mineru 2.2.2__py3-none-any.whl → 2.5.1__py3-none-any.whl

mineru/model/vlm_sglang_model/engine.py

@@ -1,264 +0,0 @@
-import asyncio
-import time
-from types import MethodType
-from typing import AsyncIterator, Dict, Iterator, List, Optional, Union
-
-import fastapi
-from sglang.srt.entrypoints.engine import Engine as _Engine
-from sglang.srt.managers.io_struct import EmbeddingReqInput, GenerateReqInput
-from sglang.srt.managers.tokenizer_manager import (
-    TokenizerManager,
-    dataclass_to_string_truncated,
-    logger,
-)
-from sglang.srt.sampling.sampling_params import SamplingParams
-from sglang.srt.server_args import ServerArgs
-
-from ...utils.run_async import run_async
-from .logit_processor import Mineru2LogitProcessor
-
-
-class BatchEngine(_Engine):
-    """
-    The engine is patched to support batch multi-modal generate, and early image preprocessing.
-    """
-
-    def __init__(self, server_args: ServerArgs, **kwargs):
-        server_args.enable_custom_logit_processor = True
-        super().__init__(server_args=server_args, **kwargs)
-        _patch_tokenizer_manager(self.tokenizer_manager)
-
-    def generate(
-        self,
-        # The input prompt. It can be a single prompt or a batch of prompts.
-        prompt: Optional[Union[List[str], str]] = None,
-        sampling_params: Optional[Union[List[Dict], Dict]] = None,
-        # The token ids for text; one can either specify text or input_ids.
-        input_ids: Optional[Union[List[List[int]], List[int]]] = None,
-        # The image input. It can be a file name, a url, or base64 encoded string.
-        # See also python/sglang/srt/utils.py:load_image.
-        image_data: Optional[Union[List[str], str]] = None,
-        return_logprob: Optional[Union[List[bool], bool]] = False,
-        logprob_start_len: Optional[Union[List[int], int]] = None,
-        top_logprobs_num: Optional[Union[List[int], int]] = None,
-        token_ids_logprob: Optional[Union[List[List[int]], List[int]]] = None,
-        lora_path: Optional[List[Optional[str]]] = None,
-        custom_logit_processor: Optional[Union[List[Optional[str]], str]] = None,
-        return_hidden_states: bool = False,
-        stream: bool = False,
-    ) -> Union[Dict, Iterator[Dict]]:
-        """
-        The arguments of this function is the same as `sglang/srt/managers/io_struct.py::GenerateReqInput`.
-        Please refer to `GenerateReqInput` for the documentation.
-        """
-        modalities_list = []
-
-        # EDIT
-        if isinstance(image_data, list):
-            for _ in range(len(image_data)):
-                modalities_list.append(["image"])
-        elif image_data is not None:
-            modalities_list.append("image")
-
-        # ADD
-        if custom_logit_processor is None:
-            custom_logit_processor = Mineru2LogitProcessor().to_str()
-
-        obj = GenerateReqInput(
-            text=prompt,
-            input_ids=input_ids,
-            sampling_params=sampling_params,
-            image_data=image_data,
-            return_logprob=return_logprob,
-            logprob_start_len=logprob_start_len,
-            top_logprobs_num=top_logprobs_num,
-            token_ids_logprob=token_ids_logprob,
-            lora_path=lora_path,
-            modalities=modalities_list,
-            custom_logit_processor=custom_logit_processor,
-            return_hidden_states=return_hidden_states,
-            stream=stream,
-        )
-        generator = _generate_request(self.tokenizer_manager, obj, None)
-
-        if stream:
-
-            def generator_wrapper():
-                while True:
-                    try:
-                        chunk = run_async(generator.__anext__())
-                        yield chunk
-                    except StopAsyncIteration:
-                        break
-
-            return generator_wrapper()
-        else:
-            ret = run_async(generator.__anext__())
-            return ret
-
-    async def async_generate(
-        self,
-        # The input prompt. It can be a single prompt or a batch of prompts.
-        prompt: Optional[Union[List[str], str]] = None,
-        sampling_params: Optional[Union[List[Dict], Dict]] = None,
-        # The token ids for text; one can either specify text or input_ids.
-        input_ids: Optional[Union[List[List[int]], List[int]]] = None,
-        # The image input. It can be a file name, a url, or base64 encoded string.
-        # See also python/sglang/srt/utils.py:load_image.
