merge remote main

pipeline/__init__.py

@@ -28,7 +28,13 @@ Usage::
 from pipeline.pipeline import Pipeline, PipelineConfig, Stage, TransformStage
 from pipeline.tokenize import AutoTokenizer, ChatTemplate, train_bpe_tokenizer
 from pipeline.text import TextNormalizer
-from pipeline.packing import SequencePacker
+from pipeline.packing import (
+    GreedyPacker,
+    FfDPacker,
+    BfdPacker,
+    BasePacker,
+    pack_tensors,
+)
 
 # I/O module
 from pipeline.io import FileScanner, HDF5Handler, export_dataset, cache_jsonl
@@ -70,7 +76,11 @@ __all__ = [
     "train_bpe_tokenizer",
     # Text processing
     "TextNormalizer",
-    "SequencePacker",
+    "GreedyPacker",
+    "FfDPacker",
+    "BfdPacker",
+    "BasePacker",
+    "pack_tensors",
     # I/O
     "FileScanner",
     "HDF5Handler",

pipeline/io/__init__.py

@@ -4,16 +4,26 @@ This module provides:
 - FileScanner: File and directory scanning utilities
 - HDF5Handler: Tensor data persistence
 - export_dataset: HuggingFace Dataset to JSONL export
-- cache_jsonl: JSONL to HDF5 tokenization and caching
+- cache_jsonl: JSONL to HDF5/binary tokenization and caching
+- dedup_jsonl: MinHash+LSH deduplication for pretraining text
+- writers: BaseWriter / H5Writer / BinWriter / TextWriter (Strategy + Factory)
 """
 
 from pipeline.io.file_scanner import FileScanner
 from pipeline.io.hdf5_handler import HDF5Handler
 from pipeline.io.export import export_dataset, cache_jsonl
+from pipeline.io.dedup import dedup_jsonl
+from pipeline.io.writers import BaseWriter, H5Writer, BinWriter, TextWriter, create_writer
 
 __all__ = [
     "FileScanner",
     "HDF5Handler",
     "export_dataset",
     "cache_jsonl",
+    "dedup_jsonl",
+    "BaseWriter",
+    "H5Writer",
+    "BinWriter",
+    "TextWriter",
+    "create_writer",
 ]

pipeline/io/dedup.py

@@ -0,0 +1,167 @@
+"""MinHash + LSH deduplication for pretraining text data."""
+
+import json
+import logging
+import os
+from pathlib import Path
+from typing import Iterator, List, Set, Tuple
+
+from tqdm import tqdm
+
+from pipeline.io.writers import TextWriter
+from pipeline.utils import error_handler
+
+logger = logging.getLogger(__name__)
+
+
+def _tokenize(text: str, ngram: int = 3) -> Set[str]:
+    return {text[i : i + ngram] for i in range(len(text) - ngram + 1)}
+
+
+def _iter_docs(input_dir: Path) -> Iterator[Tuple[str, dict]]:
+    for fpath in sorted(input_dir.glob("*.jsonl")):
+        with open(fpath, encoding="utf-8") as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                record = json.loads(line)
+                text = record.get("text", "")
+                if text:
+                    yield text, record
+
+
+def _write_h5(records: List[dict], output_dir: str, chunk_idx: int):
+    import h5py
+
+    fname = os.path.join(output_dir, f"chunk_{chunk_idx}.h5")
+    texts = [rec.get("text", "") for rec in records]
+    with h5py.File(fname, "w") as f:
+        dt = h5py.special_dtype(vlen=str)
+        ds = f.create_dataset("text", (len(texts),), dtype=dt)
+        for i, t in enumerate(texts):
+            ds[i] = t
+
+
+def _write_bin(records: List[dict], output_dir: Path, chunk_idx: int):
+    output_dir.mkdir(parents=True, exist_ok=True)
+    texts = [rec.get("text", "") + "\n" for rec in records]
+
+    meta = {"chunk": chunk_idx, "count": len(texts), "format": "text", "encoding": "utf-8"}
+    meta_path = output_dir / "meta.json"
+    existing = json.loads(meta_path.read_text()) if meta_path.exists() else {}
+    existing[str(chunk_idx)] = meta
+    meta_path.write_text(json.dumps(existing, indent=2))
+
+    (output_dir / f"text_{chunk_idx}.bin").write_bytes("".join(texts).encode("utf-8"))
+
+
+_WRITERS = {
+    "jsonl": TextWriter,
+    "h5": lambda: None,  # handled inline below
+    "bin": lambda: None,
+}
+
+
+@error_handler()
+def dedup_jsonl(
+    input_dir: str,
+    output_dir: str,
+    *,
+    threshold: float = 0.8,
+    num_perm: int = 128,
+    ngram: int = 3,
+    output_format: str = "jsonl",
+    chunk_size: int = 1_000_000,
+) -> Tuple[int, int]:
+    """Deduplicate JSONL text files using MinHash + LSH.
+
+    Args:
+        input_dir: Directory with source ``*.jsonl`` files.
+        output_dir: Directory for deduplicated output.
+        threshold: Jaccard similarity threshold (0–1).
+        num_perm: Number of MinHash permutations.
+        ngram: Character n-gram size.
+        output_format: ``"jsonl"``, ``"h5"``, or ``"bin"``.
+        chunk_size: Records per output chunk file.
+
+    Returns:
+        ``(kept, removed)`` counts.
+    """
+    from datasketch import MinHash, MinHashLSH
+
+    input_path = Path(input_dir)
+    output_path = Path(output_dir)
+    output_path.mkdir(parents=True, exist_ok=True)
+
+    logger.info(
+        f"Deduplicating {input_dir} -> {output_dir} "
+        f"(threshold={threshold}, perm={num_perm}, fmt={output_format})"
+    )
+
+    lsh = MinHashLSH(threshold=threshold, num_perm=num_perm)
+
+    kept = 0
+    removed = 0
+
+    dup_doc_ids: Set[int] = set()
+    for doc_id, (text, _record) in enumerate(tqdm(_iter_docs(input_path), desc="indexing", unit="docs")):
+        shingles = _tokenize(text, ngram=ngram)
+        if len(shingles) < ngram * 2:
+            dup_doc_ids.add(doc_id)
+            continue
+
+        m = MinHash(num_perm=num_perm)
+        for s in shingles:
+            m.update(s.encode("utf-8"))
+
+        if lsh.query(m):
+            dup_doc_ids.add(doc_id)
+        else:
+            lsh.insert(doc_id, m)
+
+    logger.info(f"Found {len(dup_doc_ids)} duplicates, writing deduplicated data")
+
+    buffer: List[dict] = []
+    chunk_idx = 0
+    writer = TextWriter(chunk_size) if output_format == "jsonl" else None
+
+    for doc_id, (_text, record) in enumerate(tqdm(_iter_docs(input_path), desc="writing", unit="docs")):
+        if doc_id in dup_doc_ids:
+            removed += 1
+            continue
+
+        kept += 1
+        buffer.append(record)
+
+        if len(buffer) >= chunk_size:
+            _flush_chunk(buffer, output_path, chunk_idx, output_format, writer)
+            chunk_idx += 1
+            buffer = []
+
+    if buffer:
+        _flush_chunk(buffer, output_path, chunk_idx, output_format, writer)
+
+    if writer:
+        writer.flush(output_path)
+
+    logger.info(f"Done. kept={kept}, removed={removed}")
+    return kept, removed
+
+
+def _flush_chunk(
+    records: List[dict],
+    output_dir: Path,
+    chunk_idx: int,
+    output_format: str,
+    writer=None,
+):
+    if output_format == "jsonl":
+        for rec in records:
+            writer.write_record(rec, output_dir)
+    elif output_format == "h5":
+        _write_h5(records, str(output_dir), chunk_idx)
+    elif output_format == "bin":
+        _write_bin(records, output_dir, chunk_idx)
+    else:
+        raise ValueError(f"Unknown output format: {output_format}")

