Skip to content

GitLab

Unverified Commit a9e879b3 authored by Cyrus Leung's avatar Cyrus Leung Committed by GitHub
Browse files

[Misc] Clean up MiniCPM-V/O code (#15337) 


Signed-off-by: default avatarDarkLight1337 <tlleungac@connect.ust.hk>
parent 3e2f37a6
Expand all Show whitespace changes
Inline Side-by-side
Showing with 531 additions and 661 deletions
examples/offline_inference/vision_language.py
View file @ a9e879b3
......@@ -361,6 +361,7 @@ def run_llava_next_video(questions: list[str],
engine_args = EngineArgs(
model="llava-hf/LLaVA-NeXT-Video-7B-hf",
max_model_len=8192,
max_num_seqs=2,
disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
)
......
tests/models/decoder_only/vision_language/test_models.py
View file @ a9e879b3
......@@ -163,24 +163,24 @@ VLM_TEST_SETTINGS = {
marks=[pytest.mark.core_model, pytest.mark.cpu_model],
),
#### Extended model tests
# "aria": VLMTestInfo(
# models=["rhymes-ai/Aria"],
# test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
# prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501
# img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
# max_model_len=4096,
# max_num_seqs=2,
# auto_cls=AutoModelForImageTextToText,
# single_image_prompts=IMAGE_ASSETS.prompts({
# "stop_sign": "<vlm_image>Please describe the image shortly.",
# "cherry_blossom": "<vlm_image>Please infer the season with reason.", # noqa: E501
# }),
# multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.", # noqa: E501
# stop_str=["<|im_end|>"],
# image_size_factors=[(0.10, 0.15)],
# max_tokens=64,
# marks=[large_gpu_mark(min_gb=64)],
# ),
"aria": VLMTestInfo(
models=["rhymes-ai/Aria"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|im_start|>user\n{img_prompt}<|im_end|>\n<|im_start|>assistant\n ", # noqa: E501
img_idx_to_prompt=lambda idx: "<fim_prefix><|img|><fim_suffix>\n",
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForImageTextToText,
single_image_prompts=IMAGE_ASSETS.prompts({
"stop_sign": "<vlm_image>Please describe the image shortly.",
"cherry_blossom": "<vlm_image>Please infer the season with reason.", # noqa: E501
}),
multi_image_prompt="<vlm_image><vlm_image>Describe the two images shortly.", # noqa: E501
stop_str=["<|im_end|>"],
image_size_factors=[(0.10, 0.15)],
max_tokens=64,
marks=[large_gpu_mark(min_gb=64)],
),
"blip2": VLMTestInfo(
models=["Salesforce/blip2-opt-2.7b"],
test_type=VLMTestType.IMAGE,
......@@ -352,6 +352,7 @@ VLM_TEST_SETTINGS = {
prompt_formatter=lambda vid_prompt: f"USER: {vid_prompt} ASSISTANT:",
num_video_frames=16,
max_model_len=4096,
max_num_seqs=2,
auto_cls=AutoModelForVision2Seq,
vllm_output_post_proc=model_utils.llava_video_vllm_to_hf_output,
),
......@@ -384,7 +385,18 @@ VLM_TEST_SETTINGS = {
),
"minicpmo_26": VLMTestInfo(
models=["openbmb/MiniCPM-o-2_6"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
test_type=(VLMTestType.IMAGE),
prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
max_model_len=4096,
max_num_seqs=2,
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(['<|im_end|>', '<|endoftext|>']), # noqa: E501
hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
),
"minicpmo_26_multi_image": VLMTestInfo(
models=["openbmb/MiniCPM-o-2_6"],
test_type=(VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
max_model_len=4096,
......@@ -392,10 +404,22 @@ VLM_TEST_SETTINGS = {
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(['<|im_end|>', '<|endoftext|>']), # noqa: E501
hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
patch_hf_runner=model_utils.minicpmo_26_patch_hf_runner,
marks=[large_gpu_mark(min_gb=32)],
),
"minicpmv_26": VLMTestInfo(
models=["openbmb/MiniCPM-V-2_6"],
test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
test_type=(VLMTestType.IMAGE),
prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
max_model_len=4096,
max_num_seqs=2,
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(['<|im_end|>', '<|endoftext|>']), # noqa: E501
hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
),
"minicpmv_26_multi_image": VLMTestInfo(
models=["openbmb/MiniCPM-V-2_6"],
test_type=(VLMTestType.MULTI_IMAGE),
prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|start_header_id|>user<|end_header_id|>\n\n{img_prompt}<|eot_id|><|start_header_id|>assistant<|end_header_id|>\n\n", # noqa: E501
img_idx_to_prompt=lambda idx: "(<image>./</image>)\n",
max_model_len=4096,
......@@ -403,6 +427,7 @@ VLM_TEST_SETTINGS = {
get_stop_token_ids=lambda tok: tok.convert_tokens_to_ids(['<|im_end|>', '<|endoftext|>']), # noqa: E501
hf_output_post_proc=model_utils.minicpmv_trunc_hf_output,
patch_hf_runner=model_utils.minicpmv_26_patch_hf_runner,
marks=[large_gpu_mark(min_gb=32)],
),
"molmo": VLMTestInfo(
models=["allenai/Molmo-7B-D-0924"],
......
tests/models/multimodal/processing/test_common.py
View file @ a9e879b3
# SPDX-License-Identifier: Apache-2.0
import copy
from functools import partial
from typing import Optional, Union
......@@ -29,7 +28,7 @@ def _test_processing_correctness(
hit_rate: float,
num_batches: int,
simplify_rate: float,
ignore_mm_keys: Optional[list[str]] = None,
ignore_mm_keys: Optional[set[str]] = None,
):
model_info = HF_EXAMPLE_MODELS.find_hf_info(model_id)
model_info.check_available_online(on_fail="skip")
......@@ -145,7 +144,7 @@ def _test_processing_correctness_hf(
baseline_processor: BaseMultiModalProcessor,
cached_processor: BaseMultiModalProcessor,
batch_idx: int,
ignore_mm_keys: Optional[list[str]] = None,
ignore_mm_keys: Optional[set[str]] = None,
):
if model_config.hf_config.model_type in ("mllama", "whisper", "ultravox"):
# For some multimodal models, tokenizer will always add bos_token
......@@ -167,11 +166,12 @@ def _test_processing_correctness_hf(
hf_processor_mm_kwargs={},
)
assert _inputs_equal(
_assert_inputs_equal(
baseline_result,
cached_result,
ignore_mm_keys,
), f"Failed ({batch_idx=}, {prompt=}, {mm_data=})"
ignore_mm_keys=ignore_mm_keys,
msg=f"Failed ({batch_idx=}, {prompt=}, {mm_data=})",
)
baseline_tokenized_result = baseline_processor.apply(
token_prompt,
......@@ -179,11 +179,12 @@ def _test_processing_correctness_hf(
hf_processor_mm_kwargs={},
)
assert _inputs_equal(
_assert_inputs_equal(
baseline_result,
baseline_tokenized_result,
ignore_mm_keys,
), f"Failed ({batch_idx=}, {prompt=}, {mm_data=})"
ignore_mm_keys=ignore_mm_keys,
msg=f"Failed ({batch_idx=}, {prompt=}, {mm_data=})",
)
cached_tokenized_result = cached_processor.apply(
token_prompt,
......@@ -191,11 +192,12 @@ def _test_processing_correctness_hf(
hf_processor_mm_kwargs={},
)
assert _inputs_equal(
_assert_inputs_equal(
cached_result,
cached_tokenized_result,
ignore_mm_keys,
), f"Failed ({batch_idx=}, {prompt=}, {mm_data=})"
ignore_mm_keys=ignore_mm_keys,
msg=f"Failed ({batch_idx=}, {prompt=}, {mm_data=})",
)
def _test_processing_correctness_mistral(
......@@ -206,7 +208,7 @@ def _test_processing_correctness_mistral(
baseline_processor: BaseMultiModalProcessor,
cached_processor: BaseMultiModalProcessor,
batch_idx: int,
ignore_mm_keys: Optional[list[str]] = None,
ignore_mm_keys: Optional[set[str]] = None,
):
images = mm_data.get("image", [])
if not isinstance(images, list):
......@@ -233,11 +235,12 @@ def _test_processing_correctness_mistral(
hf_processor_mm_kwargs={},
)
assert _inputs_equal(
_assert_inputs_equal(
baseline_tokenized_result,
cached_tokenized_result,
ignore_mm_keys,
), f"Failed ({batch_idx=}, {prompt=}, {mm_data=})"
ignore_mm_keys=ignore_mm_keys,
msg=f"Failed ({batch_idx=}, {prompt=}, {mm_data=})",
)
