Text Generation
Transformers
Safetensors
Upper Grand Valley Dani
llama
genomic
text-generation-inference
Instructions to use HuggingFaceBio/Carbon-3B with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use HuggingFaceBio/Carbon-3B with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("text-generation", model="HuggingFaceBio/Carbon-3B")# Load model directly from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("HuggingFaceBio/Carbon-3B") model = AutoModelForCausalLM.from_pretrained("HuggingFaceBio/Carbon-3B") - Inference
- Notebooks
- Google Colab
- Kaggle
- Local Apps
- vLLM
How to use HuggingFaceBio/Carbon-3B with vLLM:
Install from pip and serve model
# Install vLLM from pip: pip install vllm # Start the vLLM server: vllm serve "HuggingFaceBio/Carbon-3B" # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:8000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HuggingFaceBio/Carbon-3B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker
docker model run hf.co/HuggingFaceBio/Carbon-3B
- SGLang
How to use HuggingFaceBio/Carbon-3B with SGLang:
Install from pip and serve model
# Install SGLang from pip: pip install sglang # Start the SGLang server: python3 -m sglang.launch_server \ --model-path "HuggingFaceBio/Carbon-3B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HuggingFaceBio/Carbon-3B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }'Use Docker images
docker run --gpus all \ --shm-size 32g \ -p 30000:30000 \ -v ~/.cache/huggingface:/root/.cache/huggingface \ --env "HF_TOKEN=<secret>" \ --ipc=host \ lmsysorg/sglang:latest \ python3 -m sglang.launch_server \ --model-path "HuggingFaceBio/Carbon-3B" \ --host 0.0.0.0 \ --port 30000 # Call the server using curl (OpenAI-compatible API): curl -X POST "http://localhost:30000/v1/completions" \ -H "Content-Type: application/json" \ --data '{ "model": "HuggingFaceBio/Carbon-3B", "prompt": "Once upon a time,", "max_tokens": 512, "temperature": 0.5 }' - Docker Model Runner
How to use HuggingFaceBio/Carbon-3B with Docker Model Runner:
docker model run hf.co/HuggingFaceBio/Carbon-3B
modeling_carbon: replace _sample() override with stable LogitsProcessor API
Browse files- modeling_carbon.py +135 -353
modeling_carbon.py
CHANGED
|
@@ -1,427 +1,209 @@
|
|
| 1 |
"""
|
| 2 |
-
Carbon with
|
| 3 |
|
| 4 |
-
generate_bp()
|
| 5 |
-
|
| 6 |
-
|
| 7 |
"""
|
| 8 |
-
import os
|
| 9 |
-
from typing import Optional, Union
|
| 10 |
-
|
| 11 |
import torch
|
| 12 |
-
import torch.nn as nn
|
| 13 |
import torch.nn.functional as F
|
| 14 |
-
from transformers import LlamaForCausalLM
|
|
|
|
| 15 |
|
| 16 |
BASE_TO_IDX = {"A": 0, "T": 1, "C": 2, "G": 3, "N": -1}
|
| 17 |
IDX_TO_BASE = {0: "A", 1: "T", 2: "C", 3: "G", -1: "N"}
|
| 18 |
|
| 19 |
|
| 20 |
-
class
|
| 21 |
-
"""
|
| 22 |
|
| 23 |
-
|
| 24 |
-
|
| 25 |
"""
|
| 26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 27 |
def setup_tokenizer(self, tokenizer):
|
| 28 |
-
"""Cache tokenizer and precompute lookup tables for bp
|
| 29 |
self.tokenizer = tokenizer
|
| 30 |
k = tokenizer.k
|
| 31 |
self.k = k