pipeline/io/export.py

@@ -4,14 +4,18 @@ import json
 import logging
 import os
 from pathlib import Path
-from typing import Any, Callable, Dict, List, Optional, Union
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
 from datasets import Dataset
+from torch import Tensor
 from tqdm import tqdm
 
 from pipeline.io.file_scanner import FileScanner
 from pipeline.io.hdf5_handler import HDF5Handler
+from pipeline.io.writers import create_writer, BaseWriter
 from pipeline.processors import BaseProcessor
-from pipeline.packing import pack_tensors
+from pipeline.packing import pack_tensors, BasePacker
 from pipeline.utils import error_handler
 
 logger = logging.getLogger(__name__)
@@ -75,6 +79,32 @@ def export_dataset(
     return output_files
 
 
+def merge_tensors(
+    tensors: List[Tensor],
+    group_size: int,
+) -> List[Tensor]:
+    """Merge a list of tensors into fewer larger tensors.
+
+    Concatenates every group_size consecutive tensors into one merged
+    tensor. This reduces the number of shm blocks when loading.
+
+    Args:
+        tensors: List of 1D tensors.
+        group_size: Number of tensors to merge into each group.
+
+    Returns:
+        List of merged tensors.
+    """
+    if not tensors:
+        return []
+
+    merged: List[Tensor] = []
+    for i in range(0, len(tensors), group_size):
+        merged.append(torch.cat(tensors[i : i + group_size]))
+
+    return merged
+
+
 @error_handler()
 def cache_jsonl(
     files: List[str],
@@ -83,41 +113,85 @@ def cache_jsonl(
     *,
     pack_size: int = -1,
     pad_value: int = 0,
-    batch_size: int = 256,
+    group_size: int = 1_000,
+    pack_algo: Optional[str] = None,
+    output_format: str = "h5",
+    batch_size: int = 1000,
 ) -> List[str]:
-    """Tokenize JSONL files and pack them into HDF5 storage.
+    """Tokenize JSONL files and save as HDF5 or binary.
+
+    BFD packs in group_size-bounded batches to avoid O(N²), then all
+    packed chunks are merged and saved as one file per input file.
 
     Args:
         files: List of JSONL file paths.
-        output_dir: H5 output directory.
+        output_dir: Output directory.
         processor: Initialized Processor instance.
         pack_size: Packing length, <=0 means no packing.
         pad_value: Padding value.
-        batch_size: Number of records passed to the processor at once.
+        group_size: BFD batch granularity (token count threshold for each
+            packing batch) and merge granularity, <=0 means no merging.
+        pack_algo: Packing algorithm: 'bfd' (default), 'ffd',
+            'greedy'. Only used when pack_size > 0.
+        output_format: ``"h5"`` or ``"bin"``.
+        batch_size: Number of lines to batch-process together for parallel
+            tokenization via encode_batch (default: 1000).
 
     Returns:
-        List of generated H5 file paths.
+        List of generated file paths.
     """
     os.makedirs(output_dir, exist_ok=True)
     output_files: List[str] = []
     output_keys = processor.output_keys
 
+    dtypes = (
+        dict(processor.schema.output_fields)
+        if processor.schema is not None
+        else None
+    )
+    pad_values = {k: (0 if k == "position_ids" else (False if k.endswith("_mask") else pad_value)) for k in output_keys}
+
+    target_tokens = group_size * pack_size if group_size > 0 and pack_size > 0 else 0
+
     for file_path in files:
         file_name = Path(file_path).stem
 
-        arrows: Dict[str, List] = {key: [] for key in output_keys}
+        all_packed: Dict[str, List[Tensor]] = {key: [] for key in output_keys}
+        arrows_batch: Dict[str, List] = {key: [] for key in output_keys}
+        batch_tokens: int = 0
 
-        def append_batch(batch):
-            items = [item for _, item in batch]
+        buf: List[Tuple[int, str]] = []
+
+        def flush_buf():
+            nonlocal batch_tokens
+            if not buf:
+                return
+            samples = []
+            for line_num, line in buf:
+                try:
+                    samples.append((line_num, json.loads(line)))
+                except json.JSONDecodeError as e:
+                    logger.warning(
+                        f"JSON decode error in {file_path} line {line_num}: "
+                        f"{e}. Skipping line."
+                    )
+            buf.clear()
+            if not samples:
+                return
+            items = [item for _, item in samples]
             try:
-                results = processor.process_batch(items)
+                results = (
+                    processor.process_batch(items)
+                    if hasattr(processor, "process_batch")
+                    else [processor.process(s) for s in items]
+                )
                 if len(results) != len(items):
                     raise RuntimeError(
                         "Batch processor returned a different number of results"
                     )
             except Exception:
                 results = []
-                for line_num, item in batch:
+                for line_num, item in samples:
                     try:
                         results.append(processor.process(item))
                     except Exception as e:
@@ -126,42 +200,63 @@ def cache_jsonl(
                             f"in {file_path}: {e}. Skipping line."
                         )
                         results.append(None)
-
             for result in results:
                 if result is not None:
                     for key in output_keys:
-                        arrows[key].append(result[key])
+                        arrows_batch[key].append(result[key])
+                    if target_tokens > 0:
+                        batch_tokens += int(result[output_keys[0]].shape[0])
 
