mlx-examples/llms/speculative_decoding/model.py

from typing import List, Optional, Tuple

import mlx.core as mx
import mlx.nn as nn
import numpy as np
from mlx.utils import tree_map, tree_unflatten
from transformers import AutoTokenizer, T5Config


def _relative_position_bucket(
    relative_position, bidirectional=True, num_buckets=32, max_distance=128
):
    """
    Adapted from HF Tensorflow:
    https://github.com/huggingface/transformers/blob/main/src/transformers/models/t5/modeling_t5.py

    Translate relative position to a bucket number for relative attention. The relative position is defined as
    memory_position - query_position, i.e. the distance in tokens from the attending position to the attended-to
    position. If bidirectional=False, then positive relative positions are invalid. We use smaller buckets for
    small absolute relative_position and larger buckets for larger absolute relative_positions. All relative
    positions >=max_distance map to the same bucket. All relative positions <=-max_distance map to the same bucket.
    This should allow for more graceful generalization to longer sequences than the model has been trained on

    Args:
        relative_position: an int32 Tensor
        bidirectional: a boolean - whether the attention is bidirectional
        num_buckets: an integer
        max_distance: an integer

    Returns:
        a Tensor with the same shape as relative_position, containing int32 values in the range [0, num_buckets)
    """
    relative_buckets = 0
    if bidirectional:
        num_buckets //= 2
        relative_buckets += (relative_position > 0).astype(mx.int16) * num_buckets
        relative_position = mx.abs(relative_position)
    else:
        relative_position = -mx.minimum(
            relative_position, mx.zeros_like(relative_position)
        )
    # now relative_position is in the range [0, inf)

    # half of the buckets are for exact increments in positions
    max_exact = num_buckets // 2
    is_small = relative_position < max_exact

    # The other half of the buckets are for logarithmically bigger bins in positions up to max_distance
    scale = (num_buckets - max_exact) / np.log(max_distance / max_exact)
    relative_position_if_large = max_exact + (
        mx.log(relative_position.astype(mx.float32) / max_exact) * scale
    ).astype(mx.int16)
    relative_position_if_large = mx.minimum(relative_position_if_large, num_buckets - 1)
    relative_buckets += mx.where(
        is_small, relative_position, relative_position_if_large
    )
    return relative_buckets


class RelativePositionBias(nn.Module):
    def __init__(self, config: T5Config, bidirectional: bool):
        self.bidirectional = bidirectional
        self.num_buckets = config.relative_attention_num_buckets
        self.max_distance = config.relative_attention_max_distance
        self.n_heads = config.num_heads
        self.embeddings = nn.Embedding(
            config.relative_attention_num_buckets, config.num_heads
        )

    def __call__(self, query_length: int, key_length: int, offset: int = 0):
        """Compute binned relative position bias"""
        context_position = mx.arange(offset, query_length)[:, None]
        memory_position = mx.arange(key_length)[None, :]

        # shape (query_length, key_length)
        relative_position = memory_position - context_position
        relative_position_bucket = _relative_position_bucket(
            relative_position,
            bidirectional=self.bidirectional,
            num_buckets=self.num_buckets,
            max_distance=self.max_distance,
        )

        # shape (query_length, key_length, num_heads)
        values = self.embeddings(relative_position_bucket)

        # shape (num_heads, query_length, key_length)
        return values.transpose(2, 0, 1)


class MultiHeadAttention(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        inner_dim = config.d_kv * config.num_heads
        self.num_heads = config.num_heads
        self.query_proj = nn.Linear(config.d_model, inner_dim, bias=False)
        self.key_proj = nn.Linear(config.d_model, inner_dim, bias=False)
        self.value_proj = nn.Linear(config.d_model, inner_dim, bias=False)
        self.out_proj = nn.Linear(inner_dim, config.d_model, bias=False)

    def __call__(
        self,
        queries: mx.array,
        keys: mx.array,
        values: mx.array,
        mask: Optional[mx.array],
        cache: Optional[Tuple[mx.array, mx.array]] = None,
    ) -> [mx.array, Tuple[mx.array, mx.array]]:
        queries = self.query_proj(queries)
        keys = self.key_proj(keys)
        values = self.value_proj(values)

        num_heads = self.num_heads
        B, L, _ = queries.shape
        _, S, _ = keys.shape
        queries = queries.reshape(B, L, num_heads, -1).transpose(0, 2, 1, 3)
        keys = keys.reshape(B, S, num_heads, -1).transpose(0, 2, 1, 3)
        values = values.reshape(B, S, num_heads, -1).transpose(0, 2, 1, 3)

        if cache is not None:
            key_cache, value_cache = cache
            keys = mx.concatenate([key_cache, keys], axis=2)
            values = mx.concatenate([value_cache, values], axis=2)

        # Dimensions are [batch x num heads x sequence x hidden dim]
        scores = queries @ keys.transpose(0, 1, 3, 2)
        if mask is not None:
            scores = scores + mask.astype(scores.dtype)

        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(scores.dtype)
        values_hat = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
        return self.out_proj(values_hat), (keys, values)


class DenseActivation(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        mlp_dims = config.d_ff or config.d_model * 4
        self.gated = config.feed_forward_proj.startswith("gated")
        if self.gated:
            self.wi_0 = nn.Linear(config.d_model, mlp_dims, bias=False)
            self.wi_1 = nn.Linear(config.d_model, mlp_dims, bias=False)
        else:
            self.wi = nn.Linear(config.d_model, mlp_dims, bias=False)
        self.wo = nn.Linear(mlp_dims, config.d_model, bias=False)
        activation = config.feed_forward_proj.removeprefix("gated-")
        if activation == "relu":
            self.act = nn.relu
        elif activation == "gelu":
            self.act = nn.gelu
        elif activation == "silu":
            self.act = nn.silu
        else:
            raise ValueError(f"Unknown activation: {activation}")

    def __call__(self, x):
        if self.gated:
            hidden_act = self.act(self.wi_0(x))
            hidden_linear = self.wi_1(x)
            x = hidden_act * hidden_linear
        else:
            x = self.act(self.wi(x))
        return self.wo(x)


class TransformerEncoderLayer(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        self.attention = MultiHeadAttention(config)
        self.ln1 = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.ln2 = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dense = DenseActivation(config)

    def __call__(self, x, mask):
        y = self.ln1(x)
        y, _ = self.attention(y, y, y, mask=mask)
        x = x + y

        y = self.ln2(x)
        y = self.dense(y)
        return x + y


class TransformerEncoder(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        self.layers = [
            TransformerEncoderLayer(config) for i in range(config.num_layers)
        ]
        self.ln = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.relative_attention_bias = RelativePositionBias(config, bidirectional=True)

    def __call__(self, x: mx.array):
        pos_bias = self.relative_attention_bias(x.shape[1], x.shape[1])
        for layer in self.layers:
            x = layer(x, mask=pos_bias)
        return self.ln(x)


class TransformerDecoderLayer(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        self.self_attention = MultiHeadAttention(config)
        self.cross_attention = MultiHeadAttention(config)
        self.ln1 = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.ln2 = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.ln3 = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.dense = DenseActivation(config)

    def __call__(
        self,
        x: mx.array,
        memory: mx.array,
        mask: mx.array,
        memory_mask: mx.array,
        cache: Optional[List[Tuple[mx.array, mx.array]]] = None,
    ):
        y = self.ln1(x)
        y, cache = self.self_attention(y, y, y, mask, cache)
        x = x + y

        y = self.ln2(x)
        y, _ = self.cross_attention(y, memory, memory, memory_mask)
        x = x + y

        y = self.ln3(x)
        y = self.dense(y)
        x = x + y

        return x, cache


def create_additive_causal_mask(N: int, offset: int = 0):
    rinds = mx.arange(offset + N)
    linds = mx.arange(offset, offset + N) if offset else rinds
    mask = linds[:, None] < rinds[None]
    return mask * -1e9


class TransformerDecoder(nn.Module):
    def __init__(self, config: T5Config):
        super().__init__()
        n_layers = getattr(config, "num_decoder_layers", config.num_layers)
        self.layers = [TransformerDecoderLayer(config) for i in range(n_layers)]
        self.ln = nn.RMSNorm(config.d_model, eps=config.layer_norm_epsilon)
        self.relative_attention_bias = RelativePositionBias(config, bidirectional=False)

    def __call__(self, x, memory, cache=None):
        if cache[0] is not None:
            offset = cache[0][0].shape[2]
        else:
            offset = 0

        T = x.shape[1]
        if T > 1:
            mask = create_additive_causal_mask(T, offset)
        else:
            mask = None

        pos_bias = self.relative_attention_bias(T + offset, T + offset, offset=offset)
        if mask is not None:
            mask += pos_bias
        else:
            mask = pos_bias

        for e, layer in enumerate(self.layers):
            x, cache[e] = layer(x, memory, mask, None, cache=cache[e])
        x = self.ln(x)

        return x, cache


class OutputHead(nn.Module):
    def __init__(self, config: T5Config):
        self.linear = nn.Linear(config.d_model, config.vocab_size, bias=False)

    def __call__(self, inputs):
        return self.linear(inputs)


class Model(nn.Module):
    def __init__(self, config: T5Config):
        self.wte = nn.Embedding(config.vocab_size, config.d_model)
        self.encoder = TransformerEncoder(config)
        self.decoder = TransformerDecoder(config)
        self.tie_word_embeddings = config.tie_word_embeddings
        if not self.tie_word_embeddings:
            self.lm_head = OutputHead(config)
        self.model_dim = config.d_model
        self.reset_cache()

    def encode(self, inputs: mx.array):
        return self.encoder(self.wte(inputs))

    def truncate_cache(self, num_to_truncate):
        if num_to_truncate <= 0:
            return
        cache_length = self.cache[0][0].shape[2]
        if num_to_truncate < cache_length:
            self.cache = tree_map(lambda x: x[:, :, :-num_to_truncate, :], self.cache)
        else:
            self.reset_cache()

    def reset_cache(self):
        self.cache = [None] * len(self.decoder.layers)

    def decode(
        self,
        inputs: mx.array,
        memory: mx.array,
    ):
        inputs = self.wte(inputs)
        y, self.cache = self.decoder(inputs, memory=memory, cache=self.cache)
        if not self.tie_word_embeddings:
            y *= self.model_dim**-0.5
            y = self.lm_head(y)
        else:
            y = y @ self.wte.weight.T
        return y

    def __call__(
        self,
        inputs: mx.array,
        decoder_inputs: mx.array,
    ):
        return self.decode(decoder_inputs, self.encode(inputs))[0]