ipex-llm 2.2.0b20250105__py3-none-manylinux2010_x86_64.whl → 2.2.0b20250106__py3-none-manylinux2010_x86_64.whl

ipex_llm/transformers/npu_model.py

@@ -27,7 +27,7 @@ from transformers.configuration_utils import PretrainedConfig
 
 from ipex_llm.utils.common.log4Error import invalidInputError
 from ipex_llm.transformers.utils import logger, load_imatrix_data
-from ipex_llm.transformers.npu_models.convert import optimize_llm, optimize_llm_post
+from ipex_llm.transformers.npu_models.convert import optimize_llm
 
 
 def patch_flash_attn_import(filename: str) -> List[str]:
@@ -207,8 +207,6 @@ class _BaseAutoModelClass:
             model = model.eval()
             logger.info(f"Finish to convert model")
         else:
-            from intel_npu_acceleration_library.compiler import create_npu_kernels
-
             if optimize_model:
                 invalidInputError(
                     max_prompt_len < max_context_len,
@@ -232,11 +230,14 @@ class _BaseAutoModelClass:
                     "convert_model": convert_model,
                     "save_directory": save_directory,
                     "fuse_layers": fuse_layers,
-                    "imatrix_data": imatrix_data
+                    "imatrix_data": imatrix_data,
+                    "skip_npu_logic": mock_device == "dummy",
                 }
+                # Dummy will skip npu related logic and save the quantized model
+                if mock_device == "dummy":
+                    model.save_low_bit = types.MethodType(save_low_bit, model)
                 model = cls.optimize_npu_model(*args, **optimize_kwargs)
             else:
-                from ipex_llm.transformers.npu_models.convert import optimize_llm
                 optimize_llm(model)
                 with torch.no_grad():
                     cls.load_convert(qtype, model, "cpu", modules_to_not_convert,
@@ -258,7 +259,6 @@ class _BaseAutoModelClass:
     def optimize_npu_model(cls, *args, **kwargs):
 
         from ipex_llm.transformers.npu_models.convert_mp import optimize_llm_pre, optimize_llm
-        from intel_npu_acceleration_library.compiler import create_npu_kernels
 
         model = kwargs.pop("model")
         qtype = kwargs.pop("qtype", "sym_int4_rtn")
@@ -275,6 +275,7 @@ class _BaseAutoModelClass:
         save_directory = kwargs.pop('save_directory', None)
         fuse_layers = kwargs.pop('fuse_layers', None)
         imatrix_data = kwargs.pop('imatrix_data', None)
+        skip_npu_logic = kwargs.pop("skip_npu_logic", False)
         invalidInputError(save_directory is not None,
                           "Please provide the path to save converted model "
                           "through `save_directory`.")
@@ -294,51 +295,58 @@ class _BaseAutoModelClass:
             cls.load_convert(qtype, model, "cpu", modules_to_not_convert,
                              quantization_group_size, imatrix_data,
                              *args, **kwargs)
-            create_npu_kernels(llm)
+            if not skip_npu_logic:
+                from intel_npu_acceleration_library.compiler import create_npu_kernels
+                create_npu_kernels(llm)
         model = model.eval()
         logger.info(f"Finish to convert model")
         model.config.update({"bigdl_transformers_low_bit": qtype})
-        model.share_memory()
 
