app.py

from pathlib import Path
from typing import Any, Dict, Hashable

import streamlit as st
import streamlit.components.v1 as components
import numpy as np
import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer, BatchEncoding, GPT2LMHeadModel, PreTrainedTokenizer
from transformers import LogitsProcessorList, RepetitionPenaltyLogitsProcessor, TemperatureLogitsWarper, TopPLogitsWarper, TypicalLogitsWarper

from context_probing import estimate_unigram_logprobs
from context_probing.core import nll_score, kl_div_score
from context_probing.utils import columns_to_diagonals, get_windows, ids_to_readable_tokens

root_dir = Path(__file__).resolve().parent
highlighted_text_component = components.declare_component(
    "highlighted_text", path=root_dir / "highlighted_text" / "build"
)

compact_layout = st.experimental_get_query_params().get("compact", ["false"]) == ["true"]

if not compact_layout:
    st.title("Context length probing")
    st.markdown(
        """[📃 Paper](https://arxiv.org/abs/2212.14815) |
        [🌍 Website](https://cifkao.github.io/context-probing) |
        [🧑‍💻 Code](https://github.com/cifkao/context-probing)
        """
    )

generation_mode = st.radio(
    "Mode", ["Basic mode", "Generation mode"],
    horizontal=True, label_visibility="collapsed"
) == "Generation mode"
st.caption(
    "In basic mode, we analyze the model's one-step-ahead predictions on the input text. "
    "In generation mode, we generate a continuation of the input text (prompt) "
    "and analyze the model's predictions influencing the generated tokens."
)

model_name = st.selectbox(
    "Model",
    [
        "distilgpt2",
        "gpt2",
        "EleutherAI/gpt-neo-125m",
        "roneneldan/TinyStories-8M",
        "roneneldan/TinyStories-33M",
    ]
)
metric_name = st.radio(
    "Metric",
    (["KL divergence"] if not generation_mode else []) + ["NLL loss"],
    index=0,
    horizontal=True,
    help="**KL divergence** is computed between the predictions with the reduced context "
         "(corresponding to the highlighted token) and the predictions with the full context "
         "($c_\\text{max}$ tokens).  \n"
         "**NLL loss** is the negative log-likelihood loss (a.k.a. cross entropy) for the target "
         "token."
)

tokenizer = st.cache_resource(AutoTokenizer.from_pretrained, show_spinner=False)(model_name, use_fast=False)

# Make sure the logprobs do not use up more than ~4 GB of memory
MAX_MEM = 4e9 / (torch.finfo(torch.float16).bits / 8)
# Select window lengths such that we are allowed to fill the whole window without running out of memory
# (otherwise the window length is irrelevant); if using NLL, memory is not a consideration, but we want
# to limit runtime
multiplier = tokenizer.vocab_size if metric_name == "KL divergence" else 16384  # arbitrary number
window_len_options = [
    w for w in [8, 16, 32, 64, 128, 256, 512, 1024]
    if w == 8 or w * (2 * w) * multiplier <= MAX_MEM
]
window_len = st.select_slider(
    r"Window size ($c_\text{max}$)",
    options=window_len_options,
    value=min(128, window_len_options[-1]),
    help="The maximum context length $c_\\text{max}$ for which we compute the scores. Smaller "
         "windows are less computationally intensive, allowing for longer inputs."
)
# Now figure out how many tokens we are allowed to use:
# window_len * (num_tokens + window_len) * vocab_size <= MAX_MEM
max_tokens = int(MAX_MEM / (multiplier * window_len) - window_len)
max_tokens = min(max_tokens, 4096)

enable_null_context = st.checkbox(
    "Enable length-1 context",
    value=True,
    help="This enables computing scores for context length 1 (i.e. the previous token), which "
         "involves using an estimate of the model's unigram distribution. This is not originally "
         "proposed in the paper."
)

generate_kwargs = {}
with st.empty():
    if generation_mode:
        with st.expander("Generation options", expanded=False):
            generate_kwargs["max_new_tokens"] = st.slider(
                "Max. number of generated tokens",
                min_value=8, max_value=min(1024, max_tokens), step=8, value=min(128, max_tokens)
            )
            col1, col2, col3, col4 = st.columns(4)
            with col1:
                generate_kwargs["temperature"] = st.number_input(
                    min_value=0.01, value=0.9, step=0.05, label="`temperature`"
                )
            with col2:
                generate_kwargs["top_p"] = st.number_input(
                    min_value=0., value=0.95, max_value=1., step=0.05, label="`top_p`"
                )
            with col3:
                generate_kwargs["typical_p"] = st.number_input(
                    min_value=0., value=1., max_value=1., step=0.05, label="`typical_p`"
                )
            with col4:
                generate_kwargs["repetition_penalty"] = st.number_input(
                    min_value=1., value=1., step=0.05, label="`repetition_penalty`"
                )

DEFAULT_TEXT = """
We present context length probing, a novel explanation technique for causal
language models, based on tracking the predictions of a model as a function of the length of
available context, and allowing to assign differential importance scores to different contexts.
The technique is model-agnostic and does not rely on access to model internals beyond computing
token-level probabilities. We apply context length probing to large pre-trained language models
and offer some initial analyses and insights, including the potential for studying long-range
dependencies.
""".replace("\n", " ").strip()

with st.expander(
    f"Prompt" if generation_mode else f"Input text (≤\u2009{max_tokens} tokens)", expanded=True
):
    text = st.text_area(
        "Input text",
        st.session_state.get("input_text", DEFAULT_TEXT),
        key="input_text", label_visibility="collapsed"
    )

inputs = tokenizer([text])
[input_ids] = inputs["input_ids"]
label_ids = [*input_ids[1:], tokenizer.eos_token_id]
inputs["labels"] = [label_ids]
num_user_tokens = len(input_ids)

if num_user_tokens < 1:
    st.error("Please enter at least one token.", icon="🚨")
    st.stop()
if not generation_mode and num_user_tokens > max_tokens:
    st.error(
        f"Your input has {num_user_tokens} tokens. Please enter at most {max_tokens} tokens "
        f"or try reducing the window size.",
        icon="🚨"
    )
    st.stop()

with st.spinner("Loading model…"):
    model = st.cache_resource(AutoModelForCausalLM.from_pretrained, show_spinner=False)(model_name)

@st.cache_data(show_spinner=False)
def get_unigram_logprobs(
    _model: GPT2LMHeadModel,
    _tokenizer: PreTrainedTokenizer,
    model_name: str
):
    path = Path("data") / "unigram_logprobs" / f'{model_name.replace("/", "_")}.npy'
    if path.exists():
        return torch.as_tensor(np.load(path, allow_pickle=False))
    else:
        return estimate_unigram_logprobs(_model, _tokenizer)

if enable_null_context:
    with st.spinner("Obtaining unigram probabilities…"):
        unigram_logprobs = get_unigram_logprobs(model, tokenizer, model_name=model_name)
else:
    unigram_logprobs = torch.full((tokenizer.vocab_size,), torch.nan)
unigram_logprobs = tuple(unigram_logprobs.tolist())

@torch.inference_mode()
def get_logprobs(model, inputs, metric):
    logprobs = []
    batch_size = 8
    num_items = len(inputs["input_ids"])
    pbar = st.progress(0)
    for i in range(0, num_items, batch_size):
        pbar.progress(i / num_items, f"{i}/{num_items}")
        batch = {k: v[i:i + batch_size] for k, v in inputs.items()}
        batch_logprobs = model(**batch).logits.log_softmax(dim=-1).to(torch.float16)
        if metric != "KL divergence":
            batch_logprobs = torch.gather(
                batch_logprobs, dim=-1, index=batch["labels"][..., None]
            )
        logprobs.append(batch_logprobs)
    logprobs = torch.cat(logprobs, dim=0)
    pbar.empty()
    return logprobs

def get_logits_processor(temperature, top_p, typical_p, repetition_penalty) -> LogitsProcessorList:
    processor = LogitsProcessorList()
    if repetition_penalty != 1.0:
        processor.append(RepetitionPenaltyLogitsProcessor(penalty=repetition_penalty))
    if temperature != 1.0:
        processor.append(TemperatureLogitsWarper(temperature))
    if top_p < 1.0:
        processor.append(TopPLogitsWarper(top_p=top_p, min_tokens_to_keep=1))
    if typical_p < 1.0:
        processor.append(TypicalLogitsWarper(mass=typical_p, min_tokens_to_keep=1))
    return processor

@torch.inference_mode()
def generate(model, inputs, metric, window_len, max_new_tokens, **kwargs):
    assert metric == "NLL loss"
    start = max(0, inputs["input_ids"].shape[1] - window_len + 1)
    input_ids = inputs["input_ids"][:, start:]

    logits_warper = get_logits_processor(**kwargs)

    new_ids, logprobs = [], []
    eos_idx = None
    pbar = st.progress(0)
    max_steps = max_new_tokens + window_len - 1
    model_kwargs = dict(use_cache=True)
    for i in range(max_steps):
        pbar.progress(i / max_steps, f"{i}/{max_steps}")

        if input_ids.shape[1] == window_len:
            model_kwargs.update(use_cache=False)
            if "past_key_values" in model_kwargs:
                del model_kwargs["past_key_values"]
        model_inputs = model.prepare_inputs_for_generation(input_ids, **model_kwargs)
        model_outputs = model(**model_inputs)
        model_kwargs = model._update_model_kwargs_for_generation(model_outputs, model_kwargs, is_encoder_decoder=False)
        logits_window = model_outputs.logits.squeeze(0)
        logprobs_window = logits_window.log_softmax(dim=-1)
        if eos_idx is None:
            probs_next = logits_warper(input_ids, logits_window[[-1]]).softmax(dim=-1)
            next_token = torch.multinomial(probs_next, num_samples=1).item()
            if next_token == tokenizer.eos_token_id or i >= max_new_tokens - 1:
                eos_idx = i
        else:
            next_token = tokenizer.eos_token_id
        new_ids.append(next_token)

        input_ids = torch.cat([input_ids, torch.tensor([[next_token]])], dim=1)
        if input_ids.shape[1] > window_len:
            input_ids = input_ids[:, 1:]
        if logprobs_window.shape[0] == window_len:
            logprobs.append(
                logprobs_window[torch.arange(window_len), input_ids.squeeze(0)]
            )

        if eos_idx is not None and i - eos_idx >= window_len - 1:
            break
    pbar.empty()

    [input_ids] = input_ids.tolist()
    new_ids = new_ids[:eos_idx + 1]
    label_ids = [*input_ids, *new_ids][1:]

    return torch.as_tensor(new_ids), torch.as_tensor(label_ids), torch.stack(logprobs)[:, :, None]

@torch.inference_mode()
def run_context_length_probing(
    _model: GPT2LMHeadModel,
    _tokenizer: PreTrainedTokenizer,
    _inputs: Dict[str, torch.Tensor],
    window_len: int,
    unigram_logprobs: tuple,
    metric: str,
    generation_mode: bool,
    generate_kwargs: Dict[str, Any],
    cache_key: Hashable
):
    del cache_key

    [input_ids] = _inputs["input_ids"]
    [label_ids] = _inputs["labels"]

    with st.spinner("Running model…"):
        if generation_mode:
            new_ids, label_ids, logprobs = generate(
                model=_model,
                inputs=_inputs.convert_to_tensors("pt"), 
                metric=metric,
                window_len=window_len,
                **generate_kwargs
            )
            output_ids = [*input_ids, *new_ids]
            window_len = logprobs.shape[1]
        else:
            window_len = min(window_len, len(input_ids))
            inputs_sliding = get_windows(
                _inputs,
                window_len=window_len,
                start=0,
                pad_id=_tokenizer.eos_token_id
            ).convert_to_tensors("pt")
            logprobs = get_logprobs(model=_model, inputs=inputs_sliding, metric=metric)
            output_ids = [*input_ids, label_ids[-1]]
        num_tgt_tokens = logprobs.shape[0]

    with st.spinner("Computing scores…"):
        logprobs = logprobs.transpose(0, 1)
        logprobs = columns_to_diagonals(logprobs)
        logprobs = logprobs[:, :num_tgt_tokens]

        label_ids = label_ids[-num_tgt_tokens:]

        unigram_logprobs = torch.as_tensor(unigram_logprobs)
        unigram_logprobs[~torch.isfinite(unigram_logprobs)] = torch.nan
        if logprobs.shape[-1] == 1:
            unigram_logprobs = unigram_logprobs[label_ids].unsqueeze(-1)
        else:
            unigram_logprobs = unigram_logprobs.unsqueeze(0).repeat(num_tgt_tokens, 1)
        logprobs = torch.cat([unigram_logprobs.unsqueeze(0), logprobs], dim=0)

        if metric == "NLL loss":
            scores = nll_score(logprobs=logprobs, labels=label_ids, allow_overwrite=True)
        elif metric == "KL divergence":
            scores = kl_div_score(logprobs, labels=label_ids, allow_overwrite=True)
        del logprobs  # possibly overwritten by the score computation to save memory

        scores = (-scores).diff(dim=0).transpose(0, 1)
        scores = scores.nan_to_num()
        scores /= scores.abs().max(dim=1, keepdim=True).values + 1e-6
        scores = scores.to(torch.float16)

        if generation_mode:
            scores = F.pad(scores, (0, 0, max(0, len(input_ids) - window_len + 1), 0), value=0.)

    return output_ids, scores

if not generation_mode:
    run_context_length_probing = st.cache_data(run_context_length_probing, show_spinner=False)

if generation_mode:
    st.button("Rerun", type="primary")

output_ids, scores = run_context_length_probing(
    _model=model,
    _tokenizer=tokenizer,
    _inputs=inputs,
    window_len=window_len,
    unigram_logprobs=unigram_logprobs,
    metric=metric_name,
    generation_mode=generation_mode,
    generate_kwargs=generate_kwargs,
    cache_key=(model_name, text),
)
tokens = ids_to_readable_tokens(tokenizer, output_ids, strip_whitespace=False)

st.markdown('<label style="font-size: 14px;">Output</label>', unsafe_allow_html=True)
highlighted_text_component(
    tokens=tokens,
    scores=scores.tolist(),
    prefix_len=len(input_ids) if generation_mode else 0
)