Files
FastDeploy/fastdeploy/input/multimodal_processor.py
T

532 lines
22 KiB
Python

"""
# Copyright (c) 2025 PaddlePaddle Authors. All Rights Reserved.
#
# 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.
"""
"""Unified multimodal processor for all VL model types.
Consolidates the four separate VL processor wrappers and four separate
DataProcessor classes into a single class with pluggable Encoding strategies.
"""
import pickle
from collections.abc import Mapping
from typing import Any, Dict, Optional
import numpy as np
import zmq
from fastdeploy.entrypoints.chat_utils import parse_chat_messages
from fastdeploy.input.base_processor import BaseTextProcessor
from fastdeploy.input.encodings import EncodingRegistry
from fastdeploy.input.image_processors import ImageProcessorRegistry
from fastdeploy.input.mm_model_config import MODEL_CONFIGS
from fastdeploy.input.utils import IDS_TYPE_FLAG, process_stop_token_ids
from fastdeploy.utils import data_processor_logger
_DEFAULT_MM_LIMITS = {"image": 1, "video": 1, "audio": 1}
_SAMPLING_EPS = 1e-5
class MultiModalProcessor(BaseTextProcessor):
"""Unified multimodal processor for all supported VL model types.
Uses a composition pattern: model-type-specific encoding logic is
delegated to ``self.enc`` (an Encoding instance), while common logic
(tokenization loop, request processing, caching) lives here.
"""
def __init__(
self,
model_name_or_path: str,
model_type: str,
config=None,
limit_mm_per_prompt: Optional[Dict[str, Any]] = None,
mm_processor_kwargs: Optional[Dict[str, Any]] = None,
reasoning_parser_obj=None,
tool_parser_obj=None,
enable_processor_cache: bool = False,
):
if model_type not in MODEL_CONFIGS:
raise ValueError(f"Unsupported model_type '{model_type}'. " f"Must be one of {sorted(MODEL_CONFIGS)}.")
self.model_type = model_type
self.config = config
self.cfg = MODEL_CONFIGS[model_type]
self.enable_processor_cache = enable_processor_cache
super().__init__(
model_name_or_path,
tokenizer_type=self.cfg.tokenizer_type,
reasoning_parser_obj=reasoning_parser_obj,
tool_parser_obj=tool_parser_obj,
)
data_processor_logger.info(f"model_name_or_path: {model_name_or_path}")
processor_kwargs = self._parse_processor_kwargs(mm_processor_kwargs)
self._init_image_processor()
self._init_role_prefixes()
# Composition: create encoding strategy via registry
enc_cls = EncodingRegistry.get(self.model_type)
self.enc = enc_cls(self, processor_kwargs)
self.limit_mm_per_prompt = self._parse_limits(limit_mm_per_prompt)
def _load_tokenizer(self):
"""Load the appropriate tokenizer based on model_type."""
if self.tokenizer_type == "ernie4_5":
import os
from fastdeploy.input.ernie4_5_tokenizer import Ernie4_5Tokenizer
vocab_file_names = ["tokenizer.model", "spm.model", "ernie_token_100k.model"]
for name in vocab_file_names:
if os.path.exists(os.path.join(self.model_name_or_path, name)):
Ernie4_5Tokenizer.resource_files_names["vocab_file"] = name
break
tokenizer = Ernie4_5Tokenizer.from_pretrained(self.model_name_or_path)
else:
from paddleformers.transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained(self.model_name_or_path, padding_side="left", use_fast=True)
return tokenizer
def _init_image_processor(self):
"""Create the appropriate image processor."""
cls = ImageProcessorRegistry.get(self.model_type)
self.image_processor = cls.from_pretrained(self.model_name_or_path)
def _init_role_prefixes(self):
"""Set up role prefixes for message parsing."""
self.role_prefixes = {
"system": "",
"user": "User: ",
"bot": "Assistant: ",
"assistant": "Assistant: ",
}
if self.cfg.has_tool_role:
self.role_prefixes["tool"] = "Tool: "
def _parse_processor_kwargs(self, kwargs: Optional[dict]) -> dict:
if not kwargs:
return {}
try:
if not isinstance(kwargs, dict):
raise ValueError("mm-processor-kwargs must be a dictionary")
data_processor_logger.info(f"Processing kwargs: {kwargs}")
expected_types = self.cfg.expected_kwargs
for key, value in kwargs.items():
if key in expected_types and not isinstance(value, expected_types[key]):
raise ValueError(
f"Invalid type for {key}: expected "
f"{expected_types[key].__name__}, got {type(value).__name__}"
)
return kwargs
except Exception as e:
data_processor_logger.warning(f"Invalid mm-processor-kwargs format: {e}")
return {}
def _parse_limits(self, limits: Optional[dict]) -> dict:
if not limits:
return dict(_DEFAULT_MM_LIMITS)
try:
if not isinstance(limits, dict):
raise ValueError("limit-mm-per-prompt must be a dictionary")
data_processor_logger.info(f"_parse_limits:{limits}")
return {**_DEFAULT_MM_LIMITS, **limits}
except Exception as e:
data_processor_logger.warning(f"Invalid limit-mm-per-prompt format: {e}, using default limits")
return dict(_DEFAULT_MM_LIMITS)
def _check_mm_limits(self, item):
if isinstance(item, dict):
mm_data = item
else:
mm_data = {"image": [], "video": []}
for message in item:
if isinstance(message.get("content"), list):
for part in message["content"]:
part_type = part.get("type")
if part_type in ("image_url", "image"):
mm_data["image"].append(part)
elif part_type in ("video_url", "video"):
mm_data["video"].append(part)
for modality, data in mm_data.items():
if modality in self.limit_mm_per_prompt:
limit = self.limit_mm_per_prompt[modality]
if len(data) > limit:
raise ValueError(f"Too many {modality} items in prompt, " f"got {len(data)} but limit is {limit}")
def get_mm_max_tokens_per_item(self, seq_len: int) -> Optional[Mapping[str, int]]:
return self.enc.get_mm_max_tokens_per_item(seq_len)
def _extract_mm_items(self, request):
"""Extract images/videos from request messages, handling processor cache."""
messages = parse_chat_messages(request.get("messages"))
mm_items = []
for msg in messages:
role = msg.get("role")
if role not in self.role_prefixes:
raise ValueError(f"Unsupported role: {role}")
content = msg.get("content")
if not isinstance(content, list):
content = [content]
for item in content:
if item.get("type") in ["image", "video"]:
mm_items.append(item)
missing_hashes, missing_idx = [], []
for idx, item in enumerate(mm_items):
if not item.get("data"):
missing_hashes.append(item.get("uuid"))
missing_idx.append(idx)
if len(missing_hashes) > 0 and not self.enable_processor_cache:
raise ValueError("Missing items cannot be retrieved without processor cache.")
dealer = None
if self.enable_processor_cache:
context = zmq.Context()
dealer = context.socket(zmq.DEALER)
dealer.connect("ipc:///dev/shm/processor_cache.ipc")
missing_items = self.get_processor_cache(dealer, missing_hashes)
for idx in range(len(missing_items)):
if not missing_items[idx]:
raise ValueError(f"Missing item {idx} not found in processor cache")
mm_items[missing_idx[idx]]["data"] = missing_items[idx]
images, videos = [], []
image_uuid, video_uuid = [], []
for item in mm_items:
if item.get("type") == "image":
images.append(item["data"])
image_uuid.append(item["uuid"])
elif item.get("type") == "video":
videos.append(item["data"])
video_uuid.append(item["uuid"])
else:
raise ValueError(f"Unsupported multimodal type: {item.get('type')}")
return images, videos, image_uuid, video_uuid, dealer, missing_idx, mm_items
def text2ids(self, text, images=None, videos=None, image_uuid=None, video_uuid=None):
"""Convert text with image/video placeholders into model inputs."""
outputs = self.enc._make_outputs()
IMAGE_PLACEHOLDER = self.cfg.image_placeholder
VIDEO_PLACEHOLDER = self.cfg.video_placeholder
IMAGE_PLACEHOLDER_LEN = len(IMAGE_PLACEHOLDER)
VIDEO_PLACEHOLDER_LEN = len(VIDEO_PLACEHOLDER)
st, image_idx, video_idx = 0, 0, 0
while st < len(text):
image_pos = text.find(IMAGE_PLACEHOLDER, st)
image_pos = len(text) if image_pos == -1 else image_pos
video_pos = text.find(VIDEO_PLACEHOLDER, st)
video_pos = len(text) if video_pos == -1 else video_pos
ed = min(image_pos, video_pos)
self._add_text(text[st:ed], outputs)
if ed == len(text):
break
if ed == image_pos:
image = images[image_idx]
uuid = image_uuid[image_idx] if image_uuid else None
if not isinstance(image, tuple):
self.enc.add_image(image, outputs, uuid)
else:
self.enc.add_processed_image(image, outputs, uuid)
image_idx += 1
st = ed + IMAGE_PLACEHOLDER_LEN
else:
item = videos[video_idx]
uuid = video_uuid[video_idx] if video_uuid else None
if not isinstance(item, tuple):
if isinstance(item, dict):
frames, meta = self.enc.load_video(item["video"], item)
else:
frames, meta = self.enc.load_video(item, {})
self.enc.add_video(frames, outputs, uuid, meta=meta)
else:
self.enc.add_processed_video(item, outputs, uuid)
video_idx += 1
st = ed + VIDEO_PLACEHOLDER_LEN
return outputs
def request2ids(self, request):
"""Convert chat request with multimodal messages into model inputs."""
images, videos, image_uuid, video_uuid, dealer, missing_idx, mm_items = self._extract_mm_items(request)
if self.tokenizer.chat_template is None:
raise ValueError("This model does not support chat template.")
chat_template_kwargs = request.get("chat_template_kwargs", {})
if self.cfg.chat_template_pass_request:
# ernie: pass full request to apply_chat_template
prompt = self.tokenizer.apply_chat_template(
request,
tokenize=False,
add_generation_prompt=request.get("add_generation_prompt", True),
**chat_template_kwargs,
)
else:
messages = parse_chat_messages(request.get("messages"))
prompt = self.tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=request.get("add_generation_prompt", True),
**chat_template_kwargs,
)
request["prompt_tokens"] = prompt
outputs = self.text2ids(prompt, images, videos, image_uuid, video_uuid)
if self.enable_processor_cache:
self._update_mm_cache(dealer, missing_idx, mm_items, outputs)
return outputs
def _process_prompt_token_ids(self, request):
"""Handle the prompt_token_ids tokenisation path.
Mirrors ``request2ids`` in structure: Processor owns extract/cache,
Encoding only does pure encoding.
"""
prompt_token_ids = request.get("prompt_token_ids", [])
if not request.get("messages"):
return self.enc.prompt_token_ids2outputs(prompt_token_ids)
images, videos, image_uuid, video_uuid, dealer, missing_idx, mm_items = self._extract_mm_items(request)
outputs = self.enc.prompt_token_ids2outputs(prompt_token_ids, mm_items)
if self.enable_processor_cache:
self._update_mm_cache(dealer, missing_idx, mm_items, outputs)
return outputs
def _update_mm_cache(self, dealer, missing_idx, mm_items, outputs):
"""Write newly-processed multimodal items to the processor cache."""
missing_idx_set = set(missing_idx)
hashes_to_cache, items_to_cache = [], []
for idx in range(len(mm_items)):
if idx in missing_idx_set:
continue
meta = {}
grid_thw = np.asarray(outputs["grid_thw"][idx])
if grid_thw.ndim > 1:
t, h, w = grid_thw[0]
else:
t, h, w = grid_thw
meta["thw"] = (int(t), int(h), int(w))
if "fps" in outputs:
meta["fps"] = outputs["fps"][idx]
hashes_to_cache.append(outputs["mm_hashes"][idx])
items_to_cache.append((outputs["images"][idx], meta))
if hashes_to_cache:
self.update_processor_cache(dealer, hashes_to_cache, items_to_cache)
return outputs
def _add_text(self, tokens, outputs):
"""Add text tokens to outputs, delegating position logic to enc."""
if not tokens:
return
if isinstance(tokens, str):
tokens_str = self.tokenizer.tokenize(tokens)
tokens = self.tokenizer.convert_tokens_to_ids(tokens_str)
num_tokens = len(tokens)
outputs["input_ids"].extend(tokens)
outputs["token_type_ids"].extend([IDS_TYPE_FLAG["text"]] * num_tokens)
self.enc.add_text_positions(outputs, num_tokens)
def process_request_dict(self, request, max_model_len=None):
"""Process a request dictionary into model inputs."""
cfg = self.cfg
request = self._apply_default_parameters(request)
if not request.get("eos_token_ids"):
request["eos_token_ids"] = self.eos_token_ids
# Stop tokens
if cfg.stop_tokens_variant == "qwen3":
stop_sequences = request.get("stop", [])
if stop_sequences:
stop_seqs, stop_seqs_len = self.update_stop_seq(stop_sequences)
request["stop_token_ids"] = stop_seqs
request["stop_seqs_len"] = stop_seqs_len
else:
process_stop_token_ids(request, self.update_stop_seq)
# Bad words
if cfg.has_bad_words:
bad_words = request.get("bad_words")
bad_words_token_ids = request.get("bad_words_token_ids")
if bad_words:
bad_words_token_ids = self.update_bad_words(bad_words, bad_words_token_ids)
request["bad_words_token_ids"] = bad_words_token_ids
# Logits processor (ernie think)
if cfg.has_logits_processor_think:
logits_processors_args = self._prepare_think_stop_sentence(
request.get("logits_processors_args") or {}, max_model_len
)
request["logits_processors_args"] = logits_processors_args
# Tokenize
outputs = self._tokenize_request(request)
# Post-token handling
self._process_post_tokens(request, outputs)
# Force disable thinking for qwen_vl / qwen3_vl
if cfg.force_disable_thinking:
request["enable_thinking"] = False
# Pack outputs
outputs = self.pack_outputs(outputs)
# Assign prompt_token_ids
if cfg.preserve_prompt_token_ids and request.get("prompt_token_ids"):
pass # preserve existing
else:
request["prompt_token_ids"] = outputs["input_ids"].tolist()
request["multimodal_inputs"] = outputs