-        if not pipeline:
-            if model.config.model_type in ["qwen2", "llama", "minicpm"]:
-                from ipex_llm.transformers.npu_models.convert import optimize_llm_single_process
-                optimize_llm_single_process(
-                    llm,
-                    kv_len=max_context_len,
-                    max_prompt_len=max_prompt_len,
-                    transpose_value_cache=transpose_value_cache,
-                    group_size=quantization_group_size,
-                    qtype=qtype,
-                    save_directory=save_directory,
-                    fuse_layers=fuse_layers,
-                    has_llm=hasattr(model, "llm")
-                )
-            else:
-                optimize_llm(
-                    llm,
-                    max_context_len=max_context_len,
-                    max_prompt_len=max_prompt_len,
-                    inter_pp=inter_pp,
-                    intra_pp=intra_pp,
-                    transpose_value_cache=transpose_value_cache,
-                    group_size=quantization_group_size
-                )
+        if skip_npu_logic:
+            model.save_low_bit(model_dir=save_directory)
         else:
-            from ipex_llm.transformers.npu_pipeline_model.convert_pipeline \
-                import convert_llm
-            convert_llm(llm,
+            model.share_memory()
+
+            if not pipeline:
+                if model.config.model_type in ["qwen2", "llama", "minicpm"]:
+                    from ipex_llm.transformers.npu_models.convert import optimize_llm_single_process
+                    optimize_llm_single_process(
+                        llm,
                         kv_len=max_context_len,
                         max_prompt_len=max_prompt_len,
                         transpose_value_cache=transpose_value_cache,
                         group_size=quantization_group_size,
                         qtype=qtype,
-                        convert_model=convert_model,
                         save_directory=save_directory,
-                        fuse_layers=fuse_layers)
-        model.save_low_bit = types.MethodType(save_low_bit, model)
-        model.save_low_bit(save_directory)
-        logger.info(f"Converted model has already saved to {save_directory}.")
+                        fuse_layers=fuse_layers,
+                        has_llm=hasattr(model, "llm")
+                    )
+                else:
+                    optimize_llm(
+                        llm,
+                        max_context_len=max_context_len,
+                        max_prompt_len=max_prompt_len,
+                        inter_pp=inter_pp,
+                        intra_pp=intra_pp,
+                        transpose_value_cache=transpose_value_cache,
+                        group_size=quantization_group_size
+                    )
+            else:
+                from ipex_llm.transformers.npu_pipeline_model.convert_pipeline \
+                    import convert_llm
+                convert_llm(llm,
+                            kv_len=max_context_len,
+                            max_prompt_len=max_prompt_len,
+                            transpose_value_cache=transpose_value_cache,
+                            group_size=quantization_group_size,
+                            qtype=qtype,
+                            convert_model=convert_model,
+                            save_directory=save_directory,
+                            fuse_layers=fuse_layers)
+            model.save_low_bit = types.MethodType(save_low_bit, model)
+            model.save_low_bit(save_directory)
+            logger.info(f"Converted model has already saved to {save_directory}.")
+
         return model
 
     @classmethod
@@ -379,6 +387,7 @@ class _BaseAutoModelClass:
         intra_pp = kwargs.pop("intra_pp", None)
         transpose_value_cache = kwargs.pop("transpose_value_cache", True)
         modules_to_not_convert = kwargs.pop("modules_to_not_convert", [])
+        save_directory = kwargs.pop('save_directory', None)
 
         from transformers.models.auto.configuration_auto import AutoConfig
         from transformers.modeling_utils import no_init_weights, get_state_dict_dtype
@@ -650,16 +659,37 @@ class _BaseAutoModelClass:
             param.requires_grad_(False)
 
         if optimize_model and not pipeline:
-            from ipex_llm.transformers.npu_models.convert_mp import optimize_llm
-            optimize_llm(
-                llm,
-                max_context_len=max_context_len,
-                max_prompt_len=max_prompt_len,
-                inter_pp=inter_pp,
-                intra_pp=intra_pp,
-                transpose_value_cache=transpose_value_cache,
-                group_size=quantization_group_size
-            )
+            if model.config.model_type in ["qwen2", "llama", "minicpm"]:
+                from ipex_llm.transformers.npu_models.convert import optimize_llm_single_process
+                if save_directory is None:
+                    invalidInputError(False,
+                                      "Please specify the save_directory, the path of folder " +
+                                      "to save the compiled NPU model. If path not exists, " +
+                                      "the compiled NPU model will be saved there. " +
+                                      "Else, program will exit.")
+
+                optimize_llm_single_process(
+                    llm,
+                    kv_len=max_context_len,
+                    max_prompt_len=max_prompt_len,
+                    transpose_value_cache=transpose_value_cache,
+                    group_size=quantization_group_size,
+                    qtype=qtype,
+                    save_directory=save_directory,
+                    fuse_layers=None,
+                    has_llm=hasattr(model, "llm")
+                )
+            else:
+                from ipex_llm.transformers.npu_models.convert_mp import optimize_llm
+                optimize_llm(
+                    llm,
+                    max_context_len=max_context_len,
+                    max_prompt_len=max_prompt_len,
+                    inter_pp=inter_pp,
+                    intra_pp=intra_pp,
+                    transpose_value_cache=transpose_value_cache,
+                    group_size=quantization_group_size
+                )
         elif optimize_model and pipeline:
             from ipex_llm.transformers.npu_pipeline_model.convert_pipeline \
                 import convert_llm

ipex_llm/transformers/npu_models/convert_mp.py

@@ -18,7 +18,7 @@ import torch
 import importlib
 import numpy as np
 from ipex_llm.transformers.low_bit_linear import LowBitLinear, FP4Params
-from ipex_llm.transformers.npu_models.lm_head import LMHeadLinear, SlicedLMHead
+from ipex_llm.transformers.npu_models.lm_head import SlicedLMHead
 from ipex_llm.utils.common.log4Error import invalidInputError
 
 

ipex_llm/transformers/npu_models/linear.py

@@ -21,16 +21,25 @@
 # SPDX-License-Identifier: Apache 2.0
 #
 
-from intel_npu_acceleration_library.quantization import quantize_tensor, compress_to_i4
-from intel_npu_acceleration_library.dtypes import NPUDtype
+
 import os
 import torch
 from torch.nn import Parameter
 import uuid
 import math
-from intel_npu_acceleration_library.backend import run_matmul
 from typing import Optional, Union
 from ipex_llm.utils.common import invalidInputError
+import importlib
+
+
+def is_acclib_available():
+    return importlib.util.find_spec("intel_npu_acceleration_library") is not None
+
+
+if is_acclib_available():
+    from intel_npu_acceleration_library.quantization import quantize_tensor, compress_to_i4
+    from intel_npu_acceleration_library.dtypes import NPUDtype
+    from intel_npu_acceleration_library.backend import run_matmul
 
 
 class Linear(torch.nn.Module):
@@ -63,6 +72,7 @@ class Linear(torch.nn.Module):
         if self.training:
             out = self._mm(x, self.weight, None)
         else:
+            from intel_npu_acceleration_library.backend import run_matmul
             out = run_matmul(x, self.weight, None, self.op_id)
 
         if self.bias is None:
@@ -105,6 +115,8 @@ class Linear(torch.nn.Module):
         Returns:
             Union[Linear, QuantizedLinear]: A NPU linear layer
         """
+        from intel_npu_acceleration_library.quantization import quantize_tensor, compress_to_i4
+        from intel_npu_acceleration_library.dtypes import NPUDtype
         if dtype.is_floating_point:
             if bias is None:
                 return Linear(weight.to(dtype), None)

ipex_llm/transformers/npu_models/lm_head.py

@@ -16,96 +16,6 @@
 import torch
 from torch import nn
 import numpy as np
-from filelock import FileLock
-from intel_npu_acceleration_library.backend import NNFactory
-from intel_npu_acceleration_library.backend.bindings import lib as backend_lib
-
-
-class LMHeadLinear(NNFactory):
-    """Quantized Linear class for sliced lm_head, computing a matrix matrix multiplication
-    with weights prefetching."""
-
-    def __init__(
-        self,
-        inC: int,
-        outC: int,
-        batch: int,
-        split_num: int = 2,
-        profile: bool = False,
-        device: str = "NPU",
-        dtype: np.dtype = np.int8,
-        use_split: bool = False,
-        group_size: int = 0,
-        asym: bool = False,
-    ):
-        """Initialize the LMHeadLinear class.
-
-        Args:
-            inC (int): input channels
-            outC (int): output channels
-            batch (int): batch
-            split_num (int): split in_features of lm_head to how many parts
-            profile (bool): Enable/Disable profiling. Defaults to False.
-            device (str): Target device, default to "NPU".
-            dtype (np.dtype): weights datatype. Defaults to np.int8.
-
-        """
-        super().__init__(profile, device)
-        self.inC, self.outC = inC, outC
-        self.batch = batch
-
-        self.split_num = split_num
-        if use_split:
-            input = self.parameter((1, self.batch, self.inC))
-            res = self.dq_split_linear(input, self.split_num, self.outC, self.inC, wt_dtype=dtype,
-                                       scale_factor=(group_size == 0), asym=asym)
-        else:
-            input = self.parameter((self.batch, self.inC))
-            split_size = self.inC // split_num // 2 * 2
-
-            for i in range(self.split_num):
-                start_idx = i * split_size
-                end_idx = (i + 1) * split_size if i < self.split_num - 1 else self.inC
-                input_slice = self.slice(input, begin=[0, start_idx],
-                                         end=[self.batch, end_idx])
-                linear_slice = self.linear(input_slice, outC, split_size, bias=False,
-                                           wt_dtype=dtype, asym=asym)
-                if i == 0:
-                    res = linear_slice
-                else:
-                    res += linear_slice
-
-        print("start compiling lm_head")
-        self.compile()
-        print("end compiling lm_head")
-
-    def set_weights(self, op_id, weights):
-        self.set_weights_async(op_id, weights)
-        with FileLock(f"lmhead_run.lock"):
-            backend_lib.run(self._mm)
-
-    def set_weights_async(self, op_id, weights):
-        self.setWeights(1, op_id, *weights)
-
-    def run(
-        self, X: np.ndarray
-    ) -> np.ndarray:
-        """Run the layer:  $X * (W * S)^T$ .
-
-        Args:
-            X (np.ndarray): activation
-
-        Raises:
-            RuntimeError: Input, weights or scale shape mismatch
-
-        Returns:
-            np.ndarray: result
-        """
-        self.set_input_tensor(X, 0)
-        self.elapsed = backend_lib.run(self._mm)
-        if len(self.out) == 1:
-            return self.out[0]
-        return self.out
 
 
 class SlicedLMHead(nn.Module):
@@ -160,6 +70,7 @@ class SlicedLMHead(nn.Module):
         return self.lm_heads[0].weight.dtype
 
     def get_fused_lm_head(self):
+        from ipex_llm.transformers.npu_models.lm_head_linear import LMHeadLinear
         np_dtype = np.uint8 if self.get_weight_dtype() == torch.uint8 else np.int8
         self.fused_lm_head = LMHeadLinear(self.inC, self.outC, 1, self.split_num,
                                           False, "NPU", dtype=np_dtype, use_split=self.use_split,

ipex_llm/transformers/npu_models/lm_head_linear.py

@@ -0,0 +1,106 @@
+#
+# Copyright 2016 The BigDL Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+from filelock import FileLock
+from intel_npu_acceleration_library.backend import NNFactory
+from intel_npu_acceleration_library.backend.bindings import lib as backend_lib
+
+
+class LMHeadLinear(NNFactory):
+    """Quantized Linear class for sliced lm_head, computing a matrix matrix multiplication
+    with weights prefetching."""
+
+    def __init__(
+        self,
+        inC: int,
+        outC: int,
+        batch: int,
+        split_num: int = 2,
+        profile: bool = False,
+        device: str = "NPU",
+        dtype: np.dtype = np.int8,
+        use_split: bool = False,
+        group_size: int = 0,
+        asym: bool = False,
+    ):
+        """Initialize the LMHeadLinear class.
+
+        Args:
+            inC (int): input channels
+            outC (int): output channels
+            batch (int): batch
+            split_num (int): split in_features of lm_head to how many parts
+            profile (bool): Enable/Disable profiling. Defaults to False.
+            device (str): Target device, default to "NPU".
+            dtype (np.dtype): weights datatype. Defaults to np.int8.
+
+        """
+        super().__init__(profile, device)
+        self.inC, self.outC = inC, outC
+        self.batch = batch
+
+        self.split_num = split_num
+        if use_split:
+            input = self.parameter((1, self.batch, self.inC))
+            res = self.dq_split_linear(input, self.split_num, self.outC, self.inC, wt_dtype=dtype,
+                                       scale_factor=(group_size == 0), asym=asym)
+        else:
+            input = self.parameter((self.batch, self.inC))
+            split_size = self.inC // split_num // 2 * 2
+
+            for i in range(self.split_num):
+                start_idx = i * split_size
+                end_idx = (i + 1) * split_size if i < self.split_num - 1 else self.inC
+                input_slice = self.slice(input, begin=[0, start_idx],
+                                         end=[self.batch, end_idx])
+                linear_slice = self.linear(input_slice, outC, split_size, bias=False,
+                                           wt_dtype=dtype, asym=asym)
+                if i == 0:
+                    res = linear_slice
+                else:
+                    res += linear_slice
+
+        print("start compiling lm_head")
+        self.compile()
+        print("end compiling lm_head")
+
+    def set_weights(self, op_id, weights):
+        self.set_weights_async(op_id, weights)
+        with FileLock(f"lmhead_run.lock"):
+            backend_lib.run(self._mm)
+
+    def set_weights_async(self, op_id, weights):
+        self.setWeights(1, op_id, *weights)
+
+    def run(
+        self, X: np.ndarray
+    ) -> np.ndarray:
+        """Run the layer:  $X * (W * S)^T$ .
+
+        Args:
+            X (np.ndarray): activation
+
+        Raises:
+            RuntimeError: Input, weights or scale shape mismatch
+
+        Returns:
+            np.ndarray: result
+        """
+        self.set_input_tensor(X, 0)
+        self.elapsed = backend_lib.run(self._mm)
+        if len(self.out) == 1:
+            return self.out[0]
+        return self.out

{ipex_llm-2.2.0b20250105.dist-info → ipex_llm-2.2.0b20250106.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250105
+Version: 2.2.0b20250106
 Summary: Large Language Model Develop Toolkit
 Home-page: https://github.com/intel-analytics/ipex-llm
 Author: BigDL Authors
@@ -27,10 +27,10 @@ Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine
 Requires-Dist: torch ==2.1.2+cpu ; (platform_system == "Linux") and extra == 'all'
 Requires-Dist: torch ==2.1.2 ; (platform_system == "Windows") and extra == 'all'
 Provides-Extra: cpp
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250105 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250106 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
 Provides-Extra: cpp-arl
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250105 ; extra == 'cpp-arl'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250106 ; extra == 'cpp-arl'
 Requires-Dist: setuptools ; extra == 'cpp-arl'
 Requires-Dist: onednn-devel ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
 Requires-Dist: dpcpp-cpp-rt ==2024.2.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
@@ -65,7 +65,7 @@ Requires-Dist: transformers ==4.40.0 ; extra == 'npu'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'npu'
 Requires-Dist: torch ==2.1.2+cpu ; (platform_system == "Linux") and extra == 'npu'
 Requires-Dist: torch ==2.1.2 ; (platform_system == "Windows") and extra == 'npu'
-Requires-Dist: bigdl-core-npu ==2.6.0b20250105 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250106 ; (platform_system == "Windows") and extra == 'npu'
 Provides-Extra: serving
 Requires-Dist: py-cpuinfo ; extra == 'serving'
 Requires-Dist: fschat[model_worker,webui] ==0.2.36 ; extra == 'serving'
@@ -85,9 +85,9 @@ Requires-Dist: setuptools <70.0.0 ; extra == 'xpu'
 Requires-Dist: torch ==2.1.0a0 ; extra == 'xpu'
 Requires-Dist: torchvision ==0.16.0a0 ; extra == 'xpu'
 Requires-Dist: intel-extension-for-pytorch ==2.1.10+xpu ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250105 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250105 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250105 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250106 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250106 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250106 ; extra == 'xpu'
 Provides-Extra: xpu-2-1
 Requires-Dist: py-cpuinfo ; extra == 'xpu-2-1'
 Requires-Dist: protobuf ; extra == 'xpu-2-1'
@@ -102,9 +102,9 @@ Requires-Dist: setuptools <70.0.0 ; extra == 'xpu-2-1'
 Requires-Dist: torch ==2.1.0a0 ; extra == 'xpu-2-1'
 Requires-Dist: torchvision ==0.16.0a0 ; extra == 'xpu-2-1'
 Requires-Dist: intel-extension-for-pytorch ==2.1.10+xpu ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250105 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250105 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250105 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250106 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250106 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250106 ; extra == 'xpu-2-1'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-2-1'
 Requires-Dist: dpcpp-cpp-rt ==2024.0.2 ; (platform_system == "Windows") and extra == 'xpu-2-1'
 Requires-Dist: mkl-dpcpp ==2024.0.0 ; (platform_system == "Windows") and extra == 'xpu-2-1'
@@ -119,9 +119,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-arc'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-arc'
 Requires-Dist: tabulate ; extra == 'xpu-arc'
 Requires-Dist: setuptools ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250105 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250105 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250105 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250106 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106 ; extra == 'xpu-arc'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-arc'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arc'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arc'
@@ -141,9 +141,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-arl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-arl'
 Requires-Dist: tabulate ; extra == 'xpu-arl'
 Requires-Dist: setuptools ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250105 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250105 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250105 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250106 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106 ; extra == 'xpu-arl'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-arl'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arl'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-arl'
@@ -163,9 +163,9 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-lnl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-lnl'
 Requires-Dist: tabulate ; extra == 'xpu-lnl'
 Requires-Dist: setuptools ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250105 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250105 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250105 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250106 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106 ; extra == 'xpu-lnl'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-lnl'
 Requires-Dist: torch ==2.3.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-lnl'
 Requires-Dist: torchvision ==0.18.1+cxx11.abi ; (platform_system == "Linux") and extra == 'xpu-lnl'

{ipex_llm-2.2.0b20250105.dist-info → ipex_llm-2.2.0b20250106.dist-info}/RECORD

@@ -104,7 +104,7 @@ ipex_llm/transformers/lookup.py,sha256=c4ETIha6ZLbWvhcclSKRDdi5Ipuet4mfUnOkBa0E8
 ipex_llm/transformers/low_bit_linear.py,sha256=dyyYyCqw0GK8hzaUGanrg-uIhU1HTLEEbvbxXMlm-80,41668
 ipex_llm/transformers/model.py,sha256=KcRjkauGg48BYrUBoUZaVMpg7Piuz5JrfIpVZd3EIjs,41105
 ipex_llm/transformers/modelling_bigdl.py,sha256=7JpNVMuyq_OmtNUaMFMXdxPWZp2q0QHC02QeA-VTPOw,6709
-ipex_llm/transformers/npu_model.py,sha256=a1mkyc6EqD7AJhqbYzokGhFubNpt5trIMuZT_dQKlTk,37861
+ipex_llm/transformers/npu_model.py,sha256=YW02GeVz-9ZGqxAeSz0AOvciS-17bo9eK5ZOBrICwSQ,39508
 ipex_llm/transformers/patches.py,sha256=halPWm__ORh2fRFSIFPiCNg3LQBfrRkTPtmtRpBJCZQ,1286
 ipex_llm/transformers/pipeline_parallel.py,sha256=uNZpOXljNmdoEYnP8U-VFiN4dRZb2piQbIf2bG9LQnE,49051
 ipex_llm/transformers/qlora.py,sha256=jtPGsvWFjbTUGzDBCdfftnCis_0nJQNRpACSwXUbbGU,14943
@@ -194,13 +194,14 @@ ipex_llm/transformers/npu_models/chatglm.py,sha256=YzpGLZ7ORt6qkwW9mCwZ_xhOAI8uH
 ipex_llm/transformers/npu_models/chatglm4.py,sha256=J4523DzhIzZxIvlf1V9qU4auzEGKvC80YqyxuCJygjw,9795
 ipex_llm/transformers/npu_models/common.py,sha256=tTUJL7IxVrJSnXle6nla35wTUrBf2sOEt7Ya1qyMezY,4853
 ipex_llm/transformers/npu_models/convert.py,sha256=FILSGnoltcR9FMrCkw0eOKh6p3sbBI5i0Ms8AsJc04E,25342
-ipex_llm/transformers/npu_models/convert_mp.py,sha256=t7160V4MmYpnex2NfuLTcqoc1meGEXdYi4AAPotfbzk,24518
+ipex_llm/transformers/npu_models/convert_mp.py,sha256=ADMTnY3utRmCA9kGOCoiJ3NTI4via3TiX6i8duJ2TIE,24504
 ipex_llm/transformers/npu_models/glm_edge.py,sha256=VsJex-6530h4ZQk35TxRe1MnttAHT41omE8LV47LgBE,6723
 ipex_llm/transformers/npu_models/kv.py,sha256=2OSFO9Z6e4nGdVxXEM-Bq2qa_npYYbGmQt3lcCZxTlU,9201
-ipex_llm/transformers/npu_models/linear.py,sha256=G7W3tFXLG4FDzz-vc90_-YlEl1GxdoNz4XliqdlUb2U,10878
+ipex_llm/transformers/npu_models/linear.py,sha256=RQxL42laJTm5hz11SNl0KlJX9xM6C_0OiN6soLShDM0,11284
 ipex_llm/transformers/npu_models/llama.py,sha256=WpRcw7sLnbZeR4XoM-a6XQ-BNYeQaHBEOX1r_O5C9uo,9857
 ipex_llm/transformers/npu_models/llama_mp.py,sha256=6fyWzbFozKPOfSPDBk2x_Rsejj2P0HOR-jn8SNUWy3s,49349
-ipex_llm/transformers/npu_models/lm_head.py,sha256=ZwH23jf21pNILTriAeF46TdymoVP3_OgQArsQX6lZzI,7823
+ipex_llm/transformers/npu_models/lm_head.py,sha256=-yS0sM8905sQ2S9pwCZ6pX1vZa15aqOPGdoYEyhGbuQ,4825
+ipex_llm/transformers/npu_models/lm_head_linear.py,sha256=7iPV0UqPqEYTP4-cUtJXxgjOzAObzEY9Utuu3nTIRqI,3683
 ipex_llm/transformers/npu_models/minicpm.py,sha256=PP3nrCoHMcQ3kHrnQ0gYglCKvL0Dh5MAQX3_Yq8_Ygc,10411
 ipex_llm/transformers/npu_models/minicpm_mp.py,sha256=0iCRWN9UIUQp5tSKyu-orpGCOxEjQrJ9b8ePnG0-ZV4,41921
 ipex_llm/transformers/npu_models/minicpmv_mp.py,sha256=m11WT6s_H5wkFtlz7aHMOL9b_CoL_G5MhoL5te4la_Q,20147
@@ -253,11 +254,11 @@ ipex_llm/vllm/xpu/engine/__init__.py,sha256=pY_CpyuZd72fr6s32ejeKHKFW0K4vUU2rzZj
 ipex_llm/vllm/xpu/engine/engine.py,sha256=k4-D27WS_Gk3mA--w3HWAjPjb4Aiu043MVPi0ZoAUBc,5984
 ipex_llm/vllm/xpu/entrypoints/openai/api_server.py,sha256=GshTZFB8e4PWvqckfbmTOU6b0oLkNn7A-vzLuG9--j8,21544
 ipex_llm/vllm/xpu/entrypoints/openai/cli_args.py,sha256=2rENA2ucynMaIjiZBEh2ez1o5vR32GaP514t39CD7KM,8676
-ipex_llm-2.2.0b20250105.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
-ipex_llm-2.2.0b20250105.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
-ipex_llm-2.2.0b20250105.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
-ipex_llm-2.2.0b20250105.dist-info/METADATA,sha256=V2tvYG6plKx9JKN-Jhm0eImBpQzazdndXQ2s2OOvwAE,11374
-ipex_llm-2.2.0b20250105.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
-ipex_llm-2.2.0b20250105.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.2.0b20250105.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.2.0b20250105.dist-info/RECORD,,
+ipex_llm-2.2.0b20250106.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
+ipex_llm-2.2.0b20250106.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
+ipex_llm-2.2.0b20250106.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
+ipex_llm-2.2.0b20250106.dist-info/METADATA,sha256=RVDr0pwoPE6J0yPUZ9k7t6_jQn01wTwAXkU5ViqE-c8,11374
+ipex_llm-2.2.0b20250106.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
+ipex_llm-2.2.0b20250106.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.2.0b20250106.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.2.0b20250106.dist-info/RECORD,,