-        batch = []
         batch_size = max(1, batch_size)
         with open(file_path, "r", encoding="utf-8") as f:
             for line_num, line in enumerate(
                 tqdm(f, desc=f"Processing {file_name}", leave=False), start=1
             ):
-                try:
-                    batch.append((line_num, json.loads(line)))
-                    if len(batch) >= batch_size:
-                        append_batch(batch)
-                        batch = []
-                except json.JSONDecodeError as e:
-                    logger.warning(
-                        f"JSON decode error in {file_path} line {line_num}: {e}. Skipping line."
-                    )
-            if batch:
-                append_batch(batch)
+                buf.append((line_num, line))
+                if len(buf) >= batch_size:
+                    flush_buf()
+                    if target_tokens > 0 and batch_tokens >= target_tokens:
+                        packed = pack_tensors(
+                            arrows_batch,
+                            pack_size,
+                            pad_value,
+                            dtypes,
+                            pad_values=pad_values,
+                            algo=pack_algo,
+                        )
+                        for key in output_keys:
+                            all_packed[key].extend(packed[key])
+                            arrows_batch[key] = []
+                        batch_tokens = 0
 
-        if pack_size > 0:
-            dtypes = (
-                dict(processor.schema.output_fields)
-                if processor.schema is not None
-                else None
-            )
-            output = pack_tensors(arrows, pack_size, pad_value, dtypes)
+        flush_buf()
+
+        if arrows_batch[output_keys[0]]:
+            if pack_size > 0:
+                packed = pack_tensors(arrows_batch, pack_size, pad_value, dtypes, pad_values=pad_values, algo=pack_algo)
+                for key in output_keys:
+                    all_packed[key].extend(packed[key])
+            else:
+                for key in output_keys:
+                    all_packed[key].extend(arrows_batch[key])
+
+        if not all_packed[output_keys[0]]:
+            logger.warning(f"No valid samples in {file_path}, skipping")
+            continue
+
+        if pack_size <= 0:
+            output = all_packed
+        elif group_size > 0 and all_packed[output_keys[0]]:
+            output = {
+                key: merge_tensors(tensors, group_size)
+                for key, tensors in all_packed.items()
+            }
         else:
-            output = arrows
+            output = all_packed
 
-        h5_path = HDF5Handler.save(output_dir, file_name, output)
-        output_files.append(h5_path)
-        logger.info(f"Saved {h5_path}")
+        writer: BaseWriter = create_writer(output_format)
+        saved = writer.save(output_dir, file_name, output)
+        output_files.append(saved)
+        logger.info(f"Saved {saved}")
 
     return output_files

pipeline/io/writers.py

@@ -0,0 +1,107 @@
+"""Storage backends for tensor / text output (Strategy + Factory).
+
+Each backend implements a common ``save()`` interface so callers use
+polymorphism instead of ``if fmt == "h5" ... elif fmt == "bin" ...``.
+
+Supports:
+    - **H5Writer**: HDF5 format (via HDF5Handler)
+    - **BinWriter**: binary format – meta.json + {key}.bin (memmap-compatible)
+    - **TextWriter**: raw JSONL text (for dedup output)
+"""
+
+import json
+import os
+from abc import ABC, abstractmethod
+from pathlib import Path
+from typing import Dict, List
+
+import torch
+from torch import Tensor
+
+
+class BaseWriter(ABC):
+    """Abstract writer – call ``save(dir, name, data)`` without caring
+    about the underlying format."""
+
+    @abstractmethod
+    def save(self, output_dir: str, file_name: str, data: Dict[str, List[Tensor]]) -> str:
+        ...
+
+
+class H5Writer(BaseWriter):
+    def save(self, output_dir: str, file_name: str, data: Dict[str, List[Tensor]]) -> str:
+        from pipeline.io.hdf5_handler import HDF5Handler
+        return HDF5Handler.save(output_dir, file_name, data)
+
+
+class BinWriter(BaseWriter):
+    def save(self, output_dir: str, file_name: str, data: Dict[str, List[Tensor]]) -> str:
+        import numpy as np
+
+        os.makedirs(output_dir, exist_ok=True)
+        sub_dir = os.path.join(output_dir, file_name)
+        os.makedirs(sub_dir, exist_ok=True)
+
+        meta: Dict[str, Dict] = {}
+        for key, tensors in data.items():
+            cat = torch.cat(tensors, dim=0)
+            meta[key] = {"shape": list(cat.shape), "dtype": str(cat.dtype).split(".")[-1]}
+            np.asarray(cat.cpu().numpy()).tofile(os.path.join(sub_dir, f"{key}.bin"))
+
+        with open(os.path.join(sub_dir, "meta.json"), "w") as f:
+            json.dump(meta, f, indent=2)
+
+        return sub_dir
+
+
+class TextWriter(BaseWriter):
+    """Write raw text records as JSONL (used by dedup output)."""
+
+    def __init__(self, chunk_size: int = 1_000_000):
+        self._chunk_size = chunk_size
+        self._buffer: List[dict] = []
+        self._chunk_idx = 0
+
+    def save(self, output_dir: str, file_name: str, data: Dict[str, List[Tensor]]) -> str:
+        raise NotImplementedError("TextWriter.save_one is for tensor data; use write_record()")
+
+    def write_record(self, record: dict, output_dir: Path):
+        self._buffer.append(record)
+        if len(self._buffer) >= self._chunk_size:
+            self._flush(output_dir)
+
+    def flush(self, output_dir: Path):
+        if self._buffer:
+            self._flush(output_dir)
+
+    def _flush(self, output_dir: Path):
+        output_dir.mkdir(parents=True, exist_ok=True)
+        fpath = output_dir / f"chunk_{self._chunk_idx}.jsonl"
+        with open(fpath, "w", encoding="utf-8") as f:
+            for rec in self._buffer:
+                f.write(json.dumps(rec, ensure_ascii=False) + "\n")
+        self._chunk_idx += 1
+        self._buffer = []
+
+
+_WRITER_REGISTRY: Dict[str, type] = {}
+
+
+def register_writer(name: str):
+    def decorator(cls):
+        _WRITER_REGISTRY[name] = cls
+        return cls
+    return decorator
+
+
+def create_writer(name: str, **kwargs) -> BaseWriter:
+    cls = _WRITER_REGISTRY.get(name)
+    if cls is None:
+        raise ValueError(f"Unknown writer: {name}. Available: {list(_WRITER_REGISTRY)}")
+    return cls(**kwargs)
+
+
+# Register built-in writers
+register_writer("h5")(H5Writer)
+register_writer("bin")(BinWriter)
+register_writer("jsonl")(TextWriter)

