43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247 | class Power_Sampling:
"""
Power Sampling Class that stores the parameters and methods used for power sampling.
Attributes:
block_size (int): How many tokens to divide the total output tokens into for power sampling. number of blocks = (sampler.sampling_params.max_tokens)/block_size. Smaller block sizes = better quality but slower
MCMC_steps (int): Number of MCMC steps to perform per block. More steps = better quality but slower
token_metric (str): Metric to use for token selection. Can be "logprobs", "power_distribution", "entropy", or "likelihood_confidence"
"""
def __init__(
self,
block_size: int = 192, # How many tokens to divide the total output tokens into for power sampling. Smaller block sizes = better quality but slower
MCMC_steps: int = 8, # Number of MCMC steps to perform per block. More steps = better quality but slower
token_metric: str = "power_distribution"
):
self.block_size = block_size
self.MCMC_steps = MCMC_steps
self.token_metric = token_metric
# TODO Implement entropy as a MCMC acceptance ratio metric
# TODO Implement a PRM as a MCMC acceptance ratio metric
# TODO Implement a separate temperature for the Power Distribution metric
# Power Sampling method
def sample(
self,
sampler: AutoregressiveSampler,
prompt: str,
logging: bool = False,
log_file_path: str = None
)-> Output:
"""
Sample using power sampling.
Args:
sampler (AutoregressiveSampler): The sampler object containing sampling parameters and the LLM engine.
prompt (str): The prompt to sample from.
logging (bool, optional): Whether to log the sampling process. Defaults to False.
log_file_path (str, optional): The path to the log file. Defaults to None.
Returns:
Output (Output): The output of the sampling process.
"""
# Set the random seed for reproducibility
if sampler.sampling_params.seed is not None:
np.random.seed(sampler.sampling_params.seed)
random.seed(sampler.sampling_params.seed)
# Statistic Logging in a CSV
if(logging):
#create or overwrite log file
power_sampling_log_path = log_file_path if log_file_path is not None else f"power_sampling_log_{time.strftime('%H%M%S_%d_%m_%Y')}.csv"
with open(power_sampling_log_path, "w") as log_file:
# Extract backslash expressions for Python 3.10 compatibility
sampler_json = json.dumps(vars(sampler), default=str).replace('"', '""')
prompt_escaped = prompt.replace('"', '""')
log_file.write(f'"{sampler_json}"\n')
log_file.write(f'"{prompt_escaped}"\n')
log_file.write("proposed_target_distribution_sum,proposed_sampling_distribution_sum,current_target_distribution_sum,current_sampling_distribution_sum,new_target_distribution_normalized,new_sampling_distribution_normalized,acceptance_ratio,accepted,starting_index,tokens_generated,\n")
# Intialize arrays to store the probabilities of the current tokens
current_target_distribution = [] # Current list of unscaled log probabilities of the new sample. Length of block_size
current_sampling_distribution = [] # Current list of tokens probabilities individually scaled by temperature. Length of block_size
# New Context Window to be changed
context = []
logits = []
logprobs = []
unprocessed_log_normalization_constant = []
temp_processed_log_normalization_constant = []
entropy = []
# Number of blocks to be sampled
block_count = sampler.sampling_params.max_tokens // self.block_size
sampler_max_tokens = sampler.sampling_params.max_tokens
for block_idx in range(block_count):
# Set the max tokens for the block
sampler.sampling_params.max_tokens = self.block_size
# Sample the initial new tokens for the block
output = sampler.sample(prompt + sampler.tokenizer.decode(context, skip_special_tokens=False))
# Calculate the log probabilities of the initial new tokens for the block
target_distribution = calc_token_metric(output, sampler, self.token_metric)
sampling_distribution = calc_token_metric(output, sampler, "logprobs")
# Extend the distributions
current_target_distribution = [*current_target_distribution, *target_distribution.tolist()]
current_sampling_distribution = [*current_sampling_distribution, *sampling_distribution.tolist()]
# Extend the context with the newly generated tokens
context.extend(output.tokens)
# Extend the other Output attributes along
logits.extend(output.top_k_logits)
logprobs.extend(output.top_k_logprobs)
unprocessed_log_normalization_constant.extend(output.unprocessed_log_normalization_constant)
temp_processed_log_normalization_constant.extend(output.temp_processed_log_normalization_constant)
entropy.extend(output.entropy)
# Log Results
if(logging):
proposed_target_distribution_sum = "None"
proposed_sampling_distribution_sum = "None "
current_target_distribution_sum = "None"
current_sampling_distribution_sum = "None"
new_target_distribution_normalized = sum(target_distribution)/len(target_distribution)
new_sampling_distribution_normalized = sum(sampling_distribution)/len(sampling_distribution)
acceptance_ratio = "None"
accepted = "None"
tokens_generated = len(output.tokens)
starting_index = len(context)-tokens_generated
# Write initial generated block data to log
with open(power_sampling_log_path, "a") as log_file:
log_file.write(f"{proposed_target_distribution_sum},{proposed_sampling_distribution_sum},{current_target_distribution_sum},{current_sampling_distribution_sum},{new_target_distribution_normalized},{new_sampling_distribution_normalized},{acceptance_ratio},{accepted},{starting_index},{tokens_generated}\n")
# Perform the MCMC Steps to hone in on the target distribution
for _ in range(self.MCMC_steps):
#Find a new point to start a proposal from. Generate idx tokens for the step.
idx = random.randint(0, len(context) - 1)
#Set the new context for the proposed block
context_proposed = context[:idx]
# Set the tokens to generate
sampler.sampling_params.max_tokens = len(context) - idx
#Generate proposed block of tokens
output = sampler.sample(prompt + sampler.tokenizer.decode(context_proposed, skip_special_tokens=False))
#Find the proposed probability distributions
proposed_target_distribution = calc_token_metric(output, sampler, self.token_metric)
proposed_sampling_distribution = calc_token_metric(output, sampler, "logprobs")
#TODO Compare the log_acceptance_ratio summations to those calculated using calc_sequence_logprob
# Calculate the Metro-Hastings acceptance ratio
# Power Scaled Sequence Log Probability + Temperature Scaled Sequence Log Probability - Current Power Scaled Sequence Log Probability - Current Temperature Scaled Sequence Log Probability
log_acceptance_ratio = sum(proposed_target_distribution) + sum(current_sampling_distribution[idx:idx+len(output.tokens)]) - sum(current_target_distribution[idx:idx+len(output.tokens)]) - sum(proposed_sampling_distribution)
# Check to make sure we are comparing the correct number of elements
assert(len(proposed_target_distribution) == len(current_sampling_distribution[idx:idx+len(output.tokens)]) == len(current_target_distribution[idx:idx+len(output.tokens)]) == len(proposed_sampling_distribution))
# Log the logprobs and acceptance ratio
if(logging):
proposed_target_distribution_sum = sum(proposed_target_distribution)
proposed_sampling_distribution_sum = sum(proposed_sampling_distribution)
current_target_distribution_sum = sum(current_target_distribution[idx:idx+len(output.tokens)])
current_sampling_distribution_sum = sum(current_sampling_distribution[idx:idx+len(output.tokens)])
# Write initial generated block data to log
with open(power_sampling_log_path, "a") as log_file:
log_file.write(f"{proposed_target_distribution_sum},{proposed_sampling_distribution_sum},{current_target_distribution_sum},{current_sampling_distribution_sum},")
acceptance = False
# Accept or reject the proposed block based on the acceptance ratio
if np.random.rand() < np.exp(log_acceptance_ratio):
# Ensure the context is updated with the accepted proposal
context[idx:] = output.tokens
# Replace the tail of the other Output attributes along with the context
logits = safe_concat(logits, output.top_k_logits, idx)
logprobs = safe_concat(logprobs, output.top_k_logprobs, idx)
unprocessed_log_normalization_constant = safe_concat(
unprocessed_log_normalization_constant, output.unprocessed_log_normalization_constant, idx
)
temp_processed_log_normalization_constant = safe_concat(
temp_processed_log_normalization_constant, output.temp_processed_log_normalization_constant, idx
)
entropy = safe_concat(entropy, output.entropy, idx)
# Update the logprob lists with the accepted proposal's log probabilities
current_target_distribution = [*current_target_distribution[:idx], *proposed_target_distribution]
current_sampling_distribution = [*current_sampling_distribution[:idx], *proposed_sampling_distribution]
# Flag acceptance
acceptance = True
# Log the new distributions and acceptance ratio
if(logging):
current_target_distribution_norm = sum(current_target_distribution)/len(current_target_distribution)
current_sampling_distribution_norm = sum(current_sampling_distribution)/len(current_sampling_distribution)
acceptance_ratio = np.exp(log_acceptance_ratio)
accepted = acceptance
tokens_generated = len(output.tokens)
starting_index = idx
with open(power_sampling_log_path, "a") as log_file:
log_file.write(f"{current_target_distribution_norm},{current_sampling_distribution_norm},{acceptance_ratio},{accepted},{starting_index},{tokens_generated}\n")
# Check if an EOS token has been generated and end the process if so
if(sampler.tokenizer.eos_token_id in context):
decoded_text = sampler.tokenizer.decode(context, skip_special_tokens=False)
if logging:
with open(power_sampling_log_path, "a") as log_file:
# Extract backslash expression for Python 3.10 compatibility
decoded_escaped = decoded_text.replace('"', '""')
log_file.write(f'"{decoded_escaped}"\n')
# Set the max_new_tokens back to the original value
sampler.sampling_params.max_tokens = sampler_max_tokens
return Output(tokens=context,top_k_logits=logits,top_k_logprobs=logprobs,unprocessed_log_normalization_constant=unprocessed_log_normalization_constant,temp_processed_log_normalization_constant=temp_processed_log_normalization_constant,entropy=entropy)
# EOS never found, just return the full generated context
decoded_text = sampler.tokenizer.decode(context, skip_special_tokens=False)
if logging:
with open(power_sampling_log_path, "a") as log_file:
# Extract backslash expression for Python 3.10 compatibility
decoded_escaped = decoded_text.replace('"', '""')
log_file.write(f'"{decoded_escaped}"\n')
# Set the max_tokens back to the original value
sampler.sampling_params.max_tokens = sampler_max_tokens
return Output(tokens=context,top_k_logits=logits,top_k_logprobs=logprobs,unprocessed_log_normalization_constant=unprocessed_log_normalization_constant,temp_processed_log_normalization_constant=temp_processed_log_normalization_constant,entropy=entropy)
|