2024-09-12 00:31:19 +08:00
|
|
|
# Adapted from
|
|
|
|
# https://github.com/huggingface/transformers/blob/19e6e80e10118f855137b90740936c0b11ac397f/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py
|
|
|
|
# Copyright 2024 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.
|
|
|
|
"""Inference-only Qwen2-VL model compatible with HuggingFace weights."""
|
2024-10-18 15:29:56 -04:00
|
|
|
from functools import partial
|
2024-11-07 18:50:44 +08:00
|
|
|
from typing import (Any, Callable, Dict, Iterable, List, Literal, Mapping,
|
|
|
|
Optional, Tuple, Type, TypedDict, Union)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
import torch
|
|
|
|
import torch.nn as nn
|
|
|
|
import torch.nn.functional as F
|
|
|
|
from einops import rearrange, repeat
|
|
|
|
from PIL import Image
|
|
|
|
from transformers.image_utils import (get_image_size,
|
|
|
|
infer_channel_dimension_format,
|
|
|
|
to_numpy_array)
|
2024-10-16 13:56:17 +08:00
|
|
|
from transformers.models.qwen2_vl.configuration_qwen2_vl import (
|
|
|
|
Qwen2VLConfig, Qwen2VLVisionConfig)
|
2024-09-12 00:31:19 +08:00
|
|
|
from transformers.models.qwen2_vl.image_processing_qwen2_vl import (
|
|
|
|
make_batched_images, make_batched_videos, smart_resize)
|
|
|
|
|
|
|
|
from vllm.attention import AttentionMetadata
|
2024-10-16 19:52:01 +08:00
|
|
|
from vllm.attention.selector import _Backend
|
2024-11-08 22:17:28 -08:00
|
|
|
from vllm.config import VllmConfig
|
2024-09-23 21:46:59 +08:00
|
|
|
from vllm.distributed import get_pp_group, parallel_state
|
2024-09-12 00:31:19 +08:00
|
|
|
from vllm.distributed import utils as dist_utils
|
2024-11-01 16:21:10 -07:00
|
|
|
from vllm.inputs import (INPUT_REGISTRY, DecoderOnlyInputs, DummyData,
|
|
|
|
InputContext, token_inputs)
|
2024-09-12 00:31:19 +08:00
|
|
|
from vllm.logger import init_logger
|
|
|
|
from vllm.model_executor import SamplingMetadata
|
|
|
|
from vllm.model_executor.layers.activation import QuickGELU
|
|
|
|
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
|
|
|
|
RowParallelLinear)
|
|
|
|
from vllm.model_executor.layers.logits_processor import LogitsProcessor
|
2024-11-13 02:28:13 -06:00
|
|
|
from vllm.model_executor.layers.pooler import Pooler, PoolingType
|
2024-11-08 22:36:46 -05:00
|
|
|
from vllm.model_executor.layers.quantization import (GPTQConfig,
|
|
|
|
GPTQMarlinConfig,
|
|
|
|
QuantizationConfig)
|
2024-11-06 12:57:35 -07:00
|
|
|
from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
|
2024-09-12 00:31:19 +08:00
|
|
|
from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
|
|
|
|
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
|
|
|
from vllm.model_executor.models.qwen2 import Qwen2Model
|
2024-11-13 02:28:13 -06:00
|
|
|
from vllm.model_executor.pooling_metadata import PoolingMetadata
|
2024-09-12 00:31:19 +08:00
|
|
|
from vllm.multimodal import (MULTIMODAL_REGISTRY, MultiModalDataDict,
|
2024-11-09 11:31:02 +08:00
|
|
|
MultiModalKwargs)
|
2024-09-12 00:31:19 +08:00
|
|
|
from vllm.multimodal.base import MultiModalData
|
|
|
|
from vllm.multimodal.image import cached_get_image_processor
|
2024-10-24 14:12:05 +08:00
|
|
|
from vllm.multimodal.utils import cached_get_tokenizer
|
2024-11-13 02:28:13 -06:00
|
|
|
from vllm.sequence import IntermediateTensors, PoolerOutput, SequenceData
|
2024-10-16 13:56:17 +08:00
|
|
|
from vllm.transformers_utils.config import uses_mrope
|
2024-10-18 15:29:56 -04:00
|
|
|
from vllm.transformers_utils.processor import cached_get_processor
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-11-06 22:13:15 +08:00
|
|
|
from .interfaces import SupportsLoRA, SupportsMultiModal, SupportsPP
|
2024-10-16 19:52:01 +08:00
|
|
|
from .utils import (PPMissingLayer, get_vit_attn_backend,
|
|
|
|
is_pp_missing_parameter,
|
2024-11-10 22:41:46 -08:00
|
|
|
make_empty_intermediate_tensors_factory, maybe_prefix)
|
2024-09-23 21:46:59 +08:00
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
logger = init_logger(__name__)
|
|
|
|
|
|
|
|
# === Vision Inputs === #
|
|
|
|
|
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
class Qwen2VLImagePixelInputs(TypedDict):
|
|
|
|
type: Literal["pixel_values"]
|
2024-11-13 15:07:22 +08:00
|
|
|
pixel_values: torch.Tensor
|
2024-10-22 15:50:43 +08:00
|
|
|
"""Shape:
|
2024-09-12 00:31:19 +08:00
|
|
|
`(num_patches, num_channels * patch_size * patch_size)`
|
|
|
|
"""
|
|
|
|
|
|
|
|
image_grid_thw: torch.Tensor
|
|
|
|
"""Shape: `(num_images, 3)`
|
|
|
|
This should be in `(grid_t, grid_h, grid_w)` format.
|
|
|
|
"""
|
|
|
|
|
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
class Qwen2VLImageEmbeddingInputs(TypedDict):
|
|
|
|
type: Literal["image_embeds"]
|
2024-11-13 15:07:22 +08:00
|
|
|
image_embeds: torch.Tensor
|
|
|
|
"""Supported types:
|
|
|
|
- List[`torch.Tensor`]: A list of tensors holding all images' features.
|
|
|
|
Each tensor holds an image's features.
|
|
|
|
- `torch.Tensor`: A tensor holding all images' features
|
|
|
|
(concatenation of all images' feature tensors).
|
|
|
|
|
|
|
|
Tensor shape: `(num_image_features, hidden_size)`
|
|
|
|
- `num_image_features` varies based on
|
|
|
|
the number and resolution of the images.
|
|
|
|
- `hidden_size` must match the hidden size of language model backbone.
|
|
|
|
"""
|
|
|
|
|
|
|
|
image_grid_thw: torch.Tensor
|
|
|
|
"""Shape: `(num_images, 3)`
|
|
|
|
This should be in `(grid_t, grid_h, grid_w)` format.
|
2024-09-30 11:16:10 +08:00
|
|
|
"""
|
|
|
|
|
|
|
|
|
|
|
|
Qwen2VLImageInputs = Union[Qwen2VLImagePixelInputs,
|
|
|
|
Qwen2VLImageEmbeddingInputs]
|
|
|
|
|
|
|
|
|
2024-11-13 15:07:22 +08:00
|
|
|
class Qwen2VLVideoPixelInputs(TypedDict):
|
|
|
|
type: Literal["pixel_values_videos"]
|
2024-09-12 00:31:19 +08:00
|
|
|
pixel_values_videos: torch.Tensor
|
2024-10-22 15:50:43 +08:00
|
|
|
"""Shape:
|
|
|
|
`(num_patches,
|
2024-09-12 00:31:19 +08:00
|
|
|
num_channels * temporal_patch_size * patch_size * patch_size)`
|
|
|
|
"""
|
|
|
|
|
|
|
|
video_grid_thw: torch.Tensor
|
|
|
|
"""Shape: `(num_videos, 3)`
|
2024-10-22 15:50:43 +08:00
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
This should be in `(grid_t, grid_h, grid_w)` format.
|
|
|
|
"""
|
|
|
|
|
|
|
|
|
2024-11-13 15:07:22 +08:00
|
|
|
class Qwen2VLVideoEmbeddingInputs(TypedDict):
|
|
|
|
type: Literal["video_embeds"]
|
|
|
|
video_embeds: torch.Tensor
|
|
|
|
"""Supported types:
|
|
|
|
- List[`torch.Tensor`]: A list of tensors holding all videos' features.
|
|
|
|
Each tensor holds an video's features.
|
|
|
|
- `torch.Tensor`: A tensor holding all videos' features
|
|
|
|
(concatenation of all videos' feature tensors).
|
|
|
|
|
|
|
|
Tensor shape: `(num_image_features, hidden_size)`
|
|
|
|
- `num_image_features` varies based on
|
|
|
|
the number and resolution of the videos.
|
|
|
|
- `hidden_size` must match the hidden size of language model backbone.
|
|
|
|
"""
|
|
|
|
|
|
|
|
video_grid_thw: torch.Tensor
|
|
|
|
"""Shape: `(num_videos, 3)`
|
|
|
|
This should be in `(grid_t, grid_h, grid_w)` format.