pipeline/packing.py

@@ -1,146 +0,0 @@
-import logging
-from typing import Any, Dict, List, Optional, Tuple, Union
-
-import torch
-from torch import Tensor
-
-from pipeline.utils import error_handler
-
-logger = logging.getLogger(__name__)
-
-
-class SequencePacker:
-    """
-    Stream-concatenation packer for LLM training sequences.
-
-    Algorithm (streaming concat):
-
-        Input:  sequences = [A(len=3), B(len=5), C(len=2)], pack_size = 6
-
-        1. Validate & Normalize
-           - Check 1D dimension, unify dtype, warn on overlong sequences
-           - Result: [A, B, C]
-
-        2. Stream into buffer, slice off full chunks
-           - buffer += A(3) -> [a1 a2 a3], pos=3
-           - buffer += B(5) -> [a1 a2 a3 b1 b2 b3 b4 b5], pos=8
-             pos >= 6 -> flush [a1 a2 a3 b1 b2 b3], buffer=[b4 b5], pos=2
-           - buffer += C(2) -> [b4 b5 c1 c2], pos=4
-             loop ends -> flush tail [b4 b5 c1 c2 PAD PAD]
-
-        Output: [[a1 a2 a3 b1 b2 b3], [b4 b5 c1 c2 PAD PAD]]
-
-    Samples may be split across chunks — this is intentional and standard
-    practice in LLM training (TRL, Megatron-LM, etc.).
-
-    Cross-group consistency:
-        Different tensor groups (e.g. input_ids, loss_masks) packed with
-        separate packer instances on samples with matching lengths produce
-        identical chunk boundaries. Element-level correspondence is preserved.
-    """
-
-    def __init__(
-        self,
-        pack_size: int,
-        pad_value: Union[int, bool] = 0,
-        dtype: Optional[torch.dtype] = None,
-    ):
-        self.pack_size = pack_size
-        self.pad_value = pad_value
-        self.dtype = dtype
-        self._buffer: List = []
-        self._pos: int = 0
-        self._packages: List[Tensor] = []
-
-    def reset(self) -> None:
-        """Reset packer state for instance reuse."""
-        self._buffer = []
-        self._pos = 0
-        self._packages = []
-
-    @error_handler()
-    def pack(self, sequences: List[Tensor]) -> List[Tensor]:
-        """
-        Pack sequences via streaming concatenation into fixed-size chunks.
-
-        Sequences are concatenated in order and sliced at pack_size boundaries.
-        The final chunk is padded with pad_value.
-
-        When dtype is not set at init, it is inferred from the first input tensor.
-
-        Args:
-            sequences: List of 1D input tensors.
-
-        Returns:
-            List of packed tensors, each with length equal to pack_size.
-        """
-        if not sequences:
-            return []
-
-        # --- auto-infer dtype from first sequence ---
-        if self.dtype is None:
-            self.dtype = sequences[0].dtype
-
-        # --- validate & normalize ---
-        normalized: List[Tensor] = []
-        for i, seq in enumerate(sequences):
-            if seq.dim() != 1:
-                raise ValueError(
-                    f"Expected 1D tensor at index {i}, got {seq.dim()}D tensor with shape {seq.shape}"
-                )
-            if seq.dtype != self.dtype:
-                seq = seq.to(self.dtype)
-            normalized.append(seq)
-
-        # --- stream into buffer, slice off full chunks ---
-        self._buffer = []
-        self._packages = []
-        pack_size = self.pack_size
-        buf = self._buffer
-
-        for seq in normalized:
-            buf.extend(seq.tolist())
-            while len(buf) >= pack_size:
-                self._packages.append(torch.tensor(buf[:pack_size], dtype=self.dtype))
-                buf = buf[pack_size:]
-
-        # flush tail with padding
-        if buf:
-            padded = buf + [self.pad_value] * (pack_size - len(buf))
-            self._packages.append(torch.tensor(padded, dtype=self.dtype))
-
-        self._pos = len(buf)
-        return self._packages
-
-
-def pack_tensors(
-    tensors: Dict[str, List[Tensor]],
-    pack_size: int,
-    pad_value: Union[int, bool] = 0,
-    dtypes: Optional[Dict[str, torch.dtype]] = None,
-) -> Dict[str, List[Tensor]]:
-    """
-    Pack multiple named tensor groups in parallel.
-
-    Each group is packed independently with its own SequencePacker instance.
-    When dtypes is provided, packers use the declared dtype per key;
-    otherwise dtype is auto-inferred from the first tensor in each group.
-
-    Args:
-        tensors: Dict mapping key names to lists of 1D tensors.
-        pack_size: Fixed chunk length.
-        pad_value: Padding value for non-bool tensors.
-        dtypes: Optional per-key dtype declarations.
-
-    Returns:
-        Dict mapping key names to lists of packed tensors.
-    """
-    if dtypes is None:
-        dtypes = {}
-
-    output: Dict[str, List[Tensor]] = {}
-    for key, seqs in tensors.items():
-        dtype = dtypes.get(key)
-        packer = SequencePacker(pack_size, pad_value, dtype=dtype)
-        output[key] = packer.pack(seqs)
-    return output

pipeline/packing/__init__.py

@@ -0,0 +1,84 @@
+"""Sequence packing algorithms for LLM training data.
+
+Available packers:
+    - BfdPacker:        Best-Fit Decreasing, samples never split (default)
+    - FfDPacker:        First-Fit Decreasing, samples never split
+    - GreedyPacker:     First-fit in input order, samples never split
+"""
+
+from typing import Dict, List, Optional, Union
+
+import torch
+
+from pipeline.packing.base import BasePacker
+from pipeline.packing.binpack import GreedyPacker, FfDPacker, BfdPacker
+
+
+def pack_tensors(
+    tensors: Dict[str, List[torch.Tensor]],
+    pack_size: int,
+    pad_value: Union[int, bool] = 0,
+    dtypes: Optional[Dict[str, torch.dtype]] = None,
+    pad_values: Optional[Dict[str, Union[int, bool]]] = None,
+    algo: Optional[Union[str, BasePacker]] = None,
+) -> Dict[str, List[torch.Tensor]]:
+    """Pack multiple named tensor groups in parallel.
+
+    Each group is packed independently with its own packer instance.
+
+    Args:
+        tensors: Dict mapping key names to lists of 1D tensors.
+        pack_size: Fixed chunk length.
+        pad_value: Default padding value, used for keys not in pad_values.
+        dtypes: Optional per-key dtype declarations.
+        pad_values: Optional per-key padding values (e.g. pad_token_id for
+            'sequence', False for 'loss_mask', 0 for 'position_ids').
+        algo: Packing algorithm to use. Can be 'bfd' (default),
+              'ffd', 'greedy', or a BasePacker instance.
+
+    Returns:
+        Dict mapping key names to lists of packed tensors.
+    """
+    if dtypes is None:
+        dtypes = {}
+    if pad_values is None:
+        pad_values = {}
+
+    output: Dict[str, List[torch.Tensor]] = {}
+    for key, seqs in tensors.items():
+        key_pad = pad_values.get(key, pad_value)
+        actual_packer = _resolve_algo(algo, pack_size, key_pad)
+        dtype = dtypes.get(key)
+        if dtype is not None:
+            actual_packer.dtype = dtype
+        output[key] = actual_packer.pack(seqs)
+    return output
+
+
+def _resolve_algo(
+    algo: Optional[Union[str, BasePacker]],
+    pack_size: int,
+    pad_value: Union[int, bool],
+) -> BasePacker:
+    if algo is None or algo == "bfd":
+        return BfdPacker(pack_size, pad_value)
+    if isinstance(algo, BasePacker):
+        cls = type(algo)
+        return cls(pack_size, pad_value)
+    if algo == "ffd":
+        return FfDPacker(pack_size, pad_value)
+    if algo == "greedy":
+        return GreedyPacker(pack_size, pad_value)
+    raise ValueError(
+        f"Unknown packing algorithm: {algo}. "
+        f"Choose from: bfd, ffd, greedy"
+    )
+
+
+__all__ = [
+    "BasePacker",
+    "BfdPacker",
+    "FfDPacker",
+    "GreedyPacker",
+    "pack_tensors",
+]

