| 1 |
"""Tokenizer load + fixup. |
| 2 |
|
| 3 |
Three invariants enforced here (see CLAUDE.md pitfall #4): |
| 4 |
|
| 5 |
1. **pad_token != eos_token.** HF defaults `pad_token` to `eos_token` |
| 6 |
on most bases; if `pad_token is None`, we MUST pick a different |
| 7 |
token, or labels get corrupted by mid-sequence EOS masking. |
| 8 |
Fallback order: `unk_token` → else add `<|pad|>` as a new special |
| 9 |
token (which grows the vocab and sets `tokenizer_grew=True` for |
| 10 |
the caller to propagate into the LoRA config). |
| 11 |
2. **chat_template must be present.** Without it, SFTTrainer can't |
| 12 |
render `messages`-shaped rows. We surface a typed |
| 13 |
`TokenizerBringupError` rather than letting SFT fail deep inside |
| 14 |
TRL with an opaque message. |
| 15 |
3. **Revision pinning.** Every load goes through the base model's |
| 16 |
40-char revision SHA — never a branch — so retrains under the |
| 17 |
same spec reproduce. |
| 18 |
|
| 19 |
Returns a `TokenizerBringup` dataclass rather than a bare tokenizer so |
| 20 |
the `tokenizer_grew` flag travels with the object. |
| 21 |
""" |
| 22 |
|
| 23 |
from __future__ import annotations |
| 24 |
|
| 25 |
from dataclasses import dataclass |
| 26 |
from typing import TYPE_CHECKING, Any |
| 27 |
|
| 28 |
from dlm.data.errors import TokenizerBringupError |
| 29 |
|
| 30 |
if TYPE_CHECKING: |
| 31 |
from transformers import PreTrainedTokenizerBase |
| 32 |
|
| 33 |
_PAD_TOKEN_LITERAL = "<|pad|>" |
| 34 |
|
| 35 |
|
| 36 |
@dataclass(frozen=True) |
| 37 |
class TokenizerBringup: |
| 38 |
"""Result of `prepare_tokenizer`. |
| 39 |
|
| 40 |
`tokenizer_grew=True` means a new `<|pad|>` token was added to the |
| 41 |
vocab. The LoRA config MUST set |
| 42 |
`modules_to_save=["embed_tokens","lm_head"]` in that case — |
| 43 |
otherwise the new embedding row will not be trained and its |
| 44 |
output distribution is undefined. |
| 45 |
""" |
| 46 |
|
| 47 |
tokenizer: PreTrainedTokenizerBase |
| 48 |
tokenizer_grew: bool |
| 49 |
pad_token: str |
| 50 |
chat_template: str |
| 51 |
|
| 52 |
|
| 53 |
def prepare_tokenizer(hf_id: str, revision: str) -> TokenizerBringup: |
| 54 |
"""Load the tokenizer for `hf_id` at `revision`, apply pad/template fixups.""" |
| 55 |
from transformers import AutoTokenizer |
| 56 |
|
| 57 |
tok: Any = AutoTokenizer.from_pretrained(hf_id, revision=revision, use_fast=True) |
| 58 |
grew = _ensure_pad_token(tok) |
| 59 |
_ensure_chat_template(tok, hf_id=hf_id) |
| 60 |
|
| 61 |
pad = tok.pad_token |
| 62 |
chat_template = tok.chat_template |
| 63 |
assert isinstance(pad, str) |
| 64 |
assert isinstance(chat_template, str) |
| 65 |
|
| 66 |
return TokenizerBringup( |
| 67 |
tokenizer=tok, |
| 68 |
tokenizer_grew=grew, |
| 69 |
pad_token=pad, |
| 70 |
chat_template=chat_template, |
| 71 |
) |
| 72 |
|
| 73 |
|
| 74 |
def _ensure_pad_token(tok: Any) -> bool: |
| 75 |
"""Guarantee `tok.pad_token` is set AND distinct from `tok.eos_token`. |
| 76 |
|
| 77 |
Returns True iff a new special token was added to the vocab. |
| 78 |
""" |
| 79 |
eos = getattr(tok, "eos_token", None) |
| 80 |
current_pad = getattr(tok, "pad_token", None) |
| 81 |
|
| 82 |
if current_pad is not None and current_pad != eos: |
| 83 |
return False |
| 84 |
|
| 85 |
# Either pad is unset, or it equals eos (the HF default we must override). |
| 86 |
unk = getattr(tok, "unk_token", None) |
| 87 |
if unk is not None and unk != eos: |
| 88 |
tok.pad_token = unk |
| 89 |
return False |
| 90 |
|
| 91 |
# Last resort: add a new pad token. This grows the vocab, which |
| 92 |
# forces training to update embed_tokens + lm_head. |
| 93 |
tok.add_special_tokens({"pad_token": _PAD_TOKEN_LITERAL}) |
| 94 |
return True |
| 95 |
|
| 96 |
|
| 97 |
def _ensure_chat_template(tok: Any, *, hf_id: str) -> None: |
| 98 |
template = getattr(tok, "chat_template", None) |
| 99 |
if template is None or not str(template).strip(): |
| 100 |
raise TokenizerBringupError( |
| 101 |
f"base model {hf_id!r} has no chat_template; " |
| 102 |
"supply one via --chat-template or pick a registry base" |
| 103 |
) |