|
|
|
|
"""
|
|
|
|
|
|
|
|
|
|
|
|
Qwen2VLVideoInputs = Union[Qwen2VLVideoPixelInputs,
|
|
|
|
Qwen2VLVideoEmbeddingInputs]
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
# === Vision Encoder === #
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionMLP(nn.Module):
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
in_features: int,
|
|
|
|
hidden_features: int = None,
|
|
|
|
act_layer: Type[nn.Module] = QuickGELU,
|
|
|
|
quant_config: Optional[QuantizationConfig] = None,
|
2024-10-31 01:41:20 -04:00
|
|
|
prefix: str = "",
|
2024-09-12 00:31:19 +08:00
|
|
|
):
|
|
|
|
super().__init__()
|
|
|
|
self.fc1 = ColumnParallelLinear(in_features,
|
|
|
|
hidden_features,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.fc1")
|
2024-09-12 00:31:19 +08:00
|
|
|
self.act = act_layer()
|
|
|
|
self.fc2 = RowParallelLinear(hidden_features,
|
|
|
|
in_features,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.fc2")
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
x_parallel, _ = self.fc1(x)
|
|
|
|
x_parallel = self.act(x_parallel)
|
|
|
|
x, _ = self.fc2(x_parallel)
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
def rotate_half(x: torch.Tensor, interleaved: bool = False) -> torch.Tensor:
|
|
|
|
if not interleaved:
|
|
|
|
x1, x2 = x.chunk(2, dim=-1)
|
|
|
|
return torch.cat((-x2, x1), dim=-1)
|
|
|
|
else:
|
|
|
|
x1, x2 = x[..., ::2], x[..., 1::2]
|
|
|
|
return rearrange(torch.stack((-x2, x1), dim=-1),
|
|
|
|
"... d two -> ... (d two)",
|
|
|
|
two=2)
|
|
|
|
|
|
|
|
|
|
|
|
def apply_rotary_emb_torch(x: torch.Tensor,
|
|
|
|
cos: torch.Tensor,
|
|
|
|
sin: torch.Tensor,
|
|
|
|
interleaved: bool = False) -> torch.Tensor:
|
|
|
|
"""
|
|
|
|
x: (batch_size, seqlen, nheads, headdim)
|
|
|
|
cos, sin: (seqlen, rotary_dim / 2) or (batch_size, seqlen, rotary_dim / 2)
|
|
|
|
"""
|
|
|
|
ro_dim = cos.shape[-1] * 2
|
|
|
|
assert ro_dim <= x.shape[-1]
|
|
|
|
cos = repeat(
|
|
|
|
cos,
|
|
|
|
"... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
|
|
|
sin = repeat(
|
|
|
|
sin,
|
|
|
|
"... d -> ... 1 (2 d)" if not interleaved else "... d -> ... 1 (d 2)")
|
|
|
|
return torch.cat(
|
|
|
|
[
|
|
|
|
x[..., :ro_dim] * cos +
|
|
|
|
rotate_half(x[..., :ro_dim], interleaved) * sin, x[..., ro_dim:]
|
|
|
|
],
|
|
|
|
dim=-1,
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
def apply_rotary_pos_emb_vision(t: torch.Tensor,
|
|
|
|
freqs: torch.Tensor) -> torch.Tensor:
|
|
|
|
t_ = t.float()
|
|
|
|
cos = freqs.cos()
|
|
|
|
sin = freqs.sin()
|
|
|
|
output = apply_rotary_emb_torch(t_, cos, sin).type_as(t)
|
|
|
|
return output
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionAttention(nn.Module):
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
embed_dim: Optional[int] = None,
|
|
|
|
num_heads: Optional[int] = None,
|
|
|
|
projection_size: Optional[int] = None,
|
|
|
|
quant_config: Optional[QuantizationConfig] = None,
|
2024-10-31 01:41:20 -04:00
|
|
|
prefix: str = "",
|
2024-09-12 00:31:19 +08:00
|
|
|
) -> None:
|
|
|
|
super().__init__()
|
|
|
|
# Per attention head and per partition values.
|
|
|
|
world_size = parallel_state.get_tensor_model_parallel_world_size()
|
|
|
|
self.hidden_size_per_attention_head = dist_utils.divide(
|
|
|
|
projection_size, num_heads)
|
|
|
|
self.num_attention_heads_per_partition = dist_utils.divide(
|
|
|
|
num_heads, world_size)
|
|
|
|
|
|
|
|
self.qkv = ColumnParallelLinear(input_size=embed_dim,
|
|
|
|
output_size=3 * projection_size,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.qkv")
|
2024-09-12 00:31:19 +08:00
|
|
|
self.proj = RowParallelLinear(input_size=projection_size,
|
|
|
|
output_size=embed_dim,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.proj")
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
# Detect attention implementation.
|
2024-10-16 19:52:01 +08:00
|
|
|
self.attn_backend: _Backend = get_vit_attn_backend()
|
|
|
|
if self.attn_backend not in {
|
|
|
|
_Backend.FLASH_ATTN, _Backend.TORCH_SDPA, _Backend.XFORMERS
|
|
|
|
}:
|
|
|
|
raise RuntimeError(
|
|
|
|
f"Qwen2-VL does not support {self.attn_backend} backend now.")
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
def forward(
|
|
|
|
self,
|
|
|
|
x: torch.Tensor,
|
|
|
|
cu_seqlens: torch.Tensor,
|
|
|
|
rotary_pos_emb: torch.Tensor = None,
|
|
|
|
) -> torch.Tensor:
|
|
|
|
# [s, b, c] --> [s, b, head * 3 * head_dim]
|
|
|
|
x, _ = self.qkv(x)
|
|
|
|
|
|
|
|
# [s, b, head * 3 * head_dim] --> [s, b, head, 3 * head_dim]
|
|
|
|
new_x_shape = x.size()[:-1] + (
|
|
|
|
self.num_attention_heads_per_partition,
|
|
|
|
3 * self.hidden_size_per_attention_head,
|
|
|
|
)
|
|
|
|
x = x.view(*new_x_shape)
|
|
|
|
|
|
|
|
# [s, b, head, 3 * head_dim] --> 3 [s, b, head, head_dim]
|
|
|
|
q, k, v = dist_utils.split_tensor_along_last_dim(x, 3)
|
|
|
|
batch_size = q.shape[1]
|
|
|
|
|
2024-11-06 02:11:55 -05:00
|
|
|
q, k, v = (rearrange(x, "s b ... -> b s ...").contiguous()
|
|
|
|
for x in (q, k, v))
|
2024-09-12 00:31:19 +08:00
|
|
|
if rotary_pos_emb is not None:
|
|
|
|
q = apply_rotary_pos_emb_vision(q, rotary_pos_emb)
|
|
|
|
k = apply_rotary_pos_emb_vision(k, rotary_pos_emb)
|
|
|
|
|
2024-10-16 19:52:01 +08:00
|
|
|
if self.attn_backend == _Backend.FLASH_ATTN:
|
2024-09-12 00:31:19 +08:00
|
|
|
# from vllm_flash_attn.flash_attn_interface import (
|
|
|
|
# flash_attn_varlen_func)
|
|
|
|
from flash_attn import flash_attn_varlen_func
|
|
|
|
|
2024-11-06 02:11:55 -05:00
|
|
|
q, k, v = (rearrange(x, "b s ... -> (b s) ...") for x in [q, k, v])
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
|
|
|
|
output = flash_attn_varlen_func(q,
|
|
|
|
k,
|
|
|
|
v,
|
|
|
|
cu_seqlens_q=cu_seqlens,
|
|
|
|
cu_seqlens_k=cu_seqlens,
|
|
|
|
max_seqlen_q=max_seqlen,
|
|
|
|
max_seqlen_k=max_seqlen,
|
|
|
|
dropout_p=0,
|
|
|
|
causal=False)
|
|
|
|
|
|
|
|
context_layer = rearrange(output,
|
|
|
|
"(b s) ... -> b s ...",
|
|
|
|
b=batch_size)
|
2024-10-16 19:52:01 +08:00
|
|
|
elif self.attn_backend == _Backend.TORCH_SDPA:
|
2024-09-25 14:16:11 +08:00
|
|
|
seq_length = q.size(1)
|
2024-11-06 02:11:55 -05:00
|
|
|
q, k, v = (rearrange(x, "b s h d -> b h s d") for x in [q, k, v])
|
2024-09-25 14:16:11 +08:00
|
|
|
attention_mask = torch.zeros([1, seq_length, seq_length],
|
|
|
|
device=q.device,
|
|
|
|
dtype=torch.bool)
|
|
|
|
for i in range(1, len(cu_seqlens)):
|
|
|
|
attention_mask[..., cu_seqlens[i - 1]:cu_seqlens[i],
|
|
|
|
cu_seqlens[i - 1]:cu_seqlens[i]] = True
|
|
|
|
output = F.scaled_dot_product_attention(q,
|
|
|
|
k,
|
|
|
|
v,
|
|
|
|
attention_mask,
|
|
|
|
dropout_p=0.0)
|
|
|
|
context_layer = rearrange(output, "b h s d -> b s h d ")
|
2024-10-16 19:52:01 +08:00
|
|
|
elif self.attn_backend == _Backend.XFORMERS:
|
2024-09-12 00:31:19 +08:00
|
|
|
from xformers import ops as xops
|
|
|
|
from xformers.ops.fmha.attn_bias import BlockDiagonalMask
|
|
|
|
|
|
|
|
seqlens = (cu_seqlens[1:] - cu_seqlens[:-1]).tolist()
|
|
|
|
attn_bias = BlockDiagonalMask.from_seqlens(q_seqlen=seqlens,
|
|
|
|
kv_seqlen=None)
|
|
|
|
|
|
|
|
context_layer = xops.memory_efficient_attention_forward(
|
|
|
|
q, k, v, attn_bias=attn_bias, p=0, scale=None)
|
|
|
|
context_layer = rearrange(context_layer,
|
|
|
|
"b s h d -> s b (h d)").contiguous()
|
|
|
|
|
|
|
|
output, _ = self.proj(context_layer)
|
|
|
|
return output
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionBlock(nn.Module):
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
dim: int,
|
|
|
|
num_heads: int,
|
|
|
|
mlp_ratio: float,
|
|
|
|
act_layer: Type[nn.Module] = QuickGELU,
|
|
|
|
norm_layer: Type[nn.Module] = None,
|
|
|
|
quant_config: Optional[QuantizationConfig] = None,
|
2024-10-31 01:41:20 -04:00
|
|
|
prefix: str = "",
|
2024-09-12 00:31:19 +08:00
|
|
|
) -> None:
|
|
|
|
super().__init__()
|
|
|
|
if norm_layer is None:
|
|
|
|
norm_layer = partial(nn.LayerNorm, eps=1e-6)
|
|
|
|