|
| 32 |
-
num_special = len(tokenizer.special_tokens)
|
| 33 |
-
num_kmers = 4 ** k
|
| 34 |
|
| 35 |
-
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
|
|
|
|
|
|
|
|
|
| 38 |
bp_base_index = torch.zeros(k, num_kmers, dtype=torch.long)
|
| 39 |
-
for j in
|
| 40 |
-
|
| 41 |
-
|
| 42 |
self.register_buffer("_bp_base_index", bp_base_index.to(device), persistent=False)
|
| 43 |
|
| 44 |
self._bp_powers = torch.tensor(
|
| 45 |
[4 ** i for i in range(k - 1, -1, -1)], dtype=torch.long, device=device
|
| 46 |
)
|
|
|
|
|
|
|
| 47 |
flat_to_tid = torch.zeros(num_kmers, dtype=torch.long, device=device)
|
| 48 |
-
for kmer, tid in
|
| 49 |
-
|
| 50 |
-
|
| 51 |
self.register_buffer("_flat_idx_to_token_id", flat_to_tid, persistent=False)
|
| 52 |
|
| 53 |
def compute_bp_probs(self, logits):
|
| 54 |
-
"""Compute per-base marginal probabilities from token logits
|
| 55 |
|
| 56 |
Args:
|
| 57 |
-
logits: [B, V] or [B, L, V]
|
| 58 |
Returns:
|
| 59 |
bp_probs: [B, k, 4] or [B, L, k, 4]
|
| 60 |
"""
|
| 61 |
-
squeeze =
|
| 62 |
-
if
|
| 63 |
-
logits = logits.unsqueeze(1)
|
| 64 |
-
squeeze = True
|
| 65 |
|
| 66 |
-
kmer_logits = logits[:, :, self._kmer_ids]
|
| 67 |
kmer_probs = F.softmax(kmer_logits.float(), dim=-1)
|
| 68 |
B, L, _ = kmer_probs.shape
|
| 69 |
bp_probs = torch.zeros(B, L, self.k, 4, device=logits.device, dtype=kmer_probs.dtype)
|
| 70 |
for pos in range(self.k):
|
| 71 |
-
idx = self._bp_base_index[pos]
|
| 72 |
for nt in range(4):
|
| 73 |
bp_probs[:, :, pos, nt] = kmer_probs[:, :, idx == nt].sum(dim=-1)
|
| 74 |
|
| 75 |
-
if squeeze
|
| 76 |
-
bp_probs = bp_probs.squeeze(1) # [B, k, 4]
|
| 77 |
-
return bp_probs
|
| 78 |
|
| 79 |
-
# -------------------------------------------------------------------------
|
| 80 |
-
# generate_bp: sets a flag then delegates to the standard generate()
|
| 81 |
-
# -------------------------------------------------------------------------
|
| 82 |
-
@torch.no_grad()
|
| 83 |
def generate_bp(self, inputs=None, generation_config=None, **kwargs):
|
| 84 |
-
"""
|
| 85 |
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
supported — they are processed by the HF generate pipeline as usual.
|
| 90 |
-
|
| 91 |
-
Returns:
|
| 92 |
-
Same as generate() — token ids tensor or GenerateOutput.
|
| 93 |
"""
|
| 94 |
-
assert hasattr(self, "_bp_base_index"), "Call setup_tokenizer() first"
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
| 99 |
-
|
| 100 |
-
|
| 101 |
-
self.
|
| 102 |
-
|
| 103 |
-
|
| 104 |
-
|
| 105 |
-
|
| 106 |
-
def _sample(
|
| 107 |
-
self,
|
| 108 |
-
input_ids,
|
| 109 |
-
logits_processor,
|
| 110 |
-
stopping_criteria,
|
| 111 |
-
generation_config,
|
| 112 |
-
synced_gpus,
|
| 113 |
-
streamer,
|
| 114 |
-
**model_kwargs,
|
| 115 |
-
):
|
| 116 |
-
if not getattr(self, "_bp_generation", False):
|
| 117 |
-
return super()._sample(
|
| 118 |
-
input_ids,
|
| 119 |
-
logits_processor,
|
| 120 |
-
stopping_criteria,
|
| 121 |
-
generation_config,
|
| 122 |
-
synced_gpus,
|
| 123 |
-
streamer,
|
| 124 |
-
**model_kwargs,
|
| 125 |
-
)
