verl/models/transformers/qwen2_vl.py

# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# 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.

import inspect
import os
from typing import Optional, Tuple

import torch
from transformers.modeling_flash_attention_utils import _flash_attention_forward
from transformers.utils import is_flash_attn_greater_or_equal

from verl.utils.ulysses import (
    gather_heads_scatter_seq,
    gather_seq_scatter_heads,
    get_ulysses_sequence_parallel_world_size,
    validate_ulysses_config,
)

try:
    from flash_attn import flash_attn_func, flash_attn_varlen_func

    _flash_supports_window_size = "window_size" in list(inspect.signature(flash_attn_func).parameters)
except ImportError:
    flash_attn_varlen_func = None


def get_rope_index(
    processor,
    input_ids: torch.Tensor,
    image_grid_thw: Optional[torch.Tensor] = None,
    video_grid_thw: Optional[torch.Tensor] = None,
    second_per_grid_ts: Optional[torch.Tensor] = None,
    attention_mask: Optional[torch.Tensor] = None,
) -> torch.Tensor:
    """
    Gets the position ids for Qwen2-VL, it should be generated before sharding the sequence.
    The batch dim has been removed and the input_ids should be a 1D tensor representing a single example.
    https://github.com/huggingface/transformers/blob/v4.49.0/src/transformers/models/qwen2_5_vl/modeling_qwen2_5_vl.py#L1546
    """
    spatial_merge_size = processor.image_processor.merge_size
    tokens_per_second = 2
    image_token_id = processor.tokenizer.convert_tokens_to_ids("<|image_pad|>")
    video_token_id = processor.tokenizer.convert_tokens_to_ids("<|video_pad|>")
    vision_start_token_id = processor.tokenizer.convert_tokens_to_ids("<|vision_start|>")
    if input_ids is not None and (image_grid_thw is not None or video_grid_thw is not None):
        if attention_mask is None:
            attention_mask = torch.ones_like(input_ids)

        position_ids = torch.ones(3, input_ids.size(0), dtype=input_ids.dtype, device=input_ids.device)  # (3, seqlen)
        image_index, video_index = 0, 0
        input_ids = input_ids[attention_mask == 1]
        image_nums, video_nums = 0, 0
        vision_start_indices = torch.argwhere(input_ids == vision_start_token_id)
        vision_tokens = input_ids[vision_start_indices + 1]
        image_nums = (vision_tokens == image_token_id).sum()
        video_nums = (vision_tokens == video_token_id).sum()
        input_tokens = input_ids.tolist()
        llm_pos_ids_list: list = []
        st = 0
        remain_images, remain_videos = image_nums, video_nums
        for _ in range(image_nums + video_nums):
            if image_token_id in input_tokens and remain_images > 0:
                ed_image = input_tokens.index(image_token_id, st)
            else:
                ed_image = len(input_tokens) + 1
            if video_token_id in input_tokens and remain_videos > 0:
                ed_video = input_tokens.index(video_token_id, st)
            else:
                ed_video = len(input_tokens) + 1
            if ed_image < ed_video:
                t, h, w = (
                    image_grid_thw[image_index][0],
                    image_grid_thw[image_index][1],
                    image_grid_thw[image_index][2],
                )
                second_per_grid_t = 0
                image_index += 1
                remain_images -= 1
                ed = ed_image
            else:
                t, h, w = (
                    video_grid_thw[video_index][0],
                    video_grid_thw[video_index][1],
                    video_grid_thw[video_index][2],
                )
                second_per_grid_t = second_per_grid_ts[video_index] if second_per_grid_ts is not None else 1.0

                video_index += 1
                remain_videos -= 1
                ed = ed_video

            llm_grid_t, llm_grid_h, llm_grid_w = (
                t.item(),
                h.item() // spatial_merge_size,
                w.item() // spatial_merge_size,
            )
            text_len = ed - st

            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

            t_index = torch.arange(llm_grid_t).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w)
            t_index = (t_index * second_per_grid_t * tokens_per_second).long().flatten()
            h_index = torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
            w_index = torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
            llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
            st = ed + llm_grid_t * llm_grid_h * llm_grid_w

        if st < len(input_tokens):
            st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
            text_len = len(input_tokens) - st
            llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)

        llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
        position_ids[..., attention_mask == 1] = llm_positions.to(position_ids.device)
    else:
        if attention_mask is not None:
            position_ids = attention_mask.long().cumsum(-1) - 1
            position_ids.masked_fill_(attention_mask == 0, 1)
            position_ids = position_ids.unsqueeze(0).expand(3, -1).to(input_ids.device)
        else:
            position_ids = torch.arange(input_ids.shape[1], device=input_ids.device).view(1, -1).expand(3, -1)

    return position_ids


def prepare_fa2_from_position_ids(query: torch.Tensor, key: torch.Tensor, value: torch.Tensor, position_ids: torch.Tensor):
    query = query.view(-1, query.size(-2), query.size(-1))
    key = key.view(-1, key.size(-2), key.size(-1))
    value = value.view(-1, value.size(-2), value.size(-1))
    position_ids = position_ids.flatten()
    indices_q = torch.arange(position_ids.size(0), device=position_ids.device, dtype=torch.int32)
    cu_seqlens = torch.cat(
        (
            indices_q[position_ids == 0],
            torch.tensor(position_ids.size(), device=position_ids.device, dtype=torch.int32),
        )
    )
    max_length = cu_seqlens.diff().max()  # use cu_seqlens to infer max_length for qwen2vl mrope
    return (query, key, value, indices_q, (cu_seqlens, cu_seqlens), (max_length, max_length))


def flash_attention_forward(
    query_states: torch.Tensor,
    key_states: torch.Tensor,
    value_states: torch.Tensor,
    attention_mask: torch.Tensor,
    query_length: int,
    is_causal: bool = True,
    position_ids: Optional[torch.Tensor] = None,
    sliding_window: Optional[int] = None,
    use_top_left_mask: bool = False,
    deterministic: Optional[bool] = None,
    **kwargs,
):
    """
    Patches flash attention forward to handle 3D position ids in mrope. (3, batch_size, seq_length)
    """
    causal = is_causal if not use_top_left_mask else is_causal and query_length != 1

