C2C_demo / rosetta /model /aligner.py
fuvty's picture
[init] demo
5ccf219
"""
Token Aligner for handling different tokenizers between SLM and LLM models.
This module provides functionality to align tokens between two different tokenizers,
handling cases where the same text is tokenized differently.
"""
from typing import List, Tuple, Optional, Dict, Literal, Union
import torch
from transformers import PreTrainedTokenizerBase
from enum import Enum
class AlignmentStrategy(Enum):
"""Strategies for handling 1-to-many token alignments"""
FIRST = "first" # Always take the first LLM token
LONGEST = "longest" # Take the LLM token with the longest string
class TokenAligner:
"""
Aligns tokens between SLM (Small Language Model) and LLM (Large Language Model) tokenizers.
This class handles the case where the same text sequence is tokenized differently
by different tokenizers, using the SLM tokenization as the base and finding
corresponding LLM tokens for each SLM token.
"""
def __init__(
self,
slm_tokenizer: PreTrainedTokenizerBase,
llm_tokenizer: PreTrainedTokenizerBase,
strategy: Union[AlignmentStrategy, str] = AlignmentStrategy.FIRST,
verbose: bool = False
):
"""
Initialize the TokenAligner.
Args:
slm_tokenizer: The tokenizer for the Small Language Model (base)
llm_tokenizer: The tokenizer for the Large Language Model
strategy: Strategy for handling 1-to-many token mappings
Either AlignmentStrategy enum or string ('first' or 'longest')
verbose: Whether to print debug information during alignment
"""
self.slm_tokenizer = slm_tokenizer
self.llm_tokenizer = llm_tokenizer
if self.slm_tokenizer.pad_token is None:
self.slm_tokenizer.pad_token = self.slm_tokenizer.eos_token
self.slm_tokenizer.pad_token_id = self.slm_tokenizer.eos_token_id
if self.llm_tokenizer.pad_token is None:
self.llm_tokenizer.pad_token = self.llm_tokenizer.eos_token
self.llm_tokenizer.pad_token_id = self.llm_tokenizer.eos_token_id
# Handle string strategy input
if isinstance(strategy, str):
strategy = AlignmentStrategy(strategy.lower())
self.strategy = strategy
self.verbose = verbose
# Cache for token mappings to improve performance
self._alignment_cache: Dict[Tuple[int, ...], List[int]] = {}
def align_tokens(
self,
slm_token_ids: Union[List[int], torch.Tensor],
return_mapping: bool = False
) -> Union[List[int], Tuple[List[int], List[Tuple[int, List[int]]]]]:
"""
Align SLM tokens to LLM tokens.
Args:
slm_token_ids: Token IDs from the SLM tokenizer
return_mapping: If True, also return the detailed mapping
Returns:
If return_mapping is False: List of aligned LLM token IDs
If return_mapping is True: Tuple of (aligned_llm_token_ids, mapping_details)
where mapping_details is a list of (slm_token_id, [candidate_llm_token_ids])
"""
# Convert to list if tensor
if isinstance(slm_token_ids, torch.Tensor):
slm_token_ids = slm_token_ids.tolist()
# Check cache
cache_key = tuple(slm_token_ids)
if cache_key in self._alignment_cache and not return_mapping:
return self._alignment_cache[cache_key]
aligned_llm_tokens = []
mapping_details = []
for slm_token_id in slm_token_ids:
# Decode SLM token to string (without special token processing)
slm_token_str = self.slm_tokenizer.decode(
[slm_token_id],
skip_special_tokens=False,
clean_up_tokenization_spaces=False
)
# Handle special tokens
if slm_token_id in self.slm_tokenizer.all_special_ids:
# Try to find corresponding special token in LLM tokenizer
llm_token_id = self._map_special_token(slm_token_id, slm_token_str)
aligned_llm_tokens.append(llm_token_id)
mapping_details.append((slm_token_id, [llm_token_id]))
continue
# Tokenize the string with LLM tokenizer
llm_token_ids = self.llm_tokenizer.encode(
slm_token_str,
add_special_tokens=False,
return_tensors=None
)
if len(llm_token_ids) == 0:
# Handle empty tokenization (shouldn't normally happen)
if self.verbose:
print(f"Warning: SLM token {slm_token_id} ('{slm_token_str}') "
f"resulted in empty LLM tokenization")
# Use unknown token as fallback
llm_token_id = self.llm_tokenizer.unk_token_id or 0
aligned_llm_tokens.append(llm_token_id)
mapping_details.append((slm_token_id, [llm_token_id]))
elif len(llm_token_ids) == 1:
# Perfect 1-to-1 mapping
aligned_llm_tokens.append(llm_token_ids[0])
mapping_details.append((slm_token_id, llm_token_ids))
else:
# 1-to-many mapping, apply strategy
selected_token = self._apply_strategy(llm_token_ids, slm_token_str)
aligned_llm_tokens.append(selected_token)
mapping_details.append((slm_token_id, llm_token_ids))
if self.verbose:
selected_str = self.llm_tokenizer.decode(
[selected_token],
skip_special_tokens=False,
clean_up_tokenization_spaces=False
)
print(f"SLM token {slm_token_id} ('{slm_token_str}') -> "
f"LLM tokens {llm_token_ids}, selected {selected_token} ('{selected_str}')")
# Cache the result
self._alignment_cache[cache_key] = aligned_llm_tokens
if return_mapping:
return aligned_llm_tokens, mapping_details
return aligned_llm_tokens
def _map_special_token(self, slm_token_id: int, slm_token_str: str) -> int:
"""
Map special tokens between tokenizers.
