xpu_model_runner.py

from typing import List, Optional, Tuple

import torch
import torch.nn as nn

from vllm.attention import get_attn_backend
from vllm.config import (CacheConfig, DeviceConfig, LoadConfig, LoRAConfig,
                         ModelConfig, ParallelConfig, SchedulerConfig,
                         VisionLanguageConfig)
from vllm.distributed import broadcast_tensor_dict
from vllm.logger import init_logger
from vllm.model_executor.model_loader import get_model
from vllm.sampling_params import SamplingParams
from vllm.sequence import SamplerOutput, SequenceData, SequenceGroupMetadata
from vllm.utils import CudaMemoryProfiler, make_tensor_with_pad
from vllm.worker.model_runner import AttentionMetadata, SamplingMetadata

logger = init_logger(__name__)

_PAD_SLOT_ID = -1
_BATCH_SIZE_ALIGNMENT = 8
_BATCH_SIZES_TO_CAPTURE = [1, 2, 4] + [
    _BATCH_SIZE_ALIGNMENT * i for i in range(1, 33)
]


class XPUModelRunner:

    def __init__(
        self,
        model_config: ModelConfig,
        parallel_config: ParallelConfig,
        scheduler_config: SchedulerConfig,
        device_config: DeviceConfig,
        cache_config: CacheConfig,
        load_config: LoadConfig,
        lora_config: Optional[LoRAConfig],
        vision_language_config: Optional[VisionLanguageConfig],
        kv_cache_dtype: Optional[str] = "auto",
        is_driver_worker: bool = False,
        *args,
        **kwargs,
    ):
        self.model_config = model_config
        self.parallel_config = parallel_config
        self.scheduler_config = scheduler_config
        self.lora_config = lora_config
        self.load_config = load_config
        self.cache_config = cache_config
        self.vision_language_config = vision_language_config
        self.is_driver_worker = is_driver_worker

        self.sliding_window = model_config.get_sliding_window()
        self.device_config = device_config
        self.device = self.device_config.device

        self.kv_cache_dtype = kv_cache_dtype
        self.block_size = cache_config.block_size
        self.max_context_len_to_capture = (
            self.model_config.max_context_len_to_capture
            if self.model_config is not None else 0)

        self.attn_backend = get_attn_backend(
            self.model_config.get_num_attention_heads(self.parallel_config),
            self.model_config.get_head_size(),
            self.model_config.get_num_kv_heads(self.parallel_config),
            self.model_config.get_sliding_window(),
            self.model_config.dtype,
            self.kv_cache_dtype,
            self.block_size,
        )

        # Lazy initialization.
        self.model: nn.Module  # Set after init_Model

    def load_model(self) -> None:
        with CudaMemoryProfiler() as m:
            self.model = get_model(
                model_config=self.model_config,
                device_config=self.device_config,
                load_config=self.load_config,
                lora_config=self.lora_config,
                vision_language_config=self.vision_language_config,
                parallel_config=self.parallel_config,
                scheduler_config=self.scheduler_config,
                cache_config=self.cache_config,
            )

        self.model_memory_usage = m.consumed_memory
        logger.info("Loading model weights took %.4f GB",
                    self.model_memory_usage / float(2**30))

    @property
    def vocab_size(self) -> int:
        return self.model_config.get_vocab_size()

    @torch.inference_mode()
    def profile_run(self) -> None:
        # Enable top-k sampling to reflect the accurate memory usage.
        sampling_params = SamplingParams(top_p=0.99, top_k=self.vocab_size - 1)
        max_num_batched_tokens = self.scheduler_config.max_num_batched_tokens
        max_num_seqs = self.scheduler_config.max_num_seqs

        # Profile memory usage with max_num_sequences sequences and the total
        # number of tokens equal to max_num_batched_tokens.
        seqs: List[SequenceGroupMetadata] = []
        # Additional GPU memory may be needed for vision encoding, which needs
        # to be accounted for when calculating the GPU blocks for
        # vLLM blocker manager.
        # To exercise the worst scenario for GPU memory consumption,
        # the number of seqs (batch_size) is chosen to maximize the number
        # of images processed.
        for group_id in range(max_num_seqs):
            seq_len = (max_num_batched_tokens // max_num_seqs +
                       (group_id < max_num_batched_tokens % max_num_seqs))

            seq_data = SequenceData([0] * seq_len)
            dummy_multi_modal_data = None
            seq = SequenceGroupMetadata(
                request_id=str(group_id),
                is_prompt=True,
                seq_data={group_id: seq_data},
                sampling_params=sampling_params,
                block_tables=None,
                lora_request=None,
                multi_modal_data=dummy_multi_modal_data,
            )
            seqs.append(seq)

        # Run the model with the dummy inputs.
        num_layers = self.model_config.get_num_layers(self.parallel_config)
        kv_caches = [None] * num_layers
        self.execute_model(seqs, kv_caches)
        torch.xpu.synchronize()
        return

    def prepare_input_tensors(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
    ) -> Tuple[torch.Tensor, torch.Tensor, AttentionMetadata, SamplingMetadata,
               Optional[torch.Tensor]]:
        multi_modal_input = None
        if self.is_driver_worker:
            # NOTE: We assume that all sequences in the group are all prompts or
            # all decodes.
            is_prompt = seq_group_metadata_list[0].is_prompt
            # Prepare input tensors.
            if is_prompt:
                (input_tokens, input_positions, attn_metadata, seq_lens,
                 multi_modal_input
                 ) = self._prepare_prompt(seq_group_metadata_list)
            else:
                (input_tokens, input_positions,
                 attn_metadata) = self._prepare_decode(seq_group_metadata_list)
                seq_lens = []
            sampling_metadata = SamplingMetadata.prepare(
                seq_group_metadata_list,
                seq_lens,
                # subquery_lens is not needed if chunked prefill is not
                # supported. Since CPU worker doesn't support chunked prefill
                # just use seq_lens instead.
                seq_lens,
                self.device,
                pin_memory=False)
            # Broadcast the metadata.
            metadata_dict = {
                "input_tokens": input_tokens,
                "input_positions": input_positions,
                "selected_token_indices":
                sampling_metadata.selected_token_indices,
            }
            metadata_dict.update(attn_metadata.asdict_zerocopy())
            broadcast_tensor_dict(metadata_dict, src=0)
        else:
            metadata_dict = broadcast_tensor_dict(src=0)
            input_tokens = metadata_dict.pop("input_tokens")
            input_positions = metadata_dict.pop("input_positions")
            selected_token_indices = metadata_dict.pop(
                "selected_token_indices")
            attn_metadata = self.attn_backend.make_metadata(**metadata_dict)
            sampling_metadata = SamplingMetadata(
                seq_groups=None,
                selected_token_indices=selected_token_indices,
                categorized_sample_indices=None,
                num_prompts=0,
            )

        return (input_tokens, input_positions, attn_metadata,
                sampling_metadata, multi_modal_input)

    def _prepare_decode(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
    ) -> Tuple[torch.Tensor, torch.Tensor, AttentionMetadata]:
        assert len(seq_group_metadata_list) > 0
        input_tokens: List[int] = []
        input_positions: List[int] = []
        slot_mapping: List[int] = []
        seq_lens: List[int] = []
        block_tables: List[List[int]] = []

        for seq_group_metadata in seq_group_metadata_list:
            assert not seq_group_metadata.is_prompt
            assert seq_group_metadata.token_chunk_size == 1

            seq_ids = list(seq_group_metadata.seq_data.keys())

            for seq_id in seq_ids:
                seq_data = seq_group_metadata.seq_data[seq_id]
                generation_token = seq_data.get_last_token_id()
                input_tokens.append(generation_token)

                seq_len = seq_data.get_len()
                position = seq_len - 1
                input_positions.append(position)

                seq_len = seq_len if self.sliding_window is None else min(
                    seq_len, self.sliding_window)
                seq_lens.append(seq_len)

                block_table = seq_group_metadata.block_tables[seq_id]
                block_number = block_table[position // self.block_size]
                block_offset = position % self.block_size
                slot = block_number * self.block_size + block_offset
                slot_mapping.append(slot)

                if self.sliding_window is not None:
                    sliding_window_blocks = (self.sliding_window //
                                             self.block_size)
                    block_table = block_table[-sliding_window_blocks:]
                block_tables.append(block_table)

        max_decode_seq_len = max(seq_lens)

        input_tokens = torch.tensor(input_tokens,
                                    dtype=torch.long,
                                    device=self.device)
        input_positions = torch.tensor(input_positions,
                                       dtype=torch.long,
                                       device=self.device)
        slot_mapping = torch.tensor(slot_mapping,
                                    dtype=torch.long,
                                    device=self.device)
        seq_lens_tensor = torch.tensor(seq_lens,
                                       dtype=torch.int,
                                       device=self.device)

        max_block_table_len = max(
            len(block_table) for block_table in block_tables)
        block_tables = make_tensor_with_pad(
            block_tables,
            max_len=max_block_table_len,
            pad=0,
            dtype=torch.int,
            device=self.device,
        )

        attn_metadata = self.attn_backend.make_metadata(
            is_prompt=False,
            slot_mapping=slot_mapping,
            seq_lens=seq_lens,
            seqlen_q=None,
            max_seqlen=None,
            seq_lens_tensor=seq_lens_tensor,
            max_decode_seq_len=max_decode_seq_len,
            num_prefill_tokens=0,
            num_decode_tokens=len(input_tokens),
            num_prefills=0,
            block_tables=block_tables,
        )
        return (
            input_tokens,
            input_positions,
            attn_metadata,
        )

    @torch.inference_mode()
    def execute_model(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
        kv_caches: List[torch.Tensor],
    ) -> Optional[SamplerOutput]:
        (input_tokens, input_positions, attn_metadata, sampling_metadata,
         multi_modal_input
         ) = self.prepare_input_tensors(seq_group_metadata_list)

        model_executable = self.model
        execute_model_kwargs = {
            "input_ids": input_tokens,
            "positions": input_positions,
            "kv_caches": kv_caches,
            "attn_metadata": attn_metadata,
        }
        if self.vision_language_config:
            execute_model_kwargs.update({"image_input": multi_modal_input})

        hidden_states = model_executable(**execute_model_kwargs)

        # Compute the logits.
        logits = self.model.compute_logits(hidden_states, sampling_metadata)

        # Only perform sampling in the driver worker.
        if not self.is_driver_worker:
            return None

        # Sample the next token.
        output = self.model.sample(
            logits=logits,
            sampling_metadata=sampling_metadata,
        )
        return output

    def _prepare_prompt(
        self,
        seq_group_metadata_list: List[SequenceGroupMetadata],
    ) -> Tuple[torch.Tensor, torch.Tensor, AttentionMetadata, List[int],
               Optional[torch.Tensor]]:
        assert len(seq_group_metadata_list) > 0
        input_tokens: List[int] = []
        input_positions: List[int] = []
        slot_mapping: List[int] = []
        seq_lens: List[int] = []
        multi_modal_input_list: List[torch.Tensor] = []

        for seq_group_metadata in seq_group_metadata_list:
            assert seq_group_metadata.is_prompt
            seq_ids = list(seq_group_metadata.seq_data.keys())
            assert len(seq_ids) == 1
            seq_id = seq_ids[0]

            seq_data = seq_group_metadata.seq_data[seq_id]
            prompt_tokens = seq_data.get_token_ids()
            computed_len = seq_data.get_num_computed_tokens()
            seq_len = len(prompt_tokens)

            seq_lens.append(seq_len)  # Prompt token num
            input_tokens.extend(prompt_tokens)  # Token ids

            # Token position ids
            # NOTE(woosuk): Here we assume that the first token in the prompt
            # is always the first token in the sequence.
            input_positions.extend(list(range(computed_len, seq_len)))

            if seq_group_metadata.multi_modal_data:
                multi_modal_input_list.append(
                    seq_group_metadata.multi_modal_data.data)

            if seq_group_metadata.block_tables is None:
                # During memory profiling, the block tables are not initialized
                # yet. In this case, we just use a dummy slot mapping.
                slot_mapping.extend([_PAD_SLOT_ID] * seq_len)
                continue

            # Compute the slot mapping.
            block_table = seq_group_metadata.block_tables[seq_id]
            # Mask the [0, start_idx) tokens of the prompt with _PAD_SLOT_ID,
            # where start_idx is max(0, seq_len - sliding_window).
            # For example, if the prompt len is 10, sliding window is 8, and
            # block size is 4, the first two tokens are masked and the slot
            # mapping will be [-1, -1, 2, 3, 4, 5, 6, 7, 0, 1].
            start_idx = 0
            if self.sliding_window is not None:
                start_idx = max(0, seq_len - self.sliding_window)

            for i in range(computed_len, seq_len):
                if i < start_idx:
                    slot_mapping.append(_PAD_SLOT_ID)
                    continue

                block_number = block_table[i //
                                           self.block_size]  # type: ignore
                block_offset = i % self.block_size  # type: ignore
                slot = block_number * self.block_size + block_offset
                slot_mapping.append(slot)

        if multi_modal_input_list:
            assert self.vision_language_config, (
                "Multi-modal inputs are only supported by "
                "vision language models.")
            multi_modal_input = torch.cat(multi_modal_input_list,
                                          dim=0).to(self.device)
        else:
            multi_modal_input = None

        num_prompt_tokens = len(input_tokens)

        input_tokens = torch.tensor(input_tokens,
                                    dtype=torch.long,
                                    device=self.device)  # type: ignore
        input_positions = torch.tensor(input_positions,
                                       dtype=torch.long,
                                       device=self.device)  # type: ignore
        slot_mapping = torch.tensor(slot_mapping,
                                    dtype=torch.long,
                                    device=self.device)  # type: ignore

        max_seqlen = max(seq_lens)
        tmp = [0]
        tmp.extend(seq_lens)
        seqlen = torch.tensor(tmp)
        seqlen_q = torch.cumsum(seqlen, dim=0).to(device=self.device)

        attn_metadata = self.attn_backend.make_metadata(
            is_prompt=True,
            slot_mapping=slot_mapping,
            seq_lens=seq_lens,
            seqlen_q=seqlen_q,
            max_seqlen=max_seqlen,
            seq_lens_tensor=None,
            max_decode_seq_len=None,
            num_prefills=len(seq_lens),
            num_prefill_tokens=num_prompt_tokens,
            num_decode_tokens=0,
            block_tables=torch.tensor([], device=self.device, dtype=torch.int),
        )
        return (input_tokens, input_positions, attn_metadata, seq_lens,
                multi_modal_input)