llama_attn_replace_sft.py

# Modified based on https://github.com/lm-sys/FastChat

import warnings
from typing import Optional, Tuple

import torch
from torch import nn
import transformers
from einops import rearrange
from flash_attn import __version__ as flash_attn_version
from flash_attn.bert_padding import pad_input, unpad_input
from flash_attn.flash_attn_interface import (
    flash_attn_func,
    flash_attn_varlen_kvpacked_func,
    flash_attn_varlen_qkvpacked_func
)
from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv, rotate_half
from flash_attn.bert_padding import unpad_input, pad_input
import math

group_size_ratio = 1/4
sft_group_size = 8192

def forward_flashattn(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
    padding_mask: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    """Input shape: Batch x Time x Channel

    attention_mask: [bsz, q_len]
    """
    if not self.training:
        warnings.warn("This function should be used just for training as it may exhibit reduced inference performance. For inference, please use forward_flashattn_inference.")

    if output_attentions:
        warnings.warn(
            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
        )

    bsz, q_len, _ = hidden_states.size()

    query_states = (
        self.q_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    key_states = (
        self.k_proj(hidden_states)
        .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
        .transpose(1, 2)
    )
    value_states = (
        self.v_proj(hidden_states)
        .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
        .transpose(1, 2)
    )
    # [bsz, q_len, nh, hd]
    # [bsz, nh, q_len, hd]

    kv_seq_len = key_states.shape[-2]
    if past_key_value is not None:
        kv_seq_len += past_key_value[0].shape[-2]
    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
    query_states, key_states = apply_rotary_pos_emb(
        query_states, key_states, cos, sin, position_ids
    )

    # Past Key value support
    if past_key_value is not None:
        # reuse k, v, self_attention
        key_states = torch.cat([past_key_value[0], key_states], dim=2)
        value_states = torch.cat([past_key_value[1], value_states], dim=2)

    past_key_value = (key_states, value_states) if use_cache else None

    # repeat k/v heads if n_kv_heads < n_heads
    key_states = repeat_kv(key_states, self.num_key_value_groups)
    value_states = repeat_kv(value_states, self.num_key_value_groups)

    # Flash attention codes from
    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py

    # transform the data into the format required by flash attention
    qkv = torch.stack(
        [query_states, key_states, value_states], dim=2
    )  # [bsz, nh, 3, q_len, hd]
    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]

    # We have disabled _prepare_decoder_attention_mask in LlamaModel
    # the attention_mask should be the same as the key_padding_mask

    key_padding_mask = attention_mask.repeat(2, 1)
    nheads = qkv.shape[-2]
    # shift

    if q_len % 4096 == 0:
        group_size = int(q_len * group_size_ratio)
    else:
        group_size = sft_group_size

    qkv = qkv.reshape(bsz, q_len, 3, 2, self.num_heads // 2, self.head_dim).permute(0, 3, 1, 2, 4, 5).reshape(bsz * 2,
                                                                                                              q_len, 3,
                                                                                                              self.num_heads // 2,
                                                                                                              self.head_dim)

    x = rearrange(qkv, "b s three h d -> b s (three h d)")
    x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
    cu_q_len_tmp = torch.arange(0, max_s, group_size, device=key_padding_mask.device, dtype=cu_q_lens.dtype)
    cu_q_len_tmp2 = cu_q_len_tmp + group_size // 2
    cu_q_len_tmp2[cu_q_len_tmp2 >= max_s] = torch.iinfo(cu_q_len_tmp2.dtype).min
    cu_q_len_tmp = torch.stack([cu_q_len_tmp, cu_q_len_tmp2]).repeat(bsz, 1) + cu_q_lens[:-1].unsqueeze(-1)
    cu_q_lens = torch.cat([cu_q_len_tmp, cu_q_lens[1:].unsqueeze(-1)], dim=-1).view(-1)
    cu_q_lens = cu_q_lens[cu_q_lens >= 0]
    x_unpad = rearrange(
        x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=nheads // 2
    )
    output_unpad = flash_attn_varlen_qkvpacked_func(
        x_unpad, cu_q_lens, group_size, 0.0, softmax_scale=None, causal=True
    )
    output = rearrange(
        pad_input(
            rearrange(output_unpad, "nnz h d -> nnz (h d)"), indices, bsz * 2, q_len
        ),
        "b s (h d) -> b s h d",
        h=nheads // 2,
    )
    output = output.reshape(bsz, 2, q_len, nheads // 2, self.head_dim).transpose(1, 2).reshape(bsz, q_len, nheads,
                                                                                               self.head_dim)