|
| 126 |
-
|
| 127 |
-
# ==================================================================
|
| 128 |
-
# BP generation mode — copied from transformers 4.56.0 _sample(),
|
| 129 |
-
# with ONLY the token selection block replaced by bp marginalization.
|
| 130 |
-
# ==================================================================
|
| 131 |
-
from transformers.generation.utils import (
|
| 132 |
-
GenerateDecoderOnlyOutput,
|
| 133 |
)
|
|
|
|
|
|
|
| 134 |
|
| 135 |
-
|
| 136 |
-
pad_token_id = generation_config._pad_token_tensor
|
| 137 |
-
output_attentions = generation_config.output_attentions
|
| 138 |
-
output_hidden_states = generation_config.output_hidden_states
|
| 139 |
-
output_scores = generation_config.output_scores
|
| 140 |
-
output_logits = generation_config.output_logits
|
| 141 |
-
return_dict_in_generate = generation_config.return_dict_in_generate
|
| 142 |
-
has_eos_stopping_criteria = any(
|
| 143 |
-
hasattr(criteria, "eos_token_id") for criteria in stopping_criteria
|
| 144 |
-
)
|
| 145 |
-
do_sample = generation_config.do_sample
|
| 146 |
-
|
| 147 |
-
# init attention / hidden states / scores tuples
|
| 148 |
-
scores = () if (return_dict_in_generate and output_scores) else None
|
| 149 |
-
raw_logits = () if (return_dict_in_generate and output_logits) else None
|
| 150 |
-
decoder_attentions = (
|
| 151 |
-
() if (return_dict_in_generate and output_attentions) else None
|
| 152 |
-
)
|
| 153 |
-
decoder_hidden_states = (
|
| 154 |
-
() if (return_dict_in_generate and output_hidden_states) else None
|
| 155 |
-
)
|
| 156 |
-
|
| 157 |
-
# keep track of which sequences are already finished
|
| 158 |
-
batch_size, cur_len = input_ids.shape[:2]
|
| 159 |
-
this_peer_finished = False
|
| 160 |
-
unfinished_sequences = torch.ones(
|
| 161 |
-
batch_size, dtype=torch.long, device=input_ids.device
|
| 162 |
-
)
|
| 163 |
-
model_kwargs = self._get_initial_cache_position(
|
| 164 |
-
cur_len, input_ids.device, model_kwargs
|
| 165 |
-
)
|
| 166 |
-
|
| 167 |
-
model_forward = self.__call__
|
| 168 |
-
compile_forward = self._valid_auto_compile_criteria(
|
| 169 |
-
model_kwargs, generation_config
|
| 170 |
-
)
|
| 171 |
-
if compile_forward:
|
| 172 |
-
os.environ["TOKENIZERS_PARALLELISM"] = "0"
|
| 173 |
-
if self.config._attn_implementation == "flash_attention_2":
|
| 174 |
-
if (
|
| 175 |
-
generation_config.compile_config is not None
|
| 176 |
-
and generation_config.compile_config.fullgraph
|
| 177 |
-
):
|
| 178 |
-
generation_config.compile_config.fullgraph = False
|
| 179 |
-
model_forward = self.get_compiled_call(generation_config.compile_config)
|
| 180 |
-
|
| 181 |
-
if generation_config.prefill_chunk_size is not None:
|
| 182 |
-
model_kwargs = self._prefill_chunking(
|
| 183 |
-
input_ids, generation_config, **model_kwargs
|
| 184 |
-
)
|
| 185 |
-
is_prefill = False
|
| 186 |
-
else:
|
| 187 |
-
is_prefill = True
|
| 188 |
-
|
| 189 |
-
while self._has_unfinished_sequences(
|
| 190 |
-
this_peer_finished, synced_gpus, device=input_ids.device
|
| 191 |
-
):
|
| 192 |
-
# prepare model inputs
|
| 193 |
-
model_inputs = self.prepare_inputs_for_generation(input_ids, **model_kwargs)
|
| 194 |
-
|
| 195 |
-
# prepare variable output controls
|
| 196 |
-
model_inputs.update(
|
| 197 |
-
{"output_attentions": output_attentions} if output_attentions else {}
|
| 198 |
-
)
|
| 199 |
-
model_inputs.update(
|
| 200 |
-
{"output_hidden_states": output_hidden_states}
|
| 201 |
-
if output_hidden_states
|
| 202 |
-
else {}
|
| 203 |
-
)
|
| 204 |
-
|
| 205 |
-
if is_prefill:
|
| 206 |
-
outputs = self(**model_inputs, return_dict=True)
|
| 207 |
-
is_prefill = False
|
| 208 |
-
else:
|
| 209 |
-
outputs = model_forward(**model_inputs, return_dict=True)
|
| 210 |
-
|
| 211 |
-
# update model kwargs for next step (handles cache, attention_mask, etc.)
|
| 212 |
-
model_kwargs = self._update_model_kwargs_for_generation(
|
| 213 |
-
outputs,
|
| 214 |
-
model_kwargs,
|
| 215 |
-
is_encoder_decoder=self.config.is_encoder_decoder,
|
| 216 |
-
)
|
| 217 |
-
if synced_gpus and this_peer_finished:
|
| 218 |
-
continue
|
| 219 |
-
|
| 220 |
-
next_token_logits = outputs.logits[:, -1, :].to(
|
| 221 |
-
copy=True, dtype=torch.float32, device=input_ids.device
|
| 222 |
-
)
