fix: keep offline GGUF export off the Hub for VLM tokenizers (#7481)
Resolve cached snapshot directories before loading PreTrainedTokenizerFast during VLM processor fallback so transformers does not call is_base_mistral() -> model_info() when HF_HUB_OFFLINE is set. Also probe the local cache in _has_tokenizer_model instead of model_info when offline. Fixes unslothai/unsloth#7481
This commit is contained in:
Souravrajvi0
parent
0807d03ed0
commit
63aa532402
4 changed files with 266 additions and 10 deletions
119
tests/saving/test_offline_gguf_vlm_tokenizer_7481.py
Normal file
119
tests/saving/test_offline_gguf_vlm_tokenizer_7481.py
Normal file
|
|
@ -0,0 +1,119 @@
|
|||
"""Offline GGUF export must not probe the Hub for VLM tokenizer metadata (issue #7481).
|
||||
|
||||
Regression for ``PreTrainedTokenizerFast.from_pretrained`` on a repo id calling
|
||||
``is_base_mistral()`` -> ``model_info()`` even with ``TRANSFORMERS_OFFLINE=1``.
|
||||
Pure CPU, no network, no GPU.
|
||||
"""
|
||||
|
||||
import json
|
||||
import os
|
||||
from types import SimpleNamespace
|
||||
from unittest.mock import patch
|
||||
|
||||
import pytest
|
||||
|
||||
from unsloth.models import loader_utils as L
|
||||
|
||||
|
||||
_REPO = "llmfan46/gemma-4-E4B-it-ultra-uncensored-heretic"
|
||||
_COMMIT = "5964fe4c7339c5974e879baba8982a09616f68ca"
|
||||
|
||||
|
||||
def _write_gemma4_cache(root, repo_id=_REPO, commit=_COMMIT):
|
||||
"""Minimal cached snapshot matching the reporter's layout."""
|
||||
org, name = repo_id.split("/")
|
||||
repo_root = root / f"models--{org}--{name}"
|
||||
snap = repo_root / "snapshots" / commit
|
||||
snap.mkdir(parents=True)
|
||||
refs = repo_root / "refs"
|
||||
refs.mkdir(parents=True, exist_ok=True)
|
||||
(refs / "main").write_text(commit, encoding="utf-8")
|
||||
(snap / "tokenizer_config.json").write_text(
|
||||
json.dumps({"tokenizer_class": "GemmaTokenizer", "model_max_length": 8192}),
|
||||
encoding="utf-8",
|
||||
)
|
||||
(snap / "tokenizer.json").write_text(
|
||||
json.dumps(
|
||||
{
|
||||
"version": "1.0",
|
||||
"truncation": None,
|
||||
"padding": None,
|
||||
"added_tokens": [],
|
||||
"normalizer": None,
|
||||
"pre_tokenizer": None,
|
||||
"post_processor": None,
|
||||
"decoder": None,
|
||||
"model": {"type": "BPE", "vocab": {"<pad>": 0}, "merges": []},
|
||||
}
|
||||
),
|
||||
encoding="utf-8",
|
||||
)
|
||||
(snap / "processor_config.json").write_text("{}", encoding="utf-8")
|
||||
(snap / "config.json").write_text(
|
||||
json.dumps({"model_type": "gemma4"}),
|
||||
encoding="utf-8",
|
||||
)
|
||||
return snap
|
||||
|
||||
|
||||
def _offline_env(monkeypatch, cache_root):
|
||||
monkeypatch.setenv("HF_HUB_OFFLINE", "1")
|
||||
monkeypatch.setenv("TRANSFORMERS_OFFLINE", "1")
|
||||
monkeypatch.setenv("HF_HUB_CACHE", str(cache_root))
|
||||
|
||||
|
||||
def test_resolve_hub_repo_local_dir_from_cached_snapshot(tmp_path, monkeypatch):
|
||||
snap = _write_gemma4_cache(tmp_path)
|
||||
_offline_env(monkeypatch, tmp_path)
|
||||
|
||||
got = L._resolve_hub_repo_local_dir(_REPO, local_files_only=True, cache_dir=str(tmp_path))
|
||||
assert got == str(snap)
|
||||
|
||||
|
||||
def test_hub_repo_or_local_path_prefers_snapshot_over_repo_id(tmp_path, monkeypatch):
|
||||
snap = _write_gemma4_cache(tmp_path)
|
||||