    return self.o_proj(rearrange(output, "b s h d -> b s (h d)")), None, past_key_value

def forward_flashattn_full(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
    padding_mask: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    """Input shape: Batch x Time x Channel

    attention_mask: [bsz, q_len]
    """
    if output_attentions:
        warnings.warn(
            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
        )

    bsz, q_len, _ = hidden_states.size()

    query_states = (
        self.q_proj(hidden_states)
        .view(bsz, q_len, self.num_heads, self.head_dim)
        .transpose(1, 2)
    )
    key_states = (
        self.k_proj(hidden_states)
        .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
        .transpose(1, 2)
    )
    value_states = (
        self.v_proj(hidden_states)
        .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
        .transpose(1, 2)
    )
    # [bsz, q_len, nh, hd]
    # [bsz, nh, q_len, hd]

    kv_seq_len = key_states.shape[-2]
    if past_key_value is not None:
        kv_seq_len += past_key_value[0].shape[-2]
    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
    query_states, key_states = apply_rotary_pos_emb(
        query_states, key_states, cos, sin, position_ids
    )

    # Past Key value support
    if past_key_value is not None:
        # reuse k, v, self_attention
        key_states = torch.cat([past_key_value[0], key_states], dim=2)
        value_states = torch.cat([past_key_value[1], value_states], dim=2)

    past_key_value = (key_states, value_states) if use_cache else None

    # repeat k/v heads if n_kv_heads < n_heads
    key_states = repeat_kv(key_states, self.num_key_value_groups)
    value_states = repeat_kv(value_states, self.num_key_value_groups)

    # Flash attention codes from
    # https://github.com/HazyResearch/flash-attention/blob/main/flash_attn/flash_attention.py

    # transform the data into the format required by flash attention
    qkv = torch.stack(
        [query_states, key_states, value_states], dim=2
    )  # [bsz, nh, 3, q_len, hd]
    qkv = qkv.transpose(1, 3)  # [bsz, q_len, 3, nh, hd]

    # We have disabled _prepare_decoder_attention_mask in LlamaModel
    # the attention_mask should be the same as the key_padding_mask

    key_padding_mask = attention_mask
    nheads = qkv.shape[-2]
    x = rearrange(qkv, "b s three h d -> b s (three h d)")
    x_unpad, indices, cu_q_lens, max_s = unpad_input(x, key_padding_mask)
    x_unpad = rearrange(
        x_unpad, "nnz (three h d) -> nnz three h d", three=3, h=nheads
    )
    output_unpad = flash_attn_varlen_qkvpacked_func(
        x_unpad, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
    )
    output = rearrange(
        pad_input(
            rearrange(output_unpad, "nnz h d -> nnz (h d)"), indices, bsz, q_len
        ),
        "b s (h d) -> b s h d",
        h=nheads,
    )
    output = output.reshape(bsz, q_len, self.num_heads, self.head_dim)

    return self.o_proj(rearrange(output, "b s h d -> b s (h d)")), None, past_key_value


def forward_noflashattn(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.LongTensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
    padding_mask: Optional[torch.LongTensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    bsz, q_len, _ = hidden_states.size()

    group_size = int(q_len * group_size_ratio)

    if q_len % group_size > 0:
        raise ValueError("q_len %d should be divisible by group size %d."%(q_len, group_size))
    num_group = q_len // group_size

    if self.config.pretraining_tp > 1:
        key_value_slicing = (self.num_key_value_heads * self.head_dim) // self.config.pretraining_tp
        query_slices = self.q_proj.weight.split(
            (self.num_heads * self.head_dim) // self.config.pretraining_tp, dim=0
        )
        key_slices = self.k_proj.weight.split(key_value_slicing, dim=0)
        value_slices = self.v_proj.weight.split(key_value_slicing, dim=0)

        query_states = [F.linear(hidden_states, query_slices[i]) for i in range(self.config.pretraining_tp)]
        query_states = torch.cat(query_states, dim=-1)

        key_states = [F.linear(hidden_states, key_slices[i]) for i in range(self.config.pretraining_tp)]
        key_states = torch.cat(key_states, dim=-1)

        value_states = [F.linear(hidden_states, value_slices[i]) for i in range(self.config.pretraining_tp)]
        value_states = torch.cat(value_states, dim=-1)

    else:
        query_states = self.q_proj(hidden_states)
        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)

    kv_seq_len = key_states.shape[-2]
    if past_key_value is not None:
        kv_seq_len += past_key_value[0].shape[-2]
    cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
    query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)

    if past_key_value is not None:
        # reuse k, v, self_attention
        key_states = torch.cat([past_key_value[0], key_states], dim=2)
        value_states = torch.cat([past_key_value[1], value_states], dim=2)

    past_key_value = (key_states, value_states) if use_cache else None

    # repeat k/v heads if n_kv_heads < n_heads
    key_states = repeat_kv(key_states, self.num_key_value_groups)
    value_states = repeat_kv(value_states, self.num_key_value_groups)

    # shift
    def shift(qkv, bsz, q_len, group_size, num_heads, head_dim):
        qkv[:, num_heads // 2:] = qkv[:, num_heads // 2:].roll(-group_size // 2, dims=2)
        qkv = qkv.transpose(1, 2).reshape(bsz * (q_len // group_size), group_size, num_heads, head_dim).transpose(1, 2)
        return qkv

    query_states = shift(query_states, bsz, q_len, group_size, self.num_heads, self.head_dim)
    key_states = shift(key_states, bsz, q_len, group_size, self.num_heads, self.head_dim)
    value_states = shift(value_states, bsz, q_len, group_size, self.num_heads, self.head_dim)

    attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)

    if attn_weights.size() != (bsz * num_group, self.num_heads, group_size, group_size):
        raise ValueError(
            f"Attention weights should be of size {(bsz * num_group, self.num_heads, group_size, group_size)}, but is"
            f" {attn_weights.size()}"
        )

    attention_mask = attention_mask[:, :, :group_size, :group_size].repeat(num_group, 1, 1, 1)
    if attention_mask is not None:
        if attention_mask.size() != (bsz * num_group, 1, group_size, group_size):
            raise ValueError(
                f"Attention mask should be of size {(bsz * num_group, 1, group_size, group_size)}, but is {attention_mask.size()}"
            )
        attn_weights = attn_weights + attention_mask

    # upcast attention to fp32
    attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
    attn_output = torch.matmul(attn_weights, value_states)

    if attn_output.size() != (bsz * num_group, self.num_heads, group_size, self.head_dim):
        raise ValueError(
            f"`attn_output` should be of size {(bsz * num_group, self.num_heads, group_size, self.head_dim)}, but is"
            f" {attn_output.size()}"
        )
    attn_output = attn_output.transpose(1, 2).contiguous()

    attn_output = attn_output.reshape(bsz, q_len, self.num_heads, self.head_dim)

    # shift back
    attn_output[:, :, self.num_heads//2:] = attn_output[:, :, self.num_heads//2:].roll(group_size//2, dims=1)

    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)

    if self.config.pretraining_tp > 1:
        attn_output = attn_output.split(self.hidden_size // self.config.pretraining_tp, dim=2)
        o_proj_slices = self.o_proj.weight.split(self.hidden_size // self.config.pretraining_tp, dim=1)
        attn_output = sum([F.linear(attn_output[i], o_proj_slices[i]) for i in range(self.config.pretraining_tp)])
    else:
        attn_output = self.o_proj(attn_output)

    if not output_attentions:
        attn_weights = None

    return attn_output, attn_weights, past_key_value

# Disable the transformation of the attention mask in LlamaModel as the flash attention
# requires the attention mask to be the same as the key_padding_mask
def _prepare_decoder_attention_mask(
    self, attention_mask, input_shape, inputs_embeds, past_key_values_length
):
    # [bsz, seq_len]
    return attention_mask

