Deprecate quanto and switch to optimum-quanto (huggingface#35001)

* deprecate quanto

* fix style

src/transformers/cache_utils.py

@@ -12,7 +12,6 @@
 from .utils import (
     is_hqq_available,
     is_optimum_quanto_available,
-    is_quanto_available,
     is_torchdynamo_compiling,
     logging,
 )
@@ -790,17 +789,6 @@ def __init__(self, cache_config: CacheConfig) -> None:
                     f"You need optimum-quanto package version to be greater or equal than 0.2.5 to use `QuantoQuantizedCache`. Detected version {optimum_quanto_version}."
                 )
             from optimum.quanto import MaxOptimizer, qint2, qint4
-        elif is_quanto_available():
-            logger.warning_once(
-                "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-            )
-            quanto_version = version.parse(importlib.metadata.version("quanto"))
-            if quanto_version < version.parse("0.2.0"):
-                raise ImportError(
-                    f"You need quanto package version to be greater or equal than 0.2.0 to use `QuantoQuantizedCache`. Detected version {quanto_version}. "
-                    f"Since quanto will be deprecated, please install optimum-quanto instead with `pip install -U optimum-quanto`"
-                )
-            from quanto import MaxOptimizer, qint2, qint4
 
         if self.nbits not in [2, 4]:
             raise ValueError(f"`nbits` for `quanto` backend has to be one of [`2`, `4`] but got {self.nbits}")
@@ -824,16 +812,6 @@ def _quantize(self, tensor, axis):
             scale, zeropoint = self.optimizer(tensor, self.qtype, axis, self.q_group_size)
             qtensor = quantize_weight(tensor, self.qtype, axis, scale, zeropoint, self.q_group_size)
             return qtensor
-        elif is_quanto_available():
-            logger.warning_once(
-                "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-            )
-            from quanto import AffineQuantizer
-
-            scale, zeropoint = self.optimizer(tensor, self.qtype.bits, axis, self.q_group_size)
-            qtensor = AffineQuantizer.apply(tensor, self.qtype, axis, self.q_group_size, scale, zeropoint)
-
-        return qtensor
 
     def _dequantize(self, qtensor):
         return qtensor.dequantize()

src/transformers/generation/utils.py

@@ -45,7 +45,6 @@
     is_accelerate_available,
     is_hqq_available,
     is_optimum_quanto_available,
-    is_quanto_available,
     is_torchdynamo_compiling,
     logging,
 )
@@ -1787,7 +1786,7 @@ def _prepare_cache_for_generation(
                 )
                 cache_class = QUANT_BACKEND_CLASSES_MAPPING[cache_config.backend]
 
-                if cache_config.backend == "quanto" and not (is_optimum_quanto_available() or is_quanto_available()):
+                if cache_config.backend == "quanto" and not is_optimum_quanto_available():
                     raise ImportError(
                         "You need to install optimum-quanto in order to use KV cache quantization with optimum-quanto backend. "
                         "Please install it via  with `pip install optimum-quanto`"

src/transformers/integrations/quanto.py

@@ -12,7 +12,7 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from ..utils import is_optimum_quanto_available, is_quanto_available, is_torch_available, logging
+from ..utils import is_optimum_quanto_available, is_torch_available, logging
 
 
 if is_torch_available():
@@ -50,11 +50,6 @@ def replace_with_quanto_layers(
 
     if is_optimum_quanto_available():
         from optimum.quanto import QLayerNorm, QLinear, qfloat8, qint2, qint4, qint8
-    elif is_quanto_available():
-        logger.warning_once(
-            "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-        )
-        from quanto import QLayerNorm, QLinear, qfloat8, qint2, qint4, qint8
 
     w_mapping = {"float8": qfloat8, "int8": qint8, "int4": qint4, "int2": qint2}
     a_mapping = {None: None, "float8": qfloat8, "int8": qint8}

src/transformers/quantizers/quantizer_quanto.py

@@ -26,7 +26,6 @@
 from ..utils import (
     is_accelerate_available,
     is_optimum_quanto_available,
-    is_quanto_available,
     is_torch_available,
     logging,
 )
@@ -63,7 +62,7 @@ def post_init(self):
             )
 
     def validate_environment(self, *args, **kwargs):
-        if not (is_optimum_quanto_available() or is_quanto_available()):
+        if not is_optimum_quanto_available():
             raise ImportError(
                 "Loading an optimum-quanto quantized model requires optimum-quanto library (`pip install optimum-quanto`)"
             )
@@ -91,11 +90,6 @@ def update_torch_dtype(self, torch_dtype: "torch.dtype") -> "torch.dtype":
     def update_missing_keys(self, model, missing_keys: List[str], prefix: str) -> List[str]:
         if is_optimum_quanto_available():
             from optimum.quanto import QModuleMixin
-        elif is_quanto_available():
-            logger.warning_once(
-                "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-            )
-            from quanto import QModuleMixin
 
         not_missing_keys = []
         for name, module in model.named_modules():
@@ -122,11 +116,6 @@ def check_quantized_param(
         """
         if is_optimum_quanto_available():
             from optimum.quanto import QModuleMixin
-        elif is_quanto_available():
-            logger.warning_once(
-                "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-            )
-            from quanto import QModuleMixin
 
         device_map = kwargs.get("device_map", None)
         param_device = kwargs.get("param_device", None)

src/transformers/utils/__init__.py

@@ -175,7 +175,6 @@
     is_pytesseract_available,
     is_pytest_available,
     is_pytorch_quantization_available,
-    is_quanto_available,
     is_rjieba_available,
     is_sacremoses_available,
     is_safetensors_available,

src/transformers/utils/import_utils.py

@@ -997,13 +997,6 @@ def is_auto_awq_available():
     return _auto_awq_available
 
 
-def is_quanto_available():
-    logger.warning_once(
-        "Importing from quanto will be deprecated in v4.47. Please install optimum-quanto instead `pip install optimum-quanto`"
-    )
-    return _quanto_available
-
-
 def is_optimum_quanto_available():
     # `importlib.metadata.version` doesn't work with `optimum.quanto`, need to put `optimum_quanto`
     return _is_optimum_quanto_available