|
| 223 |
-
|
| 224 |
-
# pre-process distribution (temperature, top_k, top_p, repetition_penalty, etc.)
|
| 225 |
-
next_token_scores = logits_processor(input_ids, next_token_logits)
|
| 226 |
-
|
| 227 |
-
# Store scores, attentions and hidden_states when required
|
| 228 |
-
if return_dict_in_generate:
|
| 229 |
-
if output_scores:
|
| 230 |
-
scores += (next_token_scores,)
|
| 231 |
-
if output_logits:
|
| 232 |
-
raw_logits += (next_token_logits,)
|
| 233 |
-
if output_attentions:
|
| 234 |
-
decoder_attentions += ((outputs.attentions,),)
|
| 235 |
-
if output_hidden_states:
|
| 236 |
-
decoder_hidden_states += ((outputs.hidden_states,),)
|
| 237 |
-
|
| 238 |
-
# =============================================================
|
| 239 |
-
# BP-LEVEL TOKEN SELECTION (vectorized, the ONLY change)
|
| 240 |
-
# =============================================================
|
| 241 |
-
# [B, V] -> [B, k, 4] marginal bp probabilities
|
| 242 |
-
bp_probs = self.compute_bp_probs(next_token_scores) # [B, k, 4]
|
| 243 |
-
|
| 244 |
-
if do_sample:
|
| 245 |
-
# [B*k, 4] -> multinomial -> [B, k]
|
| 246 |
-
base_indices = torch.multinomial(
|
| 247 |
-
bp_probs.view(-1, 4), 1
|
| 248 |
-
).view(batch_size, self.k)
|
| 249 |
-
else:
|
| 250 |
-
base_indices = bp_probs.argmax(dim=-1) # [B, k]
|
| 251 |
-
|
| 252 |
-
# base_indices [B, k] -> flat kmer index -> token_id [B]
|
| 253 |
-
flat_idx = (base_indices * self._bp_powers).sum(dim=-1) # [B]
|
| 254 |
-
next_tokens = self._flat_idx_to_token_id[flat_idx] # [B]
|
| 255 |
-
# =============================================================
|
| 256 |
-
|
| 257 |
-
# finished sentences should have their next token be a padding token
|
| 258 |
-
if has_eos_stopping_criteria:
|
| 259 |
-
next_tokens = next_tokens * unfinished_sequences + pad_token_id * (
|
| 260 |
-
1 - unfinished_sequences
|
| 261 |
-
)
|
| 262 |
-
|
| 263 |
-
# update generated ids, model inputs, and length for next step
|
| 264 |
-
input_ids = torch.cat([input_ids, next_tokens[:, None]], dim=-1)
|
| 265 |
-
if streamer is not None:
|
| 266 |
-
streamer.put(next_tokens.cpu())
|
| 267 |
-
|
| 268 |
-
unfinished_sequences = unfinished_sequences & ~stopping_criteria(
|
| 269 |
-
input_ids, scores
|
| 270 |
-
)
|
| 271 |
-
this_peer_finished = unfinished_sequences.max() == 0
|
| 272 |
-
cur_len += 1
|
| 273 |
-
|
| 274 |
-
del outputs
|
| 275 |
-
|
| 276 |
-
if streamer is not None:
|
| 277 |
-
streamer.end()
|
| 278 |
-
|
| 279 |
-
if return_dict_in_generate:
|
| 280 |
-
return GenerateDecoderOnlyOutput(
|
| 281 |
-
sequences=input_ids,
|
| 282 |
-
scores=scores,
|
| 283 |
-
logits=raw_logits,
|
| 284 |
-
attentions=decoder_attentions,
|
| 285 |
-
hidden_states=decoder_hidden_states,
|
| 286 |
-
past_key_values=model_kwargs.get("past_key_values"),
|
| 287 |
-
)
|
| 288 |
-
else:
|
| 289 |
-
return input_ids
|
| 290 |
|
| 291 |
@torch.no_grad()
|
| 292 |
-
def score_sequence(self, sequences: Union[str, list
|
| 293 |
-
"""Score DNA sequence(s)
|
| 294 |
|
| 295 |
-
|
| 296 |
-
|
| 297 |
-
original sequences only (excluding padding).
