FujitsuResearch · aki916f · Aug 24, 2026 · Aug 6, 2026 · Aug 18, 2026 · Aug 18, 2026
diff --git a/.gitignore b/.gitignore
@@ -12,6 +12,9 @@ models
 .cursor/rules/gitlab-integration.mdc
 .cursor/rules/slurm-submit.mdc
 .cursor/rules/run-tests-examples.mdc
+.github/prompts
+.github/copilot-instructions.md
+.github/instructions
 .hydra/
 *.out
 *.err
@@ -22,6 +25,7 @@ site/
 debug_code/
 .pytest_cache/
 *.pt
+**/.env
 # Large test data (download separately if needed for regression tests)
 tests/onecomp/quantizer/jointq/data/model_layers_0_self_attn_k_proj.pth
 .uv-sync.lock

diff --git a/.gitlab-ci.yml b/.gitlab-ci.yml
@@ -23,6 +23,18 @@ variables:
   GIT_DEPTH: "1"
   PYTEST_MARKERS: "not slow"
 
+lint:conflict-markers:
+  stage: lint
+  tags:
+    - lint
+  variables:
+    # The MR diff base may be outside the default shallow clone.
+    GIT_DEPTH: "0"
+  rules:
+    - if: $CI_PIPELINE_SOURCE == "merge_request_event"
+  script:
+    - git diff --check "${CI_MERGE_REQUEST_DIFF_BASE_SHA}" "${CI_COMMIT_SHA}"
+
 lint:format:
   extends: .skip_docs_only
   stage: lint

diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -1,5 +1,19 @@
 # Change log
 
+## [v1.3.2] 2026-08-24
+
+### Bug Fix
+
+- Fix WikiText dataset loading in clean environments by using the canonical
+  `Salesforce/wikitext` dataset ID for perplexity evaluation and LoRA SFT examples.
+- Fix warning regarding the data type arguments of Transformers.
+- Fixed GemLite AssertionError during CPU inference execution for OneComp with Llama.cpp.
+- Fixed `QuantizedModelLoader.load_quantized_model()` calling `unfuse_moe_experts()` twice on non-fused MoE checkpoints. An earlier unconditional unfuse (run right after building the empty model) was left in place when the checkpoint-aware unfuse (added for gpt-oss fused-MoE support) was introduced. The redundant unconditional call was removed, and the checkpoint-aware unfuse now runs before `_remap_state_dict_keys()` so key remapping still aligns against the unfused per-expert module paths. This also prevents wrongly unfusing gpt-oss fused-MoE checkpoints, whose fused 3D expert tensors must be loaded as-is (`quantized_model_loader.py`).
+
+### Documentation
+
+- Add troubleshooting information for running OneComp with Llama.cpp on macOS.
+
 ## [v1.3.1] 2026-08-06
 
 ### Bug Fix
@@ -94,11 +108,11 @@
 
 ### Security
 
-- **Unsafe deserialization hardening (CWE-502)**: `QuantizedModelLoader.load_quantized_model_pt()` (alias `onecomp.load_quantized_model_pt()`) previously called `torch.load(model.pt, weights_only=False)` unconditionally, allowing arbitrary code execution when loading a malicious `.pt` checkpoint. It now refuses to load unless the caller explicitly opts in via `allow_unsafe_deserialization=True`, and emits a strong warning when it does load. For untrusted models, use the safetensors-based `load_quantized_model()`, which does not execute code.
+- **Unsafe deserialization hardening (CVE-2026-73325, CWE-502)**: `QuantizedModelLoader.load_quantized_model_pt()` (alias `onecomp.load_quantized_model_pt()`) previously called `torch.load(model.pt, weights_only=False)` unconditionally, allowing arbitrary code execution when loading a malicious `.pt` checkpoint. It now refuses to load unless the caller explicitly opts in via `allow_unsafe_deserialization=True`, and emits a strong warning when it does load. For untrusted models, use the safetensors-based `load_quantized_model()`, which does not execute code.
   - **Breaking change**: existing callers of `load_quantized_model_pt()` must pass `allow_unsafe_deserialization=True` for trusted `.pt` files.
 - **`Quantizer.load_results()` / `ResultLoader`**: same hardening applied. Loading with `weights_only=False` now requires `allow_unsafe_deserialization=True` (added as a `ResultLoader` field), and logs a warning. The safe `weights_only=True` path is unchanged.
 - Updated docstrings, docs, and the LoRA SFT example to document the risk and the required opt-in.
-- **Credit**: this unsafe deserialization issue (CWE-502) was responsibly disclosed by **Nir Yehoshua, Cipher Security Labs**. Thank you for the report.
+- **Credit**: this unsafe deserialization issue (CVE-2026-73325, CWE-502) was responsibly disclosed by **Nir Yehoshua, Cipher Security Labs**. Thank you for the report.
 
 ## [v1.2.0] 2026-06-08
 

diff --git a/SECURITY.md b/SECURITY.md
@@ -6,7 +6,7 @@ Security updates are provided for the latest release line of Fujitsu One Compres
 
 | Version | Supported          |
 | ------- | ------------------ |
-| 1.2.2   | :white_check_mark: |
+| 1.2.2 >= | :white_check_mark: |
 | < 1.2.2 | :x:                |
 
 We recommend always upgrading to the latest release before reporting an issue.
@@ -38,8 +38,20 @@ Please make a good-faith effort to avoid privacy violations, data destruction, a
 
 Thank you for helping keep OneComp and its users safe.
 
+## Resolved Security Issues
+
+### CVE-2026-73325 — Unsafe Deserialization
+
+An unsafe deserialization vulnerability (CWE-502) was identified in
+`QuantizedModelLoader.load_quantized_model_pt()`.
+
+- **Affected versions:** v1.2.0 and earlier
+- **Fixed version:** v1.2.1
+- **CVE:** CVE-2026-73325
+- **Resolution:** The issue was addressed in v1.2.1. Users should upgrade to the latest supported release.
+
 ## Security Acknowledgments
 
 We thank the following researchers for responsibly disclosing security issues in OneComp:
 
-- **Nir Yehoshua, Cipher Security Labs** — unsafe deserialization in `QuantizedModelLoader.load_quantized_model_pt()` (CWE-502), fixed in v1.2.1.
+- **Nir Yehoshua, Cipher Security Labs** — unsafe deserialization in `QuantizedModelLoader.load_quantized_model_pt()` (CVE-2026-73325, CWE-502), affecting OneComp versions through v1.2.0 and fixed in v1.2.1.
diff --git a/docs/user-guide/cpu-inference.md b/docs/user-guide/cpu-inference.md
@@ -19,6 +19,59 @@ for inference. The direct export path additionally uses llama.cpp's pure-Python
 `convert_hf_to_gguf.py` to build the model metadata/tokenizer; it is fetched
 automatically (a shallow `git clone`) or taken from `$LLAMA_CPP_DIR` if set.
 
+
+### macOS
+
+#### Symptom
+
+If you have installed gcc or clang on macOS using a tool like Homebrew, the OpenMP dynamic library with them can conflict with the OpenMP dynamic library used by OneComp's PyTorch backend.
+
+- OneComp's PyTorch backend is often configured to use an OpenMP library located within the .venv directory.
+- Depending on your environment, Llama.cpp is configured to use an OpenMP library associated with gcc or clang.
+
+If the following problems is occurring , this conflict may be occurring.
+
+- Your Python interpreter shows the following warning message.
+
+```python
+.../multiprocessing/resource_tracker.py:279: UserWarning: resource_tracker: There appear to be 1 leaked semaphore objects to clean up at shutdown
+  warnings.warn('resource_tracker: There appear to be %d '
+```
+
+- The following message is shown, when you import PyTorch and Llama.cpp in your script.
+
+```python
+OMP: Error #15: Initializing libomp.dylib, but found libomp.dylib already initialized.
+OMP: Hint This means that multiple copies of the OpenMP runtime have been linked into the program.
+```
+
+#### Solution
+
+##### 1. Search for OpenMP associated with PyTorch.
+
+Search for OpenMP library associated with PyTorch backend (using find command).
+
+```bash
+$ find $PWD/.venv -type f \( -name 'libomp.dylib' -o -name 'libgomp*.dylib' \) -print | grep torch
+/to/path/.venv/lib/python3.<PYVERSION>/site-packages/torch/lib/libomp.dylib
+```
+
+##### 2. Create an .env
+
+Create a .env file and add the OpenMP path as an environment variable, as shown below.
+
+```bash
+DYLD_LIBRARY_PATH="/to/path/.venv/lib/python3.<PYVERSION>/site-packages/torch/lib"
+```
+
+##### 3. Run uv
+
+Pass the following `--env-file` options when running uv run.
+
+```bash
+uv run --env-file /path/to/.env ... python your_script.py
+```
+
 ## One entry point: `export_to_gguf`
 
 You do not need to know which path a checkpoint requires. `export_to_gguf`

diff --git a/docs/user-guide/examples.md b/docs/user-guide/examples.md
@@ -578,7 +578,7 @@ model_config = ModelConfig(
 gptq = GPTQ(wbits=4, groupsize=128)
 
 post_process = PostProcessLoraSFT(
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config_name="wikitext-2-raw-v1",
     train_split="train",
     text_column="text",

diff --git a/docs/user-guide/post-process.md b/docs/user-guide/post-process.md
@@ -354,7 +354,7 @@ model_config = ModelConfig(
 gptq = GPTQ(wbits=4, groupsize=128)
 
 post_process = PostProcessLoraSFT(
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config_name="wikitext-2-raw-v1",
     train_split="train",
     text_column="text",
@@ -512,7 +512,7 @@ model, tokenizer = load_quantized_model_pt(
 
 ```python
 PostProcessLoraSFT(
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config_name="wikitext-2-raw-v1",
     train_split="train",
     text_column="text",
@@ -560,7 +560,7 @@ Teacher distillation aligns the quantized model's output distribution with a ful
 
 ```python
 post_process = PostProcessLoraSFT(
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config_name="wikitext-2-raw-v1",
     train_split="train",
     text_column="text",
@@ -591,7 +591,7 @@ Intermediate block alignment adds a loss term that aligns hidden states at selec
 
 ```python
 post_process = PostProcessLoraSFT(
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config_name="wikitext-2-raw-v1",
     train_split="train",
     text_column="text",

diff --git a/example/post_process/example_lora_sft.py b/example/post_process/example_lora_sft.py
@@ -66,7 +66,7 @@ def generate_text(model, tokenizer, prompt, device, max_new_tokens=64):
 gptq = GPTQ(wbits=4, groupsize=128)
 
 post_process = PostProcessLoraSFT(
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config_name="wikitext-2-raw-v1",
     train_split="train",
     text_column="text",

diff --git a/notebook/01_tutorial.ipynb b/notebook/01_tutorial.ipynb
@@ -76,7 +76,7 @@
         "DEVICE = \"cuda:0\" if torch.cuda.is_available() else \"cpu\"\n",
         "\n",
         "tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)\n",
-        "model = AutoModelForCausalLM.from_pretrained(MODEL_ID, torch_dtype=torch.float16).to(DEVICE)\n",
+        "model = AutoModelForCausalLM.from_pretrained(MODEL_ID, dtype=torch.float16).to(DEVICE)\n",
         "model.eval()\n",
         "\n",
         "n_params = sum(p.numel() for p in model.parameters())\n",

diff --git a/onecomp/__version__.py b/onecomp/__version__.py
@@ -6,4 +6,4 @@
 
 """
 
-__version__ = "1.3.1"
+__version__ = "1.3.2"
diff --git a/onecomp/cpu/export/dequantize.py b/onecomp/cpu/export/dequantize.py
@@ -98,7 +98,7 @@ def dequantize_to_hf(
         pass
 
     logger.info("Building empty dense model from config (%s)", config.model_type)
-    model = AutoModelForCausalLM.from_config(config, torch_dtype=torch_dtype)
+    model = AutoModelForCausalLM.from_config(config, dtype=torch_dtype)
 
     state: Dict[str, torch.Tensor] = {}
     for shard in sorted(glob(os.path.join(save_directory, "*.safetensors"))):

diff --git a/onecomp/pre_process/prepare_rotated_model.py b/onecomp/pre_process/prepare_rotated_model.py
@@ -319,7 +319,7 @@ def prepare_rotated_model(
         t0 = time.time()
         model = AutoModelForCausalLM.from_pretrained(
             model_path,
-            torch_dtype="auto",
+            dtype="auto",
             device_map="cpu",
             low_cpu_mem_usage=True,
         )

diff --git a/onecomp/pre_process/train_rotation.py b/onecomp/pre_process/train_rotation.py
@@ -321,15 +321,15 @@ def _convert_model_structure(config, model_type, model_path, use_sdpa=False):
     if is_llama:
         quant_model = QLlamaFC.from_pretrained(
             model_path,
-            torch_dtype="auto",
+            dtype="auto",
             device_map="cpu",
             low_cpu_mem_usage=True,
             config=config,
         )
     elif is_qwen3:
         quant_model = QQwen3FC.from_pretrained(
             model_path,
-            torch_dtype="auto",
+            dtype="auto",
             device_map="cpu",
             low_cpu_mem_usage=True,
             config=config,

diff --git a/onecomp/quantized_model_loader.py b/onecomp/quantized_model_loader.py
@@ -101,22 +101,14 @@ def load_quantized_model(
             torch_dtype = torch.bfloat16
         model = cls._build_empty_model_from_config(config_dict, torch_dtype)
 
-        # Mirror the unfuse step performed before quantization/save (see
-        # Runner.save_quantized_model) so per-expert module names such as
-        # "model.layers.0.mlp.experts.0.down_proj" resolve against the
-        # freshly-built model instead of its fused 3D expert parameters.
-        if unfuse_moe_experts(model, logger):
-            logger.info("Unfused MoE expert tensors for quantized model load")
-
         # Load state_dict from safetensors
         state_dict = cls._load_state_dict_from_dir(save_directory)
 
-        # Align checkpoint key prefixes with the empty model built from config.
-        # Gemma3 VLMs are a common case: weights saved from from_pretrained
-        # use model.language_model.model.layers. (language_model is a
-        # ForCausalLM wrapper) while from_config exposes
-        # model.language_model.layers.* directly.
-        state_dict = cls._remap_state_dict_keys(state_dict, model)
+        # Decide, from the checkpoint alone, whether the empty model must be
+        # unfused into per-expert nn.Linear modules.  This must happen before
+        # _remap_state_dict_keys so remapping aligns checkpoint keys against the
+        # unfused module paths; fused-MoE checkpoints (e.g. gpt-oss) keep the
+        # fused 3D parameters and skip unfuse.
         from .utils.unfuse_moe import (
             _checkpoint_uses_fused_moe,
             _expand_deduped_moe_keys,
@@ -131,6 +123,13 @@ def load_quantized_model(
         elif unfuse_moe_experts(model, logger):
             logger.info("Unfused MoE expert tensors for quantized model load")
 
+        # Align checkpoint key prefixes with the empty model built from config.
+        # Gemma3 VLMs are a common case: weights saved from from_pretrained
+        # use model.language_model.model.layers. (language_model is a
+        # ForCausalLM wrapper) while from_config exposes
+        # model.language_model.layers.* directly.
+        state_dict = cls._remap_state_dict_keys(state_dict, model)
+
         # Replace quantized layers with empty modules and align quantized
         # tensor keys with the actual module names in the model built from
         # config.  This is required when the saved checkpoint and the
@@ -568,11 +567,11 @@ def _build_empty_model_from_config(
         config_cls = CONFIG_MAPPING[model_type]
         model_config = config_cls.from_dict(clean_config)
         try:
-            return AutoModelForCausalLM.from_config(model_config, torch_dtype=dtype)
+            return AutoModelForCausalLM.from_config(model_config, dtype=dtype)
         except (ValueError, KeyError):
             from transformers import AutoModelForImageTextToText
 
-            return AutoModelForImageTextToText.from_config(model_config, torch_dtype=dtype)
+            return AutoModelForImageTextToText.from_config(model_config, dtype=dtype)
 
     @staticmethod
     def _set_module_by_name(

diff --git a/onecomp/quantizer/gemlite.py b/onecomp/quantizer/gemlite.py
@@ -16,13 +16,15 @@
 import torch.nn.functional as F
 
 # Optional GemLite/HQQ imports
-try:
-    from gemlite.core import DType, GemLiteLinearTriton
-    from hqq.core.quantize import BaseQuantizeConfig, HQQLinear
+HAS_GEMLITE = False
+if torch.cuda.is_available():
+    try:
+        from gemlite.core import DType, GemLiteLinearTriton
+        from hqq.core.quantize import BaseQuantizeConfig, HQQLinear
 
-    HAS_GEMLITE = True
-except (ImportError, AttributeError):
-    HAS_GEMLITE = False
+        HAS_GEMLITE = True
+    except (ImportError, AttributeError, AssertionError):
+        pass
 
 
 # Constants

diff --git a/onecomp/runner.py b/onecomp/runner.py
@@ -1260,7 +1260,7 @@ def calculate_perplexity(
         original_model=False,
         dequantized_model=False,
         quantized_model=True,
-        dataset_name="wikitext",
+        dataset_name="Salesforce/wikitext",
         dataset_config="wikitext-2-raw-v1",
         split="test",
         max_samples=None,
@@ -1342,7 +1342,7 @@ def benchmark_perplexity(
         original_model=True,
         dequantized_model=False,
         quantized_model=True,
-        dataset_name="wikitext",
+        dataset_name="Salesforce/wikitext",
         dataset_config="wikitext-2-raw-v1",
         split="test",
         max_samples=None,

diff --git a/onecomp/utils/perplexity.py b/onecomp/utils/perplexity.py
@@ -41,7 +41,7 @@ def calculate_perplexity(
     model=None,
     tokenizer=None,
     model_config=None,
-    dataset_name="wikitext",
+    dataset_name="Salesforce/wikitext",
     dataset_config="wikitext-2-raw-v1",
     split="test",
     max_samples=None,
@@ -53,7 +53,7 @@ def calculate_perplexity(
     Based on https://huggingface.co/docs/transformers/perplexity
 
     Args:
-        dataset_name (str): Dataset name (e.g. "wikitext", "allenai/c4").
+        dataset_name (str): Dataset name (e.g. "Salesforce/wikitext", "allenai/c4").
         dataset_config (str): Dataset configuration.
             - For WikiText: "wikitext-2-raw-v1"
             - For C4: "en/c4-train.00001-of-01024.json.gz" (treated as data_files)

diff --git a/onecomp/utils/vram_estimator.py b/onecomp/utils/vram_estimator.py
@@ -312,7 +312,7 @@ def estimate_wbits_from_vram(
 
     config = AutoConfig.from_pretrained(model_id)
     with torch.device("meta"):
-        model = AutoModelForCausalLM.from_config(config, torch_dtype=torch.float16)
+        model = AutoModelForCausalLM.from_config(config, dtype=torch.float16)
 
     return estimate_target_bitwidth(
         model,
Original file line number	Diff line number	Diff line change
Expand Up		@@ -6,4 +6,4 @@

		"""

		__version__ = "1.3.1"
		__version__ = "1.3.2"