-        image_data: Optional[Union[List[str], str]] = None,
-        return_logprob: Optional[Union[List[bool], bool]] = False,
-        logprob_start_len: Optional[Union[List[int], int]] = None,
-        top_logprobs_num: Optional[Union[List[int], int]] = None,
-        token_ids_logprob: Optional[Union[List[List[int]], List[int]]] = None,
-        lora_path: Optional[List[Optional[str]]] = None,
-        custom_logit_processor: Optional[Union[List[Optional[str]], str]] = None,
-        return_hidden_states: bool = False,
-        stream: bool = False,
-    ) -> Union[Dict, AsyncIterator[Dict], Iterator[Dict]]:
-        """
-        The arguments of this function is the same as `sglang/srt/managers/io_struct.py::GenerateReqInput`.
-        Please refer to `GenerateReqInput` for the documentation.
-        """
-        modalities_list = []
-
-        # EDIT
-        if isinstance(image_data, list):
-            for _ in range(len(image_data)):
-                modalities_list.append(["image"])
-        elif image_data is not None:
-            modalities_list.append("image")
-
-        # ADD
-        if custom_logit_processor is None:
-            custom_logit_processor = Mineru2LogitProcessor().to_str()
-
-        obj = GenerateReqInput(
-            text=prompt,
-            input_ids=input_ids,
-            sampling_params=sampling_params,
-            image_data=image_data,
-            return_logprob=return_logprob,
-            logprob_start_len=logprob_start_len,
-            top_logprobs_num=top_logprobs_num,
-            token_ids_logprob=token_ids_logprob,
-            lora_path=lora_path,
-            modalities=modalities_list,
-            custom_logit_processor=custom_logit_processor,
-            return_hidden_states=return_hidden_states,
-            stream=stream,
-        )
-        generator = _generate_request(self.tokenizer_manager, obj, None)
-
-        if stream is True:
-            return generator
-        else:
-            return await generator.__anext__()
-
-
-def _auto_create_handle_loop(self: TokenizerManager):
-    """
-    patch the original `auto_create_handle_loop()` method to reset `no_create_loop`
-    when the event loop changes.
-    """
-    try:
-        curr_handle_loop = asyncio.get_running_loop()
-    except RuntimeError:
-        curr_handle_loop = None
-
-    last_handle_loop = getattr(self, "_last_handle_loop", None)
-    if last_handle_loop != curr_handle_loop:
-        self.no_create_loop = False
-        setattr(self, "_last_handle_loop", curr_handle_loop)
-    return TokenizerManager.auto_create_handle_loop(self)
-
-
-def _patch_tokenizer_manager(self: TokenizerManager):
-    self.auto_create_handle_loop = MethodType(_auto_create_handle_loop, self)
-
-
-async def _one_request(
-    self: TokenizerManager,
-    obj: Union[GenerateReqInput, EmbeddingReqInput],
-    request: Optional[fastapi.Request],
-    created_time: Optional[float],
-):
-    tokenized_obj = await self._tokenize_one_request(obj)
-    state = self._send_one_request(obj, tokenized_obj, created_time)
-    async for out in self._wait_one_response(obj, state, request):
-        yield out
-
-
-async def _handle_batch_request(
-    self: TokenizerManager,
-    obj: Union[GenerateReqInput, EmbeddingReqInput],
-    request: Optional[fastapi.Request] = None,
-    created_time: Optional[float] = None,
-):
-    batch_size = obj.batch_size
-
-    generators = []
-    rids = []
-
-    if getattr(obj, "parallel_sample_num", 1) != 1:
-        raise Exception("parallel_sample_num != 1 is not supported in this patched code.")
-
-    # Send all requests
-    for i in range(batch_size):
-        tmp_obj = obj[i]
-        generators.append(_one_request(self, tmp_obj, request, created_time))
-        rids.append(tmp_obj.rid)
-
-    # Wait for all requests
-    is_stream = hasattr(obj, "stream") and obj.stream
-    if not is_stream:
-        outputs = await asyncio.gather(*(gen.__anext__() for gen in generators))
-        yield outputs
-    else:
-        rid_to_index = {rid: i for i, rid in enumerate(rids)}
-        task_map = {asyncio.create_task(gen.__anext__()): gen for gen in generators}
-        while task_map:
-            done, _ = await asyncio.wait(task_map.keys(), return_when=asyncio.FIRST_COMPLETED)
-
-            for task in done:
-                gen = task_map.pop(task)
-                try:
-                    result = task.result()
-                    result["index"] = rid_to_index[result["meta_info"]["id"]]
-                    yield result
-                    new_task = asyncio.create_task(gen.__anext__())
-                    task_map[new_task] = gen
-                except StopAsyncIteration:
-                    pass
-
-
-async def _generate_request(
-    self: TokenizerManager,
-    obj: Union[GenerateReqInput, EmbeddingReqInput],
-    request: Optional[fastapi.Request] = None,
-):
-    created_time = time.time()
-
-    self.auto_create_handle_loop()
-
-    if isinstance(obj, EmbeddingReqInput) and self.is_generation:
-        raise ValueError(
-            "This model does not appear to be an embedding model by default. "
-            "Please add `--is-embedding` when launching the server or try another model."