|
| 298 |
|
| 299 |
Args:
|
| 300 |
-
sequences:
|
| 301 |
|
| 302 |
Returns:
|
| 303 |
-
|
| 304 |
-
|
| 305 |
-
|
| 306 |
-
|
| 307 |
-
- actual_probs: Probability of the actual base at each position
|
| 308 |
-
* Single sequence: [seq_len] tensor
|
| 309 |
-
* Batch: list of [seq_len_i] tensors
|
| 310 |
-
|
| 311 |
-
bp_probs[i, j] = P(base at position i is nucleotide j | context)
|
| 312 |
-
actual_probs[i] = P(actual base at position i | context)
|
| 313 |
-
where j: 0=A, 1=T, 2=C, 3=G
|
| 314 |
-
|
| 315 |
-
Example:
|
| 316 |
-
# Single sequence
|
| 317 |
-
bp_probs, actual_probs = model.score_sequence("ACGT")
|
| 318 |
-
|
| 319 |
-
# Batch of sequences
|
| 320 |
-
bp_probs_list, actual_probs_list = model.score_sequence([
|
| 321 |
-
"ACGT" * 150,
|
| 322 |
-
"ACGT" * 149 + "AC",
|
| 323 |
-
])
|
| 324 |
"""
|
| 325 |
-
assert hasattr(self, "tokenizer"), "Call setup_tokenizer() first"
|
| 326 |
|
| 327 |
-
# Handle single sequence case
|
| 328 |
is_single = isinstance(sequences, str)
|
| 329 |
if is_single:
|
| 330 |
sequences = [sequences]
|
| 331 |
|
| 332 |
-
|
| 333 |
-
original_lens = [len(seq) for seq in sequences]
|
| 334 |
-
original_sequences = sequences.copy()
|
| 335 |
|
| 336 |
-
#
|
| 337 |
-
|
| 338 |
-
for
|
| 339 |
-
|
| 340 |
-
|
| 341 |
-
seq = seq + 'A' * padding_len
|
| 342 |
-
padded_sequences.append(seq)
|
| 343 |
|
| 344 |
-
#
|
| 345 |
-
|
| 346 |
|
| 347 |
-
# Tokenize batch (without add_special_tokens since we added manually)
|
| 348 |
inputs = self.tokenizer(
|
| 349 |
-
|
| 350 |
-
return_tensors="pt",
|
| 351 |
-
padding=True,
|
| 352 |
-
add_special_tokens=False
|
| 353 |
)
|
| 354 |
input_ids = inputs["input_ids"].to(self.device)
|
| 355 |
attention_mask = inputs["attention_mask"].to(self.device)
|
| 356 |
|
| 357 |
-
|
| 358 |
-
|
| 359 |
-
logits = outputs.logits # [B, max_seq_len, vocab_size]
|
| 360 |
-
|
| 361 |
-
# Compute bp probabilities for all token positions
|
| 362 |
-
bp_probs = self.compute_bp_probs(logits) # [B, max_seq_len, k, 4]
|
| 363 |
-
|
| 364 |
-
# Process each sequence in the batch
|
| 365 |
-
bp_probs_results = []
|
| 366 |
-
actual_probs_results = []
|
| 367 |
-
|
| 368 |
-
for i, (original_seq, original_len, padded_seq) in enumerate(
|
| 369 |
-
zip(original_sequences, original_lens, padded_sequences)
|
| 370 |
-
):
|
| 371 |
-
# Calculate number of actual sequence tokens (excluding BOS)
|
| 372 |
-
num_seq_tokens = len(padded_seq) // self.k
|
| 373 |
|
| 374 |
-
|
| 375 |
-
|
| 376 |
-
|
| 377 |
-
#
|
| 378 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 379 |
|
| 380 |
-
# Reshape: [num_seq_tokens, k, 4] -> [num_seq_tokens * k, 4]
|
| 381 |
-
seq_result = seq_bp_probs.reshape(-1, 4)
|
| 382 |
-
|
| 383 |
-
# Trim to original sequence length (remove padding)
|
| 384 |
-
seq_result = seq_result[:original_len]
|
| 385 |
-
|
| 386 |
-
# Extract actual base probabilities
|
| 387 |
-
actual_probs = self._extract_actual_probs(seq_result, original_seq)
|
| 388 |
-
|
| 389 |
-
bp_probs_results.append(seq_result)
|
| 390 |
-
actual_probs_results.append(actual_probs)
|
| 391 |
-
|
| 392 |
-
# Return single tensors if input was single sequence
|
| 393 |
if is_single:
|
| 394 |
-
return
|
| 395 |
-
|
| 396 |
-
return bp_probs_results, actual_probs_results
|
| 397 |
-
|
| 398 |
-
def _extract_actual_probs(self, bp_probs: torch.Tensor, sequence: str):
|
| 399 |
-
"""Extract probabilities of actual bases in the sequence.
|
| 400 |
-
|
| 401 |
-
For each position i in the sequence, returns the probability that the model
|
| 402 |
-
assigned to the actual base at that position.
|
| 403 |
-
|
| 404 |
-
For 'N' bases (unknown), returns the maximum probability across all 4 bases.
|
| 405 |
-
|
| 406 |
-
Args:
|
| 407 |
-
bp_probs: [seq_len, 4] probability distribution from logits
|
| 408 |
-
bp_probs[i] = P(position i | context before i)
|
| 409 |
-
sequence: DNA sequence string (may contain 'N')
|
| 410 |
-
|
| 411 |
-
Returns:
|
| 412 |
-
actual_probs: [seq_len] probabilities of actual bases
|
| 413 |
-
actual_probs[i] = bp_probs[i, sequence[i]] for A/T/C/G
|
| 414 |
-
actual_probs[i] = max(bp_probs[i]) for 'N'
|
| 415 |
-
"""
|
| 416 |
-
seq_len = len(sequence)
|
| 417 |
-
actual_probs = torch.zeros(seq_len, device=bp_probs.device, dtype=bp_probs.dtype)
|
| 418 |
|
|
|
|
|
|
|
| 419 |
for i, base in enumerate(sequence):
|
| 420 |
-
if base ==
|
| 421 |
-
|
| 422 |
-
actual_probs[i] = bp_probs[i].max()
|
| 423 |
-
else:
|
| 424 |
-
base_idx = BASE_TO_IDX[base]
|
| 425 |
-
actual_probs[i] = bp_probs[i, base_idx]
|
| 426 |
-
|
| 427 |
-
return actual_probs
|
|
|
|
| 1 |
"""
|
| 2 |
+
Carbon with bp-level generation and scoring.