def apply_rotary_pos_emb_inference(q, k, cos_sin, position_ids):
    gather_indices = position_ids[:, :, None, None]  # [bsz, seq_len, 1, 1]
    gather_indices = gather_indices.repeat(
        1, 1, cos_sin[0].shape[1], cos_sin[0].shape[3]
    )
    bsz = gather_indices.shape[0]
    cos, sin = (
        torch.gather(x.transpose(1, 2).repeat(bsz, 1, 1, 1), 1, gather_indices)
        for x in cos_sin
    )
    q, k = ((x * cos) + (rotate_half(x) * sin) for x in (q, k))
    return q, k


def forward_flashattn_inference(
    self,
    hidden_states: torch.Tensor,
    attention_mask: Optional[torch.Tensor] = None,
    position_ids: Optional[torch.Tensor] = None,
    past_key_value: Optional[Tuple[torch.Tensor]] = None,
    output_attentions: bool = False,
    use_cache: bool = False,
    padding_mask: Optional[torch.Tensor] = None,
) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
    if output_attentions:
        warnings.warn(
            "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
        )

    bsz, q_len, _ = hidden_states.size()
    kv_heads = getattr(self, "num_key_value_heads", self.num_heads)

    q, k, v = (
        op(hidden_states).view(bsz, q_len, nh, self.head_dim)
        for op, nh in (
            (self.q_proj, self.num_heads),
            (self.k_proj, kv_heads),
            (self.v_proj, kv_heads),
        )
    )
    # shape: (b, s, num_heads, head_dim)

    kv_seq_len = k.shape[1]
    past_kv_len = 0
    if past_key_value is not None:
        past_kv_len = past_key_value[0].shape[2]
        kv_seq_len += past_kv_len

    cos_sin = self.rotary_emb(v, seq_len=kv_seq_len)
    q, k = apply_rotary_pos_emb_inference(q, k, cos_sin, position_ids)

    if past_key_value is not None:
        assert (
            flash_attn_version >= "2.1.0"
        ), "past_key_value support requires flash-attn >= 2.1.0"
        # reuse k, v
        k = torch.cat([past_key_value[0].transpose(1, 2), k], dim=1)
        v = torch.cat([past_key_value[1].transpose(1, 2), v], dim=1)

    past_key_value = (k.transpose(1, 2), v.transpose(1, 2)) if use_cache else None

    if attention_mask is None:
        output = flash_attn_func(q, k, v, 0.0, softmax_scale=None, causal=True).view(
            bsz, q_len, -1
        )
    else:
        q, indices, cu_q_lens, max_s = unpad_input(q, attention_mask[:, -q_len:])
        # We can skip concat and call unpad twice but seems better to call unpad only once.
        kv, _, cu_k_lens, max_k = unpad_input(
            torch.stack((k, v), dim=2), attention_mask
        )
        output_unpad = flash_attn_varlen_kvpacked_func(
            q,
            kv,
            cu_q_lens,
            cu_k_lens,
            max_s,
            max_k,
            0.0,
            softmax_scale=None,
            causal=True,
        )
        output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
        output = pad_input(output_unpad, indices, bsz, q_len)

    return self.o_proj(output), None, past_key_value

def _prepare_decoder_attention_mask_inference(
    self, attention_mask, input_shape, inputs_embeds, past_key_values_length
):
    # [bsz, seq_len]
    if past_key_values_length > 0 and attention_mask is not None:
        attention_mask = torch.cat(
            (
                torch.full(
                    (input_shape[0], past_key_values_length),
                    True,
                    dtype=attention_mask.dtype,
                    device=attention_mask.device,
                ),
                attention_mask,
            ),
            dim=-1,
        )

    if attention_mask is not None and torch.all(attention_mask):
        return None  # This uses the faster call when training with full samples

    return attention_mask

def replace_llama_attn(use_flash_attn=True, use_full=False, inference=False):
    if use_flash_attn:
        cuda_major, cuda_minor = torch.cuda.get_device_capability()
        if cuda_major < 8:
            warnings.warn(
                "Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
                "ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
            )
        if inference:
            transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = _prepare_decoder_attention_mask_inference
            transformers.models.llama.modeling_llama.LlamaAttention.forward = forward_flashattn_inference
        else:
            transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
                _prepare_decoder_attention_mask
            )
            transformers.models.llama.modeling_llama.LlamaAttention.forward = forward_flashattn_full if use_full else forward_flashattn
    else:
        transformers.models.llama.modeling_llama.LlamaAttention.forward = forward_noflashattn