ipex-llm 2.2.0b20250105.post0__py3-none-win_amd64.whl → 2.2.0b20250106.post1__py3-none-win_amd64.whl

ipex_llm/transformers/convert.py

@@ -680,18 +680,9 @@ def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
                             optimize_lm_head=optimize_lm_head
                         )
                     device = module.weight.data.device
-                    from ipex_llm.transformers.utils import get_ipex_version
-                    if get_ipex_version() < "2.1.10+xpu":
-                        new_linear._parameters['weight'] = nn.Parameter(module.weight)
-                    else:
-                        # only from 2.1, ipex provides matmul_bias_out
-                        # so we need to transpose weight
-                        new_weight = module.weight.transpose(0, 1).contiguous()
-                        new_linear._parameters['weight'] = nn.Parameter(new_weight)
-                        new_linear.weight_type = 2
+                    new_linear._parameters['weight'] = nn.Parameter(module.weight)
                     if module.bias is not None:
-                        new_linear._parameters['bias'] = nn.Parameter(module.bias.data)\
-                            .to(device)
+                        new_linear._parameters['bias'] = nn.Parameter(module.bias.data).to(device)
                 elif qtype == ggml_tensor_qtype["bf16"]:
                     module.to(torch.bfloat16)
                     if _USE_VLLM:
@@ -1452,21 +1443,6 @@ def _optimize_post(model):
                             module.MultiheadAttention,
                             mpt_multihead_attention_forward
                             )
-    elif "gptj" in model.config.model_type:
-        # dolly-v1-6b
-        modeling_module_name = model.__class__.__module__
-        module = importlib.import_module(modeling_module_name)
-        from ipex_llm.transformers.models.gptj import gptj_attention_forward, gptj_model_forward,\
-            gptj_block_forward
-        convert_forward(model,
-                        module.GPTJAttention,
-                        gptj_attention_forward)
-        convert_forward(model,
-                        module.GPTJModel,
-                        gptj_model_forward)
-        convert_forward(model,
-                        module.GPTJBlock,
-                        gptj_block_forward)
     elif "bloom" in model.config.model_type:
         modeling_module_name = model.__class__.__module__
         module = importlib.import_module(modeling_module_name)

ipex_llm/transformers/loader.py

@@ -22,7 +22,7 @@ import time
 from datetime import date
 import argparse
 from ipex_llm.utils.common import invalidInputError
-from transformers import AutoTokenizer, GPTJForCausalLM, LlamaTokenizer
+from transformers import AutoTokenizer, LlamaTokenizer
 
 LLAMA_IDS = ['llama', 'vicuna', 'merged-baize']
 

ipex_llm/transformers/low_bit_linear.py

@@ -286,7 +286,7 @@ def use_batch_forward(x: torch.Tensor, qtype: int, output_len: int):
             or (
                 qtype in [SYM_INT8, FP4, FP6, Q4_K, Q6_K]
                 and batch_size <= 48
-                and device_name in ["arc", "pvc", "mtl", "lnl", "arl"]
+                and device_name in ["arc", "pvc", "mtl", "arl"]
                 and x.shape[1] % 256 == 0
                 and output_len % 32 == 0
             )
@@ -759,9 +759,9 @@ class FP16Linear(nn.Linear):
         self.weight_length = self.out_len * self.in_len
         self.qtype = ggml_tensor_qtype["fp16"]
         self.mp_group = mp_group
-        # weigh_type = 1 means original weight
-        # weigh_type = 2 means weight has been transposed
-        # weigh_type = 3 means weight has been transposed by esimd method
+        # weight_type = 1 means original weight
+        # weight_type = 2 means weight has been transposed
+        # weight_type = 3 means weight has been transposed by esimd method
         self.weight_type = 1
         self.optimize_lm_head = optimize_lm_head
         self.disable_fp16_opt = False
@@ -775,28 +775,14 @@ class FP16Linear(nn.Linear):
 
         x = x.to(torch.float16)
         if self.bias is not None and self.bias.dtype != x.dtype:
-                self.bias.data = self.bias.data.to(x.dtype)
+            self.bias.data = self.bias.data.to(x.dtype)
         if self.weight is not None and self.weight.dtype != x.dtype:
             self.weight.data = self.weight.data.to(x.dtype)
 
         if not self.use_esimd_kernel(x):
-            if (
-                get_ipex_version() < "2.1.10+xpu"
-                or get_xpu_device_name(x.device) not in ["arc", "pvc"]
-                or self.disable_fp16_opt
-            ):
-                if self.weight_type == 2:
-                    self.weight = torch.nn.Parameter(self.weight.transpose(0, 1).contiguous(),
-                                                     requires_grad=False)
-                    self.weight_type = 1
-                result = F.linear(x, self.weight, self.bias)
-            else:
-                if self.weight_type == 1:
-                    self.weight = torch.nn.Parameter(self.weight.transpose(0, 1).contiguous(),
-                                                     requires_grad=False)
-                    self.weight_type = 2
-                result = torch.ops.torch_ipex.matmul_bias_out(x.contiguous(),
-                                                              self.weight, self.bias)
+            invalidInputError(self.weight_type == 1, "weight_type should be 1")
+            result = F.linear(x, self.weight, self.bias)
+
             if self.mp_group is not None:
                 if get_use_vllm():
                     result = self.mp_group.all_reduce(result)
@@ -852,7 +838,7 @@ class FP16Linear(nn.Linear):
         if self.disable_fp16_opt:
             return False
         # esimd kernel can only be used for Arc and Flex
-        if gpu_type not in ["arc", "flex"]:
+        if gpu_type not in ["arc"]:
             return False
         # now esimd kernel can only be used for specific cases (llama2-7b shape)
         if self.in_len == 11008 and self.out_features == 4096:

ipex_llm/transformers/model.py

@@ -103,12 +103,6 @@ def save_low_bit(self, *args, **kwargs):
         self.to(origin_device)
 
 
-def _load_pre():
-    from transformers import GPTJModel
-    from ipex_llm.transformers.models.gptj import gptj_model_new_init
-    GPTJModel.__init__ = gptj_model_new_init
-
-
 class _BaseAutoModelClass:
     HF_MODEL = None
 
@@ -495,7 +489,6 @@ class _BaseAutoModelClass:
         else:
             if quant_config is not None:
                 kwargs["quantization_config"] = quant_config
-            _load_pre()
             try:
                 # To handle the input CUDA setting (such as 'device_map={"":0}'), ignore it
                 kwargs.pop('device_map', None)

ipex_llm/transformers/models/utils.py

@@ -168,7 +168,7 @@ def should_use_fuse_rope(hidden_states, position_ids, training):
 
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
     if model_family in ["llama", "baichuan", "internlm", "aquila", "gpt_neox", "mistral",
-                        "mixtral", "qwen2", "yuan", "stablelm", "qwen2_moe"]:
+                        "qwen2", "yuan", "stablelm", "qwen2_moe"]:
         # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
         cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
         sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
@@ -183,7 +183,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
         q_embed = (q * cos) + (rotate_half(q) * sin)
         k_embed = (k * cos) + (rotate_half(k) * sin)
         return q_embed, k_embed
-    elif model_family in ["gptj", "chatglm"]:
+    elif model_family in ["chatglm"]:
         q_embed = (q * cos) + (rotate_every_two(q) * sin)
         k_embed = (k * cos) + (rotate_every_two(k) * sin)
         return q_embed, k_embed
@@ -192,19 +192,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
                           f"{model_family} is not supported.")
 
 
-def apply_ipex_rotate_every_two(q, k, cos, sin):
-    # ipex's apply_rotary_embedding_two_qk can change the origin storage,
-    # so q/k will get the result directly.
-    from ipex_llm.transformers.utils import get_ipex_version
-    if get_ipex_version() >= "2.1.10+xpu":
-        torch.ops.torch_ipex.apply_rotary_embedding_two_qk(
-            q, k, sin, cos, q, k
-        )
-    else:
-        torch.ops.torch_ipex.apply_rotary_embedding(q, sin, cos, q)
-        torch.ops.torch_ipex.apply_rotary_embedding(k, sin, cos, k)
-
-
 def is_enough_kv_cache_room_4_36(past_key_value, idx, seq_len=1):
     # to determinate if is enough kv cache room in transformers==4.36
     # seq_len for current seq len

ipex_llm/transformers/speculative.py

@@ -432,8 +432,7 @@ def _check_and_extend_kv_cache(past_key_values, max_step_draft, kv_alloc_block_l
     from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_31, \
         extend_kv_cache
     enough_kv_room = True
-    if model_type not in ["chatglm", "qwen", "baichuan", "llama", "mistral",
-                          "gptj", "opt"]:
+    if model_type not in ["chatglm", "qwen", "baichuan", "llama", "mistral", "opt"]:
         return past_key_values, False
     cache_k = past_key_values[0][0]
     if model_type == "chatglm":
@@ -527,7 +526,7 @@ def _crop_past_key_values(self, past_key_values, new_cache_size, _enable_ipex=Fa
                         v[:-(new_cache_size), :, :, :])
                     for k, v in past_key_values
                 ]
-        elif self.config.model_type in ["baichuan", "gptj"]:
+        elif self.config.model_type in ["baichuan"]:
             past_key_values = [
                 (k[:, :, :-(new_cache_size), :],
                     v[:, :, :-(new_cache_size), :])
@@ -796,13 +795,6 @@ def _non_cpu_ipex_verify(self, verify_input_ids, past_key_values, cur_attention_
                                     device=verify_input_ids.device)
         position_ids = position_ids.unsqueeze(0).repeat(1, 1) + past_key_value_len
         forward_args["position_ids"] = position_ids
-    elif self.config.model_type == "gptj":
-        past_length = past_key_values[0][0].size(2)
-        input_len = verify_input_ids.shape[1]
-        position_ids = torch.arange(past_length, input_len + past_length,
-                                    dtype=torch.long, device=verify_input_ids.device)
-        position_ids = position_ids.unsqueeze(0).view(-1, input_len)
-        forward_args["position_ids"] = position_ids
 
     return self(**forward_args)
 
@@ -971,10 +963,6 @@ def speculative_generate(self,
                         past_key_value_len = past_key_values[0][0].shape[0]
                     position_ids = torch.Tensor([[past_key_value_len + step_draft]]).long()
                     forward_args["position_ids"] = position_ids
-                elif self.config.model_type == "gptj":
-                    past_length = draft_past_key_values[0][0].size(2)
-                    position_ids = torch.Tensor([[past_length]]).long().to(self.device)
-                    forward_args["position_ids"] = position_ids
 
                 if _enable_ipex:
                     if any(keyword in self.config.model_type

ipex_llm/transformers/utils.py

@@ -172,6 +172,8 @@ def get_xpu_device_name(device: torch.device):
     if device.type != "xpu":
         return device.type
     else:
+        # possiable device name:
+        # ["arc", "pvc", "mtl", "lnl", "bmg", "arl", "legacy", "unknown"]
         import xe_linear
         return xe_linear.get_xpu_device_name(device)
 

{ipex_llm-2.2.0b20250105.post0.dist-info → ipex_llm-2.2.0b20250106.post1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250105.post0
+Version: 2.2.0b20250106.post1
 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 ==2.6.0b20250106.post0 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250106.post1 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
 Provides-Extra: cpp-arl
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250106.post0 ; extra == 'cpp-arl'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250106.post1 ; 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: onednn ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
@@ -67,7 +67,7 @@ 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 ==2.6.0b20250106.post0 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250106.post1 ; (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'
@@ -87,9 +87,9 @@ 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 ==2.6.0b20250106.post0 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250106.post0 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250106.post0 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250106.post1 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250106.post1 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250106.post1 ; extra == 'xpu'
 Provides-Extra: xpu-2-1
 Requires-Dist: py-cpuinfo ; extra == 'xpu-2-1'
 Requires-Dist: protobuf ; extra == 'xpu-2-1'
@@ -104,9 +104,9 @@ 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 ==2.6.0b20250106.post0 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250106.post0 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250106.post0 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250106.post1 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250106.post1 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250106.post1 ; 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'
@@ -124,7 +124,7 @@ Requires-Dist: setuptools ; extra == 'xpu-2-6'
 Requires-Dist: torch ==2.6.0+xpu ; extra == 'xpu-2-6'
 Requires-Dist: torchvision ==0.21.0+xpu ; extra == 'xpu-2-6'
 Requires-Dist: torchaudio ==2.6.0+xpu ; extra == 'xpu-2-6'
-Requires-Dist: bigdl-core-xe-all ==2.6.0b20250106.post0 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.6.0b20250106.post1 ; extra == 'xpu-2-6'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-2-6'
 Provides-Extra: xpu-arc
 Requires-Dist: py-cpuinfo ; extra == 'xpu-arc'
@@ -137,9 +137,9 @@ 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 ==2.6.0b20250106.post0 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106.post0 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106.post0 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250106.post1 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106.post1 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106.post1 ; 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'
@@ -160,9 +160,9 @@ 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 ==2.6.0b20250106.post0 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106.post0 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106.post0 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250106.post1 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106.post1 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106.post1 ; 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'
@@ -183,9 +183,9 @@ 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 ==2.6.0b20250106.post0 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106.post0 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106.post0 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250106.post1 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250106.post1 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250106.post1 ; 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.post0.dist-info → ipex_llm-2.2.0b20250106.post1.dist-info}/RECORD