|
| 3 |
|
| 4 |
+
generate_bp() plugs into the standard HF generate() pipeline via a
|
| 5 |
+
LogitsProcessor — no internal methods are overridden, so it is compatible
|
| 6 |
+
with any transformers version.
|
| 7 |
"""
|
|
|
|
|
|
|
|
|
|
| 8 |
import torch
|
|
|
|
| 9 |
import torch.nn.functional as F
|
| 10 |
+
from transformers import LlamaForCausalLM, LogitsProcessor, LogitsProcessorList
|
| 11 |
+
from typing import Union
|
| 12 |
|
| 13 |
BASE_TO_IDX = {"A": 0, "T": 1, "C": 2, "G": 3, "N": -1}
|
| 14 |
IDX_TO_BASE = {0: "A", 1: "T", 2: "C", 3: "G", -1: "N"}
|
| 15 |
|
| 16 |
|
| 17 |
+
class _BPLogitsProcessor(LogitsProcessor):
|
| 18 |
+
"""Forces token selection to use per-base marginal probabilities.
|
| 19 |
|
| 20 |
+
Runs LAST in the logits-processor chain so that temperature / top-k /
|
| 21 |
+
top-p etc. influence the marginal distributions before base selection.
|
| 22 |
"""
|
| 23 |
|
| 24 |
+
def __init__(self, kmer_ids, bp_base_index, flat_idx_to_token_id, bp_powers, k, do_sample):
|
| 25 |
+
self.kmer_ids = kmer_ids
|
| 26 |
+
self.bp_base_index = bp_base_index
|
| 27 |
+
self.flat_idx_to_token_id = flat_idx_to_token_id
|
| 28 |
+
self.bp_powers = bp_powers
|
| 29 |
+
self.k = k
|
| 30 |
+
self.do_sample = do_sample
|
| 31 |
+
|
| 32 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
| 33 |
+
B = scores.shape[0]
|
| 34 |
+
kmer_probs = F.softmax(scores[:, self.kmer_ids].float(), dim=-1) # [B, num_kmers]
|
| 35 |
+
|
| 36 |
+
# Marginalise to per-base probabilities [B, k, 4]
|
| 37 |
+
bp_probs = torch.zeros(B, self.k, 4, device=scores.device, dtype=kmer_probs.dtype)
|
| 38 |
+
for pos in range(self.k):
|
| 39 |
+
idx = self.bp_base_index[pos] # [num_kmers] in {0,1,2,3}
|
| 40 |
+
for nt in range(4):
|
| 41 |
+
bp_probs[:, pos, nt] = kmer_probs[:, idx == nt].sum(dim=-1)
|
| 42 |
+
|
| 43 |
+
if self.do_sample:
|
| 44 |
+
base_indices = torch.multinomial(bp_probs.view(-1, 4), 1).view(B, self.k)
|
| 45 |
+
else:
|
| 46 |
+
base_indices = bp_probs.argmax(dim=-1) # [B, k]
|
| 47 |
+
|
| 48 |
+
flat_idx = (base_indices * self.bp_powers).sum(dim=-1) # [B]
|
| 49 |
+
selected = self.flat_idx_to_token_id[flat_idx] # [B]
|
| 50 |
+
|
| 51 |
+
# One-hot: both argmax and multinomial land on the bp-selected token
|
| 52 |
+
new_scores = torch.full_like(scores, float("-inf"))
|
| 53 |
+
new_scores.scatter_(1, selected.unsqueeze(1), 0.0)
|
| 54 |
+
return new_scores
|
| 55 |
+
|
| 56 |
+
|
| 57 |
+
class CarbonForCausalLM(LlamaForCausalLM):
|
| 58 |
+
"""LlamaForCausalLM with bp-level generation and sequence scoring."""
|
| 59 |
+
|
| 60 |
def setup_tokenizer(self, tokenizer):
|
| 61 |
+
"""Cache tokenizer and precompute lookup tables for bp-level operations."""