# Truncation
if max_model_len is not None and len(request["prompt_token_ids"]) > max_model_len:
request["prompt_token_ids"] = request["prompt_token_ids"][: max_model_len - 1]
request["prompt_token_ids_len"] = len(request["prompt_token_ids"])
# Ernie: update thinking prompt state
if cfg.has_logits_processor_think:
logits_processors_args = self._update_thinking_prompt_state(
request["prompt_token_ids"],
request.get("logits_processors_args") or {},
)
request["logits_processors_args"] = logits_processors_args
# max_tokens
max_tokens = max_model_len - len(request["prompt_token_ids"])
if request.get("max_tokens") is None:
request["max_tokens"] = max(1, max_tokens)
else:
request["max_tokens"] = min(max_tokens, request["max_tokens"])
# Ernie: default reasoning_max_tokens
if cfg.set_default_reasoning_max_tokens and request.get("reasoning_max_tokens") is None:
request["reasoning_max_tokens"] = max(int(request["max_tokens"] * 0.8), 1)
# Clamp top_p
if request.get("top_p") is not None and request.get("top_p") < _SAMPLING_EPS:
request["top_p"] = _SAMPLING_EPS
request["top_k"] = 1
# Reasoning parser
if self.reasoning_parser:
self._apply_reasoning_parser(request)
# Ernie: cap response_max_tokens
if cfg.cap_response_max_tokens:
if request.get("response_max_tokens") is not None and request.get("enable_thinking") is False:
request["max_tokens"] = min(request["response_max_tokens"], request["max_tokens"])
data_processor_logger.info(f"Processed request {request}")
return request
def _tokenize_request(self, request):
cfg = self.cfg
default_thinking = cfg.default_thinking
if request.get("prompt_token_ids") and cfg.supports_prompt_token_ids:
messages = request.get("messages")
if messages:
self._check_mm_limits(messages)
request.setdefault("enable_thinking", default_thinking)
return self._process_prompt_token_ids(request)
elif request.get("prompt"):
multimodal_data = request.get("multimodal_data") or {}
self._check_mm_limits(multimodal_data)
images = multimodal_data.get("image", None)
videos = multimodal_data.get("video", None)
request["prompt_tokens"] = request.get("prompt")
request.setdefault("enable_thinking", default_thinking)
return self.text2ids(request["prompt"], images, videos)
elif request.get("messages"):
messages = request["messages"]
self._check_mm_limits(messages)
chat_template_kwargs = request.get("chat_template_kwargs")
if chat_template_kwargs:
if isinstance(chat_template_kwargs, dict):
for k, v in chat_template_kwargs.items():
if k not in request or request[k] is None:
request[k] = v
else:
raise ValueError("Invalid input: chat_template_kwargs must be a dict")
request.setdefault("enable_thinking", default_thinking)
return self.request2ids(request)
else:
raise ValueError(f"Request must contain 'prompt', or 'messages': {request}")
def _process_post_tokens(self, request, outputs):
completion_token_ids = request.get("completion_token_ids") or request.get("generated_token_ids")
if completion_token_ids:
self.enc.append_completion_tokens(outputs, completion_token_ids)
def append_completion_tokens(self, multimodal_inputs, completion_token_ids):
"""Append completion tokens — delegates to enc."""
self.enc.append_completion_tokens(multimodal_inputs, completion_token_ids)
def pack_outputs(self, outputs):
"""Convert intermediate outputs to final packed format."""
if not outputs["images"]:
outputs["images"] = None
outputs["grid_thw"] = None
outputs["image_type_ids"] = None
else:
outputs["images"] = np.vstack(outputs["images"])
outputs["grid_thw"] = np.vstack(outputs["grid_thw"])
outputs["image_type_ids"] = np.array(outputs["image_type_ids"])
outputs["input_ids"] = np.array(outputs["input_ids"], dtype=np.int64)
outputs["token_type_ids"] = np.array(outputs["token_type_ids"], dtype=np.int64)
outputs["mm_num_token_func"] = self.enc.mm_num_tokens
# Position IDs: delegate to encoding strategy
self.enc.pack_position_ids(outputs)
return outputs
def get_processor_cache(self, socket, mm_hashes):
req = pickle.dumps(mm_hashes)
socket.send_multipart([b"", req])
_, resp = socket.recv_multipart()
mm_items = pickle.loads(resp)
data_processor_logger.info(f"Get cache of mm_hashes: {mm_hashes}")
return mm_items
def update_processor_cache(self, socket, mm_hashes, mm_items):
req = pickle.dumps((mm_hashes, mm_items))
socket.send_multipart([b"", req])
data_processor_logger.info(f"Update cache of mm_hashes: {mm_hashes}")