self.norm1 = norm_layer(dim)
|
|
|
|
self.norm2 = norm_layer(dim)
|
|
|
|
mlp_hidden_dim = int(dim * mlp_ratio)
|
|
|
|
|
|
|
|
self.attn = Qwen2VisionAttention(embed_dim=dim,
|
|
|
|
num_heads=num_heads,
|
|
|
|
projection_size=dim,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.attn")
|
2024-09-12 00:31:19 +08:00
|
|
|
self.mlp = Qwen2VisionMLP(dim,
|
|
|
|
mlp_hidden_dim,
|
|
|
|
act_layer=act_layer,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.mlp")
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
def forward(self, x: torch.Tensor, cu_seqlens: torch.Tensor,
|
|
|
|
rotary_pos_emb: torch.Tensor) -> torch.Tensor:
|
|
|
|
x = x + self.attn(self.norm1(x),
|
|
|
|
cu_seqlens=cu_seqlens,
|
|
|
|
rotary_pos_emb=rotary_pos_emb)
|
|
|
|
x = x + self.mlp(self.norm2(x))
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionPatchEmbed(nn.Module):
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
patch_size: int = 14,
|
|
|
|
temporal_patch_size: int = 2,
|
|
|
|
in_chans: int = 3,
|
|
|
|
embed_dim: int = 1152,
|
|
|
|
) -> None:
|
|
|
|
super().__init__()
|
|
|
|
self.patch_size = patch_size
|
|
|
|
self.temporal_patch_size = temporal_patch_size
|
|
|
|
self.embed_dim = embed_dim
|
|
|
|
|
|
|
|
kernel_size = [temporal_patch_size, patch_size, patch_size]
|
|
|
|
self.proj = nn.Conv3d(in_chans,
|
|
|
|
embed_dim,
|
|
|
|
kernel_size=kernel_size,
|
|
|
|
stride=kernel_size,
|
|
|
|
bias=False)
|
|
|
|
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
L, C = x.shape
|
|
|
|
x = x.view(L, -1, self.temporal_patch_size, self.patch_size,
|
|
|
|
self.patch_size)
|
|
|
|
x = self.proj(x).view(L, self.embed_dim)
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionPatchMerger(nn.Module):
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
d_model: int,
|
|
|
|
context_dim: int,
|
|
|
|
norm_layer: Type[nn.Module] = None,
|
|
|
|
spatial_merge_size: int = 2,
|
|
|
|
quant_config: Optional[QuantizationConfig] = None,
|
2024-10-31 01:41:20 -04:00
|
|
|
prefix: str = "",
|
2024-09-12 00:31:19 +08:00
|
|
|
) -> None:
|
|
|
|
super().__init__()
|
|
|
|
self.hidden_size = context_dim * (spatial_merge_size**2)
|
|
|
|
if norm_layer is None:
|
|
|
|
norm_layer = partial(nn.LayerNorm, eps=1e-6)
|
|
|
|
self.ln_q = norm_layer(context_dim)
|
|
|
|
self.mlp = nn.ModuleList([
|
|
|
|
ColumnParallelLinear(self.hidden_size,
|
|
|
|
self.hidden_size,
|
|
|
|
bias=True,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.mlp.0"),
|
2024-09-12 00:31:19 +08:00
|
|
|
nn.GELU(),
|
|
|
|
RowParallelLinear(self.hidden_size,
|
|
|
|
d_model,
|
|
|
|
bias=True,
|
2024-10-31 01:41:20 -04:00
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.mlp.2"),
|
2024-09-12 00:31:19 +08:00
|
|
|
])
|
|
|
|
|
|
|
|
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
|
|
|
x = self.ln_q(x)
|
|
|
|
x = x.view(-1, self.hidden_size)
|
|
|
|
|
|
|
|
mlp_fc1, mlp_act, mlp_fc2 = self.mlp
|
|
|
|
x_parallel, _ = mlp_fc1(x)
|
|
|
|
x_parallel = mlp_act(x_parallel)
|
|
|
|
out, _ = mlp_fc2(x_parallel)
|
|
|
|
return out
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionRotaryEmbedding(nn.Module):
|
|
|
|
|
|
|
|
def __init__(self, dim: int, theta: float = 10000.0) -> None:
|
|
|
|
super().__init__()
|
|
|
|
self.dim = dim
|
|
|
|
self.theta = theta
|
|
|
|
inv_freq = 1.0 / (theta
|
|
|
|
**(torch.arange(0, dim, 2, dtype=torch.float) / dim))
|
|
|
|
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
|
|
|
self._seq_len_cached = 0
|
|
|
|
self._freqs_cached = None
|
|
|
|
|
|
|
|
def update_freqs_cache(self, seqlen: int) -> None:
|
|
|
|
if seqlen > self._seq_len_cached:
|
|
|
|
seqlen *= 2
|
|
|
|
self._seq_len_cached = seqlen
|
|
|
|
self.inv_freq = 1.0 / (self.theta**(torch.arange(
|
|
|
|
0, self.dim, 2, dtype=torch.float, device=self.inv_freq.device)
|
|
|
|
/ self.dim))
|
|
|
|
seq = torch.arange(seqlen,
|
|
|
|
device=self.inv_freq.device,
|
|
|
|
dtype=self.inv_freq.dtype)
|
|
|
|
freqs = torch.outer(seq, self.inv_freq)
|
|
|
|
self._freqs_cached = freqs
|
|
|
|
|
|
|
|
def forward(self, seqlen: int) -> torch.Tensor:
|
|
|
|
self.update_freqs_cache(seqlen)
|
|
|
|
return self._freqs_cached[:seqlen]
|
|
|
|
|
|
|
|
|
|
|
|
class Qwen2VisionTransformer(nn.Module):
|
|
|
|
|
|
|
|
def __init__(
|
|
|
|
self,
|
|
|
|
vision_config: Qwen2VLVisionConfig,
|
|
|
|
norm_eps: float = 1e-6,
|
|
|
|
quant_config: Optional[QuantizationConfig] = None,
|
2024-10-31 01:41:20 -04:00
|
|
|
prefix: str = "",
|
2024-09-12 00:31:19 +08:00
|
|
|
) -> None:
|
|
|
|
super().__init__()
|
|
|
|
|
|
|
|
patch_size: int = vision_config.patch_size
|
|
|
|
temporal_patch_size: int = vision_config.temporal_patch_size
|
|
|
|
spatial_merge_size: int = vision_config.spatial_merge_size
|
|
|
|
in_chans: int = vision_config.in_chans
|
|
|
|
hidden_size: int = vision_config.hidden_size
|
|
|
|
embed_dim: int = vision_config.embed_dim
|
|
|
|
depth: int = vision_config.depth
|
|
|
|
num_heads: int = vision_config.num_heads
|
|
|
|
mlp_ratio: float = vision_config.mlp_ratio
|
|
|
|
|
|
|
|
self.spatial_merge_size = spatial_merge_size
|
|
|
|
|
|
|
|
self.patch_embed = Qwen2VisionPatchEmbed(
|
|
|
|
patch_size=patch_size,
|
|
|
|
temporal_patch_size=temporal_patch_size,
|
|
|
|
in_chans=in_chans,
|
|
|
|
embed_dim=embed_dim,
|
|
|
|
)
|
|
|
|
|
|
|
|
norm_layer = partial(nn.LayerNorm, eps=norm_eps)
|
|
|
|
head_dim = embed_dim // num_heads
|
|
|
|
self.rotary_pos_emb = Qwen2VisionRotaryEmbedding(head_dim // 2)
|
|
|
|
|
|
|
|
self.blocks = nn.ModuleList([
|
2024-10-31 01:41:20 -04:00
|
|
|
Qwen2VisionBlock(dim=embed_dim,
|
|
|
|
num_heads=num_heads,
|
|
|
|
mlp_ratio=mlp_ratio,
|
|
|
|
norm_layer=norm_layer,
|
|
|
|
quant_config=quant_config,
|
|
|
|
prefix=f"{prefix}.blocks.{layer_idx}")
|
|
|
|
for layer_idx in range(depth)
|
2024-09-12 00:31:19 +08:00
|
|
|
])
|
|
|
|
self.merger = Qwen2VisionPatchMerger(
|
|
|
|
d_model=hidden_size,
|
|
|
|
context_dim=embed_dim,
|
|
|
|
norm_layer=norm_layer,
|
|
|
|
quant_config=quant_config,
|
2024-10-31 01:41:20 -04:00
|
|
|
prefix=f"{prefix}.merger",
|
2024-09-12 00:31:19 +08:00
|
|
|
)
|
|
|
|
|
|
|
|
@property
|
|
|
|
def dtype(self) -> torch.dtype:
|
2024-10-31 01:41:20 -04:00
|
|
|
return self.patch_embed.proj.weight.dtype
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
@property
|
|
|
|
def device(self) -> torch.device:
|
2024-10-31 01:41:20 -04:00
|
|
|
return self.patch_embed.proj.weight.device
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
def rot_pos_emb(self, grid_thw: torch.Tensor) -> torch.Tensor:
|
|
|
|
pos_ids = []
|
|
|
|
for t, h, w in grid_thw:
|
|
|
|
hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
|
|
|
|
wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
|
|
|
|
hpos_ids = hpos_ids.reshape(
|
|
|
|
h // self.spatial_merge_size,
|
|
|
|
self.spatial_merge_size,
|
|
|
|
w // self.spatial_merge_size,
|
|
|
|
self.spatial_merge_size,
|
|
|
|
).permute(0, 2, 1, 3).flatten()
|
|
|
|
wpos_ids = wpos_ids.reshape(
|
|
|
|
h // self.spatial_merge_size,
|
|
|
|
self.spatial_merge_size,
|
|
|
|
w // self.spatial_merge_size,
|
|
|
|
self.spatial_merge_size,
|
|
|
|
).permute(0, 2, 1, 3).flatten()
|
|
|
|
pos_ids.append(
|
|
|
|
torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
|
|
|
|
pos_ids = torch.cat(pos_ids, dim=0)
|
|
|
|
max_grid_size = grid_thw[:, 1:].max()
|
|
|
|
rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
|
|
|
|
rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
|
|
|
|
return rotary_pos_emb
|
|
|
|
|
|
|
|
def forward(
|
|
|
|
self,
|
|
|
|
x: torch.Tensor,
|
|
|
|
grid_thw: torch.Tensor,
|
|
|
|
) -> torch.Tensor:
|
|
|
|
# patchify
|
|
|
|
x = x.to(device=self.device, dtype=self.dtype)
|
|
|
|
x = self.patch_embed(x)
|
|
|
|
|
|
|
|
# compute position embedding
|
|
|
|
rotary_pos_emb = self.rot_pos_emb(grid_thw)
|
|
|
|
|
|
|
|
# compute cu_seqlens
|
|
|
|
cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2],
|
|
|
|
grid_thw[:, 0]).cumsum(
|
|
|
|
dim=0, dtype=torch.int32)
|
|
|
|
cu_seqlens = F.pad(cu_seqlens, (1, 0), "constant", 0)
|
|
|
|
|
|
|
|
# transformers
|
|
|
|
x = x.unsqueeze(1)
|
|
|
|
for blk in self.blocks:
|
|
|
|
x = blk(x, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb)
|
|
|
|
|
|
|
|
# adapter
|
|
|
|
x = self.merger(x)
|
|
|
|
return x
|
|
|
|
|
|
|
|
|
|
|
|
# === Vision input helpers === #
|
|
|
|
|
|
|
|
|
2024-11-07 18:50:44 +08:00
|
|
|
def get_mm_processor_kwargs(
|
|
|
|
min_pixels: Optional[int] = None,
|
|
|
|
max_pixels: Optional[int] = None) -> Dict[str, int]:
|
|
|
|
mm_processor_kwargs = {}
|
|
|
|
if min_pixels:
|
|
|
|
mm_processor_kwargs["min_pixels"] = min_pixels
|
|
|
|
if max_pixels:
|
|
|
|
mm_processor_kwargs["max_pixels"] = max_pixels
|
|
|
|
return mm_processor_kwargs
|
|
|
|
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
def mm_input_mapper_for_qwen2_vl(
|
|
|
|
ctx: InputContext,
|
|
|
|
data: MultiModalData[object],
|
|
|
|
data_type_key: str,
|
2024-10-23 08:05:18 -06:00
|
|
|
*,
|
|
|
|
min_pixels: Optional[int] = None,
|
|
|
|
max_pixels: Optional[int] = None,
|
2024-11-09 11:31:02 +08:00
|
|
|
) -> MultiModalKwargs:
|
2024-09-12 00:31:19 +08:00
|
|
|
"""Input mapper for Qwen2-VL."""
|
2024-09-30 11:16:10 +08:00
|
|
|
if data_type_key == "image" and isinstance(data, dict):
|
2024-11-09 11:31:02 +08:00
|
|
|
return MultiModalKwargs({
|
2024-09-30 11:16:10 +08:00
|
|
|
"image_embeds": data.get("image_embeds"),
|
|
|
|
"image_grid_thw": data.get("image_grid_thw"),
|
|
|
|
})
|
2024-11-13 15:07:22 +08:00
|
|
|
if data_type_key == "video" and isinstance(data, dict):
|
|
|
|
return MultiModalKwargs({
|
|
|
|
"video_embeds": data.get("video_embeds"),
|
|
|
|
"video_grid_thw": data.get("video_grid_thw"),
|
|
|
|
})
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
model_config = ctx.model_config
|
2024-10-23 08:05:18 -06:00
|
|
|
# Handle mm processor kwargs; we pass these at creation time
|
|
|
|
# because preprocess() in transformers doesn't expose them
|
2024-11-07 18:50:44 +08:00
|
|
|
mm_processor_kwargs = get_mm_processor_kwargs(min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
2024-09-12 00:31:19 +08:00
|
|
|
image_processor = cached_get_image_processor(
|
2024-10-23 08:05:18 -06:00
|
|
|
model_config.model,
|
|
|
|
trust_remote_code=model_config.trust_remote_code,
|
|
|
|
**mm_processor_kwargs,
|
|
|
|
)
|
2024-09-12 00:31:19 +08:00
|
|
|
if image_processor is None:
|
|
|
|
raise RuntimeError("No HuggingFace processor is available "
|
|
|
|
"to process the image object")
|
|
|
|
|
|
|
|
images = None
|
|
|
|
videos = None
|
|
|
|
if data_type_key == "image":
|
|
|
|
images = data
|
|
|
|
else:
|
|
|
|
assert data_type_key == "video"
|
|
|
|
videos = data
|
|
|
|
|
|
|
|
try:
|
|
|
|
batch_data = image_processor \
|
|
|
|
.preprocess(images=images, videos=videos, return_tensors="pt") \
|
|
|
|
.data
|
|
|
|
except Exception:
|
|
|
|
logger.error("Failed to process image (%s)", data)
|
|
|
|
raise
|
|
|
|
|
2024-11-09 11:31:02 +08:00
|
|
|
return MultiModalKwargs(batch_data)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
|
|
|
|
image_input_mapper_for_qwen2_vl = partial(mm_input_mapper_for_qwen2_vl,
|
|
|
|
data_type_key="image")
|
|
|
|
video_input_mapper_for_qwen2_vl = partial(mm_input_mapper_for_qwen2_vl,
|
|
|
|
data_type_key="video")
|
|
|
|
|
|
|
|
|
|
|
|
def _get_vision_info(
|
|
|
|
image_processor,
|
|
|
|
height: int,
|
|
|
|
width: int,
|
|
|
|
min_pixels: int,
|
|
|
|
max_pixels: int,
|
|
|
|
do_resize: bool = True,
|
|
|
|
data_type_key: str = "image",
|
|
|
|
mm_count: int = 1,
|
|
|
|
):
|
|
|
|
"""Get information (resized height / width and number of vision tokens)
|
|
|
|
of input image / video frame."""
|
|
|
|
|
|
|
|
if do_resize:
|
|
|
|
resized_height, resized_width = smart_resize(
|
|
|
|
height=height,
|
|
|
|
width=width,
|
|
|
|
factor=image_processor.patch_size * image_processor.merge_size,
|
|
|
|
min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels,
|
|
|
|
)
|
|
|
|
else:
|
|
|
|
resized_height, resized_width = height, width
|
|
|
|
|
|
|
|
if data_type_key == "image":
|
|
|
|
grid_t = mm_count
|
|
|
|
else:
|
|
|
|
assert data_type_key == "video"
|
|
|
|
grid_t = max(mm_count // image_processor.temporal_patch_size, 1)
|
|
|
|
|
|
|
|
grid_h = resized_height // image_processor.patch_size
|
|
|
|
grid_w = resized_width // image_processor.patch_size
|
|
|
|
vision_tokens = grid_t * grid_h * grid_w
|
|
|
|
llm_num_vision_tokens = (vision_tokens // image_processor.merge_size //
|
|
|
|
image_processor.merge_size)
|
|
|
|
|
|
|
|
return resized_height, resized_width, llm_num_vision_tokens
|
|
|
|
|
|
|
|
|
|
|
|
def _get_max_image_info(
|
|
|
|
image_processor,
|
|
|
|
data_type_key: str = "image",
|
|
|
|
mm_count: int = 1,
|
2024-10-23 08:05:18 -06:00
|
|
|
min_pixels: Optional[int] = None,
|
|
|
|
max_pixels: Optional[int] = None,
|
2024-09-12 00:31:19 +08:00
|
|
|
):
|
2024-10-23 08:05:18 -06:00
|
|
|
# Limit min / max pixels unless they're explicitly provided
|
|
|
|
if min_pixels is None:
|
|
|
|
min_pixels = max(image_processor.min_pixels, 28 * 28)
|
|
|
|
if max_pixels is None:
|
|
|
|
max_pixels = min(image_processor.max_pixels, 1280 * 28 * 28)
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
return _get_vision_info(
|
|
|
|
image_processor,
|
|
|
|
height=9999999,
|
|
|
|
width=9999999,
|
2024-10-23 08:05:18 -06:00
|
|
|
min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels,
|
2024-09-12 00:31:19 +08:00
|
|
|
data_type_key=data_type_key,
|
|
|
|
mm_count=mm_count,
|
|
|
|
)
|
|
|
|
|
|
|
|
|
2024-10-23 08:05:18 -06:00
|
|
|
def get_max_qwen2_vl_mm_tokens(ctx: InputContext,
|
|
|
|
data_type_key: str,
|
|
|
|
*,
|
|
|
|
min_pixels=None,
|
|
|
|
max_pixels=None) -> int:
|
2024-11-07 18:50:44 +08:00
|
|
|
mm_processor_kwargs = get_mm_processor_kwargs(min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
|
|
|
image_processor = cached_get_image_processor(ctx.model_config.model,
|
|
|
|
**mm_processor_kwargs)
|
2024-09-12 00:31:19 +08:00
|
|
|
max_resized_height, max_resized_width, max_llm_image_tokens = \
|
|
|
|
_get_max_image_info(image_processor, data_type_key=data_type_key,
|
2024-10-23 08:05:18 -06:00
|
|
|
mm_count=1, min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
2024-09-12 00:31:19 +08:00
|
|
|
return max_llm_image_tokens
|
|
|
|
|
|
|
|
|
|
|
|
get_max_qwen2_vl_image_tokens = partial(get_max_qwen2_vl_mm_tokens,
|
|
|
|
data_type_key="image")
|
|
|
|
get_max_qwen2_vl_video_tokens = partial(get_max_qwen2_vl_mm_tokens,
|
|
|
|
data_type_key="video")
|
|
|
|
|
|
|
|
|
|
|
|
def dummy_data_for_qwen2_vl(
|
2024-10-23 08:05:18 -06:00
|
|
|
ctx: InputContext,
|
|
|
|
seq_len: int,
|
|
|
|
mm_counts: Mapping[str, int],
|
|
|
|
*,
|
|
|
|
min_pixels: Optional[int] = None,
|
|
|
|
max_pixels: Optional[int] = None
|
2024-09-12 00:31:19 +08:00
|
|
|
) -> Tuple[SequenceData, Optional[MultiModalDataDict]]:
|
2024-11-07 18:50:44 +08:00
|
|
|
mm_processor_kwargs = get_mm_processor_kwargs(min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
|
|
|
image_processor = cached_get_image_processor(ctx.model_config.model,
|
|
|
|
**mm_processor_kwargs)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
num_images = mm_counts["image"]
|
|
|
|
max_resized_height, max_resized_width, max_llm_image_tokens = \
|
|
|
|
_get_max_image_info(image_processor, data_type_key="image",
|
2024-10-23 08:05:18 -06:00
|
|
|
mm_count=num_images, min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
2024-09-12 00:31:19 +08:00
|
|
|
if seq_len - max_llm_image_tokens - 2 < 0:
|
|
|
|
raise RuntimeError(
|
|
|
|
f"Qwen2-VL cannot process {num_images} images in a prompt, "
|
|
|
|
"please increase max_model_len or reduce image limit by "
|
|
|
|
"--limit-mm-per-prompt.")