|
| 62 |
self.tokenizer = tokenizer
|
| 63 |
k = tokenizer.k
|
| 64 |
self.k = k
|
|
|
|
|
|
|
| 65 |
|
| 66 |
+
device = next(self.parameters()).device
|
| 67 |
+
|
| 68 |
+
# Build ordered kmer list from the tokenizer's DNA vocab
|
| 69 |
+
kmer_items = sorted(
|
| 70 |
+
[
|
| 71 |
+
(kmer, tid)
|
| 72 |
+
for kmer, tid in tokenizer.dna_token_to_id.items()
|
| 73 |
+
if len(kmer) == k and all(b in "ATCG" for b in kmer)
|
| 74 |
+
],
|
| 75 |
+
key=lambda x: x[1],
|
| 76 |
+
)
|
| 77 |
+
kmers = [item[0] for item in kmer_items]
|
| 78 |
+
kmer_ids = [item[1] for item in kmer_items]
|
| 79 |
+
num_kmers = len(kmer_ids)
|
| 80 |
|
| 81 |
+
self._kmer_ids = torch.tensor(kmer_ids, dtype=torch.long, device=device)
|
| 82 |
+
|
| 83 |
+
# bp_base_index[pos, j] = base index (0-3) of kmer j at position pos
|
| 84 |
bp_base_index = torch.zeros(k, num_kmers, dtype=torch.long)
|
| 85 |
+
for j, kmer in enumerate(kmers):
|
| 86 |
+
for pos, base in enumerate(kmer):
|
| 87 |
+
bp_base_index[pos, j] = BASE_TO_IDX[base]
|
| 88 |
self.register_buffer("_bp_base_index", bp_base_index.to(device), persistent=False)
|
| 89 |
|
| 90 |
self._bp_powers = torch.tensor(
|
| 91 |
[4 ** i for i in range(k - 1, -1, -1)], dtype=torch.long, device=device
|
| 92 |
)
|
| 93 |
+
|
| 94 |
+
# flat kmer index -> token id (flat index = sum base_idx[i] * 4^(k-1-i))
|
| 95 |
flat_to_tid = torch.zeros(num_kmers, dtype=torch.long, device=device)
|
| 96 |
+
for j, (kmer, tid) in enumerate(kmer_items):
|
| 97 |
+
flat_idx = sum(BASE_TO_IDX[c] * (4 ** (k - 1 - i)) for i, c in enumerate(kmer))
|
| 98 |
+
flat_to_tid[flat_idx] = tid
|
| 99 |
self.register_buffer("_flat_idx_to_token_id", flat_to_tid, persistent=False)
|
| 100 |
|
| 101 |
def compute_bp_probs(self, logits):
|
| 102 |
+
"""Compute per-base marginal probabilities from token logits.
|
| 103 |
|
| 104 |
Args:
|
| 105 |
+
logits: [B, V] or [B, L, V]
|
| 106 |
Returns:
|
| 107 |
bp_probs: [B, k, 4] or [B, L, k, 4]
|
| 108 |
"""
|
| 109 |
+
squeeze = logits.dim() == 2
|
| 110 |
+
if squeeze:
|
| 111 |
+
logits = logits.unsqueeze(1)
|
|
|
|
| 112 |
|
| 113 |
+
kmer_logits = logits[:, :, self._kmer_ids]
|
| 114 |
kmer_probs = F.softmax(kmer_logits.float(), dim=-1)
|
| 115 |
B, L, _ = kmer_probs.shape
|
| 116 |
bp_probs = torch.zeros(B, L, self.k, 4, device=logits.device, dtype=kmer_probs.dtype)
|
| 117 |
for pos in range(self.k):
|
| 118 |
+
idx = self._bp_base_index[pos]
|
| 119 |
for nt in range(4):
|
| 120 |
bp_probs[:, :, pos, nt] = kmer_probs[:, :, idx == nt].sum(dim=-1)
|
| 121 |
|
| 122 |
+
return bp_probs.squeeze(1) if squeeze else bp_probs
|
|
|
|
|
|
|
| 123 |
|
|
|
|
|
|
|
|
|
|
|
|
|
| 124 |
def generate_bp(self, inputs=None, generation_config=None, **kwargs):
|
| 125 |
+
"""Like generate(), but each token is selected base-by-base from marginal distributions.
|
| 126 |
|
| 127 |
+
Temperature, top_k, top_p, repetition_penalty etc. all apply as usual —
|
| 128 |
+
they run before the bp processor and shift the marginal distributions.
|
| 129 |
+
Output shape and type are identical to generate().
