: output dimension for each position
-    pos: a list of positions to be encoded: size (M,) / (H, W)
-    out: (M, D) / (H, W, D)
-    """
-    assert embed_dim % 2 == 0
-    omega = np.arange(embed_dim // 2, dtype=np.float32)
-    omega /= embed_dim / 2.0
-    omega = 1.0 / 10000**omega  # (D/2,)
-
-    if version == (2, 0):
-        pos = pos.reshape(-1)  # (M,)
-        out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
-        emb_sin = np.sin(out)  # (M, D/2)
-        emb_cos = np.cos(out)  # (M, D/2)
-        emb = np.concatenate([emb_sin, emb_cos], axis=1)  # (M, D)
-    else:
-        out = np.einsum("hw,d->hwd", pos, omega)  # (H, W, D/2), outer product
-        emb_sin = np.sin(out)  # (H, W, D/2)
-        emb_cos = np.cos(out)  # (H, W, D/2)
-        emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (H, W, D)
-    return emb
-
-
 class BaseResampler(nn.Module):
     """
     A 2D perceiver-resampler network with one cross attention layers by
@@ -245,62 +150,6 @@ class BaseResampler(nn.Module):
         return query.unsqueeze(1).repeat(1, N, 1)
 
 
-class Resampler2(BaseResampler):
-
-    def __init__(
-        self,
-        grid_size: int,
-        embed_dim: int,
-        num_heads: int,
-        kv_dim: Optional[int] = None,
-        norm_layer: Callable[[int], nn.LayerNorm] = DEFAULT_LN,
-        adaptive: bool = False,
-    ) -> None:
-        super().__init__(grid_size**2, embed_dim, num_heads, kv_dim,
-                         norm_layer)
-
-        self.adaptive = adaptive
-        pos_embed_arr = get_2d_sincos_pos_embed(embed_dim,
-                                                grid_size,
-                                                version=(2, 0))
-        self.pos_embed = nn.Parameter(
-            torch.from_numpy(pos_embed_arr).float()).requires_grad_(False)
-
-        self.apply(self._init_weights)
-
-    def forward(
-        self,
-        x: torch.Tensor,
-        tgt_sizes: torch.Tensor,
-        attn_mask: Optional[torch.Tensor] = None,
-    ):
-        if self.adaptive:
-            pos_embed_arr = get_2d_sincos_pos_embed(self.embed_dim,
-                                                    tgt_sizes,
-                                                    version=(2, 0))
-            pos_embed = torch.from_numpy(pos_embed_arr).to(device=x.device,
-                                                           dtype=x.dtype)
-        else:
-            pos_embed = get_abs_pos(self.pos_embed, tgt_sizes)
-
-        x, _ = self.kv_proj(x)
-        x = self.ln_kv(x).permute(1, 0, 2)
-
-        N = x.shape[1]
-        q = self.ln_q(self.query)
-        out = self.attn(
-            self._repeat(q, N) + self.pos_embed.unsqueeze(1),
-            x + pos_embed.unsqueeze(1),
-            x,
-            attn_mask=attn_mask,
-        )[0]
-        x = out.permute(1, 0, 2)
-
-        x = self.ln_post(x)
-        x = x @ self.proj
-        return x
-
-
 class Resampler2_5(BaseResampler):
 
     def __init__(
@@ -782,7 +631,8 @@ class MiniCPMV2_0(MiniCPMVBaseModel):
                 num_heads=embed_dim // 128,
                 grid_size=int(math.sqrt(self.config.query_num)),
                 kv_dim=vision_dim,
-                adaptive=True,
+                adaptive=False,
+                do_post_projection=True,
             )
 
         return resampler