|
|
|
|
|
|
|
|
# Check video counts.
|
|
|
|
num_videos = mm_counts["video"]
|
|
|
|
max_resized_height, max_resized_width, max_llm_video_tokens = \
|
|
|
|
_get_max_image_info(image_processor, data_type_key="video",
|
2024-10-23 08:05:18 -06:00
|
|
|
mm_count=num_videos, min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
2024-09-12 00:31:19 +08:00
|
|
|
if seq_len - max_llm_video_tokens - 2 < 0:
|
|
|
|
raise RuntimeError(
|
2024-10-23 08:05:18 -06:00
|
|
|
f"Qwen2-VL cannot process {num_videos} videos in a prompt, "
|
2024-09-12 00:31:19 +08:00
|
|
|
"please increase max_model_len or reduce video limit by "
|
|
|
|
"--limit-mm-per-prompt.")
|
|
|
|
|
|
|
|
hf_config = ctx.get_hf_config(Qwen2VLConfig)
|
2024-09-21 14:28:56 +08:00
|
|
|
|
2024-10-16 18:49:37 +08:00
|
|
|
dummy_seqdata = SequenceData.from_prompt_token_counts(
|
2024-09-21 14:28:56 +08:00
|
|
|
(hf_config.vision_start_token_id, 1),
|
|
|
|
(hf_config.image_token_id, max_llm_image_tokens),
|
|
|
|
(hf_config.vision_end_token_id, 1),
|
|
|
|
(0, seq_len - max_llm_image_tokens - 2),
|
|
|
|
)
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
dummy_image = Image.new("RGB", (max_resized_width, max_resized_height),
|
|
|
|
color=0)
|
|
|
|
|
2024-11-01 16:21:10 -07:00
|
|
|
return DummyData(dummy_seqdata, {
|
|
|
|
"image":
|
|
|
|
dummy_image if num_images == 1 else [dummy_image] * num_images
|
|
|
|
})
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
|
|
|
|
def _get_llm_num_vision_tokens(
|
|
|
|
mm_inputs: list,
|
|
|
|
data_type_key: str,
|
|
|
|
image_processor,
|
2024-10-23 08:05:18 -06:00
|
|
|
min_pixels: int,
|
|
|
|
max_pixels: int,
|
2024-09-12 00:31:19 +08:00
|
|
|
):
|
|
|
|
"""Get number of vision tokens of multimodal inputs.
|
|
|
|
|
|
|
|
This method is derived from `transformers.models.qwen2_vl.
|
|
|
|
image_processing_qwen2_vl.Qwen2VLImageProcessor._preprocess`.
|
|
|
|
"""
|
|
|
|
image = to_numpy_array(mm_inputs[0])
|
|
|
|
input_data_format = infer_channel_dimension_format(image)
|
|
|
|
height, width = get_image_size(image, channel_dim=input_data_format)
|
2024-10-23 08:05:18 -06:00
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
_, _, llm_num_vision_tokens = _get_vision_info(
|
|
|
|
image_processor,
|
|
|
|
height=height,
|
|
|
|
width=width,
|
2024-10-23 08:05:18 -06:00
|
|
|
min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels,
|
2024-09-12 00:31:19 +08:00
|
|
|
do_resize=image_processor.do_resize,
|
|
|
|
data_type_key=data_type_key,
|
|
|
|
mm_count=len(mm_inputs),
|
|
|
|
)
|
|
|
|
return llm_num_vision_tokens
|
|
|
|
|
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
def _expand_pad_tokens(inputs: list, token_id: int, make_batched_fn: Callable,
|
|
|
|
data_type_key: str, image_processor: Any,
|
2024-10-23 08:05:18 -06:00
|
|
|
prompt_token_ids: List[int], min_pixels: Optional[int],
|
|
|
|
max_pixels: Optional[int]) -> List[int]:
|
2024-09-30 11:16:10 +08:00
|
|
|
"""
|
|
|
|
Expand pad tokens for multi-modal inputs (e.g., images or videos).
|
|
|
|
|
|
|
|
Args:
|
|
|
|
inputs (list): The multi-modal inputs (e.g., images or videos).
|
|
|
|
token_id (int): The token ID used to represent the multi-modal input.
|
|
|
|
make_batched_fn (Callable): A function to batch the inputs.
|
|
|
|
data_type_key (str): The type of the multi-modal input.
|
|
|
|
image_processor (Any): The image processor used to process the inputs.
|
|
|
|
prompt_token_ids (List[int]): The list of token IDs in the prompt.
|
2024-10-23 08:05:18 -06:00
|
|
|
min_pixels (int): min pixels to used for img processing
|
|
|
|
max_pixels (int): max pixels to be used for img processing
|
2024-09-30 11:16:10 +08:00
|
|
|
|
|
|
|
Returns:
|
|
|
|
List[int]: The list of token IDs for the multi-modal inputs.
|
|
|
|
"""
|
|
|
|
indices = [
|
|
|
|
idx for idx, token in enumerate(prompt_token_ids) if token == token_id
|
|
|
|
]
|
|
|
|
inputs = make_batched_fn(inputs)
|
|
|
|
assert len(indices) == len(inputs)
|
|
|
|
|
|
|
|
prompt_token_ids_with_data = []
|
|
|
|
for cnt, data in enumerate(inputs):
|
|
|
|
num_tokens = _get_llm_num_vision_tokens(
|
|
|
|
[data] if data_type_key == "image" else data,
|
|
|
|
data_type_key=data_type_key,
|
|
|
|
image_processor=image_processor,
|
2024-10-23 08:05:18 -06:00
|
|
|
min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels,
|
2024-09-30 11:16:10 +08:00
|
|
|
)
|
|
|
|
if cnt == 0:
|
|
|
|
end_idx = indices[cnt]
|
|
|
|
non_data_tokens = prompt_token_ids[:end_idx]
|
|
|
|
else:
|
|
|
|
non_data_tokens = prompt_token_ids[indices[cnt - 1] +
|
|
|
|
1:indices[cnt]]
|
|
|
|
prompt_token_ids_with_data.extend(non_data_tokens)
|
|
|
|
prompt_token_ids_with_data.extend(token_id for _ in range(num_tokens))
|
|
|
|
prompt_token_ids_with_data.extend(prompt_token_ids[indices[-1] + 1:])
|
|
|
|
return prompt_token_ids_with_data
|
|
|
|
|
|
|
|
|
2024-10-16 18:49:37 +08:00
|
|
|
def input_processor_for_qwen2_vl(
|
|
|
|
ctx: InputContext,
|
|
|
|
inputs: DecoderOnlyInputs,
|
2024-10-23 08:05:18 -06:00
|
|
|
*,
|
|
|
|
min_pixels: Optional[int] = None,
|
|
|
|
max_pixels: Optional[int] = None,
|
2024-10-16 18:49:37 +08:00
|
|
|
) -> DecoderOnlyInputs:
|
2024-10-24 14:12:05 +08:00
|
|
|
multi_modal_data = inputs.get("multi_modal_data")
|
2024-09-12 00:31:19 +08:00
|
|
|
if multi_modal_data is None:
|
2024-10-16 18:49:37 +08:00
|
|
|
return inputs
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
image_inputs = multi_modal_data.get("image", None)
|
|
|
|
video_inputs = multi_modal_data.get("video", None)
|
|
|
|
|
|
|
|
processor = cached_get_processor(ctx.model_config.model)
|
|
|
|
image_processor = processor.image_processor
|
2024-10-23 08:05:18 -06:00
|
|
|
# Apply processor kwarg overrides for image processor options
|
|
|
|
min_pixels = min_pixels if min_pixels else image_processor.min_pixels
|
|
|
|
max_pixels = max_pixels if max_pixels else image_processor.max_pixels
|
|
|
|
|
2024-10-24 14:12:05 +08:00
|
|
|
model_config = ctx.model_config
|
2024-09-12 00:31:19 +08:00
|
|
|
hf_config = ctx.get_hf_config(Qwen2VLConfig)