# yapf: disable
......@@ -261,6 +264,7 @@ def _test_processing_correctness_mistral(
"TIGER-Lab/Mantis-8B-siglip-llama3",
"mistralai/Pixtral-12B-2409",
"mistral-community/pixtral-12b",
"openbmb/MiniCPM-Llama3-V-2_5",
"openbmb/MiniCPM-o-2_6",
"openbmb/MiniCPM-V-2_6",
"allenai/Molmo-7B-D-0924",
......@@ -290,7 +294,7 @@ def test_processing_correctness(
# In Ultravox, the audio_features can be different depending on padding
# The slight difference should not be a problem though, since
# attention_mask lets us ignore the difference.
ignore_mm_keys = ['audio_features']
ignore_mm_keys = {"audio_features"}
_test_processing_correctness(
model_id,
......@@ -328,38 +332,26 @@ def test_processing_correctness_phi3v(
)
def _inputs_equal(
def _assert_inputs_equal(
a: MultiModalInputs,
b: MultiModalInputs,
ignore_mm_keys: Optional[list[str]] = None,
*,
ignore_mm_keys: Optional[set[str]] = None,
msg: str = "",
):
return _drop_mm_kwargs_keys(a, ignore_mm_keys) == _drop_mm_kwargs_keys(
b, ignore_mm_keys)
def _drop_mm_kwargs_keys(
result: MultiModalInputs,
ignore_mm_keys: Optional[list[str]] = None,
) -> MultiModalInputs:
"""Drop specified keys from result['mm_kwargs'].
This is mainly to avoid doing exact match of audio_features in ultravox.
if ignore_mm_keys is None:
ignore_mm_keys = set()
Args:
result: Result to drop keys from
ignore_mm_keys: List of keys to ignore, e.g. ['audio_features']
"""
if not ignore_mm_keys:
return result
if msg is None:
assert "mm_kwargs" in a and "mm_kwargs" in b
else:
assert "mm_kwargs" in a and "mm_kwargs" in b, msg
if 'mm_kwargs' in result:
result = copy.deepcopy(result)
mm_kwargs = result['mm_kwargs']
for key in ignore_mm_keys:
mm_kwargs.pop(key, None)
for items in mm_kwargs._items_by_modality.values():
for item in items:
for key in ignore_mm_keys:
item.pop(key, None)
a["mm_kwargs"].pop(key, None)
b["mm_kwargs"].pop(key, None)
return result
if msg is None:
assert a == b
else:
assert a == b, msg
vllm/model_executor/models/gemma3_mm.py
View file @ a9e879b3
......@@ -295,8 +295,6 @@ class Gemma3MultiModalProcessor(BaseMultiModalProcessor[Gemma3ProcessingInfo]):
# HF processor pops the `num_crops` kwarg, which is needed by vLLM
if (images := mm_data.get("images")) is not None:
assert isinstance(images, list)
parsed_images = (self._get_data_parser().parse_mm_data({
"image":
images
......
vllm/model_executor/models/minicpmo.py
View file @ a9e879b3
......@@ -23,7 +23,7 @@
# limitations under the License.
"""Inference-only MiniCPM-O model compatible with HuggingFace weights."""
from collections.abc import Iterable, Mapping, Sequence
from typing import (Any, Callable, Dict, List, Literal, Optional, Set, Tuple,
from typing import (Any, Callable, Dict, Literal, Optional, Set, Tuple,
TypedDict, Union)
import torch
......@@ -43,24 +43,26 @@ from vllm.multimodal.parse import (AudioItem, AudioProcessorItems,
from vllm.multimodal.processing import PromptReplacement, PromptUpdate
from vllm.multimodal.profiling import ProcessorInputs
from vllm.sequence import IntermediateTensors
from vllm.utils import flatten_2d_lists
from .minicpmv import (MiniCPMV2_6, MiniCPMVDummyInputsBuilder,
MiniCPMVMultiModalDataParser,
MiniCPMVMultiModalProcessor, MiniCPMVProcessingInfo,
_minicpmv_field_config)
from .utils import AutoWeightsLoader, cast_overflow_tensors, maybe_prefix
from .utils import (AutoWeightsLoader, cast_overflow_tensors, flatten_bn,
maybe_prefix)
CPU_DEVICE = torch.device("cpu")
class MiniCPMOAudioFeatureInputs(TypedDict):
type: Literal["audio_features"]
data: torch.Tensor
audio_features: torch.Tensor
"""
Shape: `(batch_size * num_audios * num_slices, num_channels, length)`
Slice here means chunk. Audio that is too long will be split into slices,
which is the same as image.
