ipex-llm 2.2.0b20250114__py3-none-manylinux2010_x86_64.whl → 2.2.0b20250115__py3-none-manylinux2010_x86_64.whl

ipex_llm/ggml/quantize.py

@@ -53,6 +53,7 @@ ggml_tensor_qtype = {"sym_int4": 2,   # q4_0 in ggml
                      "sym_int4_rtn": 31,
                      "sym_int8_rtn": 32,
                      "asym_int4_rtn": 33,
+                     "woq_int4": 34,
                      }
 
 # mixed precison from llama.cpp

ipex_llm/transformers/convert.py

@@ -995,8 +995,9 @@ def _optimize_pre(model, qtype=None):
         from ipex_llm.transformers.models.gemma2 import merge_qkv
         model.apply(merge_qkv)
     elif model.config.model_type == "llama":
-        from ipex_llm.transformers.models.llama import merge_qkv
+        from ipex_llm.transformers.models.llama import merge_qkv, pre_compute_inv_freq
         model.apply(merge_qkv)
+        model.apply(pre_compute_inv_freq)
     elif model.config.model_type == "mllama":
         from ipex_llm.transformers.models.mllama import merge_qkv
         model.apply(merge_qkv)

ipex_llm/transformers/low_bit_linear.py

@@ -84,6 +84,7 @@ FP6_K = ggml_tensor_qtype["fp6_k"]
 SYM_INT4_RTN = ggml_tensor_qtype["sym_int4_rtn"]
 SYM_INT8_RTN = ggml_tensor_qtype["sym_int8_rtn"]
 ASYM_INT4_RTN = ggml_tensor_qtype["asym_int4_rtn"]
+WOQ_INT4 = ggml_tensor_qtype["woq_int4"]
 RTN_DTYPE = {
     SYM_INT4_RTN: torch.uint8,
     ASYM_INT4_RTN: torch.uint8,
@@ -187,7 +188,7 @@ def ggml_q_format_convet_cpu2xpu(tensor: torch.Tensor, num_elem: int, qtype: int
     src = ctypes.c_void_p(tensor.data.data_ptr())
 
     if qtype in [SYM_INT4, ASYM_INT4, SYM_INT8, NF4, NF3, FP4, FP6, FP8E4, FP8E5,
-                 Q4_K, Q6_K, FP6_K]:
+                 Q4_K, Q6_K, FP6_K, WOQ_INT4]:
         dst_tensor = torch.empty_like(tensor)
     elif qtype == ggml_tensor_qtype["sym_int5"]:
         QK = ggml.ggml_qk_size(qtype)
@@ -213,7 +214,7 @@ def ggml_q_format_convet_xpu2cpu(tensor: torch.Tensor, num_elem: int, qtype: int
     src = ctypes.c_void_p(tensor.data.data_ptr())
 
     if qtype in [SYM_INT4, ASYM_INT4, SYM_INT8, NF4, NF3, FP4, FP6, FP8E4, FP8E5,
-                 Q4_K, Q6_K, FP6_K]:
+                 Q4_K, Q6_K, FP6_K, WOQ_INT4]:
         dst_tensor = torch.empty_like(tensor)
     elif qtype == ggml_tensor_qtype["sym_int5"]:
         QK = ggml.ggml_qk_size(ggml_tensor_qtype["asym_int5"])

ipex_llm/transformers/models/llama.py

@@ -119,6 +119,13 @@ def merge_qkv(module: torch.nn.Module):
     merge_qkv_base(module, LlamaAttention)
 
 
+def pre_compute_inv_freq(module: torch.nn.Module):
+    if module.__class__.__name__ == "LlamaLinearScalingRotaryEmbedding":
+        if hasattr(module, "scaling_factor"):
+            module.register_buffer("inv_freq_scaled", None, persistent=False)
+            module.inv_freq_scaled = module.inv_freq / module.scaling_factor
+
+
 def llama_attention_forward(
     self,
     hidden_states: torch.Tensor,
@@ -147,8 +154,12 @@ def llama_attention_forward(
         import xe_addons
         if hasattr(self, "rotary_emb"):
             # transformers < 4.46
-            xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
-                                           query_states, key_states)
+            if hasattr(self.rotary_emb, "inv_freq_scaled"):
+                xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq_scaled, position_ids,
+                                               query_states, key_states)
+            else:
+                xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
+                                               query_states, key_states)
         else:
             # transformers >= 4.46
             cos, sin = position_embeddings

ipex_llm/transformers/models/utils.py

@@ -21,7 +21,7 @@ from ipex_llm.utils.common import invalidInputError
 from ipex_llm.ggml.quantize import ggml_tensor_qtype
 from ipex_llm.transformers.utils import get_xpu_device_name
 from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4,\
-    FP6, ASYM_INT4
+    FP6, ASYM_INT4, WOQ_INT4
 
 FP8_KV_ALLOC_LENGTH = 512
 KV_CACHE_ALLOC_BLOCK_LENGTH = int(os.environ.get("KV_CACHE_ALLOC_BLOCK_LENGTH", 256))
@@ -33,7 +33,7 @@ GELU = 1
 
 def decoding_fast_path_qtype_check(proj):
     qtype = getattr(proj, "qtype", None)
-    return qtype in [SYM_INT4, FP8E5, FP4]
+    return qtype in [SYM_INT4, FP8E5, FP4, WOQ_INT4]
 
 
 def init_kv_cache(batch_size, num_heads, head_dim, current_length, max_length, dtype, device):
@@ -248,7 +248,7 @@ def mlp_fusion_check(x, qtype, training):
         return False
     if x.device.type != 'xpu':
         return False
-    if qtype not in [SYM_INT4, FP8E5, FP4, IQ2_XXS, FP6]:
+    if qtype not in [SYM_INT4, FP8E5, FP4, IQ2_XXS, FP6, WOQ_INT4]:
         return False
     if training or x.requires_grad:
         return False
@@ -263,7 +263,7 @@ def use_xmx(x: torch.Tensor, qtype: int):
     device = get_xpu_device_name(x.device)
     return (
         device in ["arc", "pvc"]
-        and qtype in [SYM_INT4, SYM_INT8, FP8E4, FP8E5]
+        and qtype in [SYM_INT4, SYM_INT8, FP8E4, FP8E5, WOQ_INT4]
         and (
             (device == "pvc" and 1 < x.size(0) <= 16)
             or

{ipex_llm-2.2.0b20250114.dist-info → ipex_llm-2.2.0b20250115.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250114
+Version: 2.2.0b20250115
 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 ==2.6.0b20250114 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250115 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
 Provides-Extra: cpp-arl
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250114 ; extra == 'cpp-arl'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250115 ; 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 ==2.6.0b20250114 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250115 ; (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 ==2.6.0b20250114 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250114 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250114 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250115 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250115 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250115 ; 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 ==2.6.0b20250114 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250114 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250114 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250115 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250115 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250115 ; 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 ==2.6.0b20250114 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.6.0b20250115 ; 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 ==2.6.0b20250114 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250114 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250114 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250115 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250115 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250115 ; 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 ==2.6.0b20250114 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250114 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250114 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250115 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250115 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250115 ; 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 ==2.6.0b20250114 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250114 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250114 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250115 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250115 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250115 ; 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-2.2.0b20250114.dist-info → ipex_llm-2.2.0b20250115.dist-info}/RECORD

@@ -11,7 +11,7 @@ ipex_llm/cli/prompts/chat-with-llm.txt,sha256=PpSyd4FQQd-T7ptfXL9jZp7dgstevu1fsx
 ipex_llm/ggml/__init__.py,sha256=FzapYBUiTdZf0LzlN9hfJI-HE1OTi_2dzaYELJ9Mw8s,1272
 ipex_llm/ggml/convert.py,sha256=xfWH1E_hivbsxVo8h00STjH1Rlu9-dZQkCLLeIs1TWA,5286
 ipex_llm/ggml/convert_model.py,sha256=t-tGK9w8ZRi9dlDLTutput3ZBKj3ji94WUJi2KG8hkA,5955
-ipex_llm/ggml/quantize.py,sha256=TOoSe1w7tqAoV_fuqQhXIYb1S02A6udeud6IX0kLSic,6330
+ipex_llm/ggml/quantize.py,sha256=Cvk1R771rRDhSW7BRWcmb4ImY6TWDl_u9Vkdh7rYSuM,6367
 ipex_llm/ggml/model/__init__.py,sha256=T-EbRT6GJ_8RCu-iLmSzcftOimXSPQf2d5X72AUAy2Y,874
 ipex_llm/ggml/model/bloom/__init__.py,sha256=291QHI19FMw7Z1oaKBAf2YJ0M51iYqWC4IT1ejI-OGg,900
 ipex_llm/ggml/model/bloom/bloom.py,sha256=fUxgZd_Uc4RXaMC_naYdjekwNprM1TpURmQ8VbocShc,17975
@@ -44,39 +44,39 @@ ipex_llm/langchain/vllm/__init__.py,sha256=T-EbRT6GJ_8RCu-iLmSzcftOimXSPQf2d5X72
 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/libbloom-api.so,sha256=ylQsq2MM8zLHgypOPNGvQnUN2xYraaVbzimhJTKGF9s,33576
-ipex_llm/libs/libbloom_amx.so,sha256=vyxQmiMnVKu3dWXwVsiWwN69LK1WBf7DT8yc4ujxjCE,528856
-ipex_llm/libs/libbloom_avx.so,sha256=rfBudWvF7TITnEqDoDo9Cjwkr5_aWGQ3WuXA8104-jI,536152
-ipex_llm/libs/libbloom_avx2.so,sha256=dqqgS5Pic5-sdkbhjDmOqtNGPYBLzvUX-jeL6lgmLHM,519680
-ipex_llm/libs/libbloom_avx512.so,sha256=oUNIE-QA4axkWfomxfCDSIVVk2wl8YrslgQMeEfq5iY,515584
-ipex_llm/libs/libbloom_avxvnni.so,sha256=HXAjpFmWCwQFuN3-aodDskNRi4t4vLeaYpKLubViAXo,519680
+ipex_llm/libs/libbloom_amx.so,sha256=I8qcR4PdcuJsWBkJIqGrjD2x5skUQycW5LlQMyieWsU,529080
+ipex_llm/libs/libbloom_avx.so,sha256=WkM086ptULkBaPVL8aCmmyhoEspnvhtN91FMSWMMso0,540456
+ipex_llm/libs/libbloom_avx2.so,sha256=kmfEtwj8pAupZeitNbd5cCQyKmqMF--lZUxS6JSq1dg,523984
+ipex_llm/libs/libbloom_avx512.so,sha256=HM1W_Lqr-azWdGQ6-UClKFJq_1ljO68n9Z7yYh4v5ck,523984
+ipex_llm/libs/libbloom_avxvnni.so,sha256=-wgW5IVAGGOFsY5oCUO-BeROM9KdVz7KECe1jNijRwM,523984
 ipex_llm/libs/libgptneox-api.so,sha256=X6glOWTpK-bw9hmP2gGLcQpwqThWwyDuPqxdNigbp4w,28240
-ipex_llm/libs/libgptneox_amx.so,sha256=uc5spwmksH_MeZaw46jQiHci2D1wuB0YXpJ7GC9HEmU,581952
-ipex_llm/libs/libgptneox_avx.so,sha256=pA0H2xKMXxww1vDzvxedyTKBQNVAU167zZnynmFSSak,589208
-ipex_llm/libs/libgptneox_avx2.so,sha256=HklIdD-26xokiLRFZ0-rTtV5VKxUx7Tomn6zN629-Og,576832
-ipex_llm/libs/libgptneox_avx512.so,sha256=pINXzvghq0B2IrEtMudHJiug2E0swgjFvqXRbtCNKUM,572736
-ipex_llm/libs/libgptneox_avxvnni.so,sha256=NNlvVW0FJG4tynJAsvtpwStw_Ez_aE2Z8JAAQtG-BCA,572736
+ipex_llm/libs/libgptneox_amx.so,sha256=mX4FM5a1cYChwKfdQfP4yjdyRRiJb2_s5vMf6cHfjU8,586272
+ipex_llm/libs/libgptneox_avx.so,sha256=DcHCQDbF1C1th3YzZUBwXyxvBZ5PtfiGx-0Al4zZX2A,589416
+ipex_llm/libs/libgptneox_avx2.so,sha256=o1lBhyFHhkXw8H87BR1aB7LYP051Yl6wCSTrjbNG4js,577040
+ipex_llm/libs/libgptneox_avx512.so,sha256=x-E7BtKcS2YavtVX1qK15SjUSPtPbMbgE0NRk3g7h48,577040
+ipex_llm/libs/libgptneox_avxvnni.so,sha256=ca3ewcSSkllY3_Z24JMLzAZqDFnbEmvukp32nXprN9w,572944
 ipex_llm/libs/libjemalloc.so,sha256=gBgVkr0vI-5w7KdT38ez485J6RvRlZE1-Xe3JnLe4h8,5460184
 ipex_llm/libs/libllama-api.so,sha256=ffrgtLhkdGf2dv3jrTf1J06pwLQMYU57VXcX3LuU400,32688
-ipex_llm/libs/libllama_amx.so,sha256=vjViSXQJqy6zPG_fGe0YWrD00xb1UgLDBCzWJRaDvy8,581424
-ipex_llm/libs/libllama_avx.so,sha256=pySdsGHAg4ANLigbuy4HibiIPFUc6LZKkSxwLzUQlY0,592776
-ipex_llm/libs/libllama_avx2.so,sha256=Pk13Hy9SmX7EOow7mvTT_hz0FViK0qPKsM29orpiEtg,576304
-ipex_llm/libs/libllama_avx512.so,sha256=mXGgBKVEsukWVAs09fHjvRRURtYaH1wIvEDANjgAI84,576304
-ipex_llm/libs/libllama_avxvnni.so,sha256=TONV4V-sJMwU2QEGUyozTAzDOhLc40eR8Ck-F26hbrU,576304
+ipex_llm/libs/libllama_amx.so,sha256=c7A6tzGQGvoXfJHBGIaJP20uOpQHQ2Jd2BA26RcbcxA,585744
+ipex_llm/libs/libllama_avx.so,sha256=mWHbWCYH0Ey54v_YuMxLusx9Wze9F-B1bRa7iU8EvSA,592984
+ipex_llm/libs/libllama_avx2.so,sha256=b_AYdIE30vMwGtn5-pb4VLrKT9gInpxZEXv6b5tFvzg,580608
+ipex_llm/libs/libllama_avx512.so,sha256=OUQRjb0rb6MOzV1ZLs_FGscqlVtK8BfQ3b1jtxVmzs0,576512
+ipex_llm/libs/libllama_avxvnni.so,sha256=hesp1bhngQr4aEd3v48MkcUYYmzdi9gTByG6D-4kg1c,576512
 ipex_llm/libs/libstarcoder-api.so,sha256=C126ZZ0Yr80mpQx77U4egdbijpTJEa6gvzK5ASP6CcY,18368
-ipex_llm/libs/libstarcoder_amx.so,sha256=msBmnqcB3sqf2IYPBFU9A0M3mkTYeMrCi52cxSloW_c,721424
-ipex_llm/libs/libstarcoder_avx.so,sha256=otmtdNDgTa9iQL7GiJ9J3ruhqS2aD1LrtoKTWAGcPfc,728720
-ipex_llm/libs/libstarcoder_avx2.so,sha256=FAjKK0eGEdaGWIJI8i3Io1KsLuAaYR-GzQRmpFsenR4,716344
-ipex_llm/libs/libstarcoder_avx512.so,sha256=la8_1GyZv2dPMM0-GeY3MXPJiEVyC6WUghNKAd3s648,716344
-ipex_llm/libs/libstarcoder_avxvnni.so,sha256=3AdgMHBpTdLYdPr_utmNp0VnRyXeq0N8ri3O-QrtCYQ,716344
+ipex_llm/libs/libstarcoder_amx.so,sha256=dai_3SYzYVhvdbFH2nFPWuMF0IYFTmMhpCm2bNmroNM,721608
+ipex_llm/libs/libstarcoder_avx.so,sha256=Q5jV1YfcQP64ZRs8qfH2wuzqP6gwXS-x3ZLfhh-CXJI,733040
+ipex_llm/libs/libstarcoder_avx2.so,sha256=6TayQFLLVX8hAQToBC7bydHpV2uWBBcDj-uA-izYL3c,716568
+ipex_llm/libs/libstarcoder_avx512.so,sha256=xSAlfvoQS3Yg9ZMI_1G8i0JdzPlY0Rcj2gBY1SM7nw8,716568
+ipex_llm/libs/libstarcoder_avxvnni.so,sha256=jj4yGnTMCMm4ZwPuri-23o8YXypnwD3w-v6ch5iwf8s,716568
 ipex_llm/libs/libtcmalloc.so,sha256=sGmZ-oxCTzm9O7gBFY_Q5tInlih1tBIFfkOhvrQu2zU,393904
 ipex_llm/libs/main-bloom,sha256=TeVWpZ60CdFW5yN3ejZ6bHCnWxqG5tZo9pXw13ncTPc,130416
 ipex_llm/libs/main-gptneox,sha256=3Y0viffGzc0Gh3XTXHY4tvt6M4nkjHAbIJ1oDFqFqm0,116208
 ipex_llm/libs/main-llama,sha256=5Ml3WWow-tMXCbUQXqtXTnKqfbUhzMNHDo4v0DVA5ks,116560
 ipex_llm/libs/main-starcoder,sha256=ykjZ9Me3Av2FirUamVMl5rrBloLxdW_GAYQwGab5FXU,263336
-ipex_llm/libs/quantize-bloom,sha256=d7Fg26B1J8QUvc7zqhCQ8kmpEUdG2gQiHezp0sq0Bu4,560656
-ipex_llm/libs/quantize-gptneox,sha256=pzuFlKbdQBJ4Tv3CH35cfa2790h5m1nxYrZF2AATxnA,545432
-ipex_llm/libs/quantize-llama,sha256=VmORvJULYcjXkExOXTByXuU-OS-rtd1XyDjibXJNFd8,549624
-ipex_llm/libs/quantize-starcoder,sha256=D0AzAIzOeMXqxyZQ3XPh8k3pEINfBEw1gG3UBC7WUFQ,583656
+ipex_llm/libs/quantize-bloom,sha256=OuY0fpcXhnBh2BrK59Wo4P04WuL4tWvXf9BAl53r6a0,560848
+ipex_llm/libs/quantize-gptneox,sha256=Or0NOV18x0g9cK1b0K3NCgvkV30p51_jQQFi0Yr6wVE,545624
+ipex_llm/libs/quantize-llama,sha256=Dhqkh8C6RYAYSkpiB3P9BtKBZDoGXuYZyNgvJY71sLo,553912
+ipex_llm/libs/quantize-starcoder,sha256=sz6_UFEPoOQkJPPjv8fN2VzhTBGhLYbQZyuijvumEA4,587936
 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
@@ -94,14 +94,14 @@ 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=pJHs6GZXHIObVE4BUCuej-6BKBZZg9pYWKPrkhWSfB4,1192
-ipex_llm/transformers/convert.py,sha256=i2IOmDnQBKNtcfUL95l8w4rNBMiU4SqX_9uz_LtPHMI,98086
+ipex_llm/transformers/convert.py,sha256=_aTdg1c15aUXJD4ktgcaDcqrjyEIlsoPLbMT5zvXr6U,98150
 ipex_llm/transformers/convert_ipex.py,sha256=_nSnUTQy-yfkKaqGdqnBdWztZf3NGmnbZ0TKaDrF4X4,14617
 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=AwjV5RpI2t2bedlv7ZhLm8cfd-QJZm5hny-XyjIvdnk,6876
 ipex_llm/transformers/lookup.py,sha256=b6OlZ9OV10R9qeWw8mVryVpDxszkjwLkldvi7GPMJY8,19614
-ipex_llm/transformers/low_bit_linear.py,sha256=1diyN_5OTViAmi8CRIUh5j1s5Z9j3Umcz2MEiideV4U,40863
+ipex_llm/transformers/low_bit_linear.py,sha256=27xrIh-0-KAyepaT9lHz3W4kkjoRRo08zVO4C8LRJkc,40924
 ipex_llm/transformers/model.py,sha256=64eRLtbFftku9kcN5xSNerbZ3jAFUjCbNlM8Ly54ygE,41079
 ipex_llm/transformers/modelling_bigdl.py,sha256=7JpNVMuyq_OmtNUaMFMXdxPWZp2q0QHC02QeA-VTPOw,6709
 ipex_llm/transformers/npu_model.py,sha256=X8ZtvZJpzz64XrSPhUYXXZmdJcbZ9X6G3Vlzw-zgN1Q,39749
@@ -161,7 +161,7 @@ ipex_llm/transformers/models/gptbigcode.py,sha256=cP1_qGWoa43R2WacAMblShjku4Qupc
 ipex_llm/transformers/models/gptneox.py,sha256=loRh1x_5S6BCeOr_s5xr-N_1SQHL3Y5IiUBAEyoMUqQ,6172
 ipex_llm/transformers/models/internlm.py,sha256=JZFrI2HXsIAfM-6pA2RO0wcXopOliC1FggLMzNzaDZ4,17404
 ipex_llm/transformers/models/internvl.py,sha256=Vx0vENIEQLX2M6P398mw5TOhpks0U8xf8rtRQvy94go,8154
-ipex_llm/transformers/models/llama.py,sha256=NzpyQve_RC9ez1W-jWPLGZ80k_S1I5Rx5saAzCsDIoI,8558
+ipex_llm/transformers/models/llama.py,sha256=rqrNjuZb_jeb9MKx0z-FSVoGx8YDBxQzPJ9ZUvYhgx0,9138
 ipex_llm/transformers/models/minicpm.py,sha256=eaPNVNrep0_xGoELhZd886ff0ceoKqB6cusdAhd52eE,10145
 ipex_llm/transformers/models/minicpm3.py,sha256=11cYl8KM2hoIJNMAOZMxiwCu6dMhup9ric_OEn8-VrQ,9363
 ipex_llm/transformers/models/minicpmv.py,sha256=PP05b5iTnrMpiseCn8iJcxKJDnfq7WqXp9Mrch0kKZ0,9876
@@ -181,7 +181,7 @@ 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=fj-XtOnR6kggnFUQTMPCOOzolkPztN06WAv8QW-XRnI,7054
 ipex_llm/transformers/models/starcoder2.py,sha256=ONKvD7JCkRM0DI-R56x28QFBJ7CjD5hOZBQ_3WfOcNk,6626
-ipex_llm/transformers/models/utils.py,sha256=WYBc26vSiy_CzV07z-eT5ts90Kko2yUmS3DDZtfGcRk,15065
+ipex_llm/transformers/models/utils.py,sha256=Rdn9T4zk6Hz8ybJp6kvlyfPwgHwdxEZ8R4zGtMeozWg,15105
 ipex_llm/transformers/models/yuan.py,sha256=JYAn_ZaSGK0NBJLEIxCACfAq084a66GFJkdd5NbpmMA,7732
 ipex_llm/transformers/npu_models/__init__.py,sha256=ulEUGLjaP48LCrVeury3UxLjXxKzRi0UpSG4bYu-7f8,585
 ipex_llm/transformers/npu_models/baichuan.py,sha256=fJtd7fBrttySghRUgfZTAdxLjsSNC-XL08HISsXigLE,4685
@@ -250,11 +250,11 @@ 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-2.2.0b20250114.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
-ipex_llm-2.2.0b20250114.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
-ipex_llm-2.2.0b20250114.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
-ipex_llm-2.2.0b20250114.dist-info/METADATA,sha256=CsE0v0mVfDVChkqM_ovKZkcfFq2zcg1yL3xthEnO7JM,12705
-ipex_llm-2.2.0b20250114.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
-ipex_llm-2.2.0b20250114.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.2.0b20250114.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.2.0b20250114.dist-info/RECORD,,
+ipex_llm-2.2.0b20250115.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
+ipex_llm-2.2.0b20250115.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
+ipex_llm-2.2.0b20250115.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
+ipex_llm-2.2.0b20250115.dist-info/METADATA,sha256=IFEK6gXdz44NE1sgaflpY0KyvKZXn9rdnuAET380aUo,12705
+ipex_llm-2.2.0b20250115.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
+ipex_llm-2.2.0b20250115.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.2.0b20250115.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.2.0b20250115.dist-info/RECORD,,