_offline_env(monkeypatch, tmp_path)
|
||||
|
||||
got = L._hub_repo_or_local_path(_REPO, local_files_only=True, cache_dir=str(tmp_path))
|
||||
assert got == str(snap)
|
||||
assert got != _REPO
|
||||
|
||||
|
||||
def test_load_pretrained_tokenizer_fast_passes_snapshot_not_repo_id(tmp_path, monkeypatch):
|
||||
snap = _write_gemma4_cache(tmp_path)
|
||||
_offline_env(monkeypatch, tmp_path)
|
||||
|
||||
seen_paths = []
|
||||
|
||||
class _FakeFast:
|
||||
@classmethod
|
||||
def from_pretrained(cls, path, **kwargs):
|
||||
seen_paths.append(path)
|
||||
assert kwargs.get("local_files_only") is True
|
||||
return SimpleNamespace(name_or_path=path)
|
||||
|
||||
monkeypatch.setattr(
|
||||
"transformers.PreTrainedTokenizerFast",
|
||||
_FakeFast,
|
||||
raising=False,
|
||||
)
|
||||
|
||||
with patch("huggingface_hub.HfApi.model_info") as model_info:
|
||||
model_info.side_effect = AssertionError("model_info must not run offline")
|
||||
tok = L._load_pretrained_tokenizer_fast(_REPO, cache_dir=str(tmp_path))
|
||||
|
||||
assert seen_paths == [str(snap)]
|
||||
assert tok.name_or_path == str(snap)
|
||||
|
||||
|
||||
def test_has_tokenizer_model_offline_skips_model_info(tmp_path, monkeypatch):
|
||||
from unsloth.save import _has_tokenizer_model
|
||||
|
||||
_write_gemma4_cache(tmp_path)
|
||||
_offline_env(monkeypatch, tmp_path)
|
||||
|
||||
tok = SimpleNamespace(name_or_path=_REPO, tokenizer=None)
|
||||
|
||||
with patch("huggingface_hub.HfApi.model_info") as model_info:
|
||||
model_info.side_effect = AssertionError("model_info must not run offline")
|
||||
assert _has_tokenizer_model(tok, token=None) is False
|
||||
|
|
@ -1126,6 +1126,113 @@ def _has_local_processor_files(path):
|
|||
)
|
||||
|
||||
|
||||
def _resolve_hub_repo_local_dir(
|
||||
repo_id,
|
||||
*,
|
||||
token = None,
|
||||
cache_dir = None,
|
||||
local_files_only = False,
|
||||
filenames = (
|
||||
"tokenizer_config.json",
|
||||
"config.json",
|
||||
"tokenizer.json",
|
||||
"preprocessor_config.json",
|
||||
"processor_config.json",
|
||||
),
|
||||
):
|
||||
"""Return a local snapshot directory for a Hub repo id when files are cached.
|
||||
|
||||
``PreTrainedTokenizerFast.from_pretrained`` on a repo id can call
|
||||
``is_base_mistral()`` -> ``model_info()`` even with ``local_files_only=True``
|
||||
(issue #7481). Loading from the resolved snapshot dir avoids that probe.
|
||||
"""
|
||||
if not isinstance(repo_id, str) or not repo_id:
|
||||
return None
|
||||
if os.path.isdir(repo_id):
|
||||
return repo_id
|
||||
if cache_dir is None:
|
||||
cache_dir = os.environ.get("HF_HUB_CACHE")
|
||||
from huggingface_hub import hf_hub_download
|
||||
|
||||
for filename in filenames:
|
||||
try:
|
||||
path = hf_hub_download(
|
||||
repo_id = repo_id,
|
||||
filename = filename,
|
||||
token = token,
|
||||
cache_dir = cache_dir,
|
||||
local_files_only = local_files_only,
|
||||
)
|
||||
if path and os.path.isfile(path):
|
||||
return os.path.dirname(path)
|
||||
except Exception:
|
||||
continue
|
||||
return None
|
||||
|
||||
|
||||
def _hub_repo_or_local_path(
|
||||
repo_id,
|
||||
*,
|
||||
token = None,
|
||||
cache_dir = None,
|
||||
local_files_only = False,
|
||||
filenames = None,
|
||||
):
|
||||
"""Prefer a cached snapshot path over a Hub repo id when offline or ``local_files_only``."""