|
|
|
|
|
|
|
|
|
|
|
|
|
| 130 |
"""
|
| 131 |
+
assert hasattr(self, "_bp_base_index"), "Call setup_tokenizer(tokenizer) first"
|
| 132 |
+
|
| 133 |
+
gc = generation_config or self.generation_config
|
| 134 |
+
do_sample = kwargs.get("do_sample", getattr(gc, "do_sample", False))
|
| 135 |
+
|
| 136 |
+
bp_proc = _BPLogitsProcessor(
|
| 137 |
+
kmer_ids=self._kmer_ids,
|
| 138 |
+
bp_base_index=self._bp_base_index,
|
| 139 |
+
flat_idx_to_token_id=self._flat_idx_to_token_id,
|
| 140 |
+
bp_powers=self._bp_powers,
|
| 141 |
+
k=self.k,
|
| 142 |
+
do_sample=do_sample,
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 143 |
)
|
| 144 |
+
existing = list(kwargs.pop("logits_processor", None) or [])
|
| 145 |
+
kwargs["logits_processor"] = LogitsProcessorList(existing + [bp_proc])
|
| 146 |
|
| 147 |
+
return super().generate(inputs=inputs, generation_config=generation_config, **kwargs)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 148 |
|
| 149 |
@torch.no_grad()
|
| 150 |
+
def score_sequence(self, sequences: Union[str, list]):
|
| 151 |
+
"""Score DNA sequence(s) at base resolution.
|
| 152 |
|
| 153 |
+
Returns per-base probability distributions and the probability of the
|
| 154 |
+
actual base at each position, given all preceding context.
|
|
|
|
| 155 |
|
| 156 |
Args:
|
| 157 |
+
sequences: single DNA string or list of DNA strings (ACGT only)
|
| 158 |
|
| 159 |
Returns:
|
| 160 |
+
(bp_probs, actual_probs) for a single sequence, or
|
| 161 |
+
(list of bp_probs, list of actual_probs) for a batch.
|
| 162 |
+
bp_probs[i]: [seq_len_i, 4] — P(base | context) at each position
|
| 163 |
+
actual_probs[i]: [seq_len_i] — P(actual base | context)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 164 |
"""
|
| 165 |
+
assert hasattr(self, "tokenizer"), "Call setup_tokenizer(tokenizer) first"
|
| 166 |
|
|
|
|
| 167 |
is_single = isinstance(sequences, str)
|
| 168 |
if is_single:
|
| 169 |
sequences = [sequences]
|
| 170 |
|
| 171 |
+
original_lens = [len(s) for s in sequences]
|
|
|
|
|
|
|
| 172 |
|
| 173 |
+
# Right-pad to multiple of k with 'A' (matches tokenizer convention)
|
| 174 |
+
padded = []
|
| 175 |
+
for s in sequences:
|
| 176 |
+
r = len(s) % self.k
|
| 177 |
+
padded.append(s + "A" * (self.k - r) if r else s)
|
|
|
|
|
|
|
| 178 |
|
| 179 |
+
# Prepend <dna> tag manually (training format)
|
| 180 |
+
tagged = ["<dna>" + s for s in padded]
|
| 181 |
|
|
|
|
| 182 |
inputs = self.tokenizer(
|
| 183 |
+
tagged, return_tensors="pt", padding=True, add_special_tokens=False
|
|
|
|
|
|
|
|
|
|
| 184 |
)
|
| 185 |
input_ids = inputs["input_ids"].to(self.device)
|
| 186 |
attention_mask = inputs["attention_mask"].to(self.device)
|
| 187 |
|
| 188 |
+
logits = self(input_ids, attention_mask=attention_mask, return_dict=True).logits
|
| 189 |
+
bp_probs_all = self.compute_bp_probs(logits) # [B, L, k, 4]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 190 |
|
| 191 |
+
bp_results, actual_results = [], []
|
| 192 |
+
for i, (seq, orig_len, pad_seq) in enumerate(zip(sequences, original_lens, padded)):
|
| 193 |
+
num_tokens = len(pad_seq) // self.k
|
| 194 |
+
# logits[t] predicts token t+1; logits[0] (from <dna>) predicts token 1
|
| 195 |
+
seq_bp = bp_probs_all[i, :num_tokens] # [num_tokens, k, 4]
|
| 196 |
+
seq_bp = seq_bp.reshape(-1, 4)[:orig_len] # [orig_len, 4]
|
| 197 |
+
actual = self._extract_actual_probs(seq_bp, seq)
|
| 198 |
+
bp_results.append(seq_bp)
|
| 199 |
+
actual_results.append(actual)
|
| 200 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 201 |
if is_single:
|
| 202 |
+
return bp_results[0], actual_results[0]
|
| 203 |
+
return bp_results, actual_results
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 204 |
|
| 205 |
+
def _extract_actual_probs(self, bp_probs: torch.Tensor, sequence: str) -> torch.Tensor:
|
| 206 |
+
actual = torch.zeros(len(sequence), device=bp_probs.device, dtype=bp_probs.dtype)
|
| 207 |
for i, base in enumerate(sequence):
|
| 208 |
+
actual[i] = bp_probs[i].max() if base == "N" else bp_probs[i, BASE_TO_IDX[base]]
|
| 209 |
+
return actual
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|