From 94026fc8dc8e23423ec6bf232ba9ba6646cba21e Mon Sep 17 00:00:00 2001
From: Junhyuk Lee <58055473+xodn348@users.noreply.github.com>
Date: Tue, 19 May 2026 03:05:13 -0500
Subject: [PATCH] fix(loader): honour HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE in
 from_pretrained (#5598)

Reads HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE in FastLanguageModel.from_pretrained and FastModel.from_pretrained, forcing local_files_only=True so all delegation paths (load_in_4bit, load_in_8bit, full_finetuning, qat_scheme) and direct FastModel callers (FastVisionModel, FastTextModel) honour offline mode. Also gates HF_HUB_ENABLE_HF_TRANSFER in unsloth/dataprep/synthetic.py and adds an early return in get_statistics. Pairs with unslothai/unsloth-zoo#675. Fixes #5316.
---
 unsloth/dataprep/synthetic.py |  7 ++++++-
 unsloth/models/_utils.py      |  7 +++++++
 unsloth/models/loader.py      | 19 +++++++++++++++++++
 3 files changed, 32 insertions(+), 1 deletion(-)

diff --git a/unsloth/dataprep/synthetic.py b/unsloth/dataprep/synthetic.py
index 612c531f47..1be83c3bb9 100644
--- a/unsloth/dataprep/synthetic.py
+++ b/unsloth/dataprep/synthetic.py
@@ -21,7 +21,12 @@ from collections import deque
 import time
 import os
 
-os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
+_OFFLINE_VALS = {"1", "true", "yes", "on"}
+if not (
+    os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _OFFLINE_VALS
+    or os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _OFFLINE_VALS
+):
+    os.environ["HF_HUB_ENABLE_HF_TRANSFER"] = "1"
 import requests
 import torch
 import gc
diff --git a/unsloth/models/_utils.py b/unsloth/models/_utils.py
index a46d1f0c0e..f4313c4518 100644
--- a/unsloth/models/_utils.py
+++ b/unsloth/models/_utils.py
@@ -1771,6 +1771,13 @@ def get_statistics(local_files_only = False):
         return
     if local_files_only:
         return
+    # Also skip when HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE are set.
+    _offline_vals = {"1", "true", "yes", "on"}
+    if (
+        os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _offline_vals
+        or os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _offline_vals
+    ):
+        return
     from huggingface_hub.utils import (
         disable_progress_bars,
         enable_progress_bars,
diff --git a/unsloth/models/loader.py b/unsloth/models/loader.py
index fc91178d88..c10443e289 100644
--- a/unsloth/models/loader.py
+++ b/unsloth/models/loader.py
@@ -308,6 +308,16 @@ class FastLanguageModel(FastLlamaModel):
             if is_dist:
                 device_map = distributed_device_map
 
+        # Honour offline env vars BEFORE FastModel delegation so 8bit /
+        # full-finetuning / qat paths also receive local_files_only.
+        if not kwargs.get("local_files_only", False):
+            _offline = {"1", "true", "yes", "on"}
+            if (
+                os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _offline
+                or os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _offline
+            ):
+                kwargs["local_files_only"] = True
+
         if load_in_8bit or full_finetuning or qat_scheme is not None:
             return FastModel.from_pretrained(
                 model_name = model_name,
@@ -1055,6 +1065,15 @@ class FastModel(FastBaseModel):
         model_config = None
         peft_config = None
         local_files_only = kwargs.get("local_files_only", False)
+        # Mirror env-var fallback for direct callers (FastVisionModel / FastTextModel).
+        if not local_files_only:
+            _offline = {"1", "true", "yes", "on"}
+            if (
+                os.environ.get("TRANSFORMERS_OFFLINE", "").strip().lower() in _offline
+                or os.environ.get("HF_HUB_OFFLINE", "").strip().lower() in _offline
+            ):
+                local_files_only = True
+                kwargs["local_files_only"] = True
 
         try:
             model_config = AutoConfig.from_pretrained(