    # Assuming 4D tensors, key_states.shape[1] is the key/value sequence length (source length).
    use_sliding_windows = _flash_supports_window_size and sliding_window is not None and key_states.shape[1] > sliding_window
    flash_kwargs = {"window_size": (sliding_window, sliding_window)} if use_sliding_windows else {}

    if is_flash_attn_greater_or_equal("2.4.1"):
        if deterministic is None:
            deterministic = os.environ.get("FLASH_ATTENTION_DETERMINISTIC", "0") == "1"
        flash_kwargs["deterministic"] = deterministic

    if position_ids is not None and query_length != 1 and not (torch.diff(position_ids[0], dim=-1) >= 0).all():
        batch_size = query_states.size(0)
        query_states, key_states, value_states, _, cu_seq_lens, max_seq_lens = prepare_fa2_from_position_ids(query_states, key_states, value_states, position_ids[0])  # remove channel dimension
        cu_seqlens_q, cu_seqlens_k = cu_seq_lens
        max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
        attn_output = flash_attn_varlen_func(
            query_states,
            key_states,
            value_states,
            cu_seqlens_q=cu_seqlens_q,
            cu_seqlens_k=cu_seqlens_k,
            max_seqlen_q=max_seqlen_in_batch_q,
            max_seqlen_k=max_seqlen_in_batch_k,
            dropout_p=kwargs.pop("dropout", 0.0),
            softmax_scale=kwargs.pop("softmax_scale", None),
            causal=causal,
            **flash_kwargs,
        )
        attn_output = attn_output.view(batch_size, -1, attn_output.size(-2), attn_output.size(-1))
    else:
        attn_output = _flash_attention_forward(
            query_states,
            key_states,
            value_states,
            attention_mask,
            query_length,
            is_causal=is_causal,
            sliding_window=sliding_window,
            use_top_left_mask=use_top_left_mask,
            deterministic=deterministic,
            **kwargs,
        )  # do not pass position_ids to old flash_attention_forward

    return attn_output


def ulysses_flash_attn_forward(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]] = None,  # will become mandatory in v4.46
    **kwargs,
) -> Tuple[torch.Tensor, None, None]:
    from transformers.models.qwen2_vl.modeling_qwen2_vl import apply_multimodal_rotary_pos_emb, repeat_kv

    bsz, q_len, _ = hidden_states.size()  # q_len = seq_length / sp_size
    query_states = self.q_proj(hidden_states)  # (batch_size, seq_length / sp_size, num_heads * head_size)
    key_states = self.k_proj(hidden_states)
    value_states = self.v_proj(hidden_states)

    query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
    key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
    value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)

    ulysses_sp_size = get_ulysses_sequence_parallel_world_size()

    if ulysses_sp_size > 1:
        validate_ulysses_config(self.num_heads, ulysses_sp_size)

        key_states = repeat_kv(key_states, self.num_key_value_groups)
        value_states = repeat_kv(value_states, self.num_key_value_groups)
        query_states = gather_seq_scatter_heads(query_states, seq_dim=2, head_dim=1)
        key_states = gather_seq_scatter_heads(key_states, seq_dim=2, head_dim=1)
        value_states = gather_seq_scatter_heads(value_states, seq_dim=2, head_dim=1)
        # (batch_size, num_head / sp_size, seq_length, head_size)
        full_q_len = query_states.size(2)  # full_q_len = seq_length
    else:
        full_q_len = q_len

    # Because the input can be padded, the absolute sequence length depends on the max position id.
    if position_embeddings is None:
        cos, sin = self.rotary_emb(value_states, position_ids)
    else:
        cos, sin = position_embeddings

    query_states, key_states = apply_multimodal_rotary_pos_emb(query_states, key_states, cos, sin, self.rope_scaling["mrope_section"])
    dropout_rate = 0.0 if not self.training else self.attention_dropout

    # Reashape to the expected shape for Flash Attention
    query_states = query_states.transpose(1, 2)
    key_states = key_states.transpose(1, 2)
    value_states = value_states.transpose(1, 2)

    if self.config.use_sliding_window and getattr(self.config, "sliding_window", None) is not None and self.layer_idx >= self.config.max_window_layers:
        sliding_window = self.config.sliding_window
    else:
        sliding_window = None

    attn_output = flash_attention_forward(
        query_states,
        key_states,
        value_states,
        attention_mask,
        full_q_len,
        dropout=dropout_rate,
        sliding_window=sliding_window,
        is_causal=self.is_causal,
        use_top_left_mask=self._flash_attn_uses_top_left_mask,
        position_ids=position_ids,  # important: pass position ids
    )  # (batch_size, seq_length, num_head / sp_size, head_size)
    if ulysses_sp_size > 1:
        attn_output = gather_heads_scatter_seq(attn_output, head_dim=2, seq_dim=1)

    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
    attn_output = self.o_proj(attn_output)
    return attn_output, None, None