profile pic
⌘ '
raccourcis clavier
Search

Transformers

  • Étiquette

    • publié à

    • modifié à

    • durée

      6 min de lecture (1019 words)

    • source

      llms.txt

    See also: LLMs, embedding, visualisation from Brendan Bycroft

    A multi-layer perceptron (MLP) architecture built on top of a multi-head attention mechanism (Vaswani et al., 2023) to signal high entropy tokens to be amplified and less important tokens to be diminished.

    ELI5: Mom often creates a food list consists of nn of items to buy. Your job is to guess what the last item on this list would be.

    Most implementations are autoregressive. Most major SOTA are decoder-only, as encoder-decoder models has lack behind due to their expensive encoding phase.

    state-space models which address transformers’ efficiency issuesarXiv in attention layers within information-dense data

    memory limitations.

    excerpt from arxivarXiv

    inference.

    Either compute-bound (batch inference, saturated usage) or memory-bound (latency)

    speculative decoding ⇒ memory-bound (to saturate FLOPs)

    next-token prediction.

    Sampling: we essentially look forward K-tokens, and then we sample from the distribution of the next token.

    Byte-Latent Transformer

    idea: learn from raw-bytes and skip tokenizer/detokenizer protocol.

    Feynman-Kac

    Let V\mathcal{V} be the vocab of given transformers model, and S=V∗\mathcal{S} = \mathcal{V}^{*} the set of multi-token strings. Assume V\mathcal{V} contains token EOS and write F⊆S\mathcal{F} \subseteq \mathcal{S} for the set of EOS-terminated strings.

    Feynman-Kac Transformer model

    is a tuple (s0,{Mt}t≥1,{Gt}t≥1)(s_{0}, \{M_t\}_{t\ge 1}, \{G_t\}_{t\ge 1}) where:

    • s0∈Ss_{0} \in \mathcal{S} is an initial state, which will take as empty string ϵ\epsilon
    • Mt(st∣st−1,fθ)M_t(s_t \mid s_{t-1}, f_\theta) is a Markov kernel from st−1∈Fcs_{t-1} \in \mathcal{F}^c to st∈Ss_t \in \mathcal{S}, parameterised by a transformer network fθ:Fc→R∣V∣f_\theta: \mathcal{F}^c \to \mathbb{R}^{\mid \mathcal{V} \mid} mapping non-EOS-terminated strings to vectors of logits
    • Gt(st−1,st,fθ)G_t(s_{t-1}, s_t, f_\theta) is a potential function, mapping a pair (st−1,st)∈Fc×S(s_{t-1}, s_t) \in \mathcal{F}^c \times \mathcal{S} to a real-valued non-negative score.

    Goal: generate from distribution P\mathbb{P} that reweights Markove chain M\mathbb{M} by potential functions GtG_t. We define step-t filtering posteriors:

    Pt(st)=EM[∏i=1t∧TGi(Si−1,Si,fθ)⋅[St=st]]EM[∏i=1t∧TGi(Si−1,Si,fθ)]P_t(s_t) = \frac{\mathbb{E}_\mathbb{M} \left[ \prod_{i=1}^{t \wedge T} G_i(S_{i-1}, S_i, f_\theta) \cdot [S_t = s_t] \right]}{\mathbb{E}_\mathbb{M} \left[ \prod_{i=1}^{t \wedge T} G_i(S_{i-1}, S_i, f_\theta) \right]}

    Given that TT is mostly finite we can then define overall posterior P(s)=lim⁡t→∞Pt(s)\mathbb{P}(s) = \lim_{t \to \infty} \mathbb{P}_t(s) (Lew et al., 2023, p. see 2.2 for examples)

    "\\begin{algorithm}\n\\caption{Sequential Monte Carlo Transformer Steering}\n\\begin{algorithmic}\n\\State \\textbf{Input:} $N$ (\\# particles), $K$ (factor), Feynman-Kac Transformer model $\\{s_0, \\{M_t\\}_{t \\geq 1}, \\{G_t\\}_{t \\geq 1}\\}$\n\\State \\textbf{Output:} Weighted particle approximation $\\{(x_i, w_i)\\}_{i=1,\\ldots,N}$ of the posterior $\\mathbb{P}$ \\\\\n\\State \\textbf{Output:} Unbiased estimate $\\hat{Z}$ of the partition function $Z = \\mathbb{E}_\\mathbb{M}[\\prod_{t=1}^T G_t(s_t, s_{t-1}, f_\\theta)]$ \\\\\n\\State Initialize $f_\\theta \\gets \\texttt{CachedTransformer}()$\n\\State Initialize $(x_i, w_i) \\gets (s_0, 1)$ for $i = 1, \\ldots, N$\n\\State Initialize $t \\gets 1$\n\\While{$x_i \\not\\in \\mathcal{F}$ for some $i \\in \\{1, \\ldots, N\\}$}\n \\State $K_i \\gets K (1 - \\mathbb{1}_{\\mathcal{F}}(x_i)) + \\mathbb{1}_{\\mathcal{F}}(x_i)$ for $i = 1, \\ldots, N$\n \\State $N' \\gets \\sum_{i=1}^N K_i$\n \\For{$i \\in \\{1, \\ldots, N\\}$}\n \\If{$x_i \\in \\mathcal{F}$}\n \\State Set $(x_{i,1}, w_{i,1}) \\gets (x_i, w_i \\cdot \\frac{N'}{N})$\n \\Else\n \\State Generate $x_{i,k} \\sim M_t(\\cdot \\mid x_i, f_\\theta)$ for $k = 1, \\ldots, K$\n \\State Set $w_{i,k} \\gets w_i \\cdot G_t(x_i, x_{i,k}, f_\\theta) \\cdot \\frac{N'}{K N}$ for $k = 1, \\ldots, K$\n \\EndIf\n \\EndFor\n \\State Set normalized weights $\\hat{w}_{i,k} \\gets \\frac{w_{(i,k)}}{\\sum_{j=1}^N \\sum_{l=1}^{K_j} w_{(j,l)}}$ for $i = 1, \\ldots, N$ and $k = 1, \\ldots, K_i$\n \\State Set $c^* \\gets \\inf\\{c \\in \\mathbb{R}_{> 0} \\mid \\sum_{i=1}^N \\sum_{k=1}^{K_i} (\\mathbb{1} \\wedge c \\hat{w}_{(i,k)}) > N\\}$\n \\State Set $(I_\\text{det}, I_\\text{stoch}, I_\\text{strat}) \\gets (\\{(i,k) \\mid c^{*} \\hat{w}_{i,k} \\geq 1\\}, \\{(i,k) \\mid c^{*} \\cdot \\hat{w}_{i,k} < 1\\}, \\{\\})$\n \\State Set $\\alpha \\gets \\frac{\\sum_{i \\in I_\\text{stoch}} \\hat{w}_i}{|I_\\text{det}|}$ and generate $U \\sim \\text{Uniform}([0, \\alpha])$\n \\For{$i \\in I_\\text{stoch}$}\n \\State Set $U \\gets U - \\hat{w}_i$\n \\If{$U < 0$}\n \\State Set $I_\\text{strat} \\gets I_\\text{strat} \\cup \\{i\\}$\n \\State Set $U \\gets U + \\alpha$\n \\EndIf\n \\EndFor\n \\State Set particles $\\{(x_i, w_i)\\}_{i=1,\\ldots,|I_\\text{det}|} \\gets \\{(x_j, w_j \\cdot \\frac{N}{N'}) \\mid j \\in I_\\text{det}\\}$\n \\State Set particles $\\{(x_i, w_i)\\}_{i=|I_\\text{det}|+1,\\ldots,N} \\gets \\{(x_j, \\frac{N}{c^* N'} \\sum_{l=1}^{N} \\sum_{k=1}^{K_l} w_{(j,k)}) \\mid j \\in I_\\text{strat}\\}$\n\\EndWhile\n\\State \\Return $\\left((x_i, w_i)_{i=1,\\ldots,N}, \\hat{Z} = \\frac{1}{N} \\sum_{i=1}^N w_i \\right)$\n\\end{algorithmic}\n\\end{algorithm}"

    Algorithm 3 Sequential Monte Carlo Transformer Steering

    Input: NN (# particles), KK (factor), Feynman-Kac Transformer model {s0,{Mt}t≥1,{Gt}t≥1}\{s_0, \{M_t\}_{t \geq 1}, \{G_t\}_{t \geq 1}\}

    Output: Weighted particle approximation {(xi,wi)}i=1,…,N\{(x_i, w_i)\}_{i=1,\ldots,N} of the posterior P\mathbb{P}

    Output: Unbiased estimate Z^\hat{Z} of the partition function Z=EM[∏t=1TGt(st,st−1,fθ)]Z = \mathbb{E}_\mathbb{M}[\prod_{t=1}^T G_t(s_t, s_{t-1}, f_\theta)]

    Initialize fθ←CachedTransformer()f_\theta \gets \texttt{CachedTransformer}()

    Initialize (xi,wi)←(s0,1)(x_i, w_i) \gets (s_0, 1) for i=1,…,Ni = 1, \ldots, N

    Initialize t←1t \gets 1

    while xi∉Fx_i \not\in \mathcal{F} for some i∈{1,…,N}i \in \{1, \ldots, N\} do

    Ki←K(1−1F(xi))+1F(xi)K_i \gets K (1 - \mathbb{1}_{\mathcal{F}}(x_i)) + \mathbb{1}_{\mathcal{F}}(x_i) for i=1,…,Ni = 1, \ldots, N

    N′←∑i=1NKiN' \gets \sum_{i=1}^N K_i

    for i∈{1,…,N}i \in \{1, \ldots, N\} do

    if xi∈Fx_i \in \mathcal{F} then

    Set (xi,1,wi,1)←(xi,wi⋅N′N)(x_{i,1}, w_{i,1}) \gets (x_i, w_i \cdot \frac{N'}{N})

    else

    Generate xi,k∼Mt(⋅∣xi,fθ)x_{i,k} \sim M_t(\cdot \mid x_i, f_\theta) for k=1,…,Kk = 1, \ldots, K

    Set wi,k←wi⋅Gt(xi,xi,k,fθ)⋅N′KNw_{i,k} \gets w_i \cdot G_t(x_i, x_{i,k}, f_\theta) \cdot \frac{N'}{K N} for k=1,…,Kk = 1, \ldots, K

    end if

    end for

    Set normalized weights w^i,k←w(i,k)∑j=1N∑l=1Kjw(j,l)\hat{w}_{i,k} \gets \frac{w_{(i,k)}}{\sum_{j=1}^N \sum_{l=1}^{K_j} w_{(j,l)}} for i=1,…,Ni = 1, \ldots, N and k=1,…,Kik = 1, \ldots, K_i

    Set c∗←inf⁡{c∈R>0∣∑i=1N∑k=1Ki(1∧cw^(i,k))>N}c^* \gets \inf\{c \in \mathbb{R}_{> 0} \mid \sum_{i=1}^N \sum_{k=1}^{K_i} (\mathbb{1} \wedge c \hat{w}_{(i,k)}) > N\}

    Set (Idet,Istoch,Istrat)←({(i,k)∣c∗w^i,k≥1},{(i,k)∣c∗⋅w^i,k<1},{})(I_\text{det}, I_\text{stoch}, I_\text{strat}) \gets (\{(i,k) \mid c^{*} \hat{w}_{i,k} \geq 1\}, \{(i,k) \mid c^{*} \cdot \hat{w}_{i,k} < 1\}, \{\})

    Set α←∑i∈Istochw^i∣Idet∣\alpha \gets \frac{\sum_{i \in I_\text{stoch}} \hat{w}_i}{|I_\text{det}|} and generate U∼Uniform([0,α])U \sim \text{Uniform}([0, \alpha])

    for i∈Istochi \in I_\text{stoch} do

    Set U←U−w^iU \gets U - \hat{w}_i

    if U<0U < 0 then

    Set Istrat←Istrat∪{i}I_\text{strat} \gets I_\text{strat} \cup \{i\}

    Set U←U+αU \gets U + \alpha

    end if

    end for

    Set particles {(xi,wi)}i=1,…,∣Idet∣←{(xj,wj⋅NN′)∣j∈Idet}\{(x_i, w_i)\}_{i=1,\ldots,|I_\text{det}|} \gets \{(x_j, w_j \cdot \frac{N}{N'}) \mid j \in I_\text{det}\}

    Set particles {(xi,wi)}i=∣Idet∣+1,…,N←{(xj,Nc∗N′∑l=1N∑k=1Klw(j,k))∣j∈Istrat}\{(x_i, w_i)\}_{i=|I_\text{det}|+1,\ldots,N} \gets \{(x_j, \frac{N}{c^* N'} \sum_{l=1}^{N} \sum_{k=1}^{K_l} w_{(j,k)}) \mid j \in I_\text{strat}\}

    end while

    return ((xi,wi)i=1,…,N,Z^=1N∑i=1Nwi)\left((x_i, w_i)_{i=1,\ldots,N}, \hat{Z} = \frac{1}{N} \sum_{i=1}^N w_i \right)

    Bibliographie

    • Lew, A. K., Zhi-Xuan, T., Grand, G., & Mansinghka, V. K. (2023). Sequential Monte Carlo Steering of Large Language Models using Probabilistic Programs. arXiv preprint arXiv:2306.03081 [arXiv]arXiv
    • Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., & Polosukhin, I. (2023). Attention Is All You Need. arXiv preprint arXiv:1706.03762 [arXiv]arXiv

    Vous pourriez aimer ce qui suit

  • Connectionist networks
  • Phaedo and Apology
  • Algebraic geometry
  • PID Controller
  • Garbage in Garbage out
  • Very Slender Differences
  • Kolmogorov–Arnold representation theorem
    • colophon
    nobby designer word for 'the design of the current site'
    procedural notes.
    contact[at]aarnphm[dot]xyz
    On ChatGPT and its pedagogical consequences
    And its implication on how we assess learning. an overview.
    structured decoding, a guide for the impatient
    and vLLM integration with XGrammar
    Autoregressive models
    A statistical model is autoregressive if it predicts future values based on past values. For example, an autoregressive model might seek to predict a stock’s future prices based on its past performance.
    Embedding
    See also Transformers.
    Intelligence amplification
    I’m playing around with calling our tech, as it is today, IA (intelligence amplification) instead of AI. IA have the vibe of tools for thought, needing human interaction, and resemble a lot more what we actually have today.
    LLMs
    large language models, often implemented as autoregressive transformers models. GPTs and friends Most variants of LLMs are decoder-only (Radford et al., 2019) Have “capabilities” to understand natural language.
    The Llama 3 Herd of Model
    excerpt from the papers by Meta Research.
    Machine learning
    Detects pattern within data and use it to make useful prediction. Generally AI \subset ML \subset DL Some main exploration: Transformers Large language models NLP CNN Logistic regression Optimization gradient descent hyperparameter tuning ensemble learning Recommender systems Reinforcement learning Q-learning Policy Gradient Monte-Carlo Tree Search Generative Models GAN VAE Autoencoder sparse autoencoder sparse crosscoders Supervised Q-learning Low-rank adapters Fields mechanistic interpretability Related: linear algebra.
    Turing-complete Transformers
    Turing Complete Transformers: Two Transformers Are More Powerful Than One"We prove transformers are not Turing complete, propose a new architecture that is Turing complete, ...
    autoencoders
    feature extraction from neural network
    constrained decoding
    structured generations in vLLM a la carte
    craft.
    and celebration of the act of making.
    induction heads
    and mathematical framework of transformers
    mechanistic interpretability
    and reverse engineering neural networks.
    ml optimization
    A list of optimization functions that can be used in ML training to reduce loss, and more.