Padding is used therefore `data` is `torch.Tensor`.
Padding is used therefore `audio_features` is `torch.Tensor`.
"""
audio_feature_lens: torch.Tensor
......@@ -68,7 +70,7 @@ class MiniCPMOAudioFeatureInputs(TypedDict):
Shape: `(batch_size * num_audios * num_slices)`
This should be feature length of each audio slice,
which equals to `data.shape[-1]`
which equals to `audio_features.shape[-1]`
"""
audio_bounds: torch.Tensor
......@@ -81,7 +83,7 @@ class MiniCPMOAudioFeatureInputs(TypedDict):
class MiniCPMOAudioEmbeddingInputs(TypedDict):
type: Literal["audio_embeds"]
data: List[torch.Tensor]
audio_embeds: torch.Tensor
"""
Shape: `(batch_size * num_images * num_slices, hidden_size)`
......@@ -102,18 +104,11 @@ MiniCPMOAudioInputs = Union[MiniCPMOAudioFeatureInputs,
def _minicpmo_field_config(hf_inputs: Mapping[str, torch.Tensor]):
audio_num_slices = hf_inputs.get("audio_num_slices", torch.empty(0))
return dict(
**_minicpmv_field_config(hf_inputs),
audio_features=MultiModalFieldConfig.flat_from_sizes(
"audio", audio_num_slices),
audio_feature_lens=MultiModalFieldConfig.flat_from_sizes(
"audio", audio_num_slices),
audio_num_slices=MultiModalFieldConfig.batched("audio"),
audio_orders_in_mm_data=MultiModalFieldConfig.batched("audio"),
audio_embeds=MultiModalFieldConfig.flat_from_sizes(
"audio", audio_num_slices),
audio_features=MultiModalFieldConfig.batched("audio"),
audio_feature_lens=MultiModalFieldConfig.batched("audio"),
audio_embeds=MultiModalFieldConfig.batched("audio"),
)
......@@ -153,9 +148,6 @@ class MiniCPMOMultiModalDataParser(MiniCPMVMultiModalDataParser):
class MiniCPMOProcessingInfo(MiniCPMVProcessingInfo):
audio_pattern = "(<audio>./</audio>)"
def get_supported_mm_modalities(self) -> List[str]:
return ["image", "video", "audio"]
def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
return {"image": None, "video": None, "audio": None}
......@@ -277,95 +269,47 @@ class MiniCPMOMultiModalProcessor(
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
) -> Mapping[str, NestedTensors]:
mm_data = dict(mm_data)
audios = mm_data.pop("audios", [])
audio_embeds = mm_data.pop("audio_embeds", [])
if isinstance(audios, (list, torch.Tensor)) and len(audios) > 0:
audio_outputs = {
"audio_lens": [],
"audio_features": [],
"audio_feature_lens": [],
"audio_num_segments": []
}
for audio in audios:
single_audio_outputs = super().call_base_hf_processor(
prompt=self.info.audio_pattern,
mm_data={
"audios": audio,
"chunk_input": True
if (audios := mm_data.get("audios")) is None:
return {}
parsed_audios = (self._get_data_parser().parse_mm_data({
"audio": audios
}).get_items("audio", AudioProcessorItems))
audio_inputs = self._base_call_hf_processor(
prompts=[self.info.audio_pattern] * len(parsed_audios),
mm_data={"audios": [[audio] for audio in parsed_audios]},
mm_kwargs={
**mm_kwargs, "chunk_input": True
},
mm_kwargs=mm_kwargs)
audio_outputs["audio_lens"].append(len(audio))
audio_outputs["audio_features"].append(
single_audio_outputs["audio_features"])
audio_outputs["audio_num_segments"].append(
len(single_audio_outputs["audio_feature_lens"][0]))
audio_outputs["audio_feature_lens"] += \
single_audio_outputs["audio_feature_lens"]
audio_outputs["audio_features"] = [
audio_feature for single_audio_features in \
audio_outputs["audio_features"]
for audio_feature in single_audio_features
]
audio_outputs["audio_feature_lens"] = torch.cat(
audio_outputs["audio_feature_lens"])
elif len(audio_embeds):
audio_outputs = {
"audio_lens": [
self.info.get_audio_len_by_num_chunks(
sum(chunk_embeds.shape[0]
for chunk_embeds in single_audio_embeds))
for single_audio_embeds in audio_embeds
],
"audio_embeds": [
chunk_embeds for single_audio_embeds in audio_embeds
for chunk_embeds in single_audio_embeds
],
"audio_num_segments": [
len(single_audio_embeds)
for single_audio_embeds in audio_embeds
out_keys={"audio_features", "audio_feature_lens"},
)
# Avoid padding since we need the output for each audio to be
# independent of other audios for the cache to work correctly
unpadded_audio_features = [
feat[:, :feature_len] for feat, feature_len in zip(
audio_inputs["audio_features"],
audio_inputs["audio_feature_lens"],
)
]
}
else:
audio_outputs = {}
return audio_outputs
audio_inputs["audio_features"] = unpadded_audio_features
return audio_inputs
def get_placeholder_match_pattern(self) -> str:
return r"\(<(image|video|audio)>./</\1>\)"
def get_placeholder_split_pattern(self) -> str:
return r"\(<(?:image|video|audio)>./</(?:image|video|audio)>\)"
def process_mm_inputs(
self,
mm_data: Mapping[str, object],
mm_kwargs: Mapping[str, object],
) -> Mapping[str, Mapping[str, NestedTensors]]:
) -> Mapping[str, NestedTensors]:
return {
"image": self.process_images(mm_data, mm_kwargs),
"video": self.process_videos(mm_data, mm_kwargs),
"audio": self.process_audios(mm_data, mm_kwargs),
**super().process_mm_inputs(mm_data, mm_kwargs),
**self.process_audios(mm_data, mm_kwargs),
}
def get_modality_num_counter(self, modality: str) -> str:
if modality == "audio":
return "audio_lens"
return super().get_modality_num_counter(modality)
def get_num_slices_by_modality(self, inputs: Dict[str, object],
modality: str, index: int) -> int:
if modality == "audio":
return inputs["audio"]["audio_num_segments"][index]
return super().get_num_slices_by_modality(inputs, modality, index)
def get_prompt_texts_by_modality(self, inputs: Dict[str, object],
modality: str, index: int) -> str:
if modality == "audio":
return self.get_audio_prompt_texts(
inputs["audio"]["audio_lens"][index])
return super().get_prompt_texts_by_modality(inputs, modality, index)
def _get_prompt_updates(
self,
mm_items: MultiModalDataItems,
......@@ -622,15 +566,16 @@ class MiniCPMO(MiniCPMV2_6):
# Copied from HF repo of MiniCPM-o-2_6,
# designed for batched inputs and outputs
def get_audio_hidden_states(self, data: MiniCPMOAudioInputs,
chunk_length: int) -> torch.Tensor:
chunk_length: int) -> list[torch.Tensor]:
wavforms = data.get(
"data",
"audio_features",
[]) # (bs, 80, frames) or [], multi audios need filled in advance
audio_feature_lens_raw = [data.get("audio_feature_lens",
[])] # list, [[x1, x2], [y1], [z1]]
# exist audio
if len(wavforms) > 0:
if len(wavforms) == 0:
return []
audio_feature_lens = torch.hstack(audio_feature_lens_raw)
batch_size, _, max_mel_seq_len = wavforms.shape
max_seq_len = (max_mel_seq_len - 1) // 2 + 1
......@@ -648,8 +593,8 @@ class MiniCPMO(MiniCPMV2_6):
padding_mask = seq_range >= lengths_expand # 1 for padded values
audio_attention_mask_ = padding_mask.view(
batch_size, 1, 1, max_seq_len).expand(batch_size, 1,
max_seq_len, max_seq_len)
batch_size, 1, 1, max_seq_len).expand(batch_size, 1, max_seq_len,
max_seq_len)
audio_attention_mask = audio_attention_mask_.to(
dtype=self.apm.conv1.weight.dtype,
device=self.apm.conv1.weight.device)
......@@ -690,18 +635,15 @@ class MiniCPMO(MiniCPMV2_6):
idx += 1
final_audio_embeds.append(target_audio_embeds)
return final_audio_embeds
else:
return []
def get_embedding_with_audios(self, vlm_embedding: torch.Tensor,
audio_inputs: Optional[MiniCPMOAudioInputs],
audio_inputs: MiniCPMOAudioInputs,
chunk_length: int) -> torch.Tensor:
device, dtype = vlm_embedding.device, vlm_embedding.dtype
if audio_inputs["type"] == "audio_embeds":
audio_embeddings = audio_inputs["data"]
audio_embeddings = [
audio_embeddings[i].to(device=device, dtype=dtype)
for i in range(len(audio_embeddings))
item.to(device=device, dtype=dtype)
for item in audio_inputs["audio_embeds"]
]
else:
audio_embeddings = self.get_audio_hidden_states(
......@@ -746,40 +688,56 @@ class MiniCPMO(MiniCPMV2_6):
def _parse_and_validate_audio_inputs(
self, input_ids: torch.Tensor,
**kwargs: object) -> Tuple[MiniCPMOAudioInputs]:
audio_features = kwargs.pop("audio_features", [])
audio_feature_lens = kwargs.pop("audio_feature_lens", [])
**kwargs: object) -> Optional[MiniCPMOAudioInputs]:
audio_features = kwargs.pop("audio_features", None)
audio_embeds = kwargs.pop("audio_embeds", None)
audio_start_id = kwargs.pop("audio_start_id", None)
audio_end_id = kwargs.pop("audio_end_id", None)
if audio_features is None and audio_embeds is None:
return None
audio_start_id = kwargs.pop("audio_start_id")
if not isinstance(audio_start_id, torch.Tensor):
raise ValueError("Incorrect type of audio_start_id. "
f"Got type: {type(audio_start_id)}")
audio_end_id = kwargs.pop("audio_end_id")
if not isinstance(audio_end_id, torch.Tensor):
raise ValueError("Incorrect type of audio_end_id. "
f"Got type: {type(audio_end_id)}")
if audio_embeds is not None:
audio_embeds = [
audio_embeds[i][j] for i in range(len(audio_embeds))
for j in range(len(audio_embeds[i]))
]
if not isinstance(audio_embeds, (torch.Tensor, list)):
raise ValueError("Incorrect type of audio_embeds. "
f"Got type: {type(audio_embeds)}")
return MiniCPMOAudioEmbeddingInputs(
type="audio_embeds",
audio_embeds=flatten_bn(flatten_2d_lists(audio_embeds),
concat=True),
audio_bounds=self._get_audio_bounds(input_ids, audio_start_id,
audio_end_id),
data=audio_embeds,
type="audio_embeds")
if len(audio_features) > 0:
audio_features_all = [
i.permute(1, 0) for audio_feature in audio_features
for i in audio_feature
]
audio_features = torch.nn.utils.rnn.pad_sequence(
audio_features_all, batch_first=True,
padding_value=0.0).permute(0, 2, 1)
audio_feature_lens = torch.cat(
[item for item in audio_feature_lens])
)
if audio_features is not None:
if not isinstance(audio_features, (torch.Tensor, list)):
raise ValueError("Incorrect type of audio_features. "
f"Got type: {type(audio_features)}")
audio_feature_lens = kwargs.pop("audio_feature_lens")
if not isinstance(audio_feature_lens, (torch.Tensor, list)):
raise ValueError("Incorrect type of audio_feature_lens. "
f"Got type: {type(audio_feature_lens)}")
return MiniCPMOAudioFeatureInputs(
type="audio_features",
audio_features=flatten_bn(audio_features, concat=True),
audio_feature_lens=flatten_bn(
flatten_2d_lists(audio_feature_lens), concat=True),
audio_bounds=self._get_audio_bounds(input_ids, audio_start_id,
audio_end_id),
data=audio_features,
audio_feature_lens=audio_feature_lens,
type="audio_features")
return None
)
raise AssertionError("This line should be unreachable.")
def _parse_and_validate_inputs(self, input_ids: torch.Tensor,
**kwargs: object):
......@@ -803,7 +761,7 @@ class MiniCPMO(MiniCPMV2_6):
else:
image_inputs, audio_inputs = \
self._parse_and_validate_inputs(input_ids, **kwargs)
vlm_embeddings, _ = self.get_embedding_with_vision(
vlm_embeddings = self.get_embedding_with_vision(
input_ids, image_inputs)
if audio_inputs is not None:
......
vllm/model_executor/models/minicpmv.py
View file @ a9e879b3
This diff is collapsed.
vllm/multimodal/inputs.py
View file @ a9e879b3
......@@ -665,6 +665,13 @@ class MultiModalKwargs(UserDict[str, NestedTensors]):
return cast(BatchedTensorInputs, json_mapped)
def __delitem__(self, key: str) -> None:
super().__delitem__(key)
for items in self._items_by_modality.values():
for item in items:
item.pop(key, None)
def __eq__(self, other: object) -> bool:
if not isinstance(other, self.__class__):
return False
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment