[Cache] Fix environment variable handling for offline mode (#1902)

SUMMARY:

Previously, llm-compressor ignored HF_HUB_CACHE and other environment
variables when loading models and datasets, making offline mode
difficult to use with unified cache directories.

This change:
- Removes hard-coded TRANSFORMERS_CACHE in model_load/helpers.py to
respect HF_HOME, HF_HUB_CACHE environment variables
- Propagates cache_dir from model_args to dataset_args to enable unified
cache directory for both models and datasets
- Updates dataset loading to use cache_dir parameter instead of
hardcoded None

Now users can specify cache_dir parameter or use HF_HOME/HF_HUB_CACHE
environment variables for true offline operation.

Offline mode is super helpful to supply-chain security use cases. It
helps us generate trustworthy SBOMs for AI stuff. 🔐
🧠

TEST PLAN:

I start with the oneshot example from the README, and called it
`example.py`:

```python
""" This is the example from the README """

from llmcompressor.modifiers.smoothquant import SmoothQuantModifier
from llmcompressor.modifiers.quantization import GPTQModifier
from llmcompressor import oneshot

recipe = [
    SmoothQuantModifier(smoothing_strength=0.8),
    GPTQModifier(scheme="W8A8", targets="Linear", ignore=["lm_head"]),
]

oneshot(
    model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    dataset="open_platypus",
    recipe=recipe,
    output_dir="TinyLlama-1.1B-Chat-v1.0-INT8",
    max_seq_length=2048,
    num_calibration_samples=512,
)
```

Next, remove your hf local cache to ensure your system has nothing
available to it yet:

```bash
❯ rm -rf ~/.cache/huggingface
```

Then, run `example.py` with the HF_HUB_OFFLINE=1 env var. This should
fail, proving that you have nothing cached.

```bash
❯ HF_HUB_OFFLINE=1 python example.py
Traceback (most recent call last):
  File "/home/rbean/code/llm-compressor/testtest/lib64/python3.13/site-packages/transformers/utils/hub.py", line 479, in cached_files
...
<snip>
...
OSError: We couldn't connect to 'https://huggingface.co' to load the files, and couldn't find them in the cached files.
Check your internet connection or see how to run the library in offline mode at 'https://huggingface.co/docs/transformers/installation#offline-mode'.
```

Good. Now, run it with `HF_HUB=./hf-hub` which will run it in online
mode, populating the cache in a new non-standard location (just to be
sure things don't get mixed up during our test):

```bash
❯ HF_HOME=./hf-hub python example.py
<lots of downloading happens, but you can ctrl-C when it gets into the real compression work>
```

Now, finally, you can run with both HF_HOME and HF_HUB_OFFLINE=1 and
prove to yourself that llm-compressor uses that freshly-populated cache
for both the model and the dataset.

```bash
❯ HF_HOME=./hf-hub HF_HUB_OFFLINE=1 python example.py
<it works!>
```

---------

Signed-off-by: Ralph Bean <rbean@redhat.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Brian Dellabetta <brian-dellabetta@users.noreply.github.com>

Author: 3 people

src/llmcompressor/args/model_arguments.py

@@ -50,10 +50,6 @@ class ModelArguments:
             "help": "Pretrained processor name or path if not the same as model_name"
         },
     )
-    cache_dir: str | None = field(
-        default=None,
-        metadata={"help": "Where to store the pretrained data from huggingface.co"},
-    )
 
     use_auth_token: bool = field(
         default=False,

src/llmcompressor/entrypoints/oneshot.py

@@ -231,7 +231,6 @@ def oneshot(
     config_name: Optional[str] = None,
     tokenizer: Optional[Union[str, PreTrainedTokenizerBase]] = None,
     processor: Optional[Union[str, ProcessorMixin]] = None,
-    cache_dir: Optional[str] = None,
     use_auth_token: bool = False,
     precision: str = "auto",
     tie_word_embeddings: bool = False,
@@ -279,8 +278,6 @@ def oneshot(
         model_name.
     :param processor: Pretrained processor name or path if not the same as
         model_name.
-    :param cache_dir: Where to store the pretrained data from
-        huggingface.co.
     :param use_auth_token: Whether to use Hugging Face auth token for private
         models.
     :param precision: Precision to cast model weights to, default to auto.

src/llmcompressor/entrypoints/utils.py

@@ -175,7 +175,7 @@ def initialize_model_from_path(
     model_path = model_args.model
     config = AutoConfig.from_pretrained(
         model_args.config_name if model_args.config_name else model_path,
-        cache_dir=model_args.cache_dir,
+        cache_dir=None,
         revision=model_args.model_revision,
         use_auth_token=True if model_args.use_auth_token else None,
         trust_remote_code=model_args.trust_remote_code_model,
@@ -211,7 +211,7 @@ def initialize_model_from_path(
             )
             teacher_kwargs = {
                 "config": teacher_config,
-                "cache_dir": model_args.cache_dir,
+                "cache_dir": None,
                 "use_auth_token": True if model_args.use_auth_token else None,
                 "torch_dtype": parse_dtype(model_args.precision),
                 "device_map": teacher_device_map,
@@ -233,7 +233,7 @@ def initialize_model_from_path(
 
     model_kwargs = {
         "config": config,
-        "cache_dir": model_args.cache_dir,
+        "cache_dir": None,
         "revision": model_args.model_revision,
         "use_auth_token": True if model_args.use_auth_token else None,
         "torch_dtype": parse_dtype(model_args.precision),
@@ -266,7 +266,7 @@ def initialize_processor_from_path(
     try:
         processor = AutoProcessor.from_pretrained(
             processor_src,
-            cache_dir=model_args.cache_dir,
+            cache_dir=None,
             use_fast=True,
             revision=model_args.model_revision,
             use_auth_token=True if model_args.use_auth_token else None,
@@ -285,7 +285,7 @@ def initialize_processor_from_path(
         logger.debug("Could not load fast processor, loading slow processor instead")
         processor = AutoProcessor.from_pretrained(
             processor_src,
-            cache_dir=model_args.cache_dir,
+            cache_dir=None,
             use_fast=False,
             revision=model_args.model_revision,
             use_auth_token=True if model_args.use_auth_token else None,

src/llmcompressor/pytorch/model_load/helpers.py

@@ -149,16 +149,17 @@ def copy_python_files_from_model_cache(model, save_path: str):
         import shutil
 
         from huggingface_hub import hf_hub_download
-        from transformers import TRANSFORMERS_CACHE
         from transformers.utils import http_user_agent
 
         cache_path = config._name_or_path
         if not os.path.exists(cache_path):
             user_agent = http_user_agent()
+            # Use cache_dir=None to respect HF_HOME, HF_HUB_CACHE, and other
+            # environment variables for cache location
             config_file_path = hf_hub_download(
                 repo_id=cache_path,
                 filename="config.json",
-                cache_dir=TRANSFORMERS_CACHE,
+                cache_dir=None,
                 force_download=False,
                 user_agent=user_agent,
             )

src/llmcompressor/transformers/finetune/data/base.py

@@ -195,7 +195,7 @@ def load_dataset(self):
         logger.debug(f"Loading dataset {self.dataset_args.dataset}")
         return get_raw_dataset(
             self.dataset_args,
-            None,
+            cache_dir=None,
             split=self.split,
             streaming=self.dataset_args.streaming,
             **self.dataset_args.raw_kwargs,

tests/llmcompressor/transformers/finetune/data/conftest.py

@@ -18,7 +18,6 @@ def tiny_llama_model_args(tiny_llama_path):
 def tiny_llama_tokenizer(tiny_llama_model_args):
     tokenizer = AutoTokenizer.from_pretrained(
         tiny_llama_model_args.model,
-        cache_dir=tiny_llama_model_args.cache_dir,
         use_fast=True,
         revision=tiny_llama_model_args.model_revision,
         use_auth_token=True if tiny_llama_model_args.use_auth_token else None,