|
||||
lfo = bool(local_files_only) or _env_says_offline()
|
||||
if not lfo and os.path.isdir(repo_id):
|
||||
return repo_id
|
||||
local_dir = _resolve_hub_repo_local_dir(
|
||||
repo_id,
|
||||
token = token,
|
||||
cache_dir = cache_dir,
|
||||
local_files_only = lfo,
|
||||
filenames = filenames
|
||||
or (
|
||||
"tokenizer_config.json",
|
||||
"config.json",
|
||||
"tokenizer.json",
|
||||
"preprocessor_config.json",
|
||||
"processor_config.json",
|
||||
),
|
||||
)
|
||||
return local_dir if local_dir is not None else repo_id
|
||||
|
||||
|
||||
def _load_pretrained_tokenizer_fast(
|
||||
tokenizer_name,
|
||||
*,
|
||||
padding_side = "left",
|
||||
token = None,
|
||||
trust_remote_code = False,
|
||||
cache_dir = None,
|
||||
local_files_only = False,
|
||||
):
|
||||
"""Load ``PreTrainedTokenizerFast`` without Hub metadata probes when cached/offline."""
|
||||
from transformers import PreTrainedTokenizerFast
|
||||
|
||||
lfo = bool(local_files_only) or _env_says_offline()
|
||||
load_path = _hub_repo_or_local_path(
|
||||
tokenizer_name,
|
||||
token = token,
|
||||
cache_dir = cache_dir,
|
||||
local_files_only = lfo,
|
||||
filenames = (
|
||||
"tokenizer_config.json",
|
||||
"tokenizer.json",
|
||||
"tokenizer.model",
|
||||
),
|
||||
)
|
||||
return PreTrainedTokenizerFast.from_pretrained(
|
||||
load_path,
|
||||
padding_side = padding_side,
|
||||
token = token,
|
||||
trust_remote_code = trust_remote_code,
|
||||
cache_dir = cache_dir,
|
||||
local_files_only = lfo,
|
||||
)
|
||||
|
||||
|
||||
def _resolve_checkpoint_tokenizer_name(
|
||||
old_model_name,
|
||||
kwargs,
|
||||
|
|
|
|||
|
|
@ -529,7 +529,9 @@ def unsloth_base_fast_generate(self, *args, **kwargs):
|
|||
# Offline helpers live in loader_utils.py (shared canonical source).
|
||||
from .loader_utils import (
|
||||
_get_effective_local_files_only,
|
||||
_hub_repo_or_local_path,
|
||||
_is_offline_related_error,
|
||||
_load_pretrained_tokenizer_fast,
|
||||
_offline_aware_load,
|
||||
)
|
||||
|
||||
|
|
@ -565,20 +567,27 @@ def _construct_vlm_processor_fallback(
|
|||
tell an offline failure (retry from cache) from a genuine one."""
|
||||
_fb_err = None
|
||||
try:
|
||||
from transformers import AutoImageProcessor, PreTrainedTokenizerFast, AutoConfig
|
||||
from transformers import AutoImageProcessor, AutoConfig
|
||||
from transformers.models.auto.processing_auto import PROCESSOR_MAPPING_NAMES
|
||||
import json
|
||||
|
||||
load_path = _hub_repo_or_local_path(
|
||||
tokenizer_name,
|
||||
token = token,
|
||||
cache_dir = cache_dir,
|
||||
local_files_only = local_files_only,
|
||||
)
|
||||
# Load image processor
|
||||
image_processor = AutoImageProcessor.from_pretrained(
|
||||
tokenizer_name,
|
||||
load_path,
|
||||
token = token,
|
||||
trust_remote_code = trust_remote_code,
|
||||
cache_dir = cache_dir,
|
||||
local_files_only = local_files_only,
|
||||
)