Args:
slm_token_id: The SLM special token ID
slm_token_str: The string representation of the special token
Returns:
The corresponding LLM token ID
"""
# Common special token mappings
special_token_map = {
self.slm_tokenizer.pad_token_id: self.llm_tokenizer.pad_token_id,
self.slm_tokenizer.eos_token_id: self.llm_tokenizer.eos_token_id,
self.slm_tokenizer.bos_token_id: self.llm_tokenizer.bos_token_id,
self.slm_tokenizer.unk_token_id: self.llm_tokenizer.unk_token_id,
}
# Direct mapping if available
if slm_token_id in special_token_map and special_token_map[slm_token_id] is not None:
return special_token_map[slm_token_id]
# Try to find by string representation
try:
llm_token_id = self.llm_tokenizer.convert_tokens_to_ids(slm_token_str)
if llm_token_id != self.llm_tokenizer.unk_token_id:
return llm_token_id
except:
pass
# Fallback to unknown token
return self.llm_tokenizer.unk_token_id or 0
def _apply_strategy(self, llm_token_ids: List[int], original_str: str) -> int:
"""
Apply the selected strategy to choose one LLM token from multiple candidates.
Args:
llm_token_ids: List of candidate LLM token IDs
original_str: The original string from SLM token
Returns:
The selected LLM token ID
"""
if self.strategy == AlignmentStrategy.FIRST:
return llm_token_ids[0]
elif self.strategy == AlignmentStrategy.LONGEST:
# Find the token with the longest string representation
longest_token = llm_token_ids[0]
longest_length = 0
for token_id in llm_token_ids:
token_str = self.llm_tokenizer.decode(
[token_id],
skip_special_tokens=False,
clean_up_tokenization_spaces=False
)
if len(token_str) > longest_length:
longest_length = len(token_str)
longest_token = token_id
return longest_token
else:
# Default to first token if unknown strategy
return llm_token_ids[0]
def align_sequence(
self,
text: str,
return_details: bool = False
) -> Union[Tuple[List[int], List[int]], Dict[str, any]]:
"""
Tokenize text with both tokenizers and return aligned sequences.
Args:
text: The input text to tokenize and align
return_details: If True, return detailed alignment information
Returns:
If return_details is False: Tuple of (slm_token_ids, aligned_llm_token_ids)
If return_details is True: Dictionary with detailed alignment information
"""
# Tokenize with SLM
slm_tokens = self.slm_tokenizer.encode(
text,
add_special_tokens=True,
return_tensors=None
)
# Get aligned LLM tokens
if return_details:
aligned_llm_tokens, mapping = self.align_tokens(slm_tokens, return_mapping=True)
# Decode tokens for inspection
slm_decoded = [
self.slm_tokenizer.decode([tid], skip_special_tokens=False, clean_up_tokenization_spaces=False)
for tid in slm_tokens
]
llm_decoded = [
self.llm_tokenizer.decode([tid], skip_special_tokens=False, clean_up_tokenization_spaces=False)
for tid in aligned_llm_tokens
]
# Original LLM tokenization for comparison
original_llm_tokens = self.llm_tokenizer.encode(
text,
add_special_tokens=True,
return_tensors=None
)
# One-to-one mapping statistics
num_tokens = len(slm_tokens)
one_to_one_count = sum(1 for _slm_id, candidates in mapping if len(candidates) == 1)
one_to_one_rate = (one_to_one_count / num_tokens) if num_tokens > 0 else 0.0
return {
'text': text,
'slm_token_ids': slm_tokens,
'slm_decoded': slm_decoded,
'aligned_llm_token_ids': aligned_llm_tokens,
'aligned_llm_decoded': llm_decoded,
'original_llm_token_ids': original_llm_tokens,
'mapping': mapping,
'strategy': self.strategy.value,
'num_tokens': num_tokens,
'one_to_one_count': one_to_one_count,
'one_to_one_rate': one_to_one_rate
}
else:
aligned_llm_tokens = self.align_tokens(slm_tokens)
return slm_tokens, aligned_llm_tokens
def visualize_alignment(self, text: str):
"""
Print a visual representation of the token alignment.
Args:
text: The text to analyze
"""
details = self.align_sequence(text, return_details=True)
print("=" * 80)
print(f"Text: {text}")
print(f"Strategy: {details['strategy']}")
print("=" * 80)
print(f"SLM tokens ({len(details['slm_token_ids'])}): {details['slm_token_ids']}")
print(f"Aligned LLM tokens ({len(details['aligned_llm_token_ids'])}): {details['aligned_llm_token_ids']}")
print(f"Original LLM tokens ({len(details['original_llm_token_ids'])}): {details['original_llm_token_ids']}")
print("-" * 80)
print("Token-by-token alignment:")
for i, (slm_id, llm_id) in enumerate(zip(details['slm_token_ids'], details['aligned_llm_token_ids'])):
slm_str = details['slm_decoded'][i]
llm_str = details['aligned_llm_decoded'][i]
mapping_info = details['mapping'][i]
if len(mapping_info[1]) > 1:
candidates_str = ', '.join([
f"{tid}:'{self.llm_tokenizer.decode([tid], skip_special_tokens=False, clean_up_tokenization_spaces=False)}'"
for tid in mapping_info[1]
])
print(f" [{i:3d}] SLM {slm_id:6d} ('{slm_str}') -> "
f"LLM {llm_id:6d} ('{llm_str}') "
f"[candidates: {candidates_str}]")
else:
print(f" [{i:3d}] SLM {slm_id:6d} ('{slm_str}') -> "
f"LLM {llm_id:6d} ('{llm_str}')")
print("=" * 80)
def clear_cache(self):
"""Clear the alignment cache."""
self._alignment_cache.clear()
# ========================
# Chat messages alignment
# ========================
def _apply_chat_template_to_ids(
self,
tokenizer: PreTrainedTokenizerBase,
messages: List[Dict[str, str]],
add_generation_prompt: bool,
enable_thinking: bool,
remove_last_surfix: bool
) -> Tuple[str, List[int], Optional[List[Tuple[int, int]]]]:
"""
Apply chat template (no tokenization) then tokenize to ids with optional offsets.
If remove_last_surfix is True, remove the last suffix from the LLM text
Returns (templated_text, input_ids, offsets) where offsets may be None.
"""
if remove_last_surfix:
assert messages[-1]["role"] == "assistant", "Last message must be an assistant message"
templated_text = tokenizer.apply_chat_template(
messages[:-1],
tokenize=False,
add_generation_prompt=True,
enable_thinking=enable_thinking
)
templated_text += messages[-1]["content"]
else:
templated_text = tokenizer.apply_chat_template(
messages,
tokenize=False,
add_generation_prompt=add_generation_prompt,
enable_thinking=enable_thinking
)
encoded = tokenizer(
templated_text,
add_special_tokens=False,
return_offsets_mapping=True
)
input_ids: List[int] = encoded["input_ids"]
offsets = encoded.get("offset_mapping")
return templated_text, input_ids, offsets
@staticmethod
def _first_non_empty_content(messages: List[Dict[str, str]]) -> Optional[str]:
for m in messages:
content = m.get("content")
if isinstance(content, str) and len(content.strip()) > 0:
return content
return None
def _find_boundary_token_index(
self,
tokenizer: PreTrainedTokenizerBase,
templated_text: str,
offsets: Optional[List[Tuple[int, int]]],
content_text: Optional[str]
) -> int:
"""
Find token index where the first non-empty message content starts.
Falls back to 0 if not found.
"""
if not content_text:
return 0
char_idx = templated_text.find(content_text)
if char_idx < 0:
# Try a shorter probe to improve chances
probe = content_text[: min(32, len(content_text))]
if len(probe) > 0:
char_idx = templated_text.find(probe)
if char_idx < 0:
return 0
if offsets:
for idx, (start, _end) in enumerate(offsets):
if start >= char_idx:
return idx
return len(offsets)
# Fallback without offsets: tokenize prefix and count tokens
prefix = templated_text[:char_idx]
prefix_ids = tokenizer(prefix, add_special_tokens=False)["input_ids"]
return len(prefix_ids)
@staticmethod
def _compute_content_spans(templated_text: str, messages: List[Dict[str, str]]) -> List[Tuple[int, int]]:
"""
Compute character spans in templated_text that correspond to message contents.
Searches sequentially to reduce ambiguity when contents repeat.
Enhanced matching: ensures the found content is followed by '<' (special token start)
to avoid matching content inside special tokens like <begin_of_text>.
"""
spans: List[Tuple[int, int]] = []
search_from = 0
for m in messages:
content = m.get("content")
if not isinstance(content, str) or len(content) == 0:
continue
# Find all possible matches starting from search_from
idx = search_from
found_valid_match = False
while idx < len(templated_text):
idx = templated_text.find(content, idx)
if idx < 0:
break
# Check if this match is valid (followed by '<' indicating a special token)
end_pos = idx + len(content)
if end_pos < len(templated_text) and templated_text[end_pos] == '<':
# Valid match: content is followed by a special token
spans.append((idx, end_pos))
search_from = end_pos
found_valid_match = True
break
else:
# Check if this is the end of the text (also valid for last message)
if end_pos == len(templated_text):
spans.append((idx, end_pos))
search_from = end_pos
found_valid_match = True
break
# Invalid match, try next occurrence
idx += 1
# Fallback: if no valid match found with '<' requirement, use the old method
# but only as a last resort and with additional validation
if not found_valid_match:
idx = templated_text.find(content, search_from)
if idx < 0:
# Try searching from start as last resort
idx = templated_text.find(content)
if idx >= 0:
end_pos = idx + len(content)
# Additional check: avoid matching inside obvious special tokens
# Check if we're inside a special token (preceded by '<' and not followed by '>')
start_context = templated_text[max(0, idx-10):idx]
end_context = templated_text[end_pos:min(len(templated_text), end_pos+10)]
# Skip if we're clearly inside a special token
if ('<' in start_context and '>' not in start_context and
'begin_of_text' in templated_text[max(0, idx-20):idx+20]):
# This looks like we're matching inside <begin_of_text> or similar
continue
spans.append((idx, end_pos))
search_from = end_pos
return spans
@staticmethod
def _build_token_mask_from_spans(
offsets: Optional[List[Tuple[int, int]]],
num_tokens: int,
spans: List[Tuple[int, int]]
) -> List[bool]:
"""
Build a boolean mask for tokens whose offset range overlaps any span.
If offsets are missing, default to all False.
"""
if not offsets or len(offsets) != num_tokens:
return [False] * num_tokens
mask: List[bool] = []
for (start, end) in offsets:
if end <= start:
mask.append(False)
continue
is_msg = False
for s, e in spans:
# overlap check
if start < e and end > s:
is_msg = True
break
mask.append(is_msg)
return mask
@staticmethod
def _spans_to_token_ranges(
offsets: List[Tuple[int, int]],
spans: List[Tuple[int, int]]
) -> List[Tuple[int, int]]:
"""
Convert character spans to token index ranges using offsets.
start token = first token with end > span_start
end token = first token with start >= span_end
"""
ranges: List[Tuple[int, int]] = []
n = len(offsets)
for s, e in spans:
# find start index
start_idx = 0
while start_idx < n and offsets[start_idx][1] <= s:
start_idx += 1
# find end index
end_idx = start_idx
while end_idx < n and offsets[end_idx][0] < e:
end_idx += 1
ranges.append((start_idx, end_idx))
return ranges
def align_chat_messages(
self,
messages: List[Dict[str, str]],
add_generation_prompt: bool = True,
enable_thinking: bool = False,
return_details: bool = False,
remove_last_surfix: bool = False
) -> Dict[str, any]:
"""
Align chat-templated sequences by sections (template/message/template...):
- Preserve all template tokens (pad the shorter template section)
- For each message section, map SLM tokens to LLM tokens 1:1 via strategy
- If remove_last_surfix is True, remove the last suffix from the LLM text
Returns essentials: slm_ids_padded, llm_ids_padded, message_mask (shared),
slm_padding_mask, llm_padding_mask (True where token is padding inserted).
When return_details=True, also returns 'sections' with aligned ranges.
"""
assert not (add_generation_prompt and remove_last_surfix), "add_generation_prompt and remove_last_surfix cannot be True at the same time"
# Build templated sequences with offsets
slm_text, slm_ids, slm_offsets = self._apply_chat_template_to_ids(
self.slm_tokenizer, messages, add_generation_prompt, enable_thinking, remove_last_surfix
)
llm_text, llm_ids, llm_offsets = self._apply_chat_template_to_ids(
self.llm_tokenizer, messages, add_generation_prompt, enable_thinking, remove_last_surfix
)
# Required pad tokens
assert self.slm_tokenizer.pad_token_id is not None, "SLM pad_token_id required"
assert self.llm_tokenizer.pad_token_id is not None, "LLM pad_token_id required"
slm_pad_id = self.slm_tokenizer.pad_token_id
llm_pad_id = self.llm_tokenizer.pad_token_id
# Content spans (char) and token ranges
content_spans_slm = self._compute_content_spans(slm_text, messages)
content_spans_llm = self._compute_content_spans(llm_text, messages)
assert slm_offsets is not None and llm_offsets is not None, "offset_mapping required"
slm_msg_ranges = self._spans_to_token_ranges(slm_offsets, content_spans_slm)
llm_msg_ranges = self._spans_to_token_ranges(llm_offsets, content_spans_llm)
# Build section ranges (template/message alternating)
def build_sections(total_len: int, msg_ranges: List[Tuple[int,int]]):
sections: List[Tuple[str,int,int]] = []
prev = 0
for (s, e) in msg_ranges:
if prev < s:
sections.append(("template", prev, s))
sections.append(("message", s, e))
prev = e
if prev < total_len:
sections.append(("template", prev, total_len))
return sections
slm_sections = build_sections(len(slm_ids), slm_msg_ranges)
llm_sections = build_sections(len(llm_ids), llm_msg_ranges)
assert len(slm_sections) == len(llm_sections), "Section count mismatch"
slm_out: List[int] = []
llm_out: List[int] = []
mask_out: List[bool] = []
slm_pad_mask_out: List[bool] = []
llm_pad_mask_out: List[bool] = []
detailed_sections: List[Dict[str, Union[str, Tuple[int,int]]]] = []
for (stype_s, s_s, e_s), (stype_l, s_l, e_l) in zip(slm_sections, llm_sections):
assert stype_s == stype_l, "Section type mismatch"
slm_start_out = len(slm_out)
llm_start_out = len(llm_out)
if stype_s == "template":
slm_seg_len = e_s - s_s
llm_seg_len = e_l - s_l
target_len = slm_seg_len if slm_seg_len >= llm_seg_len else llm_seg_len
slm_pad_needed = target_len - slm_seg_len
llm_pad_needed = target_len - llm_seg_len
slm_seg = slm_ids[s_s:e_s] + [slm_pad_id] * slm_pad_needed
llm_seg = llm_ids[s_l:e_l] + [llm_pad_id] * llm_pad_needed
slm_out.extend(slm_seg)
llm_out.extend(llm_seg)
mask_out.extend([False] * target_len)
slm_pad_mask_out.extend([False] * slm_seg_len + [True] * slm_pad_needed)
llm_pad_mask_out.extend([False] * llm_seg_len + [True] * llm_pad_needed)
else: # message
slm_msg = slm_ids[s_s:e_s]
llm_msg = self.align_tokens(slm_msg)
assert len(llm_msg) == len(slm_msg)
slm_out.extend(slm_msg)
llm_out.extend(llm_msg)
mask_out.extend([True] * len(slm_msg))
# no padding in message sections
slm_pad_mask_out.extend([False] * len(slm_msg))
llm_pad_mask_out.extend([False] * len(slm_msg))
slm_end_out = len(slm_out)
llm_end_out = len(llm_out)
detailed_sections.append({
'type': stype_s,
'slm_range': (slm_start_out, slm_end_out),
'llm_range': (llm_start_out, llm_end_out)
})
result_min = {
'slm_ids_padded': slm_out,
'llm_ids_padded': llm_out,
'message_mask': mask_out,
'slm_padding_mask': slm_pad_mask_out,
'llm_padding_mask': llm_pad_mask_out
}
if return_details:
result_min['sections'] = detailed_sections
result_min['slm_text'] = slm_text
result_min['llm_text'] = llm_text
return result_min