|
|
|
|
|
|
|
|
# To avoid redundant processing of vision objects (resize, rescale, etc.),
|
|
|
|
# we extract code of calculating number of vision tokens from
|
|
|
|
# `transformers.models.qwen2_vl.processing_qwen2_vl.Qwen2VLProcessor`.
|
|
|
|
#
|
|
|
|
# The following code is equivalent to:
|
2024-10-16 18:49:37 +08:00
|
|
|
# prompt = inputs["prompt"]
|
2024-09-12 00:31:19 +08:00
|
|
|
# inputs = processor(text=[prompt],
|
|
|
|
# images=image_inputs,
|
|
|
|
# videos=video_inputs,
|
|
|
|
# padding=True,
|
|
|
|
# return_tensors="pt")
|
|
|
|
# prompt_token_ids = inputs["input_ids"][0].tolist()
|
|
|
|
|
2024-10-24 14:12:05 +08:00
|
|
|
tokenizer = cached_get_tokenizer(
|
|
|
|
model_config.tokenizer,
|
|
|
|
trust_remote_code=model_config.trust_remote_code)
|
|
|
|
|
|
|
|
prompt_token_ids = inputs["prompt_token_ids"]
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
# Expand image pad tokens.
|
2024-09-30 11:16:10 +08:00
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
if image_inputs is not None:
|
2024-09-30 11:16:10 +08:00
|
|
|
if isinstance(image_inputs, dict):
|
|
|
|
prompt_token_ids_with_image = []
|
|
|
|
image_indices = [
|
|
|
|
idx for idx, token in enumerate(prompt_token_ids)
|
|
|
|
if token == hf_config.image_token_id
|
|
|
|
]
|
2024-11-13 15:07:22 +08:00
|
|
|
|
|
|
|
# ensure all image tokens have grid_thw
|
|
|
|
assert \
|
|
|
|
len(image_indices) == image_inputs["image_grid_thw"].size(0), \
|
|
|
|
"image token num does not match image_grid_thw.shape"
|
|
|
|
|
|
|
|
image_counter = 0
|
|
|
|
pad_token_counter = 0
|
2024-09-30 11:16:10 +08:00
|
|
|
for idx, token in enumerate(prompt_token_ids):
|
|
|
|
if idx in image_indices:
|
2024-11-13 15:07:22 +08:00
|
|
|
grid_thw = image_inputs["image_grid_thw"][image_counter]
|
|
|
|
grid_t, grid_h, grid_w = grid_thw
|
|
|
|
num_pad_tokens = (grid_t * grid_h * grid_w //
|
|
|
|
image_processor.merge_size //
|
|
|
|
image_processor.merge_size)
|
2024-09-30 11:16:10 +08:00
|
|
|
prompt_token_ids_with_image.extend([token] *
|
|
|
|
num_pad_tokens)
|
2024-11-13 15:07:22 +08:00
|
|
|
image_counter += 1
|
|
|
|
pad_token_counter += num_pad_tokens
|
2024-09-30 11:16:10 +08:00
|
|
|
else:
|
|
|
|
prompt_token_ids_with_image.append(token)
|
2024-11-13 15:07:22 +08:00
|
|
|
|
|
|
|
# ensure all embeddings are used
|
|
|
|
assert \
|
|
|
|
pad_token_counter == image_inputs["image_embeds"].size(0), \
|
|
|
|
"image_embeds.shape does not match image_grid_thw"
|
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
prompt_token_ids = prompt_token_ids_with_image
|
|
|
|
else:
|
|
|
|
prompt_token_ids = _expand_pad_tokens(image_inputs,
|
|
|
|
hf_config.image_token_id,
|
2024-10-23 08:05:18 -06:00
|
|
|
make_batched_images,
|
|
|
|
"image",
|
2024-09-30 11:16:10 +08:00
|
|
|
image_processor,
|
2024-10-23 08:05:18 -06:00
|
|
|
prompt_token_ids,
|
|
|
|
min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
2024-09-30 11:16:10 +08:00
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
if video_inputs is not None:
|
2024-11-13 15:07:22 +08:00
|
|
|
if isinstance(video_inputs, dict):
|
|
|
|
prompt_token_ids_with_video = []
|
|
|
|
video_indices = [
|
|
|
|
idx for idx, token in enumerate(prompt_token_ids)
|
|
|
|
if token == hf_config.video_token_id
|
|
|
|
]
|
|
|
|
|
|
|
|
# ensure all video tokens have grid_thw
|
|
|
|
assert \
|
|
|
|
len(video_indices) == video_inputs["video_grid_thw"].size(0), \
|
|
|
|
"video token num does not match video_grid_thw.shape"
|
|
|
|
|
|
|
|
video_counter = 0
|
|
|
|
pad_token_counter = 0
|
|
|
|
for idx, token in enumerate(prompt_token_ids):
|
|
|
|
if idx in video_indices:
|
|
|
|
grid_thw = video_inputs["video_grid_thw"][video_counter]
|
|
|
|
grid_t, grid_h, grid_w = grid_thw
|
|
|
|
num_pad_tokens = (grid_t * grid_h * grid_w //
|
|
|
|
image_processor.merge_size //
|
|
|
|
image_processor.merge_size)
|
|
|
|
prompt_token_ids_with_video.extend([token] *
|
|
|
|
num_pad_tokens)
|
|
|
|
video_counter += 1
|
|
|
|
pad_token_counter += num_pad_tokens
|
|
|
|
else:
|
|
|
|
prompt_token_ids_with_video.append(token)
|
|
|
|
|
|
|
|
# ensure all embeddings are used
|
|
|
|
assert \
|
|
|
|
pad_token_counter == video_inputs["video_embeds"].size(0), \
|
|
|
|
"video_embeds.shape does not match video_grid_thw"
|
|
|
|
|
|
|
|
prompt_token_ids = prompt_token_ids_with_video
|
|
|
|
else:
|
|
|
|
prompt_token_ids = _expand_pad_tokens(video_inputs,
|
|
|
|
hf_config.video_token_id,
|
|
|
|
make_batched_videos,
|
|
|
|
"video",
|
|
|
|
image_processor,
|
|
|
|
prompt_token_ids,
|
|
|
|
min_pixels=min_pixels,
|
|
|
|
max_pixels=max_pixels)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-10-24 14:12:05 +08:00
|
|
|
prompt = inputs.get("prompt")
|
|
|
|
if prompt is None:
|
|
|
|
prompt = tokenizer.decode(prompt_token_ids)
|
|
|
|
|
2024-10-16 18:49:37 +08:00
|
|
|
return token_inputs(
|
2024-09-12 00:31:19 +08:00
|
|
|
prompt_token_ids=prompt_token_ids,
|
2024-10-24 14:12:05 +08:00
|
|
|
prompt=prompt,
|
2024-09-12 00:31:19 +08:00
|
|
|
multi_modal_data=multi_modal_data,
|
|
|
|
)
|
|
|
|
|
|
|
|
|
|
|
|
@MULTIMODAL_REGISTRY.register_image_input_mapper(
|
|
|
|
image_input_mapper_for_qwen2_vl)
|
|
|
|
@MULTIMODAL_REGISTRY.register_input_mapper("video",
|
|
|
|
video_input_mapper_for_qwen2_vl)
|
|
|
|
@MULTIMODAL_REGISTRY.register_max_image_tokens(get_max_qwen2_vl_image_tokens)
|
|
|
|
@MULTIMODAL_REGISTRY.register_max_multimodal_tokens(
|
|
|
|
"video", get_max_qwen2_vl_video_tokens)
|
|
|
|
@INPUT_REGISTRY.register_dummy_data(dummy_data_for_qwen2_vl)
|
|
|
|
@INPUT_REGISTRY.register_input_processor(input_processor_for_qwen2_vl)
|
2024-10-03 19:56:58 -07:00
|
|
|
class Qwen2VLForConditionalGeneration(nn.Module, SupportsMultiModal,
|
2024-11-06 22:13:15 +08:00
|
|
|
SupportsLoRA, SupportsPP):
|
|
|
|
packed_modules_mapping = {
|
|
|
|
"qkv_proj": [
|
|
|
|
"q_proj",
|
|
|
|
"k_proj",
|
|
|
|
"v_proj",
|
|
|
|
],
|
|
|
|
"gate_up_proj": [
|
|
|
|
"gate_proj",
|
|
|
|
"up_proj",
|
|
|
|
],
|
|
|
|
}
|
|
|
|
|
|
|
|
# LoRA specific attributes
|
|
|
|
# TODO Support LoRA for the visual encoder in the future.
|
|
|
|
supported_lora_modules = [
|
|
|
|
"qkv_proj",
|
|
|
|
"o_proj",
|
|
|
|
"gate_up_proj",
|
|
|
|
"down_proj",
|
|
|
|
]
|
|
|
|
embedding_modules = {}
|
|
|
|
embedding_padding_modules = []
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-11-10 22:41:46 -08:00
|
|
|
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
|
2024-09-12 00:31:19 +08:00
|
|
|
super().__init__()
|
2024-11-08 22:17:28 -08:00
|
|
|
config = vllm_config.model_config.hf_config
|
|
|
|
cache_config = vllm_config.cache_config
|
|
|
|
quant_config = vllm_config.quant_config
|
2024-11-13 02:28:13 -06:00
|
|
|
pooler_config = vllm_config.model_config.pooler_config
|
2024-11-08 22:17:28 -08:00
|
|
|
multimodal_config = vllm_config.model_config.multimodal_config
|
2024-09-12 00:31:19 +08:00
|
|
|
assert not cache_config.enable_prefix_caching, \
|
|
|
|
"Qwen2-VL currently does not support prefix caching"
|
|
|
|
|
|
|
|
self.config = config
|
|
|
|
self.multimodal_config = multimodal_config
|
|
|
|
|
|
|
|
self.visual = Qwen2VisionTransformer(
|
|
|
|
config.vision_config,
|
|
|
|
norm_eps=getattr(config, "rms_norm_eps", 1e-6),
|
2024-11-08 22:36:46 -05:00
|
|
|
quant_config=self._maybe_ignore_quant_config(quant_config),
|
2024-11-10 22:41:46 -08:00
|
|
|
prefix=maybe_prefix(prefix, "visual"),
|
2024-09-12 00:31:19 +08:00
|
|
|
)
|
|
|
|
|
2024-11-10 22:41:46 -08:00
|
|
|
self.model = Qwen2Model(vllm_config=vllm_config,
|
|
|
|
prefix=maybe_prefix(prefix, "model"))
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-09-23 21:46:59 +08:00
|
|
|
if get_pp_group().is_last_rank:
|
|
|
|
if config.tie_word_embeddings:
|
|
|
|
self.lm_head = self.model.embed_tokens
|
|
|
|
else:
|
|
|
|
self.lm_head = ParallelLMHead(config.vocab_size,
|
|
|
|
config.hidden_size,
|
2024-10-29 19:02:59 -04:00
|
|
|
quant_config=quant_config,
|
2024-11-10 22:41:46 -08:00
|
|
|
prefix=maybe_prefix(
|
|
|
|
prefix, "lm_head"))
|
2024-09-12 00:31:19 +08:00
|
|
|
else:
|
2024-09-23 21:46:59 +08:00
|
|
|
self.lm_head = PPMissingLayer()
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
self.logits_processor = LogitsProcessor(config.vocab_size)
|
2024-11-06 12:57:35 -07:00
|
|
|
self.sampler = get_sampler()
|
2024-11-13 02:28:13 -06:00
|
|
|
self._pooler = Pooler.from_config_with_defaults(
|
|
|
|
pooler_config,
|
|
|
|
pooling_type=PoolingType.LAST,
|
|
|
|
normalize=True,
|
|
|
|
softmax=False)
|
2024-09-23 21:46:59 +08:00
|
|
|
self.make_empty_intermediate_tensors = (
|
|
|
|
make_empty_intermediate_tensors_factory(
|
|
|
|
["hidden_states", "residual"], config.hidden_size))
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-11-08 22:36:46 -05:00
|
|
|
def _maybe_ignore_quant_config(self, quant_config: QuantizationConfig):
|
|
|
|
# GPTQ configs do not have a list of ignored modules, however AutoGPTQ
|
|
|
|
# seems to avoid vision encoder sections for some models.
|
|
|
|
# See: https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct-GPTQ-Int4
|
|
|
|
if isinstance(quant_config, (GPTQConfig, GPTQMarlinConfig)):
|
|
|
|
return None
|
|
|
|
return quant_config
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
def _validate_and_reshape_mm_tensor(self,
|
|
|
|
mm_input: Union[torch.Tensor,
|
|
|
|
List[torch.Tensor]],
|
|
|
|
name: str) -> torch.Tensor:
|
|
|
|
if not isinstance(mm_input, (torch.Tensor, list)):
|
|
|
|
raise ValueError(f"Incorrect type of {name}. "
|
|
|
|
f"Got type: {type(mm_input)}")
|
|
|
|
if isinstance(mm_input, torch.Tensor):
|
|
|
|
if mm_input.ndim == 2:
|
|
|
|
return mm_input
|
|
|
|
if mm_input.ndim != 3:
|
|
|
|
raise ValueError(f"{name} should be 2D or batched 3D tensor. "
|
|
|
|
f"Got ndim: {mm_input.ndim}")
|
|
|
|
return torch.concat(list(mm_input))
|
|
|
|
else:
|
|
|
|
return torch.concat(mm_input)
|
|
|
|
|
|
|
|
def _parse_and_validate_image_input(
|
|
|
|
self, **kwargs: object) -> Optional[Qwen2VLImageInputs]:
|
|
|
|
pixel_values = kwargs.pop("pixel_values", None)
|
2024-09-30 11:16:10 +08:00
|
|
|
image_embeds = kwargs.pop("image_embeds", None)
|
2024-09-12 00:31:19 +08:00
|
|
|
image_grid_thw = kwargs.pop("image_grid_thw", None)
|
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
if pixel_values is None and image_embeds is None:
|
2024-09-12 00:31:19 +08:00
|
|
|
return None
|
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
if pixel_values is not None:
|
|
|
|
pixel_values = self._validate_and_reshape_mm_tensor(
|
|
|
|
pixel_values, "image pixel values")
|
|
|
|
image_grid_thw = self._validate_and_reshape_mm_tensor(
|
|
|
|
image_grid_thw, "image grid_thw")
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
if not isinstance(pixel_values, (torch.Tensor, list)):
|
|
|
|
raise ValueError("Incorrect type of image pixel values. "
|
|
|
|
f"Got type: {type(pixel_values)}")
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
return Qwen2VLImagePixelInputs(type="pixel_values",
|
2024-11-13 15:07:22 +08:00
|
|
|
pixel_values=pixel_values,
|
2024-09-30 11:16:10 +08:00
|
|
|
image_grid_thw=image_grid_thw)
|
|
|
|
|
|
|
|
if image_embeds is not None:
|
2024-10-04 22:34:58 +08:00
|
|
|
image_embeds = self._validate_and_reshape_mm_tensor(
|
|
|
|
image_embeds, "image embeds")
|
2024-11-13 15:07:22 +08:00
|
|
|
image_grid_thw = self._validate_and_reshape_mm_tensor(
|
|
|
|
image_grid_thw, "image grid_thw")
|
2024-10-04 22:34:58 +08:00
|
|
|
|
2024-09-30 11:16:10 +08:00
|
|
|
if not isinstance(image_embeds, torch.Tensor):
|
|
|
|
raise ValueError("Incorrect type of image embeddings. "
|
|
|
|
f"Got type: {type(image_embeds)}")
|
|
|
|
return Qwen2VLImageEmbeddingInputs(type="image_embeds",
|
2024-11-13 15:07:22 +08:00
|
|
|
image_embeds=image_embeds,
|
|
|
|
image_grid_thw=image_grid_thw)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
def _parse_and_validate_video_input(
|
|
|
|
self, **kwargs: object) -> Optional[Qwen2VLVideoInputs]:
|
|
|
|
pixel_values_videos = kwargs.pop("pixel_values_videos", None)
|
2024-11-13 15:07:22 +08:00
|
|
|
video_embeds = kwargs.pop("video_embeds", None)
|
2024-09-12 00:31:19 +08:00
|
|
|
video_grid_thw = kwargs.pop("video_grid_thw", None)
|
|
|
|
|
2024-11-13 15:07:22 +08:00
|
|
|
if pixel_values_videos is None and video_embeds is None:
|
2024-09-12 00:31:19 +08:00
|
|
|
return None
|
|
|
|
|
2024-11-13 15:07:22 +08:00
|
|
|
if pixel_values_videos is not None:
|
|
|
|
pixel_values_videos = self._validate_and_reshape_mm_tensor(
|
|
|
|
pixel_values_videos, "video pixel values")
|
|
|
|
video_grid_thw = self._validate_and_reshape_mm_tensor(
|
|
|
|
video_grid_thw, "video grid_thw")
|
|
|
|
|
|
|
|
return Qwen2VLVideoPixelInputs(
|
|
|
|
type="pixel_values_videos",
|
|
|
|
pixel_values_videos=pixel_values_videos,
|
|
|
|
video_grid_thw=video_grid_thw,
|
|
|
|
)
|
|
|
|
|
|
|
|
if video_embeds is not None:
|
|
|
|
video_embeds = self._validate_and_reshape_mm_tensor(
|
|
|
|
video_embeds, "video embeds")
|
|
|
|
video_grid_thw = self._validate_and_reshape_mm_tensor(
|
|
|
|
video_grid_thw, "video grid_thw")
|
|
|
|
|
|
|
|
if not isinstance(video_embeds, torch.Tensor):
|
|
|
|
raise ValueError("Incorrect type of video embeddings. "
|
|
|
|
f"Got type: {type(video_embeds)}")
|
|
|
|
return Qwen2VLVideoEmbeddingInputs(type="video_embeds",
|
|
|
|
video_embeds=video_embeds,
|
|
|
|
video_grid_thw=video_grid_thw)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
def _process_image_input(self,
|
|
|
|
image_input: Qwen2VLImageInputs) -> torch.Tensor:
|
2024-09-30 11:16:10 +08:00
|
|
|
if image_input["type"] == "image_embeds":
|
2024-11-13 15:07:22 +08:00
|
|
|
return image_input["image_embeds"].type(self.visual.dtype)
|
2024-09-30 11:16:10 +08:00
|
|
|
|
2024-11-13 15:07:22 +08:00
|
|
|
pixel_values = image_input["pixel_values"].type(self.visual.dtype)
|
2024-09-12 00:31:19 +08:00
|
|
|
image_embeds = self.visual(pixel_values,
|
|
|
|
grid_thw=image_input["image_grid_thw"])
|
|
|
|
return image_embeds
|
|
|
|
|
|
|
|
def _process_video_input(self,
|
|
|
|
video_input: Qwen2VLVideoInputs) -> torch.Tensor:
|
2024-11-13 15:07:22 +08:00
|
|
|
if video_input["type"] == "video_embeds":
|
|
|
|
return video_input["video_embeds"].type(self.visual.dtype)
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
pixel_values_videos = video_input["pixel_values_videos"].type(
|
|
|
|
self.visual.dtype)
|
|
|
|
video_embeds = self.visual(pixel_values_videos,
|
|
|
|
grid_thw=video_input["video_grid_thw"])
|
|
|
|
return video_embeds
|
|
|
|
|
|
|
|
def _merge_multimodal_embeddings(
|
|
|
|
self,
|
|
|
|
input_ids: torch.Tensor,
|
|
|
|
inputs_embeds: torch.Tensor,
|
|
|
|
multimodal_embeddings: torch.Tensor,
|
|
|
|
placeholder_token_id: int,
|
|
|
|
) -> torch.Tensor:
|
|
|
|
mask = (input_ids == placeholder_token_id)
|
|
|
|
inputs_embeds[mask, :] = multimodal_embeddings
|
|
|
|
return inputs_embeds
|
|
|
|
|
|
|
|
def forward(
|
|
|
|
self,
|
|
|
|
input_ids: torch.Tensor,
|
|
|
|
positions: torch.Tensor,
|
|
|
|
kv_caches: List[torch.Tensor],
|
|
|
|
attn_metadata: AttentionMetadata,
|
|
|
|
intermediate_tensors: Optional[IntermediateTensors] = None,
|
|
|
|
**kwargs: object,
|
2024-10-03 19:56:58 -07:00
|
|
|
) -> Union[torch.Tensor, IntermediateTensors]:
|
2024-09-12 00:31:19 +08:00
|
|
|
"""Run forward pass for Qwen2-VL.
