inplace hadamard

auto_round/__main__.py

@@ -727,6 +727,7 @@ def tune(args):
         model_dtype=args.model_dtype,
         momentum=args.momentum,
         trust_remote_code=not args.disable_trust_remote_code,
+        hadamard_config="default",
     )
 
     model_name = args.model.rstrip("/")

auto_round/compressors/base.py

@@ -557,18 +557,32 @@ def __init__(
             except (ImportError, ModuleNotFoundError):
                 logger.error("algorithm extension import error, fallback to default mode")
 
-        # apply hadamard transform
+        hadamard_config = 1
         if hadamard_config:
-            from auto_round.experimental.transform.apply import apply_hadamard_transform
-            from auto_round.experimental.utils import check_supported_schemes, normalize_hadamard_config
-
-            check_supported_schemes(self.scheme)
-
-            self.model = apply_hadamard_transform(
-                self.model, hadamard_config, need_calibration=True if self.iters > 0 else False
-            )
-
-            self.hadamard_config = normalize_hadamard_config(hadamard_config)
+            self.enable_hadamard = True
+            from auto_round.experimental.hadamard_inplace import apply_hadamard_rotation
+
+            # from auto_round.experimental.utils import normalize_hadamard_config
+
+            #
+            # _hcfg = normalize_hadamard_config(hadamard_config)
+            # if not isinstance(_hcfg, HadamardConfig):
+            #     _hcfg = HadamardConfig(**_hcfg)
+            # group_size = _hcfg.block_size  # e.g. 32 for MXFP4
+
+            self.model.to("cuda")
+            apply_hadamard_rotation(self.model)
+
+            # from auto_round.experimental.transform.apply import apply_hadamard_transform
+            # from auto_round.experimental.utils import check_supported_schemes, normalize_hadamard_config
+            #
+            # check_supported_schemes(self.scheme)
+            #
+            # self.model = apply_hadamard_transform(
+            #     self.model, hadamard_config, need_calibration=True if self.iters > 0 else False
+            # )
+            #
+            # self.hadamard_config = normalize_hadamard_config(hadamard_config)
 
     def _gen_auto_scheme(self) -> dict[str, dict]:
         if self.mllm:

auto_round/eval/evaluation.py

@@ -428,7 +428,7 @@ def run_model_evaluation(model, tokenizer, autoround, folders, formats, device_s
     eval_gguf_model = any(file.endswith("gguf") for file in os.listdir(eval_folder))
 
     # Determine if model instance evaluation is needed
-    need_model_instance = (autoround.act_bits <= 8 and formats[-1] == "fake") or eval_gguf_model
+    need_model_instance = (autoround.act_bits <= 8) or formats[-1] == "fake" or eval_gguf_model
 
     if need_model_instance:
         # Load or prepare model instance

auto_round/experimental/hadamard_inplace/__init__.py

@@ -0,0 +1,12 @@
+# # Copyright (C) 2026 Intel Corporation
+# # SPDX-License-Identifier: Apache-2.0
+
+from auto_round.experimental.hadamard_inplace.apply_hadamard_rotation import apply_hadamard_rotation
+
+# from auto_round.experimental.hadamard_inplace.model_config import (
+#     MAPPING_REGISTRY,
+#     RotationMapping,
+#     get_mapping,
+#     infer_mapping_from_model,
+#     register_mapping,
+# )