pipeline/packing/base.py

@@ -0,0 +1,49 @@
+from abc import ABC, abstractmethod
+from typing import List, Optional, Union
+
+import torch
+from torch import Tensor
+
+
+class BasePacker(ABC):
+    """Abstract base class for sequence packing algorithms.
+
+    All packers must implement pack() and reset().
+    pack() takes a list of 1D tensors and returns a list of packed fixed-size tensors.
+    """
+
+    def __init__(
+        self,
+        pack_size: int,
+        pad_value: Union[int, bool] = 0,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        self.pack_size = pack_size
+        self.pad_value = pad_value
+        self.dtype = dtype
+
+    @abstractmethod
+    def pack(self, sequences: List[Tensor]) -> List[Tensor]:
+        """Pack sequences into fixed-size chunks."""
+        ...
+
+    @abstractmethod
+    def reset(self) -> None:
+        """Reset packer state for instance reuse."""
+        ...
+
+    def _validate_and_normalize(self, sequences: List[Tensor]) -> List[Tensor]:
+        """Validate 1D tensors and unify dtype."""
+        if self.dtype is None and sequences:
+            self.dtype = sequences[0].dtype
+
+        normalized: List[Tensor] = []
+        for i, seq in enumerate(sequences):
+            if seq.dim() != 1:
+                raise ValueError(
+                    f"Expected 1D tensor at index {i}, got {seq.dim()}D tensor with shape {seq.shape}"
+                )
+            if seq.dtype != self.dtype:
+                seq = seq.to(self.dtype)
+            normalized.append(seq)
+        return normalized

pipeline/packing/binpack.py

@@ -0,0 +1,174 @@
+from typing import List, Optional, Union
+
+import torch
+from torch import Tensor
+
+from pipeline.packing.base import BasePacker
+from pipeline.utils import error_handler
+
+
+def _truncate(tokens: List, max_len: int) -> List:
+    return tokens[:max_len]
+
+
+def _pad_bin(bin_list: List, target_len: int, pad_value: Union[int, bool], dtype: torch.dtype) -> Tensor:
+    bin_list.extend([pad_value] * (target_len - len(bin_list)))
+    return torch.tensor(bin_list, dtype=dtype)
+
+
+class GreedyPacker(BasePacker):
+    """Greedy first-fit packer (no sorting).
+
+    Sequences are packed in input order into the first bin with enough space.
+    Overlong sequences (> pack_size) are truncated to pack_size.
+    """
+
+    def __init__(
+        self,
+        pack_size: int,
+        pad_value: Union[int, bool] = 0,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__(pack_size, pad_value, dtype)
+        self._bins: List[List] = []
+
+    def reset(self) -> None:
+        self._bins = []
+
+    @error_handler()
+    def pack(self, sequences: List[Tensor]) -> List[Tensor]:
+        if not sequences:
+            return []
+
+        normalized = self._validate_and_normalize(sequences)
+        self._bins = []
+        pack_size = self.pack_size
+        pad_value = self.pad_value
+
+        for seq in normalized:
+            seq_len = int(seq.shape[0])
+            if seq_len > pack_size:
+                self._bins.append(_truncate(seq.tolist(), pack_size))
+                continue
+            placed = False
+            for bin_list in self._bins:
+                if len(bin_list) + seq_len <= pack_size:
+                    bin_list.extend(seq.tolist())
+                    placed = True
+                    break
+            if not placed:
+                self._bins.append(list(seq.tolist()))
+
+        packages: List[Tensor] = []
+        for bin_list in self._bins:
+            packages.append(_pad_bin(bin_list, pack_size, pad_value, self.dtype))
+
+        return packages
+
+
+class FfDPacker(BasePacker):
+    """First-Fit Decreasing (FFD) bin-packing packer.
+
+    Sequences are sorted by descending length, then packed into the first
+    bin with enough space. Overlong sequences are truncated to pack_size.
+    """
+
+    def __init__(
+        self,
+        pack_size: int,
+        pad_value: Union[int, bool] = 0,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__(pack_size, pad_value, dtype)
+        self._bins: List[List] = []
+
+    def reset(self) -> None:
+        self._bins = []
+
+    @error_handler()
+    def pack(self, sequences: List[Tensor]) -> List[Tensor]:
+        if not sequences:
+            return []
+
+        normalized = self._validate_and_normalize(sequences)
+        self._bins = []
+        pack_size = self.pack_size
+        pad_value = self.pad_value
+
+        indexed = [(int(s.shape[0]), s) for s in normalized]
+        indexed.sort(key=lambda x: x[0], reverse=True)
+
+        for seq_len, seq in indexed:
+            if seq_len > pack_size:
+                self._bins.append(_truncate(seq.tolist(), pack_size))
+                continue
+            placed = False
+            for bin_list in self._bins:
+                if len(bin_list) + seq_len <= pack_size:
+                    bin_list.extend(seq.tolist())
+                    placed = True
+                    break
+            if not placed:
+                self._bins.append(list(seq.tolist()))
+
+        packages: List[Tensor] = []
+        for bin_list in self._bins:
+            packages.append(_pad_bin(bin_list, pack_size, pad_value, self.dtype))
+
+        return packages
+
+
+class BfdPacker(BasePacker):
+    """Best-Fit Decreasing (BFD) bin-packing packer.
+
+    Sequences are sorted by descending length, then packed into the bin
+    that minimizes remaining space (tightest fit).
+    Overlong sequences are truncated to pack_size.
+    """
+
+    def __init__(
+        self,
+        pack_size: int,
+        pad_value: Union[int, bool] = 0,
+        dtype: Optional[torch.dtype] = None,
+    ):
+        super().__init__(pack_size, pad_value, dtype)
+        self._bins: List[List] = []
+
+    def reset(self) -> None:
+        self._bins = []
+
+    @error_handler()
+    def pack(self, sequences: List[Tensor]) -> List[Tensor]:
+        if not sequences:
+            return []
+
+        normalized = self._validate_and_normalize(sequences)
+        self._bins = []
+        pack_size = self.pack_size
+        pad_value = self.pad_value
+
+        indexed = [(int(s.shape[0]), s) for s in normalized]
+        indexed.sort(key=lambda x: x[0], reverse=True)
+
+        for seq_len, seq in indexed:
+            if seq_len > pack_size:
+                self._bins.append(_truncate(seq.tolist(), pack_size))
+                continue
+            best_idx = -1
+            best_remain = pack_size + 1
+            for i, bin_list in enumerate(self._bins):
+                remain = pack_size - len(bin_list)
+                if seq_len <= remain < best_remain:
+                    best_remain = remain
+                    best_idx = i
+            if best_idx >= 0:
+                self._bins[best_idx].extend(seq.tolist())
+            else:
+                self._bins.append(list(seq.tolist()))
+
+        packages: List[Tensor] = []
+        for bin_list in self._bins:
+            packages.append(_pad_bin(bin_list, pack_size, pad_value, self.dtype))
+
+        return packages

pipeline/processors/base.py

@@ -88,6 +88,21 @@ class BaseProcessor(ABC):
         """Return list of output tensor key names."""
         pass
 
+    def process_batch(self, input_dicts: List[Dict[str, Any]]) -> List[Dict[str, Tensor]]:
+        """Process a batch of input samples.
+
+        Default implementation calls process() for each sample.
+        Subclasses should override for efficient batch processing
+        (e.g., using tokenizer.encode_batch).
+
+        Args:
+            input_dicts: List of input dictionaries.
+
+        Returns:
+            List of output dictionaries mapping output key names to tensors.
+        """
+        return [self.process(d) for d in input_dicts]
+
     def validate_input(self, input_dict: Dict[str, Any]) -> None:
         """Validate input against schema before processing.
 

pipeline/processors/pretrain.py

@@ -44,11 +44,11 @@ class PreTrainProcessor(BaseProcessor):
         return {"sequence": torch.tensor(tokens, dtype=torch.int32)}
 
     def process_batch(self, input_dicts: List[Dict[str, Any]]) -> List[Dict[str, Tensor]]:
-        texts = [f"{item['text']}{self._eos_token}" for item in input_dicts]
-        encoded = self.tokenizer.encode(texts)
+        texts = [f"{d['text']}{self._eos_token}" for d in input_dicts]
+        batch_tokens = self.tokenizer.encode(texts)
         return [
             {"sequence": torch.tensor(tokens, dtype=torch.int32)}
-            for tokens in encoded
+            for tokens in batch_tokens
         ]
 
     @property

pipeline/processors/sft.py

@@ -7,7 +7,7 @@ from torch import Tensor
 
 from pipeline.tokenize import AutoTokenizer
 from pipeline.strategies import PromptStrategy, ChatMLStrategy
-from pipeline.processors.base import BaseProcessor, ProcessorSchema, encode_with_mask
+from pipeline.processors.base import BaseProcessor, ProcessorSchema
 from pipeline.processors.factory import ProcessorFactory
 
 
@@ -15,15 +15,15 @@ from pipeline.processors.factory import ProcessorFactory
 class SFTProcessor(BaseProcessor):
     """Supervised fine-tuning data processor.
 
-    Supports two input formats:
+    Input formats:
       1. messages (recommended):
          ``{"messages": [{"role": "user", "content": "..."},
                          {"role": "assistant", "content": "..."}]}``
-         Multi-turn and system prompts are supported.
-         The tokenizer's ``apply_chat_template`` is used for rendering.
+         Multi-turn and system prompts are supported.  Each assistant
+         turn gets ``loss_mask = 1``; all other roles get 0.
       2. legacy query/response:
          ``{"query": "...", "response": "..."}``
-         Falls back to the configured PromptStrategy (ChatML by default).
+         Internally converted to messages.
 
     Output schema:
         - sequence: int32 tensor - Combined token IDs (prompt + response)
@@ -63,39 +63,23 @@ class SFTProcessor(BaseProcessor):
         if "messages" in input_dict:
             return self._process_messages(input_dict["messages"])
         if "query" in input_dict and "response" in input_dict:
-            return self._process_legacy(input_dict)
+            return self._process_messages([
+                {"role": "user", "content": input_dict["query"]},
+                {"role": "assistant", "content": input_dict["response"]},
+            ])
         raise KeyError(
             "Input must contain 'messages' or 'query'/'response' pair"
         )
 
-    def process_batch(self, input_dicts: List[Dict[str, Any]]) -> List[Dict[str, Tensor]]:
-        results: List[Optional[Dict[str, Tensor]]] = [None] * len(input_dicts)
-        message_indices = [i for i, item in enumerate(input_dicts) if "messages" in item]
-        legacy_indices = [
-            i
-            for i, item in enumerate(input_dicts)
-            if "messages" not in item and "query" in item and "response" in item
-        ]
-        if len(message_indices) + len(legacy_indices) != len(input_dicts):
-            raise KeyError("Input must contain 'messages' or 'query'/'response' pair")
-
-        if message_indices:
-            items = [input_dicts[i] for i in message_indices]
-            batch_results = self._process_messages_batch(
-                [item["messages"] for item in items]
-            )
-            for index, result in zip(message_indices, batch_results):
-                results[index] = result
-
-        if legacy_indices:
-            items = [input_dicts[i] for i in legacy_indices]
-            batch_results = self._process_legacy_batch(items)
-            for index, result in zip(legacy_indices, batch_results):
-                results[index] = result
-
-        if any(result is None for result in results):
-            raise RuntimeError("Batch processing did not produce all results")
-        return results
+    def _extract_messages(self, input_dict: Dict[str, Any]) -> Optional[List[Dict[str, str]]]:
+        if "messages" in input_dict:
+            return input_dict["messages"]
+        if "query" in input_dict and "response" in input_dict:
+            return [
+                {"role": "user", "content": input_dict["query"]},
+                {"role": "assistant", "content": input_dict["response"]},
+            ]
+        return None
 
     def _process_messages(self, messages: List[Dict[str, str]]) -> Dict[str, Tensor]:
         if not messages:
@@ -103,118 +87,72 @@ class SFTProcessor(BaseProcessor):
         if messages[-1]["role"] != "assistant":
             raise ValueError("Last message must have role 'assistant'")
 
-        last_asst_idx = max(
-            i for i, m in enumerate(messages) if m["role"] == "assistant"
-        )
+        strategy = self.strategy or ChatMLStrategy(self.tokenizer)
 
-        full_text = self.tokenizer.apply_chat_template(
-            messages, tokenize=False, add_generation_prompt=False
-        )
-        full_ids = self.tokenizer.encode(full_text, add_special_tokens=False)
+        prompt, resp = strategy.format_messages(messages)
 
-        prompt_text = self.tokenizer.apply_chat_template(
-            messages[:last_asst_idx],
-            tokenize=False,
-            add_generation_prompt=True,
-        )
-        prompt_ids = self.tokenizer.encode(prompt_text, add_special_tokens=False)
-
-        resp_ids = full_ids[len(prompt_ids) :]
-        if not resp_ids:
-            raise ValueError("Empty assistant response")
-
-        tokens, loss_mask = encode_with_mask(prompt_ids, list(resp_ids))
-
-        if self.max_seq_len and len(tokens) > self.max_seq_len:
-            tokens = tokens[: self.max_seq_len]
+        sequence = torch.tensor(prompt + resp, dtype=torch.int32)
+        loss_mask = torch.zeros(len(sequence), dtype=torch.bool)
+        loss_mask[len(prompt) :] = True
+        if self.max_seq_len and len(sequence) > self.max_seq_len:
+            sequence = sequence[: self.max_seq_len]
             loss_mask = loss_mask[: self.max_seq_len]
-
-        position_ids = torch.arange(len(tokens), dtype=torch.int32)
+        position_ids = torch.arange(len(sequence), dtype=torch.int32)
         return {
-            "sequence": tokens,
+            "sequence": sequence,
             "loss_mask": loss_mask,
             "position_ids": position_ids,
         }
 
-    def _process_messages_batch(
-        self, conversations: List[List[Dict[str, str]]]
-    ) -> List[Dict[str, Tensor]]:
-        for messages in conversations:
-            if not messages:
-                raise ValueError("Messages list is empty")
-            if messages[-1]["role"] != "assistant":
-                raise ValueError("Last message must have role 'assistant'")
+    def process_batch(self, input_dicts: List[Dict[str, Any]]) -> List[Optional[Dict[str, Tensor]]]:
+        strategy = self.strategy or ChatMLStrategy(self.tokenizer)
 
-        assistant_indices = [
-            max(i for i, message in enumerate(messages) if message["role"] == "assistant")
-            for messages in conversations
-        ]
-        full_texts = [
-            self.tokenizer.apply_chat_template(
-                messages, tokenize=False, add_generation_prompt=False
-            )
-            for messages in conversations
-        ]
-        prompt_texts = [
-            self.tokenizer.apply_chat_template(
-                messages[:assistant_idx], tokenize=False, add_generation_prompt=True
-            )
-            for messages, assistant_idx in zip(conversations, assistant_indices)
-        ]
-        full_ids_batch = self.tokenizer.encode(full_texts, add_special_tokens=False)
-        prompt_ids_batch = self.tokenizer.encode(prompt_texts, add_special_tokens=False)
+        prompts_text: List[str] = []
+        fulls_text: List[str] = []
+        indices: List[int] = []
+        results: List[Optional[Dict[str, Tensor]]] = [None] * len(input_dicts)
 
-        results = []
-        for full_ids, prompt_ids in zip(full_ids_batch, prompt_ids_batch):
-            resp_ids = full_ids[len(prompt_ids) :]
-            if not resp_ids:
-                raise ValueError("Empty assistant response")
-            tokens, loss_mask = encode_with_mask(prompt_ids, list(resp_ids))
-            if self.max_seq_len and len(tokens) > self.max_seq_len:
-                tokens = tokens[: self.max_seq_len]
+        for i, d in enumerate(input_dicts):
+            try:
+                messages = self._extract_messages(d)
+                if not messages or messages[-1]["role"] != "assistant":
+                    continue
+                last_asst = max(j for j, m in enumerate(messages) if m["role"] == "assistant")
+                prompt_text = self.tokenizer.apply_chat_template(
+                    messages[:last_asst], add_generation_prompt=True, tokenize=False
+                )
+                full_text = self.tokenizer.apply_chat_template(
+                    messages[: last_asst + 1], add_generation_prompt=False, tokenize=False
+                )
+                prompts_text.append(prompt_text)
+                fulls_text.append(full_text)
+                indices.append(i)
+            except Exception:
+                continue
+
+        if not prompts_text:
+            return results
+
+        prompt_tokens_list = self.tokenizer.encode(prompts_text)
+        full_tokens_list = self.tokenizer.encode(fulls_text)
+
+        for j, idx in enumerate(indices):
+            prompt_tokens = prompt_tokens_list[j]
+            full_tokens = full_tokens_list[j]
+            resp_tokens = full_tokens[len(prompt_tokens):]
+            sequence = torch.tensor(prompt_tokens + resp_tokens, dtype=torch.int32)
+            loss_mask = torch.zeros(len(sequence), dtype=torch.bool)
+            loss_mask[len(prompt_tokens):] = True
+            if self.max_seq_len and len(sequence) > self.max_seq_len:
+                sequence = sequence[: self.max_seq_len]
                 loss_mask = loss_mask[: self.max_seq_len]
-            results.append(
-                {
-                    "sequence": tokens,
-                    "loss_mask": loss_mask,
-                    "position_ids": torch.arange(len(tokens), dtype=torch.int32),
-                }
-            )
-        return results
+            position_ids = torch.arange(len(sequence), dtype=torch.int32)
+            results[idx] = {
+                "sequence": sequence,
+                "loss_mask": loss_mask,
+                "position_ids": position_ids,
+            }
 
-    def _process_legacy(self, input_dict: Dict[str, Any]) -> Dict[str, Tensor]:
-        strategy = self.strategy or ChatMLStrategy(self.tokenizer)
-
-        query_tokens = self.tokenizer.encode(input_dict["query"])
-        response_tokens = self.tokenizer.encode(input_dict["response"])
-
-        prompt = strategy.assemble_prompt(query_tokens)
-        response = strategy.assemble_response(response_tokens)
-
-        tokens, loss_mask = encode_with_mask(prompt, response)
-        position_ids = torch.arange(len(tokens), dtype=torch.int32)
-        return {"sequence": tokens, "loss_mask": loss_mask, "position_ids": position_ids}
-
-    def _process_legacy_batch(
-        self, input_dicts: List[Dict[str, Any]]
-    ) -> List[Dict[str, Tensor]]:
-        strategy = self.strategy or ChatMLStrategy(self.tokenizer)
-        query_batch = self.tokenizer.encode([item["query"] for item in input_dicts])
-        response_batch = self.tokenizer.encode(
-            [item["response"] for item in input_dicts]
-        )
-        results = []
-        for query_tokens, response_tokens in zip(query_batch, response_batch):
-            prompt = strategy.assemble_prompt(query_tokens)
-            response = strategy.assemble_response(response_tokens)
-            tokens, loss_mask = encode_with_mask(prompt, response)
-            results.append(
-                {
-                    "sequence": tokens,
-                    "loss_mask": loss_mask,
-                    "position_ids": torch.arange(len(tokens), dtype=torch.int32),
-                }
-            )
         return results
 
     @property

pipeline/strategies/chatml.py

@@ -1,42 +1,91 @@
 """ChatML format strategy."""
 
-from typing import List
+from typing import Dict, List, Tuple
 
 from pipeline.tokenize import AutoTokenizer
 from pipeline.strategies.base import PromptStrategy
 from pipeline.strategies.factory import StrategyFactory
 
+DEFAULT_CHATML_TEMPLATE = (
+    "{% for message in messages %}"
+    "{% if message['role'] == 'system' %}"
+    "{{ '<|im_start|>system\n' + message['content'] + '<|im_end|>\n' }}"
+    "{% elif message['role'] == 'user' %}"
+    "{{ '<|im_start|>user\n' + message['content'] + '<|im_end|>\n' }}"
+    "{% elif message['role'] == 'assistant' %}"
+    "{{ '<|im_start|>assistant\n' + message['content'] + '<|im_end|>\n' }}"
+    "{% endif %}"
+    "{% endfor %}"
+    "{% if add_generation_prompt %}"
+    "{{ '<|im_start|>assistant\n' }}"
+    "{% endif %}"
+)
+
 
 @StrategyFactory.register("chatml")
 class ChatMLStrategy(PromptStrategy):
-    """ChatML format strategy."""
+    """ChatML format strategy.
 
-    def __init__(
-        self,
-        tokenizer: AutoTokenizer,
-        user_start: str = "<｜im▁start｜>user\n",
-        user_end: str = "<｜im▁end｜>\n",
-        assistant_start: str = "<｜im▁start｜>assistant\n",
-        assistant_end: str = "<｜im▁end｜>\n",
-    ):
+    Renders messages using the tokenizer's jinja chat_template from
+    ``tokenizer_config.json``.  Falls back to DEFAULT_CHATML_TEMPLATE
+    when no template is configured.
+
+    The strategy does **not** hard-code any special tokens – all
+    formatting is driven by the jinja template.
+    """
+
+    def __init__(self, tokenizer: AutoTokenizer):
         super().__init__(tokenizer)
-
-        self._user_start_ids = self._encode_format(user_start)
-        self._user_end_ids = self._encode_format(user_end)
-        self._assistant_start_ids = self._encode_format(assistant_start)
-        self._assistant_end_ids = self._encode_format(assistant_end)
+        if tokenizer._chat_template is None:
+            tokenizer.set_chat_template(DEFAULT_CHATML_TEMPLATE)
 
     @property
     def name(self) -> str:
         return "chatml"
 
+    def format_messages(
+        self,
+        messages: List[Dict[str, str]],
+    ) -> Tuple[List[int], List[int]]:
+        """Render a single-turn messages conversation.
+
+        Returns ``(prompt_tokens, response_tokens)`` where
+        *prompt_tokens* contains everything up to (and including) the
+        last assistant start marker, and *response_tokens* is the
+        assistant content plus the closing markers.
+        """
+        last_asst = max(
+            i for i, m in enumerate(messages) if m["role"] == "assistant"
+        )
+
+        prompt = self.tokenizer.apply_chat_template(
+            messages[:last_asst],
+            add_generation_prompt=True,
+            tokenize=True,
+        )
+        full = self.tokenizer.apply_chat_template(
+            messages[: last_asst + 1],
+            add_generation_prompt=False,
+            tokenize=True,
+        )
+        return prompt, full[len(prompt) :]
+
     def assemble_prompt(self, query_tokens: List[int]) -> List[int]:
-        return (
-            self._user_start_ids
-            + query_tokens
-            + self._user_end_ids
-            + self._assistant_start_ids
+        text = self.tokenizer.decode(query_tokens)
+        return self.tokenizer.apply_chat_template(
+            [{"role": "user", "content": text}],
+            add_generation_prompt=True,
+            tokenize=True,
         )
 
     def assemble_response(self, response_tokens: List[int]) -> List[int]:
-        return response_tokens + self._assistant_end_ids
+        text = self.tokenizer.decode(response_tokens)
+        full = self.tokenizer.apply_chat_template(
+            [{"role": "assistant", "content": text}],
+            add_generation_prompt=False,
+            tokenize=True,
+        )
+        opening = self.tokenizer.apply_chat_template(
+            [], add_generation_prompt=True, tokenize=True
+        )
+        return full[len(opening) :]

pipeline/tokenize/tokenizer.py

@@ -244,22 +244,18 @@ class AutoTokenizer:
                 "Tokenizer not initialized. Load or create a tokenizer first."
             )
 
-        if isinstance(tokens, str):
-            encoded = self._tokenizer.encode(
-                tokens,
-                is_pretokenized=is_pretokenized,
-                add_special_tokens=add_special_tokens,
-            )
-            return encoded.ids if out_ids else encoded.tokens
-        else:
-            encoded_list = self._tokenizer.encode_batch(
-                tokens,
-                is_pretokenized=is_pretokenized,
-                add_special_tokens=add_special_tokens,
-            )
-            return [
-                encoded.ids if out_ids else encoded.tokens for encoded in encoded_list
-            ]
+        single = isinstance(tokens, str)
+        if single:
+            tokens = [tokens]
+        encoded_list = self._tokenizer.encode_batch(
+            tokens,
+            is_pretokenized=is_pretokenized,
+            add_special_tokens=add_special_tokens,
+        )
+        result = [
+            encoded.ids if out_ids else encoded.tokens for encoded in encoded_list
+        ]
+        return result[0] if single else result
 
     def decode(self, tokens: List[int], skip_special_tokens: bool = True) -> str:
         """Decode token IDs to text."""
@@ -270,6 +266,12 @@ class AutoTokenizer:
 
         return self._tokenizer.decode(tokens, skip_special_tokens=skip_special_tokens)
 
+    def token_to_id(self, token: str) -> Optional[int]:
+        """Convert a token string to its integer ID."""
+        if self._tokenizer is None:
+            raise RuntimeError("Tokenizer not initialized.")
+        return self._tokenizer.token_to_id(token)
+
     def __len__(self) -> int:
         if self._tokenizer is None:
             return 0