@@ -41,35 +41,35 @@ ipex_llm/langchain/llms/transformerspipelinellm.py,sha256=vm522YPPwWxxAPVvQBtxRf
 ipex_llm/langchain/vllm/__init__.py,sha256=T-EbRT6GJ_8RCu-iLmSzcftOimXSPQf2d5X72AUAy2Y,874
 ipex_llm/langchain/vllm/vllm.py,sha256=6dxc-ZISZQrJilEa_HA827l75Dv9rcHpY_G6FdJ8BVs,7793
 ipex_llm/libs/__init__.py,sha256=47DEQpj8HBSa-_TImW-5JCeuQeRkm5NMpJWZG3hSuFU,0
-ipex_llm/libs/bloom-api.dll,sha256=91zD_S-epu_Gp_dAblCkxDwZFjSBhhfho6to8r3K3Ac,36352
-ipex_llm/libs/bloom.dll,sha256=wxdywP2aMiNAKEw3uU31aAe9RR2r1JjhIaqPuslYBZU,506880
-ipex_llm/libs/gptneox-api.dll,sha256=MNnJv8cjFUg_siCIlKPSCg8_kTVHrTFTvDPKu8SAHZw,24576
-ipex_llm/libs/gptneox.dll,sha256=d-3KIFy39GehluUVnpipGeN1qTqVkKWrz5dNT6j9SPk,567296
-ipex_llm/libs/libbloom_avx.dll,sha256=iwht0gpe7yTRMF3F5ZQ8DRs26vQi7i2D8Mafg_Yyt5Q,535040
-ipex_llm/libs/libbloom_vnni.dll,sha256=VJxjLCbNP0wZBW6Rw0owR_rc9z3PKPONHLHOIZYiZiM,506880
-ipex_llm/libs/libgptneox_avx.dll,sha256=3Ja3SkOp6wCtA1NB6GRBQLseB91bzgbhKOWRKh0o01A,595456
-ipex_llm/libs/libgptneox_vnni.dll,sha256=8hinv9vxdzjsW7PJhukAxKeLAtF2f2PQOmzwnNEzDD8,567808
-ipex_llm/libs/libllama_avx.dll,sha256=R7yRH3t4q0CXy0Sv91m8j2z0wnG6Qa_-JxEp6kTcgt0,589824
-ipex_llm/libs/libllama_vnni.dll,sha256=Cg9vrS5J4Uev0iW7p5346-flWKLjNJqyItv89g3qwRs,561664
-ipex_llm/libs/libstarcoder_avx.dll,sha256=dCgKfpmh57NN4sjs6j7bQFtO-URXcTmPS6oktgLHwsU,626688
-ipex_llm/libs/libstarcoder_vnni.dll,sha256=KZufV9B-riiEIitRMxv9nmGWxXM6Qbrv4E8ybustdQA,598528
-ipex_llm/libs/llama-api.dll,sha256=lS0rqYlsIUJMOBi1rkLqurGnCsCaDCAe4QZCHZGGeGI,25600
-ipex_llm/libs/llama.dll,sha256=2exLzMA9NFPkMldvfZeT-Ua6moJjcUDsTwv-hyLEjFg,561152
-ipex_llm/libs/main-bloom.exe,sha256=Rjc3aUPGdbJ1nkRNWZ9j637pA4nZ6fC4EBXJE3vhKkM,103424
-ipex_llm/libs/main-gptneox.exe,sha256=eIq223joAOSPSi0nKxNI0_z9CjC_9f0XGx5ZEFdMX8w,98816
-ipex_llm/libs/main-llama.exe,sha256=-jaXiwAERqW8QYjOB1VYs2-hU-HSLYoheA9R7hQnqPg,99840
-ipex_llm/libs/main-starcoder.exe,sha256=-2M-nKhMW8AMhUvwno-iN-x32G_Csb6TCK4b45-ywQE,157696
-ipex_llm/libs/pipeline.dll,sha256=J_rocSlOut7S0JZQnZlUD-IQ92D5P4pAvzbrPZFG7us,72704
-ipex_llm/libs/quantize-bloom.exe,sha256=XYSsf4HvaYKc0J8022uz6nz6AifT-8FhB7dqJRuEGoE,126464
-ipex_llm/libs/quantize-bloom_vnni.exe,sha256=4mpem2yXPOPKoSsezR32Ckng-_mVE37yHWxY4PG48mc,127488
-ipex_llm/libs/quantize-gptneox.exe,sha256=ZqX02ScTq8X90MRYbJVo5NFdI8E0MzwCpS2UGGK6mO8,104448
-ipex_llm/libs/quantize-gptneox_vnni.exe,sha256=rr-sll3eMtxB0Rg-tXQsFUhcrHkkLnAu8HZWzeqVPJ4,104960
-ipex_llm/libs/quantize-llama.exe,sha256=XrWIjKL-5XaqThU-fNnfVS4pTc1Yp_pGlhwALHU4c-k,109568
-ipex_llm/libs/quantize-llama_vnni.exe,sha256=AuQqKGxs2U3S2Zyk-GMDFFpaTKeQpGR3ydW5gZ8FbWA,110592
-ipex_llm/libs/quantize-starcoder.exe,sha256=X3TroCzyM_i8El6D0-V2mufnRgZc5kBsW1qNoJKIdcE,127488
-ipex_llm/libs/quantize-starcoder_vnni.exe,sha256=K1NgSQOr5-xEEOHqD25qxJe5Z6esvUivIJEkBYto6KM,128512
-ipex_llm/libs/starcoder-api.dll,sha256=YTFtKQ0jSo2cgwO7HkIfjPvD5jANUaDWSa9ykCLfU5s,21504
-ipex_llm/libs/starcoder.dll,sha256=KWX7QLhGFBGUBea-Od0udwKxnJnblsi4y8TSJl9p67g,598016
+ipex_llm/libs/bloom-api.dll,sha256=MSVCD9yi2nlgksGF-5u7tKE660pzfRthAuTWK_7ozbE,36352
+ipex_llm/libs/bloom.dll,sha256=DMJiNnnBbuZArLpPrSIeVgBQtgHRKK0gfqF3e_4N4Ak,506880
+ipex_llm/libs/gptneox-api.dll,sha256=B9yLQPHTYv0o88cZ3pYCfY5ksnetr0xcWC4wKYXhsKs,24576
+ipex_llm/libs/gptneox.dll,sha256=B8H5jlqDLVMSi4wqdPZ0ozI1smbuY7jRN-9YQXSoDQI,567296
+ipex_llm/libs/libbloom_avx.dll,sha256=syY7Lc2kwoM7D57WPaj2XIMJCXyY1IfXClgkY8Lrs1w,535040
+ipex_llm/libs/libbloom_vnni.dll,sha256=RtkwaaPVlbWv6OtxClKILy5CYlL98olnJYdFbe95tEM,506880
+ipex_llm/libs/libgptneox_avx.dll,sha256=1gmdd3DdFrMJBI4pXyI0_703ndz2kopDkK8vz7uPyKI,595456
+ipex_llm/libs/libgptneox_vnni.dll,sha256=bUKisE3_Cg4rZ2AtjGhVoXtKqNVAo4MZI7N7Q4Q_cf0,567808
+ipex_llm/libs/libllama_avx.dll,sha256=lfcj2lw9lIgmDifxIzNnzcz6fHuYUIVYvdvwoM8ijOE,589824
+ipex_llm/libs/libllama_vnni.dll,sha256=5xY7lEJ-sW6owTNcBO8SWRCoweo7k8N4FS4v9-lQcdc,561664
+ipex_llm/libs/libstarcoder_avx.dll,sha256=z8P_H55-1VQFB9KcBxnBrSlkLlsLpxPghN9VfKpQulA,626688
+ipex_llm/libs/libstarcoder_vnni.dll,sha256=rPEGDTZlUTl0hUP40uRe8idjkVhNkyYk818RYmONL1Q,598528
+ipex_llm/libs/llama-api.dll,sha256=XJP636817YbUBOneSVGLqr9-qefV4DV6vR_8P7e8Q7k,25600
+ipex_llm/libs/llama.dll,sha256=YF2m8rwTy7hRbCo31FG-fo1q6DStBAxenNgkz4_gkr0,561152
+ipex_llm/libs/main-bloom.exe,sha256=lWEzFmr2zWPLnDUy3l-FBv5k-xLn7BXm1A3RSkZx890,103424
+ipex_llm/libs/main-gptneox.exe,sha256=ex0FAncs6bLngz7-jmtCxuDNEWlZCXThBMCdveDJnEs,98816
+ipex_llm/libs/main-llama.exe,sha256=ooEHT6YMGZTCKo3h-m3E7f8vkc6aMJPZiRH9dU_4Wu0,99840
+ipex_llm/libs/main-starcoder.exe,sha256=r9rZZhk0V4_6L7EH7evcBdmRyWSqiZ_oNon3mjSCXEY,157696
+ipex_llm/libs/pipeline.dll,sha256=5rhnCddgAUozj2ZQox5ykjsWivaNsK-jVcvArO9Hfy8,72704
+ipex_llm/libs/quantize-bloom.exe,sha256=V-ht-dVMJ9dq4O9vokGlBoOyTCCIgZoZMC-Cn8UdPL4,126464
+ipex_llm/libs/quantize-bloom_vnni.exe,sha256=q14nAfmyRJW57xLG10wKCODG-cSOZ6KP7VCIVPAqslM,127488
+ipex_llm/libs/quantize-gptneox.exe,sha256=Mjhxhcb8FXwrGI0dMRrHXzOOWkLZwtYQkpj4i2gSTYk,104448
+ipex_llm/libs/quantize-gptneox_vnni.exe,sha256=_9D8mqr8zCI7sAzRZ_KlCayqDq0ctQWcwAzI4NWdH7U,104960
+ipex_llm/libs/quantize-llama.exe,sha256=TsPfUBMnQNcKA44_QG9wrmRSEcs2bAmiSOD-ll57g6w,109568
+ipex_llm/libs/quantize-llama_vnni.exe,sha256=1fUjt6woxEAzYQ1b0Ok3JRo1YcFcTiogB-l0RdOs_iY,110592
+ipex_llm/libs/quantize-starcoder.exe,sha256=-wvnnZhGw_eKPnbhZiiaepceFbd_zMUY7qQwrKg4miw,127488
+ipex_llm/libs/quantize-starcoder_vnni.exe,sha256=iPbhV8PTL7J3JXJnJkRw1tsvaHA0kGxgwpttSzVkUHQ,128512
+ipex_llm/libs/starcoder-api.dll,sha256=bpzxqpZZE9Q7Yghy-33_Hkvo-OLvKntfqH-DJLWfXjQ,21504
+ipex_llm/libs/starcoder.dll,sha256=QXucnOagZLcxQF8PG_nvQ9kW4XvexngcqrmJE5N5o9s,598016
 ipex_llm/llamaindex/__init__.py,sha256=T-EbRT6GJ_8RCu-iLmSzcftOimXSPQf2d5X72AUAy2Y,874
 ipex_llm/llamaindex/llms/__init__.py,sha256=KP1lEdGqDuxPoxL1ZSH25Pm2kKMPJBWUTLR0ckSLMIU,1139
 ipex_llm/llamaindex/llms/bigdlllm.py,sha256=FQBzq1KOjfc6uofTXAha3O7TqpJkNfOFepXQmOVlbnI,26314
@@ -87,25 +87,25 @@ ipex_llm/serving/fastchat/tgi_api_protocol.py,sha256=brT3k3-V0NJrU4fRqUwWjC0O3iO
 ipex_llm/serving/fastchat/tgi_api_server.py,sha256=agNTAEiZPSuj3dEdIdYKwkoY0cXOUDX06DiM9VP2knQ,24418
 ipex_llm/serving/fastchat/vllm_worker.py,sha256=ZLz2Q9GxJO6r_LOiP6epgCRjBGk-K4EB1SNEWSJp5DA,11091
 ipex_llm/transformers/__init__.py,sha256=l4KkMkLe-pRC7b_kj6LCfeifgE-Uo33_Av_FwN9HnFA,1074
-ipex_llm/transformers/convert.py,sha256=pFm6VlU84u_Llr2sp6-gRrEYDeNgIk2QPukolq4IE1s,99947
+ipex_llm/transformers/convert.py,sha256=TxWdTTOSvh-j5jqokQJVWykta4U4LHupE1QJ-9udzwc,98733
 ipex_llm/transformers/convert_ipex.py,sha256=iKXo0n8fVFTOA2fNYYrByMFK0dovL-kLd2sVDk88AlQ,14334
 ipex_llm/transformers/embedding.py,sha256=bdgk59DvD4ZZyxRzewXOR7g56nThgO6uhIwk8QL7f-s,9299
 ipex_llm/transformers/kv.py,sha256=k4TU18LlA-Sbq9WNNQnfuzu3RSFBwFhmaV3BcGN5bAo,19191
 ipex_llm/transformers/lisa.py,sha256=F5WxbtXQ7RdKulj83h_2DnEIgKiKGZf7zvOmg6QBl2s,3289
-ipex_llm/transformers/loader.py,sha256=cOgX93xOC-4dt01GTJ5wyd7PjZ8S43r4mctkR2YxVuw,6893
+ipex_llm/transformers/loader.py,sha256=AwjV5RpI2t2bedlv7ZhLm8cfd-QJZm5hny-XyjIvdnk,6876
 ipex_llm/transformers/lookup.py,sha256=b6OlZ9OV10R9qeWw8mVryVpDxszkjwLkldvi7GPMJY8,19614
-ipex_llm/transformers/low_bit_linear.py,sha256=nKraUvZJ7UdXP29HSE4CJPIVxmN-TvG8dpT4gpleuyQ,41688
-ipex_llm/transformers/model.py,sha256=KcRjkauGg48BYrUBoUZaVMpg7Piuz5JrfIpVZd3EIjs,41105
+ipex_llm/transformers/low_bit_linear.py,sha256=lPIvDuRoS0zusiJ6vw_fOTJgK5ylh4CuD3U-qs8ih4Y,40869
+ipex_llm/transformers/model.py,sha256=fj7LBjrWtWwDJJYXnWiXsLGS4ayqqHfnh0p51dSDssE,40908
 ipex_llm/transformers/modelling_bigdl.py,sha256=7JpNVMuyq_OmtNUaMFMXdxPWZp2q0QHC02QeA-VTPOw,6709
 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
 ipex_llm/transformers/relora.py,sha256=-dYzUV0P-IhO2jFdnzN9-v_sFzJpRj3ZwN9eCJzOoCw,16567
-ipex_llm/transformers/speculative.py,sha256=Zf1nQb5GXpJQrUHBTL-H4RUBfdv3lGhfehzudHimhYk,64109
+ipex_llm/transformers/speculative.py,sha256=0XNLgc9dGswJHVPrXo4iM7pPxkWwfFfJMECcivJSnIc,63368
 ipex_llm/transformers/streamer.py,sha256=RrVlLblzCOtABRUpaMXAyaMnCGgLUtAi_YesLumRbww,4842
 ipex_llm/transformers/training_patch.py,sha256=oxMkUtqyvqJiprw6dE3skkYfD1HOmUlH9N0hBkbn0G0,10799
-ipex_llm/transformers/utils.py,sha256=fXLIlr9hoBr27p3w3xzczZGPk2cCTIRbUKBkiVCGYbc,16889
+ipex_llm/transformers/utils.py,sha256=7syzq4jnEo-mWmS9E2VZ2GcFrjojWI8E7Hcx0tloifg,16996
 ipex_llm/transformers/xpu_customize_fwd.py,sha256=wFpIhs5F6tkNs8gBOrLxWdhLzO3EDHovVkERPIAoAvg,7611
 ipex_llm/transformers/xpu_ops.py,sha256=H46-69pMRQhekbAEoDfNacCInLWycMHDqrgMGLvFYfI,4362
 ipex_llm/transformers/awq/__init__.py,sha256=Du5gu3-eeAkeDO_dEMBTzrDBA66DSN3uL3-rn8WGXQw,875
@@ -151,7 +151,6 @@ ipex_llm/transformers/models/gemma2.py,sha256=2WZuv-FLzJyTJFaYxOuzJt47QE64M0lHnz
 ipex_llm/transformers/models/glm.py,sha256=gHYgfn20jPRL-ElXy-rUqMh6_LQcc5x7DEXSZuRA4E0,7094
 ipex_llm/transformers/models/gpt2.py,sha256=YSaNgK1uLCFDuIFqnKO0Mi-AsOZsYav-7pNf_NpKGdM,3445
 ipex_llm/transformers/models/gptbigcode.py,sha256=cP1_qGWoa43R2WacAMblShjku4QupcCZiLaPPAoOUs4,9101
-ipex_llm/transformers/models/gptj.py,sha256=TTIx461X2nOcIkrAcZhEf7d7mjJ3yvEC9KLVc1-hrpc,17973
 ipex_llm/transformers/models/gptneox.py,sha256=loRh1x_5S6BCeOr_s5xr-N_1SQHL3Y5IiUBAEyoMUqQ,6172
 ipex_llm/transformers/models/internlm.py,sha256=ZbIUMDwNRcrCeduXfbA_uq1AUEWawEt6CJRvQl3LkAg,17832
 ipex_llm/transformers/models/internvl.py,sha256=Vx0vENIEQLX2M6P398mw5TOhpks0U8xf8rtRQvy94go,8154
@@ -175,7 +174,7 @@ ipex_llm/transformers/models/rwkv5.py,sha256=OkRNj1pCAZg1z2Fw-I0DEnxLEdZyPeRSQ6m
 ipex_llm/transformers/models/sd.py,sha256=VvHV5u-0k2MgHu3NL9113hPj7DgfxqctuKzEEeNfRDU,5981
 ipex_llm/transformers/models/stablelm.py,sha256=RGQCYuQhYqtZ1j3RZkYi0_QvCRnUgUIPYxfBcLnElzg,6885
 ipex_llm/transformers/models/starcoder2.py,sha256=4P3mhRYf2Kreb1ESjrQGfy1puLMmZXgV35zf-Tksvao,6462
-ipex_llm/transformers/models/utils.py,sha256=Qbz7UkYSbsM5bodH2445O0-JF50Mu3UEwW0j2ZNxHSU,15997
+ipex_llm/transformers/models/utils.py,sha256=85rGIzGZvWe3JjYpWcuc1nfzI_tn_zFcdZpWivxJkl0,15457
 ipex_llm/transformers/models/yuan.py,sha256=1jRPebwAK2ENbyYokOmb4LSVo-szucWiygz9zTv-scs,7656
 ipex_llm/transformers/npu_models/__init__.py,sha256=ulEUGLjaP48LCrVeury3UxLjXxKzRi0UpSG4bYu-7f8,585
 ipex_llm/transformers/npu_models/baichuan.py,sha256=fJtd7fBrttySghRUgfZTAdxLjsSNC-XL08HISsXigLE,4685
@@ -244,11 +243,11 @@ 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.post0.data/scripts/ipex-llm-init.bat,sha256=HPtCYuDYwEatq7dAwOvdfVcHYCpAVdbj75K1qh0vQek,2578
-ipex_llm-2.2.0b20250105.post0.data/scripts/llm-chat.ps1,sha256=6qrs-hGVAV8IKh7Jx8nq_XrnZcjd7qGU5wndArM7Yag,2769
-ipex_llm-2.2.0b20250105.post0.data/scripts/llm-cli.ps1,sha256=3qBtTLs_EjYDnM8YyCpJhzLnGCKTEGssu9UNqfkjVXs,3009
-ipex_llm-2.2.0b20250105.post0.dist-info/METADATA,sha256=-eNpo4zm9w1DQqVCTFi228urj8ylbTuXml4uNwlEP3E,12825
-ipex_llm-2.2.0b20250105.post0.dist-info/WHEEL,sha256=6iYPr8vTHsyDK75jr9X0V3I9wPSVmtwr_8fdATBciGk,98
-ipex_llm-2.2.0b20250105.post0.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.2.0b20250105.post0.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.2.0b20250105.post0.dist-info/RECORD,,
+ipex_llm-2.2.0b20250106.post1.data/scripts/ipex-llm-init.bat,sha256=HPtCYuDYwEatq7dAwOvdfVcHYCpAVdbj75K1qh0vQek,2578
+ipex_llm-2.2.0b20250106.post1.data/scripts/llm-chat.ps1,sha256=6qrs-hGVAV8IKh7Jx8nq_XrnZcjd7qGU5wndArM7Yag,2769
+ipex_llm-2.2.0b20250106.post1.data/scripts/llm-cli.ps1,sha256=3qBtTLs_EjYDnM8YyCpJhzLnGCKTEGssu9UNqfkjVXs,3009
+ipex_llm-2.2.0b20250106.post1.dist-info/METADATA,sha256=I0vPU5mDtPZR3wpY87fYHdn6r14U0T50NWGP7EsF5s8,12825
+ipex_llm-2.2.0b20250106.post1.dist-info/WHEEL,sha256=6iYPr8vTHsyDK75jr9X0V3I9wPSVmtwr_8fdATBciGk,98
+ipex_llm-2.2.0b20250106.post1.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.2.0b20250106.post1.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.2.0b20250106.post1.dist-info/RECORD,,

ipex_llm/transformers/models/gptj.py

@@ -1,441 +0,0 @@
-#
-# 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.
-#
-# This file is adapted from
-# https://github.com/huggingface/transformers/blob/main/src/transformers/models/gptj/modeling_gptj.py
-#
-
-import torch
-from typing import Optional, Tuple, Union
-from ipex_llm.transformers.models.utils import init_kv_cache, extend_kv_cache, \
-    apply_rotary_pos_emb, append_kv_cache, apply_ipex_rotate_every_two
-from transformers.utils.import_utils import is_torch_fx_proxy
-from transformers.modeling_outputs import BaseModelOutputWithPast
-from transformers.models.gptj.modeling_gptj import GPTJModel
-from ipex_llm.utils.common import invalidInputError
-
-import os
-
-KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
-
-
-def _get_embed_positions(self, position_ids):
-    embed_positions = self.embed_positions
-    if embed_positions.device != position_ids.device:
-        embed_positions = embed_positions.to(position_ids.device)
-        self.embed_positions = embed_positions
-    return embed_positions.repeat(position_ids.shape[0], 1, 1)
-
-
-def _attn(
-    self,
-    query,
-    key,
-    value,
-    attention_mask=None,
-    head_mask=None,
-):
-    # compute causal mask from causal mask buffer
-    query_length, key_length = query.size(-2), key.size(-2)
-    causal_mask = self.bias[:, :, key_length - query_length: key_length, :key_length]
-
-    # Keep the attention weights computation in fp32 to avoid overflow issues
-    query = query.to(torch.float32)
-    key = key.to(torch.float32)
-
-    attn_weights = torch.matmul(query, key.transpose(-1, -2))
-
-    mask_value = torch.finfo(attn_weights.dtype).min
-    # Need to be a tensor, otherwise we get error:
-    # `RuntimeError: expected scalar type float but found double`.
-    # Need to be on the same device, otherwise `RuntimeError: ..., x and y to be on the same device`
-    mask_value = torch.tensor(mask_value, dtype=attn_weights.dtype).to(attn_weights.device)
-    attn_weights = torch.where(causal_mask, attn_weights, mask_value)
-
-    attn_weights = attn_weights / self.scale_attn
-
-    if attention_mask is not None:
-        # Apply the attention mask
-        attn_weights = attn_weights + attention_mask
-
-    attn_weights = nn.functional.softmax(attn_weights, dim=-1)
-    attn_weights = attn_weights.to(value.dtype)
-    attn_weights = self.attn_dropout(attn_weights)
-
-    # Mask heads if we want to
-    if head_mask is not None:
-        attn_weights = attn_weights * head_mask
-
-    attn_output = torch.matmul(attn_weights, value)
-
-    return attn_output, attn_weights
-
-
-def gptj_attention_forward(
-    self,
-    hidden_states: torch.FloatTensor,
-    layer_past: Optional[Tuple[torch.Tensor]] = None,
-    attention_mask: Optional[torch.FloatTensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    head_mask: Optional[torch.FloatTensor] = None,
-    use_cache: Optional[bool] = False,
-    rotary_emb: Optional[Tuple]=None,
-    output_attentions: Optional[bool] = False,
-) -> Union[
-    Tuple[torch.Tensor, Tuple[torch.Tensor]],
-    Optional[Tuple[torch.Tensor, Tuple[torch.Tensor], Tuple[torch.Tensor, ...]]],
-]:
-    query = self.q_proj(hidden_states)
-    key = self.k_proj(hidden_states)
-    value = self.v_proj(hidden_states)
-
-    query = self._split_heads(query, self.num_attention_heads, self.head_dim, True)
-    key = self._split_heads(key, self.num_attention_heads, self.head_dim, True)
-    value = self._split_heads(value, self.num_attention_heads, self.head_dim, False)
-
-    sin, cos = rotary_emb
-    use_fuse_rope = hidden_states.device.type == "xpu" and not self.training
-
-    if self.rotary_dim is not None:
-        k_rot = key[:, :, :, : self.rotary_dim]
-        q_rot = query[:, :, :, : self.rotary_dim]
-
-        if use_fuse_rope:
-            apply_ipex_rotate_every_two(q_rot, k_rot, cos, sin)
-        else:
-            k_pass = key[:, :, :, self.rotary_dim:]
-            q_pass = query[:, :, :, self.rotary_dim:]
-            q_rot, k_rot = apply_rotary_pos_emb(q_rot, k_rot, cos, sin, position_ids, "gptj")
-            key = torch.cat([k_rot, k_pass], dim=-1)
-            query = torch.cat([q_rot, q_pass], dim=-1)
-    else:
-        if use_fuse_rope:
-            apply_ipex_rotate_every_two(query, key, cos, sin)
-        else:
-            query, key = apply_rotary_pos_emb(query, key, cos, sin, position_ids, "gptj")
-
-    batch_size, q_len, _ = hidden_states.size()
-
-    key = key.permute(0, 2, 1, 3).contiguous()
-    query = query.permute(0, 2, 1, 3).contiguous()
-
-    kv_seq_len = key.size(-2)
-    device = hidden_states.device
-
-    if layer_past is not None:
-        kv_seq_len += layer_past[0].size(2)
-
-    if layer_past is not None:
-        cache_k = layer_past[0]
-        cache_v = layer_past[1]
-        past_length = cache_k.size(2)
-        if cache_k.stride()[1] < kv_seq_len * cache_k.size(3):
-            new_cache_k, new_cache_v = extend_kv_cache(batch_size,
-                                                       self.num_attention_heads,
-                                                       self.head_dim,
-                                                       past_length,
-                                                       kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
-                                                       dtype=cache_v.dtype,
-                                                       device=device)
-            new_cache_k[:] = cache_k
-            new_cache_v[:] = cache_v
-            cache_k = new_cache_k
-            cache_v = new_cache_v
-        key, value = append_kv_cache(cache_k, cache_v, key, value)
-
-    elif use_cache:
-        key_cache, value_cache = init_kv_cache(batch_size,
-                                               self.num_attention_heads,
-                                               self.head_dim,
-                                               kv_seq_len,
-                                               kv_seq_len + KV_CACHE_ALLOC_BLOCK_LENGTH,
-                                               dtype=value.dtype,
-                                               device=device)
-        key_cache[:] = key
-        value_cache[:] = value
-        key = key_cache
-        value = value_cache
-
-    if use_cache is True:
-        present = (key, value)
-    else:
-        present = None
-
-    # compute self-attention: V x Softmax(QK^T)
-    attn_output, attn_weights = self._attn(query, key, value, attention_mask, head_mask)
-
-    attn_output = self._merge_heads(attn_output, self.num_attention_heads, self.head_dim)
-    attn_output = self.out_proj(attn_output)
-    attn_output = self.resid_dropout(attn_output)
-
-    outputs = (attn_output, present)
-    if output_attentions:
-        outputs += (attn_weights,)
-
-    return outputs  # a, present, (attentions)
-
-
-def gptj_block_forward(
-    self,
-    hidden_states: Optional[torch.FloatTensor],
-    layer_past: Optional[Tuple[torch.Tensor]] = None,
-    attention_mask: Optional[torch.FloatTensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    head_mask: Optional[torch.FloatTensor] = None,
-    use_cache: Optional[bool] = False,
-    rotary_emb: Optional[Tuple]=None,
-    output_attentions: Optional[bool] = False,
-) -> Union[Tuple[torch.Tensor], Optional[Tuple[torch.Tensor, Tuple[torch.FloatTensor, ...]]]]:
-    residual = hidden_states
-    hidden_states = self.ln_1(hidden_states)
-    attn_outputs = self.attn(
-        hidden_states=hidden_states,
-        layer_past=layer_past,
-        attention_mask=attention_mask,
-        position_ids=position_ids,
-        head_mask=head_mask,
-        use_cache=use_cache,
-        rotary_emb=rotary_emb,
-        output_attentions=output_attentions,
-    )
-    attn_output = attn_outputs[0]  # output_attn: a, present, (attentions)
-    outputs = attn_outputs[1:]
-
-    feed_forward_hidden_states = self.mlp(hidden_states)
-    hidden_states = attn_output + feed_forward_hidden_states + residual
-
-    if use_cache:
-        outputs = (hidden_states,) + outputs
-    else:
-        outputs = (hidden_states,) + outputs[1:]
-
-    return outputs  # hidden_states, present, (attentions)
-
-
-def create_sinusoidal_positions(num_pos: int, dim: int) -> torch.Tensor:
-    inv_freq = 1.0 / (10000 ** (torch.arange(0, dim, 2) / dim))
-    sinusoid_inp = torch.einsum("i , j -> i j",
-                                torch.arange(num_pos, dtype=torch.float), inv_freq).float()
-    return torch.cat((torch.sin(sinusoid_inp), torch.cos(sinusoid_inp)), dim=1)
-
-
-old_init = GPTJModel.__init__
-
-
-def gptj_model_new_init(self, config):
-    old_init(self, config)
-    embed_dim = config.hidden_size
-    rotary_dim = config.rotary_dim
-    pos_embd_dim = rotary_dim or embed_dim
-    max_positions = config.max_position_embeddings
-    self.embed_positions = create_sinusoidal_positions(max_positions, pos_embd_dim)
-
-
-def get_new_embed_positions(position_ids, prev_embed_positions):
-    embed_positions = prev_embed_positions
-    if embed_positions.device != position_ids.device:
-        embed_positions = embed_positions.to(position_ids.device)
-        prev_embed_positions = embed_positions
-    return embed_positions.repeat(position_ids.shape[0], 1, 1), prev_embed_positions
-
-
-def gptj_model_forward(
-    self,
-    input_ids: Optional[torch.LongTensor] = None,
-    past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None,
-    attention_mask: Optional[torch.FloatTensor] = None,
-    token_type_ids: Optional[torch.LongTensor] = None,
-    position_ids: Optional[torch.LongTensor] = None,
-    head_mask: Optional[torch.FloatTensor] = None,
-    inputs_embeds: Optional[torch.FloatTensor] = None,
-    use_cache: Optional[bool] = None,
-    output_attentions: Optional[bool] = None,
-    output_hidden_states: Optional[bool] = None,
-    return_dict: Optional[bool] = None,
-) -> Union[Tuple, BaseModelOutputWithPast]:
-    output_attentions = output_attentions if output_attentions is not None \
-        else self.config.output_attentions
-    output_hidden_states = (
-        output_hidden_states if output_hidden_states is not None
-        else self.config.output_hidden_states
-    )
-    use_cache = use_cache if use_cache is not None else self.config.use_cache
-    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
-
-    if input_ids is not None and inputs_embeds is not None:
-        invalidInputError(False,
-                          "You cannot specify both input_ids and inputs_embeds at the same time")
-    elif input_ids is not None:
-        self.warn_if_padding_and_no_attention_mask(input_ids, attention_mask)
-        input_shape = input_ids.size()
-        input_ids = input_ids.view(-1, input_shape[-1])
-        batch_size = input_ids.shape[0]
-    elif inputs_embeds is not None:
-        input_shape = inputs_embeds.size()[:-1]
-        batch_size = inputs_embeds.shape[0]
-    else:
-        invalidInputError(False, "You have to specify either input_ids or inputs_embeds")
-
-    device = input_ids.device if input_ids is not None else inputs_embeds.device
-
-    if token_type_ids is not None:
-        token_type_ids = token_type_ids.view(-1, input_shape[-1])
-
-    if past_key_values is None:
-        past_length = 0
-        past_key_values = tuple([None] * len(self.h))
-    else:
-        past_length = past_key_values[0][0].size(-2)
-
-    if position_ids is None:
-        position_ids = torch.arange(past_length, input_shape[-1] + past_length,
-                                    dtype=torch.long, device=device)
-        position_ids = position_ids.unsqueeze(0)
-
-    # Attention mask.
-    if attention_mask is not None:
-        if batch_size <= 0:
-            invalidInputError(False, "batch_size has to be defined and > 0")
-        attention_mask = attention_mask.view(batch_size, -1)
-        # We create a 3D attention mask from a 2D tensor mask.
-        # Sizes are [batch_size, 1, 1, to_seq_length]
-        # So we can broadcast to [batch_size, num_heads, from_seq_length, to_seq_length]
-        # this attention mask is more simple than the triangular masking of causal attention
-        # used in OpenAI GPT, we just need to prepare the broadcast dimension here.
-        attention_mask = attention_mask[:, None, None, :]
-
-        # Since attention_mask is 1.0 for positions we want to attend and 0.0 for
-        # masked positions, this operation will create a tensor which is 0.0 for
-        # positions we want to attend and the dtype's smallest value for masked positions.
-        # Since we are adding it to the raw scores before the softmax, this is
-        # effectively the same as removing these entirely.
-        attention_mask = attention_mask.to(dtype=self.dtype)  # fp16 compatibility
-        attention_mask = (1.0 - attention_mask) * torch.finfo(self.dtype).min
-
-    # Prepare head mask if needed
-    # 1.0 in head_mask indicate we keep the head
-    # attention_probs has shape bsz x num_attention_heads x N x N
-    # head_mask has shape n_layer x batch x num_attention_heads x N x N
-    head_mask = self.get_head_mask(head_mask, self.config.n_layer)
-
-    if inputs_embeds is None:
-        inputs_embeds = self.wte(input_ids)
-
-    hidden_states = inputs_embeds
-
-    if token_type_ids is not None:
-        token_type_embeds = self.wte(token_type_ids)
-        hidden_states = hidden_states + token_type_embeds
-
-    hidden_states = self.drop(hidden_states)
-
-    output_shape = (-1,) + input_shape[1:] + (hidden_states.size(-1),)
-
-    if self.gradient_checkpointing and self.training:
-        if use_cache:
-            logger.warning_once(
-                "`use_cache=True` is incompatible with gradient checkpointing."
-                "Setting `use_cache=False`..."
-            )
-            use_cache = False
-
-    presents = () if use_cache else None
-    all_self_attentions = () if output_attentions else None
-    all_hidden_states = () if output_hidden_states else None
-
-    # Repeat cos sin here, call only once for each token.
-    # If put this to attension forward, it will generate too many times.
-    if is_torch_fx_proxy(position_ids) or torch.jit.is_tracing():
-        # The logic to conditionally copy to GPU could not be traced, so we do this
-        # every time in the torch.fx case
-        embed_positions = get_embed_positions(self.embed_positions, position_ids)
-    else:
-        embed_positions, self.embed_positions = get_new_embed_positions(position_ids,
-                                                                        self.embed_positions)
-
-    repeated_position_ids = position_ids.unsqueeze(-1).repeat(1, 1, embed_positions.shape[-1])
-    sincos = torch.gather(embed_positions, 1, repeated_position_ids)
-    sin, cos = torch.split(sincos, sincos.shape[-1] // 2, dim=-1)
-    sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
-    cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
-
-    for i, (block, layer_past) in enumerate(zip(self.h, past_key_values)):
-        # Model parallel
-        if self.model_parallel:
-            torch.cuda.set_device(hidden_states.device)
-            # Ensure layer_past is on same device as hidden_states (might not be correct)
-            if layer_past is not None:
-                layer_past = tuple(past_state.to(hidden_states.device) for past_state in layer_past)
-            # Ensure that attention_mask is always on the same device as hidden_states
-            if attention_mask is not None:
-                attention_mask = attention_mask.to(hidden_states.device)
-            if isinstance(head_mask, torch.Tensor):
-                head_mask = head_mask.to(hidden_states.device)
-        if output_hidden_states:
-            all_hidden_states = all_hidden_states + (hidden_states,)
-
-        if self.gradient_checkpointing and self.training:
-            outputs = self._gradient_checkpointing_func(
-                block.__call__,
-                hidden_states,
-                None,
-                attention_mask,
-                position_ids,
-                head_mask[i],
-                use_cache,
-                output_attentions,
-            )
-        else:
-            outputs = block(
-                hidden_states=hidden_states,
-                layer_past=layer_past,
-                attention_mask=attention_mask,
-                position_ids=position_ids,
-                head_mask=head_mask[i],
-                use_cache=use_cache,
-                rotary_emb=(sin, cos),
-                output_attentions=output_attentions,
-            )
-
-        hidden_states = outputs[0]
-        if use_cache is True:
-            presents = presents + (outputs[1],)
-
-        if output_attentions:
-            all_self_attentions = all_self_attentions + (outputs[2 if use_cache else 1],)
-
-        # Model Parallel: If it's the last layer for that device, put things on the next device
-        if self.model_parallel:
-            for k, v in self.device_map.items():
-                if i == v[-1] and "cuda:" + str(k) != self.last_device:
-                    hidden_states = hidden_states.to("cuda:" + str(k + 1))
-
-    hidden_states = self.ln_f(hidden_states)
-
-    hidden_states = hidden_states.view(output_shape)
-    # Add last hidden state
-    if output_hidden_states:
-        all_hidden_states = all_hidden_states + (hidden_states,)
-
-    if not return_dict:
-        return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions]
-                     if v is not None)
-
-    return BaseModelOutputWithPast(
-        last_hidden_state=hidden_states,
-        past_key_values=presents,
-        hidden_states=all_hidden_states,
-        attentions=all_self_attentions,
-    )