-        )
-
-    obj.normalize_batch_and_arguments()
-
-    if self.log_requests:
-        max_length, skip_names, _ = self.log_request_metadata
-        logger.info(f"Receive: obj={dataclass_to_string_truncated(obj, max_length, skip_names=skip_names)}")
-
-    async with self.model_update_lock.reader_lock:
-        is_single = obj.is_single
-        if is_single:
-            tokenized_obj = await self._tokenize_one_request(obj)
-            state = self._send_one_request(obj, tokenized_obj, created_time)
-            async for response in self._wait_one_response(obj, state, request):
-                yield response
-        else:
-            async for response in _handle_batch_request(self, obj, request, created_time):
-                yield response

mineru/model/vlm_sglang_model/image_processor.py

@@ -1,213 +0,0 @@
-import ast
-import asyncio
-import re
-from typing import List, Optional, Union
-
-import numpy as np
-
-from sglang.version import __version__ as sglang_version
-from packaging import version
-if version.parse(sglang_version) >= version.parse("0.4.9"):
-    # sglang >= 0.4.9
-    from sglang.srt.multimodal.processors.base_processor import (
-        BaseMultimodalProcessor as BaseProcessor,
-    )
-    from sglang.srt.multimodal.mm_utils import divide_to_patches, expand2square, select_best_resolution
-else:
-    # 0.4.7 <= sglang < 0.4.9
-    from sglang.srt.managers.multimodal_processors.base_processor import (
-        BaseMultimodalProcessor as BaseProcessor,
-    )
-    from sglang.srt.mm_utils import divide_to_patches, expand2square, select_best_resolution
-
-get_global_processor = None
-from sglang.srt.utils import load_image, logger
-from sglang.utils import get_exception_traceback
-
-from .model import Mineru2QwenForCausalLM
-
-
-# image_best_res is only resized (not padded).
-def process_anyres_image(image, processor, grid_pinpoints):
-    if isinstance(grid_pinpoints, str) and "x" in grid_pinpoints:
-        patch_size = processor.crop_size["height"]
-        assert patch_size in [224, 336, 384, 448, 512], "patch_size should be in [224, 336, 384, 448, 512]"
-        matches = re.findall(r"\((\d+)x(\d+)\)", grid_pinpoints)
-        range_start = tuple(map(int, matches[0]))
-        range_end = tuple(map(int, matches[-1]))
-        grid_pinpoints = [
-            (i, j) for i in range(range_start[0], range_end[0] + 1) for j in range(range_start[1], range_end[1] + 1)
-        ]
-        grid_pinpoints = [[dim * patch_size for dim in pair] for pair in grid_pinpoints]
-
-    if type(grid_pinpoints) is list:
-        possible_resolutions = grid_pinpoints
-    else:
-        possible_resolutions = ast.literal_eval(grid_pinpoints)
-    best_resolution = select_best_resolution(image.size, possible_resolutions)
-
-    image_best_res = image.resize(best_resolution)  # <<<<<<< Here changed
-    patches = divide_to_patches(image_best_res, processor.crop_size["height"])
-    image_original_resize = image.resize((processor.crop_size["height"], processor.crop_size["height"]))
-
-    image_patches = [image_original_resize] + patches
-    image_patches = [processor.preprocess(image_patch)["pixel_values"][0] for image_patch in image_patches]
-    return np.stack(image_patches, axis=0)
-
-
-class Mineru2ImageProcessor(BaseProcessor):
-    def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
-        super().__init__(hf_config, server_args, _processor, *args, **kwargs)
-
-    @staticmethod
-    def _process_single_image_task(
-        image_data: Union[str, bytes],
-        image_aspect_ratio: Optional[str] = None,
-        image_grid_pinpoints: Optional[str] = None,
-        image_processor=None,
-    ):
-        if image_processor is None:
-            assert get_global_processor is not None
-            image_processor = get_global_processor().image_processor
-
-        try:
-            image, image_size = load_image(image_data)
-            if image_size is not None:
-                # It is a video with multiple images
-                image_hash = hash(image_data)
-                pixel_values = image_processor(image)["pixel_values"]
-                pixel_values = np.stack(pixel_values, axis=0)
-                return pixel_values, image_hash, image_size
-            else:
-                # It is an image
-                image_hash = hash(image_data)
-                if image_aspect_ratio == "pad":
-                    image = expand2square(
-                        image,
-                        tuple(int(x * 255) for x in image_processor.image_mean),
-                    )
-                    pixel_values = image_processor(image.convert("RGB"))["pixel_values"][0]
-                elif image_aspect_ratio == "anyres" or (image_aspect_ratio is not None and "anyres_max" in image_aspect_ratio):
-                    pixel_values = process_anyres_image(image, image_processor, image_grid_pinpoints)
-                else:
-                    pixel_values = image_processor(image)["pixel_values"][0]
-                return pixel_values, image_hash, image.size
-        except Exception:
-            logger.error("Exception in TokenizerManager:\n" + get_exception_traceback())
-
-    async def _process_single_image(self, image_data: Union[bytes, str], aspect_ratio: str, grid_pinpoints: str):
-        if hasattr(self, "cpu_executor"):
-            executor = self.cpu_executor
-        else:
-            executor = self.executor
-
-        if get_global_processor is not None:
-            image_processor = None  # save ipc cost
-        else:
-            image_processor = self._processor.image_processor
-
-        if executor is not None:
-            loop = asyncio.get_running_loop()
-            return await loop.run_in_executor(
-                executor,
-                Mineru2ImageProcessor._process_single_image_task,
-                image_data,
-                aspect_ratio,
-                grid_pinpoints,
-                image_processor,
-            )
-        else:
-            return self._process_single_image_task(
-                image_data,
-                aspect_ratio,
-                grid_pinpoints,
-                image_processor,
-            )
-
-    async def process_mm_data_async(
-        self,
-        image_data: List[Union[str, bytes]],
-        input_text,
-        request_obj,
-        *args,
-        **kwargs,
-    ):
-        from sglang.srt.managers.schedule_batch import Modality, MultimodalDataItem
-
-        if not image_data:
-            return None
-
-        modalities = request_obj.modalities or ["image"]
-        aspect_ratio = getattr(self.hf_config, "image_aspect_ratio", None)
-        grid_pinpoints = (
-            self.hf_config.image_grid_pinpoints
-            if hasattr(self.hf_config, "image_grid_pinpoints")
-               and "anyres" in aspect_ratio
-            else None
-        )
-
-        if isinstance(image_data, str):
-            image_data = [image_data]
-
-        if isinstance(image_data, list) and len(image_data) > 0:
-            if "multi-images" in modalities or "video" in modalities:
-                # Multiple images
-                aspect_ratio = "pad"  # LLaVA OneVision Handling: more than one image --> interleaved image mode or video mode. We do not use anyres
-                pixel_values, data_hashes, image_sizes = [], [], []
-                res = []
-                for img_data in image_data:
-                    res.append(
-                        self._process_single_image(
-                            img_data, aspect_ratio, grid_pinpoints
-                        )
-                    )
-
-                res = await asyncio.gather(*res)
-                for pixel_v, image_h, image_s in res:
-                    pixel_values.append(pixel_v)
-                    data_hashes.append(image_h)
-                    image_sizes.append(image_s)
-
-                if isinstance(pixel_values[0], np.ndarray):
-                    pixel_values = np.stack(pixel_values, axis=0)
-            else:
-                # A single image
-                pixel_values, image_hash, image_size = await self._process_single_image(
-                    image_data[0], aspect_ratio, grid_pinpoints
-                )
-                image_sizes = [image_size]
-        else:
-            raise ValueError(f"Invalid image data: {image_data}")
-        modality = Modality.IMAGE
-        if isinstance(request_obj.modalities, list):
-            if request_obj.modalities[0] == "multi-images":
-                modality = Modality.MULTI_IMAGES
-            elif request_obj.modalities[0] == "video":
-                modality = Modality.VIDEO
-
-        if version.parse(sglang_version) >= version.parse("0.4.9.post3"):
-            # sglang >= 0.4.9.post3
-            return {
-                "mm_items": [
-                    MultimodalDataItem(
-                        feature=pixel_values,
-                        model_specific_data={
-                            "image_sizes": image_sizes,
-                        },
-                        modality=modality,
-                    )
-                ],
-            }
-        else:
-            # 0.4.7 <= sglang <= 0.4.9.post2
-            return {
-                "mm_items": [
-                    MultimodalDataItem(
-                        pixel_values=pixel_values,
-                        image_sizes=image_sizes,
-                        modality=modality,
-                    )
-                ],
-            }
-
-ImageProcessorMapping = {Mineru2QwenForCausalLM: Mineru2ImageProcessor}

mineru/model/vlm_sglang_model/logit_processor.py

@@ -1,90 +0,0 @@
-from typing import List
-
-from sglang.srt.sampling.custom_logit_processor import CustomLogitProcessor
-
-
-class Mineru2LogitProcessor(CustomLogitProcessor):
-    """
-    Stateless logit processor for Mineru2.