|
|
|
|
|
|
|
|
Args:
|
|
|
|
input_ids: Flattened (concatenated) input_ids corresponding to a
|
|
|
|
batch.
|
|
|
|
positions: Flattened (concatenated) position ids corresponding to a
|
|
|
|
batch.
|
|
|
|
**NOTE**: If mrope is enabled (default setting for Qwen2-VL
|
|
|
|
opensource models), the shape will be `(3, seq_len)`,
|
|
|
|
otherwise it will be `(seq_len,).
|
|
|
|
pixel_values: Pixel values to be fed to a model.
|
|
|
|
`None` if no images are passed.
|
|
|
|
image_grid_thw: Tensor `(n_images, 3)` of image 3D grid in LLM.
|
|
|
|
`None` if no images are passed.
|
|
|
|
pixel_values_videos: Pixel values of videos to be fed to a model.
|
|
|
|
`None` if no videos are passed.
|
|
|
|
video_grid_thw: Tensor `(n_videos, 3)` of video 3D grid in LLM.
|
|
|
|
`None` if no videos are passed.
|
|
|
|
"""
|
2024-10-03 19:56:58 -07:00
|
|
|
if intermediate_tensors is not None:
|
|
|
|
input_ids = None
|
2024-09-12 00:31:19 +08:00
|
|
|
inputs_embeds = None
|
|
|
|
else:
|
2024-10-03 19:56:58 -07:00
|
|
|
image_input = self._parse_and_validate_image_input(**kwargs)
|
|
|
|
video_input = self._parse_and_validate_video_input(**kwargs)
|
2024-09-12 00:31:19 +08:00
|
|
|
|
2024-10-03 19:56:58 -07:00
|
|
|
if image_input is None and video_input is None:
|
|
|
|
inputs_embeds = None
|
|
|
|
else:
|
2024-10-16 13:56:17 +08:00
|
|
|
if uses_mrope(self.config):
|
2024-10-03 19:56:58 -07:00
|
|
|
assert positions.ndim == 2 and positions.size(0) == 3, (
|
|
|
|
"multimodal section rotary embedding requires "
|
|
|
|
f"(3, seq_len) positions, but got {positions.size()}")
|
|
|
|
|
|
|
|
inputs_embeds = self.model.embed_tokens(input_ids)
|
|
|
|
|
|
|
|
if image_input is not None:
|
|
|
|
image_embeds = self._process_image_input(image_input)
|
|
|
|
inputs_embeds = self._merge_multimodal_embeddings(
|
|
|
|
input_ids,
|
|
|
|
inputs_embeds,
|
|
|
|
image_embeds,
|
|
|
|
placeholder_token_id=self.config.image_token_id,
|
|
|
|
)
|
|
|
|
|
|
|
|
if video_input is not None:
|
|
|
|
video_embeds = self._process_video_input(video_input)
|
|
|
|
inputs_embeds = self._merge_multimodal_embeddings(
|
|
|
|
input_ids,
|
|
|
|
inputs_embeds,
|
|
|
|
video_embeds,
|
|
|
|
placeholder_token_id=self.config.video_token_id,
|
|
|
|
)
|
|
|
|
|
|
|
|
input_ids = None
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
hidden_states = self.model(
|
|
|
|
input_ids=input_ids,
|
|
|
|
positions=positions,
|
|
|
|
kv_caches=kv_caches,
|
|
|
|
attn_metadata=attn_metadata,
|
2024-09-23 21:46:59 +08:00
|
|
|
intermediate_tensors=intermediate_tensors,
|
2024-09-12 00:31:19 +08:00
|
|
|
inputs_embeds=inputs_embeds,
|
|
|
|
)
|
|
|
|
return hidden_states
|
|
|
|
|
|
|
|
def compute_logits(self, hidden_states: torch.Tensor,
|
|
|
|
sampling_metadata: SamplingMetadata) -> torch.Tensor:
|
|
|
|
logits = self.logits_processor(self.lm_head, hidden_states,
|
|
|
|
sampling_metadata)
|
|
|
|
return logits
|
|
|
|
|
|
|
|
def sample(
|
|
|
|
self,
|
|
|
|
logits: torch.Tensor,
|
|
|
|
sampling_metadata: SamplingMetadata,
|
|
|
|
) -> Optional[SamplerOutput]:
|
|
|
|
next_tokens = self.sampler(logits, sampling_metadata)
|
|
|
|
return next_tokens
|
|
|
|
|
2024-11-13 02:28:13 -06:00
|
|
|
def pooler(
|
|
|
|
self,
|
|
|
|
hidden_states: torch.Tensor,
|
|
|
|
pooling_metadata: PoolingMetadata,
|
|
|
|
) -> Optional[PoolerOutput]:
|
|
|
|
return self._pooler(hidden_states, pooling_metadata)
|
|
|
|
|
2024-09-12 00:31:19 +08:00
|
|
|
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"),
|
|
|
|
("gate_up_proj", "up_proj", 1),
|
|
|
|
("gate_up_proj", "gate_proj", 0),
|
|
|
|
]
|
|
|
|
params_dict = dict(self.named_parameters(remove_duplicate=False))
|
|
|
|
for name, loaded_weight in weights:
|
|
|
|
if "rotary_emb.inv_freq" in name:
|
|
|
|
continue
|
|
|
|
if self.config.tie_word_embeddings and "lm_head.weight" in name:
|
|
|
|
continue
|
|
|
|
for (param_name, weight_name, shard_id) in stacked_params_mapping:
|
|
|
|
if weight_name not in name:
|
|
|
|
continue
|
|
|
|
name = name.replace(weight_name, param_name)
|
2024-09-13 15:58:28 +08:00
|
|
|
# Skip loading extra bias for GPTQ models.
|
|
|
|
if name.endswith(".bias") and name not in params_dict:
|
|
|
|
continue
|
2024-09-23 21:46:59 +08:00
|
|
|
if is_pp_missing_parameter(name, self):
|
|
|
|
continue
|
2024-09-12 00:31:19 +08:00
|
|
|
param = params_dict[name]
|
|
|
|
weight_loader = param.weight_loader
|
|
|
|
weight_loader(param, loaded_weight, shard_id)
|
|
|
|
break
|
|
|
|
else:
|
2024-10-31 01:41:20 -04:00
|
|
|
if "visual" in name and name.endswith("qkv.weight"):
|
2024-09-12 00:31:19 +08:00
|
|
|
visual_num_heads = self.config.vision_config.num_heads
|
|
|
|
visual_embed_dim = self.config.vision_config.embed_dim
|
|
|
|
head_size = visual_embed_dim // visual_num_heads
|
|
|
|
loaded_weight = loaded_weight.view(3, visual_num_heads,
|
|
|
|
head_size,
|
|
|
|
visual_embed_dim)
|
|
|
|
loaded_weight = loaded_weight.transpose(0, 1)
|
|
|
|
loaded_weight = loaded_weight.reshape(-1, visual_embed_dim)
|
2024-10-31 01:41:20 -04:00
|
|
|
elif "visual" in name and name.endswith("qkv.bias"):
|
2024-09-12 00:31:19 +08:00
|
|
|
visual_num_heads = self.config.vision_config.num_heads
|
|
|
|
visual_embed_dim = self.config.vision_config.embed_dim
|
|
|
|
head_size = visual_embed_dim // visual_num_heads
|
|
|
|
loaded_weight = loaded_weight.view(3, visual_num_heads,
|
|
|
|
head_size)
|
|
|
|
loaded_weight = loaded_weight.transpose(0, 1)
|
|
|
|
loaded_weight = loaded_weight.reshape(-1)
|
|
|
|
try:
|
2024-09-13 15:58:28 +08:00
|
|
|
# Skip loading extra bias for GPTQ models.
|
|
|
|
if name.endswith(".bias") and name not in params_dict:
|
|
|
|
continue
|
2024-09-23 21:46:59 +08:00
|
|
|
if is_pp_missing_parameter(name, self):
|
|
|
|
continue
|
2024-09-12 00:31:19 +08:00
|
|
|
param = params_dict[name]
|
|
|
|
except KeyError:
|
2024-10-14 07:56:24 -07:00
|
|
|
raise ValueError(f"Unexpected weight: {name}") from None
|
2024-09-12 00:31:19 +08:00
|
|
|
|
|
|
|
weight_loader = getattr(param, "weight_loader",
|
|
|
|
default_weight_loader)
|
|
|
|
weight_loader(param, loaded_weight)
|