vall-e/vall_e/samplers.py

523 lines
18 KiB
Python

import math
import torch
import torch.nn.functional as F
import numpy as np
import time
from torch import Tensor, einsum, nn
from einops import rearrange
from dataclasses import asdict, dataclass, field
# Simple filter to modify a token's probability if it shows up in the past
# `one_time` will only apply the penalty once
# `decay` is a factor that will exponentially apply to how far away it is
def reptition_penalize( logits, previous, factor=1.0, decay=0.0, one_time=False, limit=75 ):
if factor == 1.0 or previous is None:
return logits
unique = set()
priors = reversed(previous)
for distance, token in enumerate(priors):
# rep-pen range
if limit and distance >= limit:
continue
# skip if we're only applying the decay once
if one_time and token in unique:
continue
distance += 1
logits[:, token] /= factor * (distance ** decay)
# add to set if we care about it
if one_time:
unique.add(token)
return logits
# Simple "filter" that modifies the logit for the stop token, based on the sequence length
# `length` is the length of the sequence currently
# `factor` is the power the length is raised to, so values > 0 will yield longer sequences, values < 0 will yield shorter sequences
# `token` is the stop token.
def length_penalize( logits, length, factor=0.0, token=-1 ):
if factor == 0.0:
return logits
logits[:, token] /= (length ** factor)
return logits
# Simple way to ban tokens
def ban_tokens( logits, tokens ):
for token in tokens:
# token not in logits
if logits.shape[-1] >= token:
continue
logits[:, token] = -float("inf")
return logits
# Performs min_p filtering
# From https://github.com/huggingface/transformers/blob/v4.45.2/src/transformers/generation/logits_process.py#L537
def min_p_filtering( logits, min_p=0.0, min_tokens_to_keep=32 ):
if min_p <= 0.0:
return logits
# Convert logits to probabilities
probs = torch.softmax(logits, dim=-1)
# Get the probability of the top token for each sequence in the batch
top_probs, _ = probs.max(dim=-1, keepdim=True)
# Calculate the actual min_p threshold by scaling min_p with the top token's probability
scaled_min_p = min_p * top_probs
sorted_indices = torch.argsort(logits, descending=True, dim=-1)
sorted_indices_to_remove = torch.gather(probs < scaled_min_p, dim=-1, index=sorted_indices)
sorted_indices_to_remove[..., :min_tokens_to_keep] = False
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
return logits.masked_fill(indices_to_remove, -float("inf"))
# Credit to https://github.com/microsoft/unilm/blob/master/xtune/src/transformers/modeling_utils.py#L1145 / https://gist.github.com/thomwolf/1a5a29f6962089e871b94cbd09daf317
def top_k_top_p_filtering( logits, top_k=0, top_p=1.0, filter_value=-float("Inf"), min_tokens=1 ):
"""Filter a distribution of logits using top-k and/or nucleus (top-p) filtering
Args:
logits: logits distribution shape (batch size, vocabulary size)
if top_k > 0: keep only top k tokens with highest probability (top-k filtering).
if top_p < 1.0: keep the top tokens with cumulative probability >= top_p (nucleus filtering).
Nucleus filtering is described in Holtzman et al. (http://arxiv.org/abs/1904.09751)
Make sure we keep at least min_tokens per batch example in the output
"""
if top_k > 0:
top_k = min(max(top_k, min_tokens), logits.size(-1)) # Safety check
# Remove all tokens with a probability less than the last token of the top-k
indices_to_remove = logits < torch.topk(logits, top_k)[0][..., -1, None]
logits[indices_to_remove] = filter_value
if top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(logits, descending=True)
cumulative_probs = torch.cumsum(F.softmax(sorted_logits, dim=-1), dim=-1)
# Remove tokens with cumulative probability above the threshold (token with 0 are kept)
sorted_indices_to_remove = cumulative_probs > top_p
if min_tokens > 1:
# Keep at least min_tokens (set to min_tokens-1 because we add the first one below)
sorted_indices_to_remove[..., :min_tokens] = 0
# Shift the indices to the right to keep also the first token above the threshold
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
# scatter sorted tensors to original indexing
indices_to_remove = sorted_indices_to_remove.scatter(1, sorted_indices, sorted_indices_to_remove)
logits[indices_to_remove] = filter_value
return logits
# credit to https://github.com/LostRuins/koboldcpp/pull/464 // https://github.com/kalomaze/koboldcpp/tree/dynamic-temp
def dynamic_temperature( logits, temperature=1.0, min_temperature = 0.0, k = 10, sigmoidCenterPoint = 0.5 ):
# loop over logits[:], as the NAR will have logits.shape[0] > 1
for i in range(logits.shape[0]):
sum_exp = 0.0
maximum = torch.max( logits[i] )
for logit in logits[i]:
sum_exp += math.exp( logit - maximum )
prob_max_token_before_temp = 1.0 / sum_exp
dynamic_temperature = temperature - (temperature - min_temperature) / (1 + math.exp(-k * (prob_max_token_before_temp - sigmoidCenterPoint)))
logits[i] /= dynamic_temperature
return logits
# picks the top K tokens amongst a batch of logits
# logits: [Tensor] list of logits
# candidates: [(batch, token)] list, where batch indicates the index of the logits the given token is from
def top_k_logits_list( logits_list, k ):
# ( batch, tokens ) => ( batch x tokens )
logits = torch.cat( logits_list )
candidates = list(torch.topk(logits.flatten(), k).indices.tolist()) # perform top-k across all logits
for i, index in enumerate(candidates):
t = []
N = np.prod(logits.size())
for n in logits.size():
N //= n
t.append(index // N)
index %= N
candidates[i] = tuple(t)
return candidates
# Credit to: https://github.com/basusourya/mirostat/
# performs mirostat-based sampling
# logits: Tensor of logit probabilities
# state: the mirostat state
def mirostat_sample( logits, state = None ):
def compute_k(prob, n, tau):
num = 0
den = 0
for i in range(100):
b = prob[i]/prob[i+1]
t = (i+2)/(i+1)
num += math.log(b)*math.log(t)
den += math.log(t)**2
s = num/den
eps = s-1
k = ((eps*(2**(tau)))/(1-n**(-eps)))**(1/s)
k = round(k)
return k
if "max_surprise" not in state:
state["max_surprise"] = state["tau"] * 2
if "error_surprise" not in state:
state["error_surprise"] = 0
if "running_total_surprise" not in state:
state["running_total_surprise"] = 0
sorted_logits, sorted_indices = torch.sort( logits[-1, :], descending=True )
prob_original = torch.softmax( sorted_logits, dim=-1 ).tolist()
k = compute_k(prob_original, state["n"], state["max_surprise"]) + 1
sorted_logits = sorted_logits[0:k]
sorted_indices = sorted_indices[0:k]
prob_topk = torch.softmax(sorted_logits, dim = 0)
prev_i = torch.multinomial(prob_topk, num_samples=1, replacement=True)
state["index_surprise"] = math.log2(1/prob_original[prev_i])
state["running_total_surprise"] += state["index_surprise"]
state["error_surprise"] = state["index_surprise"] - state["tau"]
state["max_surprise"] -= state["eta"] * state["error_surprise"]
state["token"] = sorted_indices[prev_i]
return state
# Credits to: https://github.com/oobabooga/text-generation-webui/pull/5677
# performs DRY sampling
# * (honestly it looks close to rep pen anyways but what do I know)
# `logits` are the scores used to sample against
# `previous` are the prior tokens to penalize with
# `factor` is the scalar multiplier
# `base` is the base number to raise to the (length - allowed_length)th power
# `allowed_length` limits the range to apply DRY to
def dry_sampling( logits, previous=None, factor=0.0, base=1.75, allowed_length=2 ):
if factor == 0.0 or previous is None:
return logits
lengths = {}
for i, token in enumerate( previous ):
length = 1
while length < max(allowed_length, 50):
j = i - length
# Start of input reached.
if j < 0:
break
# Start of match reached.
if previous[j] != previous[-length-1]:
break
length += 1
lengths[token] = max(length, lengths[token]) if token in lengths else length
for token, length in lengths.items():
if length < allowed_length:
break
logits[:, token] -= factor * base ** (length - allowed_length)
return logits
LN_2 = 0.69314718056 # ln(2) = 1.0 / LOG2_E
# Grabbed from https://github.com/xjdr-alt/entropix/blob/main/entropix/sampler.py
def calculate_entropix_metrics( logits, attentions=None, dim=-1, use_stats=False ):
"""Calculate the entropy and varentropy of the probability distribution using logsoftmax."""
log_probs = F.log_softmax(logits, dim=dim)
probs = torch.exp(log_probs)
entropy = -torch.sum(probs * log_probs, dim=dim) / LN_2 # Convert to base-2
varentropy = torch.sum(probs * (log_probs / LN_2 + entropy.unsqueeze(-1))**2, dim=dim)
if attentions is None:
return {
"logits_entropy": torch.mean(entropy).item(),
"logits_varentropy": torch.mean(varentropy).item(),
}
last_attention_scores = attentions[-1].unsqueeze(0) # ( bsz, heads, seq_len, seq_len )
attention_probs = F.softmax(last_attention_scores, dim=-1)
if use_stats:
attn_stats = AttnStats.new( 1, attentions.shape[0], attentions.shape[1], logits.device )
for idx, attn in enumerate( attentions ):
attn_stats.update( attn.unsqueeze(0)[:, :, -1, :], idx ) # (bsz, heads, last_token, seq_len)
attn_entropy = attn_stats.entropy
attn_varentropy = attn_stats.varentropy
else:
attn_entropy = -torch.sum(attention_probs * torch.log2(torch.clamp(attention_probs, 1e-10, 1.0)), dim=-1)
attn_varentropy = torch.var(attn_entropy, dim=1)
# Add a small epsilon to avoid NaN when all values are the same
attn_varentropy = torch.where(torch.isnan(attn_varentropy), torch.zeros_like(attn_varentropy), attn_varentropy)
mean_attention = torch.mean(attention_probs, dim=1)
agreement = torch.mean(torch.abs(attention_probs - mean_attention.unsqueeze(1)), dim=(1, 2))
interaction_strength = torch.mean(torch.abs(last_attention_scores), dim=(1, 2, 3))
return {
"logits_entropy": torch.mean(entropy).item(),
"logits_varentropy": torch.mean(varentropy).item(),
"attn_entropy": torch.mean(attn_entropy).item(),
"attn_varentropy": torch.mean(attn_varentropy).item(),
"agreement": torch.mean(agreement).item(),
"interaction_strength": interaction_strength.item(), # torch.mean(interaction_strength).item(),
"action": -1
}
from typing import NamedTuple
class AttnStats(NamedTuple):
entropy: torch.Tensor # (bsz, n_layers, num_heads)
varentropy: torch.Tensor # (bsz, n_layers, num_heads)
n_layers: int
n_heads: int
@classmethod
def new(cls, bsz: int, n_layers: int, n_heads: int, device = "cuda") -> 'AttnStats':
return cls(
entropy=torch.zeros((bsz, n_layers, n_heads), dtype=torch.float32, device=device),
varentropy=torch.zeros((bsz, n_layers, n_heads), dtype=torch.float32, device=device),
n_layers=n_layers,
n_heads=n_heads
)
@property
def avg_entropy(self):
return self.entropy.sum(dim=-1, keepdim=False) # Average across heads
@property
def avg_varentropy(self):
return self.varentropy.sum(dim=-1, keepdim=False) # Average across heads
@property
def std_error(self):
return torch.sqrt(torch.mean(self.varentropy)) / (self.n_heads * self.n_layers)
def update(self, scores: torch.Tensor, layer_idx: int):
# scores shape: (bsz, n_heads, seqlen, n_words)
probs = torch.nn.functional.softmax(scores, dim=-1)
new_entropy = -torch.sum(torch.where(probs > 0, probs * torch.log(probs), torch.tensor(0.0)), dim=-1)
new_varentropy = torch.sum(probs * (torch.log(probs) + new_entropy.unsqueeze(-1))**2, dim=-1)
# Update entropy and varentropy tensors
self.entropy[:, layer_idx, :] = new_entropy
self.varentropy[:, layer_idx, :] = new_varentropy
return self
# to-do: play around with these values
@dataclass()
class EntropixSamplerConfig:
temp: float = 0.666
top_p: float = 0.90
top_k: int = 27
min_p: float = 0.01 # was 0.03 # Turn this down to 0.01 to reduce the shoggoth
low_ent_thresh: float = 0.1 # 3.0
low_vent_thresh: float = 0.1 # 3.0
med_ent_thresh: float = 3.0 # 6.0
high_ent_thresh: float = 5.0 # 9.0
high_vent_thresh: float = 5.0 # 9.0
# TODO this is a bit of a nasty mess, but also makes all the hyperparameters visible
helv_attn_ent_offset: float = 1.3
helv_attn_ent_coef: float = 0.2
lehv_interaction_strength_offset: float = 1.2
lehv_interaction_strength_coef: float = 0.3
hehv_attn_ent_coef: float = 0.2
hehv_attn_vent_offset: float = 2.0
hehv_attn_vent_coef: float = 0.5
# TODO not convinced this should
n_adaptive_samples: int = 5
# Adaptive sampling parameters
ada_temp_logits: float = 0.3
ada_temp_attn: float = 0.2
ada_temp_agree: float = 0.2
ada_top_p: float = 0.1
ada_top_k_int: float = 0.3
ada_top_k_agree: float = 0.2
ada_min_p: float = 0.5
ada_score_logits_ent: float = 0.1
ada_score_attn_ent: float = 0.2
ada_score_logits_vent: float = 0.3
ada_score_attn_vent: float = 0.4
ada_score_agree: float = 0.5
ada_score_int: float = 0.6
# extra stuff
temperature_max: float = 1.25
temperature_min: float = 0.5
top_k_min: int = 1
top_k_max: int = 1024
top_p_min: int = 0.1
top_p_max: int = 1.0
min_p_min: int = 0.01
min_p_max: int = 0.5
Exponential = torch.distributions.exponential.Exponential(1.0)
# Doing as close to the original sampling method just to reduce variance
def _sample_entropix(
logits,
temperature=1.0,
top_k=0,
top_p=1.0,
min_p=0.0,
cfg=EntropixSamplerConfig(),
):
def clamp(n, lo, hi):
return max(lo, min(n, hi))
if top_k == 0:
top_k = logits.shape[-1]
logit = logits[-1, :]
temperature = clamp( float(temperature), cfg.temperature_min, cfg.temperature_max )
top_p = clamp( float(top_p), cfg.top_p_min, cfg.top_p_max )
top_k = clamp( int(top_k), cfg.top_k_min, cfg.top_k_max )
min_p = clamp( float(min_p), cfg.min_p_min, cfg.min_p_max )
probs = F.softmax(logit / temperature, dim=-1)
# Apply min_p sampling
if min_p > 0.0:
p_max = float(torch.max(probs, dim=-1, keepdim=True).values)
indices_to_remove = probs < (min_p * p_max)
logit = torch.where(indices_to_remove, torch.full_like(logit, float('-inf')), logit)
# Apply top-k sampling
top_k_probs, top_k_indices = torch.topk(probs, k=min(top_k, probs.shape[-1]))
probs_sort = torch.flip(top_k_probs, dims=[-1])
probs_idx = torch.flip(top_k_indices, dims=[-1])
probs_sum = torch.cumsum(probs_sort, dim=-1)
# Apply top-p sampling
mask = torch.where(probs_sum - probs_sort > top_p, torch.tensor(1.0, device=logit.device), torch.tensor(0.0, device=logit.device))
probs_sort = probs_sort * (1 - mask)
probs_sort = probs_sort / torch.sum(probs_sort, dim=-1, keepdim=True)
q = Exponential.sample(probs_sort.shape)
"""
# q = torch.rand(probs_sort.shape, generator=generator, device=probs_sort.device)
"""
next_token = torch.argmax(probs_sort / q, dim=-1, keepdim=True)
next_token_g = torch.take_along_dim(probs_idx, next_token, dim=-1)
return next_token_g
def sample_entropix(
logits,
attentions,
temperature=1.0,
top_k=27,
top_p=1.0,
min_p=0.0,
cfg=EntropixSamplerConfig(),
metrics_only=False,
):
"""
temperature = cfg.temp
top_k = cfg.top_k
top_p = cfg.top_p
"""
# logits: ( seq_len, vocab )
# attentions: ( layer, heads, seq_len, seq_len )
metrics = calculate_entropix_metrics( logits[-1:, :], attentions[:, :, -1:, :] )
ent, vent = metrics["logits_entropy"], metrics["logits_varentropy"]
attn_ent, attn_vent = metrics["attn_entropy"], metrics["attn_varentropy"]
agreement = metrics["agreement"]
interaction_strength = metrics["interaction_strength"]
# Low Entropy, Low Varentropy: "flowing with unspoken intent"
if ent < cfg.low_ent_thresh and vent < cfg.low_vent_thresh:
metrics["action"] = 0
res = logits[-1, :].argmax(dim=1)
# High Entropy, Low Varentropy: "treading carefully, asking clarifying questions"
elif ent > cfg.high_ent_thresh and vent < cfg.low_vent_thresh:
metrics["action"] = 1
# sample with slightly higher temperature
temperature *= cfg.helv_attn_ent_offset + cfg.helv_attn_ent_coef * attn_ent # Increase temperature based on attention entropy
res = _sample_entropix( logits, temperature, top_k, top_p, min_p, cfg=cfg )
# Low Entropy, High Varentropy: "exploring forks in the path"
elif ent < cfg.high_ent_thresh and vent > cfg.high_vent_thresh:
metrics["action"] = 2
temperature *= cfg.lehv_interaction_strength_offset + cfg.lehv_interaction_strength_coef * interaction_strength # Increase temperature based on interaction strength
top_k = max(5, int(top_k * (1 + 0.5 * (1 - agreement)))) # Increase top_k when agreement is low
res = _sample_entropix( logits, temperature, top_k, top_p, min_p, cfg=cfg )
# High Entropy, High Varentropy: "resampling in the mist"
elif ent > cfg.med_ent_thresh and vent > cfg.high_vent_thresh:
metrics["action"] = 3
# Use high temperature and adjusted top_p based on attention metrics
temperature *= cfg.hehv_attn_vent_offset + cfg.hehv_attn_vent_coef * attn_vent # Increase temperature based on attention varentropy
top_p = max(0.5, top_p - cfg.hehv_attn_ent_coef * attn_ent) # Decrease top_p when attention entropy is high
res = _sample_entropix( logits, temperature, top_k, top_p, min_p, cfg=cfg )
# Middle ground: use adaptive sampling
else:
metrics["action"] = 4
log_softmax = F.log_softmax(logits, dim=-1)
logits_uncertainty = ent + vent
attn_uncertainty = attn_ent + attn_vent
temperature *= 1 + cfg.ada_temp_logits * logits_uncertainty + cfg.ada_temp_attn * attn_uncertainty - cfg.ada_temp_agree * agreement
top_p = top_p * (1 + cfg.ada_top_p * attn_vent)
top_k = round(float(top_k * (1 + cfg.ada_top_k_int * interaction_strength - cfg.ada_top_k_agree * agreement)))
min_p = cfg.min_p * (1 - cfg.ada_min_p * logits_uncertainty)
samples = [ _sample_entropix( logits.clone(), temperature, top_k, top_p, min_p, cfg=cfg ) for _ in range(cfg.n_adaptive_samples) ]
def score_sample(sample):
one_hot = F.one_hot( sample, logits.shape[-1] )
log_prob = torch.sum(log_softmax * one_hot)
confidence_score = (
(1 - ent) * cfg.ada_score_logits_ent +
(1 - attn_ent) * cfg.ada_score_attn_ent +
(1 - vent) * cfg.ada_score_logits_vent +
(1 - attn_vent) * cfg.ada_score_attn_vent +
agreement * cfg.ada_score_agree +
interaction_strength * cfg.ada_score_int
)
"""
if 1024 in sample:
return 1000
"""
return log_prob + confidence_score
sample_scores = [ score_sample(sample) for sample in samples ]
best_sample_idx = torch.argmax(torch.asarray(sample_scores))
res = samples[best_sample_idx]
"""
metrics = {
"attn_entropy": metrics["attn_entropy"],
"attn_varentropy": metrics["attn_varentropy"],
}
"""
"""
metrics["temperature"] = temperature
metrics["top_k"] = top_k
metrics["top_p"] = top_p
metrics["min_p"] = min_p
"""
return res, metrics