-
-    (base-class: sglang.srt.sampling.custom_logit_processor.CustomLogitProcessor)
-
-    This processor applies token-level constraints to prevent repetition during generation.
-    It supports two main constraints:
-
-    - no_repeat_ngram_size (int):
-        Prevents repeating the same n-gram of specified size in the output.
-        Inspired by Hugging Face's NoRepeatNGramLogitsProcessor.
-        This implementation is slower due to its lack of specialized optimization.
-
-    - no_repeat_token_count (int):
-        (Placeholder for future logic)
-        Intended to prevent repeating the same token multiple times.
-        Not yet implemented in this version.
-    """
-
-    def __init__(self) -> None:
-        super().__init__()
-        self._generated_ngrams = {}  # Cache of generated n-grams by request ID
-        self._time = {}  # Timestamp of the last update for each request
-        self._gen_step = 0  # Global generation step counter
-
-    def __call__(self, logits, batch_info: List[dict]):
-        """
-        Applies repetition constraints to the logits before sampling tokens.
-
-        Args:
-            logits (FloatTensor): A tensor of shape (batch_size, vocab_size) containing raw token logits.
-            batch_info (List[dict]): A list of metadata dicts for each sample in the batch. Each dict must include:
-                - "__req__": Request object containing request ID and output_ids.
-                - "no_repeat_ngram_size": Size of n-gram to avoid repeating.
-
-        Returns:
-            FloatTensor: The modified logits tensor with banned token logits set to -inf.
-        """
-        from sglang.srt.managers.schedule_batch import Req
-
-        self._gen_step += 1  # Update global generation step
-
-        for idx, info in enumerate(batch_info):
-            if not isinstance(info, dict) or "__req__" not in info:
-                continue
-
-            req: Req = info["__req__"]
-            rid = req.rid
-            output_ids = req.output_ids
-            ngram_size = info.get("no_repeat_ngram_size", 0)
-
-            # Skip if there are not enough tokens to form an n-gram
-            if ngram_size <= 0 or len(output_ids) < ngram_size:
-                continue
-
-            # Record the current step for cache cleanup tracking
-            self._time[rid] = self._gen_step
-
-            # Initialize n-gram cache for this request if it doesn't exist
-            if rid not in self._generated_ngrams:
-                self._generated_ngrams[rid] = {}
-
-            # Get the n-gram prefix (all but the last token)
-            prev_ngram = tuple(output_ids[-ngram_size:-1])
-            last_token = output_ids[-1]
-
-            # Store this n-gram occurrence
-            self._generated_ngrams[rid][prev_ngram] = self._generated_ngrams[rid].get(prev_ngram, []) + [last_token]
-
-            # Get the next-token candidates to ban based on current prefix
-            current_prefix = tuple(output_ids[-ngram_size + 1 :])
-            banned_tokens = self._generated_ngrams[rid].get(current_prefix, [])
-
-            # Set the logits of banned tokens to negative infinity
-            for token in banned_tokens:
-                logits[idx][token] = -float("inf")
-
-        # Clean up cache for expired requests
-        expired_rids = [rid for rid, last_used in self._time.items() if last_used < self._gen_step]
-        for rid in expired_rids:
-            self._generated_ngrams.pop(rid, None)
-            self._time.pop(rid, None)
-
-        return logits