|
||||
# Load tokenizer via PreTrainedTokenizerFast (bypasses tokenizer_class check)
|
||||
tok = PreTrainedTokenizerFast.from_pretrained(
|
||||
# Load tokenizer via PreTrainedTokenizerFast (bypasses tokenizer_class check).
|
||||
# Resolve the cached snapshot first so transformers does not call model_info (#7481).
|
||||
tok = _load_pretrained_tokenizer_fast(
|
||||
tokenizer_name,
|
||||
padding_side = "left",
|
||||
token = token,
|
||||
|
|
@ -638,7 +647,7 @@ def _construct_vlm_processor_fallback(
|
|||
# Try the top-level config.model_type which often has the processor mapping.
|
||||
try:
|
||||
config = AutoConfig.from_pretrained(
|
||||
tokenizer_name,
|
||||
load_path,
|
||||
token = token,
|
||||
trust_remote_code = trust_remote_code,
|
||||
cache_dir = cache_dir,
|
||||
|
|
@ -1551,9 +1560,16 @@ class FastBaseModel:
|
|||
# Last resort: AutoTokenizer, then PreTrainedTokenizerFast (raise on network failure to retry).
|
||||
def _last_resort_tokenizer(lfo):
|
||||
from transformers import AutoTokenizer as _AutoTokenizer
|
||||
|
||||
load_path = _hub_repo_or_local_path(
|
||||
tokenizer_name,
|
||||
token = token,
|
||||
cache_dir = kwargs.get("cache_dir"),
|
||||
local_files_only = lfo,
|
||||
)
|
||||
try:
|
||||
return _AutoTokenizer.from_pretrained(
|
||||
tokenizer_name,
|
||||
load_path,
|
||||
padding_side = "left",
|
||||
token = token,
|
||||
trust_remote_code = trust_remote_code,
|
||||
|
|
@ -1561,8 +1577,7 @@ class FastBaseModel:
|
|||
local_files_only = lfo,
|
||||
)
|
||||
except Exception:
|
||||
from transformers import PreTrainedTokenizerFast
|
||||
return PreTrainedTokenizerFast.from_pretrained(
|
||||
return _load_pretrained_tokenizer_fast(
|
||||
tokenizer_name,
|
||||
padding_side = "left",
|
||||
token = token,
|
||||
|
|
|
|||
|
|
@ -51,7 +51,7 @@ import subprocess
|
|||
import psutil
|
||||
import re
|
||||
from transformers.models.llama.modeling_llama import logger
|
||||
from .models.loader_utils import get_model_name
|
||||
from .models.loader_utils import get_model_name, _env_says_offline, _resolve_hub_repo_local_dir
|
||||
from .models._utils import _convert_torchao_model
|
||||
from .ollama_template_mappers import OLLAMA_TEMPLATES, MODEL_TO_OLLAMA_TEMPLATE_MAPPER
|
||||
from transformers import ProcessorMixin, PreTrainedTokenizerBase
|
||||
|
|
@ -445,6 +445,21 @@ def _has_tokenizer_model(tokenizer, token = None):
|
|||
if source in _TOKENIZER_MODEL_CACHE:
|
||||
return _TOKENIZER_MODEL_CACHE[source]
|
||||
|
||||
# Offline: probe the local cache instead of model_info (issue #7481).
|
||||
if _env_says_offline():
|
||||
local_dir = _resolve_hub_repo_local_dir(
|
||||
source,
|
||||
token = token,
|
||||
local_files_only = True,
|
||||
filenames = ("tokenizer.model", "tokenizer.json", "tokenizer_config.json"),
|
||||
cache_dir = os.environ.get("HF_HUB_CACHE"),
|
||||
)
|
||||
if local_dir is not None:
|
||||
has_tokenizer_model = os.path.isfile(os.path.join(local_dir, "tokenizer.model"))
|
||||
_TOKENIZER_MODEL_CACHE[source] = has_tokenizer_model
|
||||
return has_tokenizer_model
|
||||
return False
|
||||
|
||||
try:
|
||||
repo_info = HfApi(token = token).model_info(source, files_metadata = False)
|
||||
except Exception:
|
||||
|
|
@ -3773,7 +3788,7 @@ def unsloth_convert_lora_to_ggml_and_save_locally(
|
|||
return _unsloth_save_lora_gguf(self, tokenizer, save_directory, outtype = outtype)
|
||||
|
||||
|
||||
from .models.loader_utils import get_model_name
|
||||
from .models.loader_utils import get_model_name, _env_says_offline, _resolve_hub_repo_local_dir
|
||||
from unsloth_zoo.saving_utils import (
|
||||
merge_and_overwrite_lora,
|
||||
prepare_saving,
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue