ipex-llm 2.3.0b20250603__py3-none-win_amd64.whl → 2.3.0b20250605__py3-none-win_amd64.whl

ipex_llm/ggml/quantize.py

@@ -54,6 +54,9 @@ ggml_tensor_qtype = {"sym_int4": 2,   # q4_0 in ggml
                      "sym_int8_rtn": 32,
                      "asym_int4_rtn": 33,
                      "woq_int4": 34,
+                     "torch_fp8_e5m2": 35,
+                     "torch_fp8": 35,
+                     "torch_fp8_e4m3": 36
                      }
 
 # mixed precison from llama.cpp

ipex_llm/transformers/low_bit_linear.py

@@ -86,6 +86,8 @@ 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"]
+TORCH_FP8E5 = ggml_tensor_qtype["torch_fp8_e5m2"]
+TORCH_FP8E4 = ggml_tensor_qtype["torch_fp8_e4m3"]
 RTN_DTYPE = {
     SYM_INT4_RTN: torch.uint8,
     ASYM_INT4_RTN: torch.uint8,
@@ -106,39 +108,44 @@ def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
                        imatrix: torch.Tensor=None,
                        in_features: int=None,
                        enable_scale_search: bool=False):
-    QK = ggml.ggml_qk_size(qtype)
-    block_size_in_bytes = ggml.ggml_type_size(qtype)
-
-    invalidInputError(tensor.dtype == torch.float,
-                      "Input tensor must be float32")
-    src = tensor.data.data_ptr()
-    src = ctypes.cast(src, ctypes.POINTER(ctypes.c_float))
-    n = tensor.numel()  # all elements
-    k = tensor.shape[-1]
-    invalidInputError(k % QK == 0,
-                      f"Last dim of input tensor must be multiple of {QK}")
-
-    dst_size = (n // QK) * block_size_in_bytes
-    if qtype in [SYM_INT8_RTN, SYM_INT4_RTN, ASYM_INT4_RTN]:
-        dst_tensor = torch.empty(dst_size, dtype=RTN_DTYPE[qtype],
-                                 device=device)
-        dst_tensor = dst_tensor.reshape(tensor.shape[0], tensor.shape[-1] // QK)
-        if qtype == ASYM_INT4_RTN:
-            scale = torch.empty((n // k) * 2, dtype=torch.float32,
-                                device=device)
-        else:
-            scale = torch.empty(n // k, dtype=torch.float32,
-                                device=device)
-    elif qtype == NF4:
-        # Deepspeed zero3 requires unified dtype,
-        # thus here uses bfloat16 consistent to other layers
-        # dst_size above is computed based on uint8, and for bfloat16,
-        # buffer size should be half
-        dst_tensor = torch.empty(dst_size // 2, dtype=torch.bfloat16,
-                                 device=device)
+    if qtype in [TORCH_FP8E5, TORCH_FP8E4]:
+        fp8_dtype = torch.float8_e5m2 if qtype == TORCH_FP8E5 else torch.float8_e4m3fn
+        dst_tensor = torch.empty(tensor.shape, device=device, dtype=fp8_dtype)
+        scale = torch.zeros(1, device=device, dtype=torch.float32)
     else:
-        dst_tensor = torch.empty(dst_size, dtype=torch.uint8,
-                                 device=device)
+        QK = ggml.ggml_qk_size(qtype)
+        block_size_in_bytes = ggml.ggml_type_size(qtype)
+
+        invalidInputError(tensor.dtype == torch.float,
+                          "Input tensor must be float32")
+        src = tensor.data.data_ptr()
+        src = ctypes.cast(src, ctypes.POINTER(ctypes.c_float))
+        n = tensor.numel()  # all elements
+        k = tensor.shape[-1]
+        invalidInputError(k % QK == 0,
+                          f"Last dim of input tensor must be multiple of {QK}")
+
+        dst_size = (n // QK) * block_size_in_bytes
+        if qtype in [SYM_INT8_RTN, SYM_INT4_RTN, ASYM_INT4_RTN]:
+            dst_tensor = torch.empty(dst_size, dtype=RTN_DTYPE[qtype],
+                                     device=device)
+            dst_tensor = dst_tensor.reshape(tensor.shape[0], tensor.shape[-1] // QK)
+            if qtype == ASYM_INT4_RTN:
+                scale = torch.empty((n // k) * 2, dtype=torch.float32,
+                                    device=device)
+            else:
+                scale = torch.empty(n // k, dtype=torch.float32,
+                                    device=device)
+        elif qtype == NF4:
+            # Deepspeed zero3 requires unified dtype,
+            # thus here uses bfloat16 consistent to other layers
+            # dst_size above is computed based on uint8, and for bfloat16,
+            # buffer size should be half
+            dst_tensor = torch.empty(dst_size // 2, dtype=torch.bfloat16,
+                                     device=device)
+        else:
+            dst_tensor = torch.empty(dst_size, dtype=torch.uint8,
+                                     device=device)
 
     if not convert_shape_only and device != 'meta':
         dst = ctypes.c_void_p(dst_tensor.data.data_ptr())
@@ -158,6 +165,17 @@ def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
                                                                enable_scale_search,
                                                                imatrix)
                 return dst_tensor, scale.type(torch.float16)
+            elif qtype in [TORCH_FP8E5, TORCH_FP8E4]:
+                import xe_linear
+                tensor_device = tensor.device
+                tensor_xpu = tensor.to("xpu")
+                dst_tensor = dst_tensor.to("xpu")
+                scale = scale.to("xpu")
+
+                xe_linear.dynamic_scaled_fp8_quant(dst_tensor, tensor_xpu, scale)
+
+                # scale = scale.to(tensor_device)
+                dst_tensor = dst_tensor.to(tensor_device)
             else:
                 ggml.ggml_quantize_tensor(src, dst, qtype, n, k, hist, enable_scale_search)
         else:
@@ -171,6 +189,8 @@ def ggml_convert_qtype(tensor: torch.Tensor, qtype: int,
                                                    hist, imatrix)
     if qtype in [SYM_INT8_RTN, SYM_INT4_RTN, ASYM_INT4_RTN]:
         return dst_tensor, scale.type(torch.float16)
+    elif qtype in [TORCH_FP8E5, TORCH_FP8E4]:
+        return dst_tensor, scale
     else:
         return dst_tensor
 
@@ -179,7 +199,7 @@ def ggml_q_format_convet_cpu2xpu(tensor: torch.Tensor, num_elem: int, qtype: int
     if qtype == NF4:
         invalidInputError(tensor.dtype == torch.bfloat16,
                           "NF4 Input tensor must be bfloat16")
-    else:
+    elif qtype not in [TORCH_FP8E5, TORCH_FP8E4]:
         invalidInputError(tensor.dtype == torch.uint8,
                           "Input tensor except NF4 must be uint8")
 
@@ -208,7 +228,7 @@ def ggml_q_format_convet_xpu2cpu(tensor: torch.Tensor, num_elem: int, qtype: int
     if qtype == NF4:
         invalidInputError(tensor.dtype == torch.bfloat16,
                           "NF4 Input tensor must be bfloat16")
-    else:
+    elif qtype not in [TORCH_FP8E5, TORCH_FP8E4]:
         invalidInputError(tensor.dtype == torch.uint8,
                           "Input tensor must be uint8")
 
@@ -319,7 +339,8 @@ class FP4Params(torch.nn.Parameter):
                 qtype=None,
                 imatrix=None,
                 in_features=None,
-                enable_scale_search=False):
+                enable_scale_search=False,
+                torch_fp8_scale=None):
         if data is None:
             data = torch.empty(0)
 
@@ -332,6 +353,7 @@ class FP4Params(torch.nn.Parameter):
         self.imatrix = imatrix
         self.in_features = in_features
         self.enable_scale_search = enable_scale_search
+        self.torch_fp8_scale = torch_fp8_scale
         return self
 
     def ggml_mse(self, w, ggml_qtype, device):
@@ -391,7 +413,11 @@ class FP4Params(torch.nn.Parameter):
                                                  imatrix=self.imatrix,
                                                  in_features=self.in_features,
                                                  enable_scale_search=self.enable_scale_search)
-                self.data = w_quantized
+                if self.qtype in [TORCH_FP8E5, TORCH_FP8E4]:
+                    self.data = w_quantized[0]
+                    self.torch_fp8_scale = w_quantized[1]
+                else:
+                    self.data = w_quantized
             self.quantized = True
             self._shape = w.shape
         return self
@@ -414,6 +440,8 @@ class FP4Params(torch.nn.Parameter):
 
     def to(self, *args, **kwargs):
         device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
+        if self.qtype in [TORCH_FP8E5, TORCH_FP8E4]:
+            dtype = None
         if (device is not None and device.type == "cpu" and self.data.device.type == "cpu"):
             return self.quantize(device.type)
         elif device is not None and device.type == "meta" and self.data.device.type == "meta":
@@ -424,6 +452,7 @@ class FP4Params(torch.nn.Parameter):
             self.data = ggml_q_format_convet_cpu2xpu(self.data,
                                                      reduce(mul, self._shape, 1),
                                                      self.qtype)
+            fp8_scale = None if self.torch_fp8_scale is None else self.torch_fp8_scale.to(device)
             new_param = FP4Params(super().to(device=device,
                                              dtype=dtype,
                                              non_blocking=non_blocking),
@@ -431,9 +460,11 @@ class FP4Params(torch.nn.Parameter):
                                   quantized=self.quantized,
                                   _shape=self._shape,
                                   qtype=self.qtype,
-                                  enable_scale_search=self.enable_scale_search)
+                                  enable_scale_search=self.enable_scale_search,
+                                  torch_fp8_scale=fp8_scale)
             return new_param
         elif (device is not None and device.type == "cpu" and self.data.device.type == "xpu"):
+            fp8_scale = None if self.torch_fp8_scale is None else self.torch_fp8_scale.to(device)
             new_param = FP4Params(super().to(device=device,
                                              dtype=dtype,
                                              non_blocking=non_blocking),
@@ -441,7 +472,8 @@ class FP4Params(torch.nn.Parameter):
                                   quantized=self.quantized,
                                   _shape=self._shape,
                                   qtype=self.qtype,
-                                  enable_scale_search=self.enable_scale_search)
+                                  enable_scale_search=self.enable_scale_search,
+                                  torch_fp8_scale=fp8_scale)
             ggml_xpu = new_param.data
             new_param.data = ggml_q_format_convet_xpu2cpu(ggml_xpu,
                                                           reduce(mul, new_param._shape, 1),
@@ -614,6 +646,7 @@ class LowBitLinear(nn.Linear):
         # Due to inconsistent training status in some models like Baichuan-7b-Chat,
         # we should check both self.training and torch.is_inference_mode_enabled().
         is_training = self.training and not torch.is_inference_mode_enabled()
+
         if is_training:
             # below logic is only for training
             autocast_dtype = get_autocast_dtype(x.device.type)
@@ -643,6 +676,8 @@ class LowBitLinear(nn.Linear):
 
         if self.weight.device.type == "xpu":
             if is_training and x_2d.requires_grad:
+                invalidInputError(self.weight.qtype not in [TORCH_FP8E5, TORCH_FP8E4],
+                                  "TORCH_FP8 training is not supported.")
                 result = MatMulLowBit.apply(x_2d, self.weight, self.out_len)
             else:
                 do_empty_cache = self.low_memory_mode and x_2d.shape[0] >= 1024
@@ -654,7 +689,11 @@ class LowBitLinear(nn.Linear):
                 else:
                     w = self.weight.data
 
-                if use_batch_forward(x_2d, self.weight.qtype, self.out_len) and \
+                if self.weight.qtype in [TORCH_FP8E5, TORCH_FP8E4]:
+                    import xe_linear
+                    result = xe_linear.run_linear_fp8(x_2d, w, self.bias,
+                                                      self.weight.torch_fp8_scale)
+                elif use_batch_forward(x_2d, self.weight.qtype, self.out_len) and \
                         (x_2d.dtype == torch.half or self.conver_to_half):
                     import xe_batch
                     result = xe_batch.batch_forward(x_2d, w, self.qtype)
@@ -682,13 +721,13 @@ class LowBitLinear(nn.Linear):
                 else:
                     invalidInputError(False, "mp_group is not None, but no supported backend found")
 
-            if self.bias is not None:
+            if self.bias is not None and self.weight.qtype not in [TORCH_FP8E5, TORCH_FP8E4]:
                 result += self.bias
         else:
             # CPU logic
             # todo may need to set a different number on different platforms
-            invalidInputError(self.qtype != NF3 and self.qtype != NF4 and self.qtype != FP8E4
-                              and self.qtype != FP4 and self.qtype != FP8E5,
+            invalidInputError(self.qtype not in [NF3, NF4, FP8E4, FP4, FP8E5,
+                                                 TORCH_FP8E5, TORCH_FP8E4],
                               "NF3, NF4, FP4 and FP8 quantization are currently not"
                               " supported on CPU")
             if self.training and x.requires_grad:

ipex_llm/vllm/xpu/model_convert.py

@@ -129,9 +129,15 @@ def get_load_function(low_bit):
                 if "glm-4v" in self.vllm_config.model_config.model.lower() and \
                         low_bit in ("sym_int4", "woq_int4"):
                     modules = ["dense_4h_to_h"]
+                if "phi4mm" in self.vllm_config.model_config.hf_config.model_type:
+                    modules = ["vision_encoder", "embed_tokens_extend"]
                 if low_bit == "fp16":
                     # to fix qwen2.5-vl and glm-4v
-                    modules = ["vision", "visual"]
+                    if modules is None:
+                        modules = ["vision", "visual"]
+                    else:
+                        modules.append("vision")
+                        modules.append("visual")
                 optimize_model(self.model,
                                low_bit=low_bit,
                                torch_dtype=self.vllm_config.model_config.dtype,

{ipex_llm-2.3.0b20250603.dist-info → ipex_llm-2.3.0b20250605.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.3.0b20250603
+Version: 2.3.0b20250605
 Summary: Large Language Model Develop Toolkit
 Home-page: https://github.com/intel-analytics/ipex-llm
 Author: BigDL Authors
@@ -27,7 +27,7 @@ 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 ==2.7.0b20250603 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.7.0b20250605 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
 Requires-Dist: onednn-devel ==2025.0.1 ; (platform_system == "Windows") and extra == 'cpp'
 Requires-Dist: onednn ==2025.0.1 ; (platform_system == "Windows") and extra == 'cpp'
@@ -60,7 +60,7 @@ 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 ==2.7.0b20250603 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.7.0b20250605 ; (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'
@@ -80,9 +80,9 @@ 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 ==2.7.0b20250603 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250603 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250603 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.7.0b20250605 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250605 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250605 ; extra == 'xpu'
 Provides-Extra: xpu-2-1
 Requires-Dist: py-cpuinfo ; extra == 'xpu-2-1'
 Requires-Dist: protobuf ; extra == 'xpu-2-1'
@@ -97,9 +97,9 @@ 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 ==2.7.0b20250603 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250603 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250603 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.7.0b20250605 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250605 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250605 ; 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'
@@ -117,7 +117,7 @@ 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 ==2.7.0b20250603 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.7.0b20250605 ; extra == 'xpu-2-6'
 Requires-Dist: onednn-devel ==2025.0.1 ; extra == 'xpu-2-6'
 Requires-Dist: onednn ==2025.0.1 ; extra == 'xpu-2-6'
 Requires-Dist: dpcpp-cpp-rt ==2025.0.2 ; extra == 'xpu-2-6'
@@ -132,7 +132,7 @@ Requires-Dist: tokenizers ==0.15.2 ; extra == 'xpu-2-6-arl'
 Requires-Dist: accelerate ==0.23.0 ; extra == 'xpu-2-6-arl'
 Requires-Dist: tabulate ; extra == 'xpu-2-6-arl'
 Requires-Dist: setuptools ; extra == 'xpu-2-6-arl'
-Requires-Dist: bigdl-core-xe-all ==2.7.0b20250603 ; extra == 'xpu-2-6-arl'
+Requires-Dist: bigdl-core-xe-all ==2.7.0b20250605 ; extra == 'xpu-2-6-arl'
 Requires-Dist: onednn-devel ==2025.0.1 ; extra == 'xpu-2-6-arl'
 Requires-Dist: onednn ==2025.0.1 ; extra == 'xpu-2-6-arl'
 Requires-Dist: dpcpp-cpp-rt ==2025.0.2 ; extra == 'xpu-2-6-arl'

{ipex_llm-2.3.0b20250603.dist-info → ipex_llm-2.3.0b20250605.dist-info}/RECORD

@@ -9,7 +9,7 @@ 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=Cvk1R771rRDhSW7BRWcmb4ImY6TWDl_u9Vkdh7rYSuM,6367
+ipex_llm/ggml/quantize.py,sha256=3RQvkCvYz6rTRXAaXMK854JA6g1d8uq5JXZ7OZcj1eg,6490
 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
@@ -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=2iiONXdYbNVFXpjWnSg6ALdpUC_vXYDpDZ3kypbZkMI,36352
-ipex_llm/libs/bloom.dll,sha256=ef7cAOoa7FMrArAntBQcZ-cXdQgwkPk1Ewb8tEjNMAU,507904
-ipex_llm/libs/gptneox-api.dll,sha256=xV_cwbAdOJgbaXBeKHYPFKST91PvIubVFKPIju3dLqM,24576
-ipex_llm/libs/gptneox.dll,sha256=byWmqMhwZkeFHbSuq6uAtT3pNw3CFtJ9-EZki5qxmok,568320
-ipex_llm/libs/libbloom_avx.dll,sha256=RwPN6SlxgYSfV_IoBR0057GEAqn9qFah391CFf9_6Oc,536576
-ipex_llm/libs/libbloom_vnni.dll,sha256=CSlK6fIHt_LWUNXSDcmr07Oc3N5-I_X83v9WRXQdTMY,508416
-ipex_llm/libs/libgptneox_avx.dll,sha256=ewwZC9V6SVtqWURxBi6Ci-oMYaaC5493RNBsY2WLp8w,596992
-ipex_llm/libs/libgptneox_vnni.dll,sha256=q9isGSvnkBc0WlRj2VvseTGTSwPXZ7lKnyivodAe4DA,568832
-ipex_llm/libs/libllama_avx.dll,sha256=rA-alWTcLOP7-fHg3Ei6UxvFi4vZ2jSKTilexk0-lRU,591360
-ipex_llm/libs/libllama_vnni.dll,sha256=5TVLhryFFizH1Q-1Lt8fmwF8U2cWPHLTer1SAm-8w_k,563200
-ipex_llm/libs/libstarcoder_avx.dll,sha256=CD3s7jzP25G-pOSqZs1nE8t4ONHzHQxQflqqeKtmGBU,627712
-ipex_llm/libs/libstarcoder_vnni.dll,sha256=elHNYTRjU4Q8-IiGUKZu2AjaXpoFnCRRttSDu3fDQeE,599552
-ipex_llm/libs/llama-api.dll,sha256=pn160dmbwa7bUgA1A1rSjCNLY80GVsQlkRlPEiDe-IM,25600
-ipex_llm/libs/llama.dll,sha256=CQccqnFLy28phB3JLUwUF40FNJTC2eFaSFaam0V76bA,562688
-ipex_llm/libs/main-bloom.exe,sha256=5kZPqipliB4_a_stE1yEKD-xpMLgi0qqS-odEnR1arU,103424
-ipex_llm/libs/main-gptneox.exe,sha256=RtSnKro6TYvCxT4ibRVVqTicIQ8l1uEvdh5qLPwKrAg,98816
-ipex_llm/libs/main-llama.exe,sha256=WCAD0IsQqpB7kE4MZ-Rve0v044M1uUPEvaNwiVL8Ww8,99840
-ipex_llm/libs/main-starcoder.exe,sha256=klpzDUV0n7AKrVdMg3wvkOb5BNvZBb7ubU0FBLioYzA,157696
-ipex_llm/libs/pipeline.dll,sha256=H04OpIP5tsRZX_pxzlTN9fOhQeCfRy3VYiorrqu9-fA,73216
-ipex_llm/libs/quantize-bloom.exe,sha256=BTCZTMYlQ1h8_r2qwGa3haYmZ43KiVp0CLjAlzi4a6E,126464
-ipex_llm/libs/quantize-bloom_vnni.exe,sha256=eWeX0oiA7AeEh0uPEye-dHxutkp5bJvY78Kcc9DC0Mg,128000
-ipex_llm/libs/quantize-gptneox.exe,sha256=lZaClPPwAzgZ2CG1BO3nJQgLAfiNFVGEiHAVfLTgkIs,104448
-ipex_llm/libs/quantize-gptneox_vnni.exe,sha256=ZV2tctKKCtlX-SEfLzpVIKJPz-3xXRPX2PEpR8Xab08,104960
-ipex_llm/libs/quantize-llama.exe,sha256=pYHtPC90SMF6667uN4291-5PMCXI5fHqhaG6AAYaw8g,110080
-ipex_llm/libs/quantize-llama_vnni.exe,sha256=pMlorWCDsD-rzeewNAH-P7NifwW7Fcu-grPJwvqw1Yo,110592
-ipex_llm/libs/quantize-starcoder.exe,sha256=1PZhBJcP3fzeGeZ9NXV53UkpWO287HgE5zaNK4zfuEM,127488
-ipex_llm/libs/quantize-starcoder_vnni.exe,sha256=VX7MpJ--Pn-8BbFLv-C773eNXtzYQsXlg5KZVZxcsaE,128512
-ipex_llm/libs/starcoder-api.dll,sha256=0vCKcVFlGiBtDsubvgOkWHYNS57oMIj-3Zp5_ZThsvo,21504
-ipex_llm/libs/starcoder.dll,sha256=CT2jIqGrkRHQwo7WNBme0TNTBjvpJl518jlKVe68wpg,599040
+ipex_llm/libs/bloom-api.dll,sha256=C2Gy5HeAYJzx9slLgEXvTEtYSbnF0NecHsyKCKR8to4,36352
+ipex_llm/libs/bloom.dll,sha256=YYDzX4Z5hnTQyZjFXdzZoFIkplvpxwT56PrXiQnNZcE,507904
+ipex_llm/libs/gptneox-api.dll,sha256=AeWoztSA6tms5sqE4daiExLgBwzC9pKDW36U8Wy38a4,24576
+ipex_llm/libs/gptneox.dll,sha256=VihlfvAOtf32NqVl59LsIioACyt0q5YNXQMsGfez92o,568320
+ipex_llm/libs/libbloom_avx.dll,sha256=oYoPfOZ8qtCcelb9IUPEdK958_EjLqULMn0UvH9p0UM,536576
+ipex_llm/libs/libbloom_vnni.dll,sha256=RSGJVyxMpNWML4evaCVoPQFijungzIAj67FPPaJKYKQ,508416
+ipex_llm/libs/libgptneox_avx.dll,sha256=ZLH-nd2MEokE52jTNjNIGzQ6mSVlp4dRnkyXTNT8VPk,596992
+ipex_llm/libs/libgptneox_vnni.dll,sha256=j4YHCGvcyR78WZ70F9EsR9QWhK5ztOI6IgaPlo1GXI8,568832
+ipex_llm/libs/libllama_avx.dll,sha256=2jGixTIolwhmc_7H1Apm2Hwoq_yyHE4emejSFVoUq78,591360
+ipex_llm/libs/libllama_vnni.dll,sha256=8LAx0h40W4F4caV6ADFgQRc4tVfnLLPC6owjg6lZLlY,563200
+ipex_llm/libs/libstarcoder_avx.dll,sha256=bOyDjdRFnpPFmoL7bEaqE1nE_m3J_HtUJkAO5vZPM_I,627712
+ipex_llm/libs/libstarcoder_vnni.dll,sha256=CHWEQ-ApRnHc1LovlXwjO63UzfhRKCnTrUOhHxmY65o,599552
+ipex_llm/libs/llama-api.dll,sha256=Hd_JfzQjtNfd0fnnNFLO-UHo-exlFxMoxW2NgV7y1co,25600
+ipex_llm/libs/llama.dll,sha256=mn2QSGZ6IY6Z5G_m-qNO_wrwOy3eQJ1AlQ4roNpu3PE,562688
+ipex_llm/libs/main-bloom.exe,sha256=cHNHHlKQEzrHgm0QpYqjwhuBFZ6huMjbdRurTJtRtB8,103424
+ipex_llm/libs/main-gptneox.exe,sha256=MpiUlvrfmLtWbcBfEmfisU8T5FbxXD0Uk2acx7yACig,98816
+ipex_llm/libs/main-llama.exe,sha256=nqBLLBVjNnZCh72KL3XmmS4tTKiKuiTFB-vxPSFURpE,99840
+ipex_llm/libs/main-starcoder.exe,sha256=aDOKY_AhaqAhAgMJXmbAF1aIv6Pow6-mnW65R5ojwcc,157696
+ipex_llm/libs/pipeline.dll,sha256=y5x6scPUEi-8pV2SqCo6_x87WGxV_jxwRRMvLU4XoLU,73216
+ipex_llm/libs/quantize-bloom.exe,sha256=7dghzsKCkrchRMhme5PabOkuBvkOIAV191T_6Eo3FVA,126464
+ipex_llm/libs/quantize-bloom_vnni.exe,sha256=ioOn79_l1u7m1QUpvb0DWhQIVdoBdXMxAsA9mDKn_W8,128000
+ipex_llm/libs/quantize-gptneox.exe,sha256=6_wZBbgeaYie8T7LL37nh9mOCOyLc4sDi8T0oFZgFG0,104448
+ipex_llm/libs/quantize-gptneox_vnni.exe,sha256=ZJXaAOygZRYS8efiPFc2iJ42BWKeJAQ21Gyb_qYCPAM,104960
+ipex_llm/libs/quantize-llama.exe,sha256=VRiSH4HzV8laqb2EIi6DJ07XMVy4J-N8Mx2gf8wk1Ww,110080
+ipex_llm/libs/quantize-llama_vnni.exe,sha256=kW2sdD6eRXS6IkYXQkZb0fdCHrJs_voUVlqy3pz1xMs,110592
+ipex_llm/libs/quantize-starcoder.exe,sha256=HXD4f_muhQnlqzbFpeHNbvPPf0qXk3D69bPhFps6t7I,127488
+ipex_llm/libs/quantize-starcoder_vnni.exe,sha256=oCK-30YAEKF0DkwR1to8F97LzvWuPNeiEKqtfJ0yKCg,128512
+ipex_llm/libs/starcoder-api.dll,sha256=b8v58Pgz-gxyqSDo670JIZs7EadA2MymaARbmSFfYlk,21504
+ipex_llm/libs/starcoder.dll,sha256=IFRyTp4l9XfRgaB5yVik1EM2qHgpTEg3GQVNSciScXI,599040
 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
@@ -94,7 +94,7 @@ ipex_llm/transformers/kv.py,sha256=src_HcVDKFwQ1V8hdTrFQw5RIwUewM9VOR47GVTPJG4,2
 ipex_llm/transformers/lisa.py,sha256=F5WxbtXQ7RdKulj83h_2DnEIgKiKGZf7zvOmg6QBl2s,3289
 ipex_llm/transformers/loader.py,sha256=c9qfJSC6-in-mkd-iKb1igk3nHWUYS3QtyH2cOazmKc,6825
 ipex_llm/transformers/lookup.py,sha256=b6OlZ9OV10R9qeWw8mVryVpDxszkjwLkldvi7GPMJY8,19614
-ipex_llm/transformers/low_bit_linear.py,sha256=f47v3w3DUG0G65RawgiL5y9N8l_GRRz6uaCSTMga2zM,39281
+ipex_llm/transformers/low_bit_linear.py,sha256=03TMG4GZsgRPvchQC2h7eMU9IQ9XCyVcdh3Pvi7_Rew,41550
 ipex_llm/transformers/model.py,sha256=tWTzKsCz8A1P5gYEeG9KZgpxQgbP9hQ-TWAdkebA6Jg,40886
 ipex_llm/transformers/modelling_bigdl.py,sha256=7JpNVMuyq_OmtNUaMFMXdxPWZp2q0QHC02QeA-VTPOw,6709
 ipex_llm/transformers/npu_model.py,sha256=X8mdY6N9TYlxG41wmFloX44ZUjyitFzdKbhzO7TToFY,40309
@@ -253,16 +253,16 @@ ipex_llm/vllm/cpu/entrypoints/openai/cli_args.py,sha256=hB398yYtKauASRzevctScdbF
 ipex_llm/vllm/xpu/__init__.py,sha256=zBSG6nzrVF5QnpR6_f7kPhBFeowTE9gaZ7D5m98E7_w,585
 ipex_llm/vllm/xpu/ipex_llm_v1_wrapper.py,sha256=pd939vFomKIg9Qn2NO4u0OF6hPgvQpqcfJSxqBzcqhA,825
 ipex_llm/vllm/xpu/ipex_llm_wrapper.py,sha256=_CbhvBuf_KPnmLfngYKtJl5gPAHVsG2mWth3wSeaH3M,892
-ipex_llm/vllm/xpu/model_convert.py,sha256=oedafTsnysTi78PGYcjn1w5rnIBfBx4_mpZp2fF6z44,10093
+ipex_llm/vllm/xpu/model_convert.py,sha256=HZeTrQHMYfgXlz1b9KiKdAUZ57nLgpv6VhM5CkiSrUc,10416
 ipex_llm/vllm/xpu/engine/__init__.py,sha256=sOvwLx_Zj0jiRCGj9W3DgGTfcSU3hABYhgIQI7T6cxU,879
 ipex_llm/vllm/xpu/engine/engine.py,sha256=XAprw7VifjfnR915TZOaKcxe3QCFsVBgxzS8qOdn1yg,14462
 ipex_llm/vllm/xpu/entrypoints/openai/api_server.py,sha256=VlmS56hBHBZTIZ5Jhvb4TZN-h28O7uMn33hX8NiJXKk,45719
 ipex_llm/vllm/xpu/entrypoints/openai/cli_args.py,sha256=hB398yYtKauASRzevctScdbFIjiiSGMAe1bwEuIHrhY,10893
-ipex_llm-2.3.0b20250603.data/scripts/ipex-llm-init.bat,sha256=HPtCYuDYwEatq7dAwOvdfVcHYCpAVdbj75K1qh0vQek,2578
-ipex_llm-2.3.0b20250603.data/scripts/llm-chat.ps1,sha256=6qrs-hGVAV8IKh7Jx8nq_XrnZcjd7qGU5wndArM7Yag,2769
-ipex_llm-2.3.0b20250603.data/scripts/llm-cli.ps1,sha256=3qBtTLs_EjYDnM8YyCpJhzLnGCKTEGssu9UNqfkjVXs,3009
-ipex_llm-2.3.0b20250603.dist-info/METADATA,sha256=ksoQDNkxXOZiquKGwSnn-LS4DfdIbFQ9mCbCzDg4AH4,8865
-ipex_llm-2.3.0b20250603.dist-info/WHEEL,sha256=6iYPr8vTHsyDK75jr9X0V3I9wPSVmtwr_8fdATBciGk,98
-ipex_llm-2.3.0b20250603.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.3.0b20250603.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.3.0b20250603.dist-info/RECORD,,
+ipex_llm-2.3.0b20250605.data/scripts/ipex-llm-init.bat,sha256=HPtCYuDYwEatq7dAwOvdfVcHYCpAVdbj75K1qh0vQek,2578
+ipex_llm-2.3.0b20250605.data/scripts/llm-chat.ps1,sha256=6qrs-hGVAV8IKh7Jx8nq_XrnZcjd7qGU5wndArM7Yag,2769
+ipex_llm-2.3.0b20250605.data/scripts/llm-cli.ps1,sha256=3qBtTLs_EjYDnM8YyCpJhzLnGCKTEGssu9UNqfkjVXs,3009
+ipex_llm-2.3.0b20250605.dist-info/METADATA,sha256=2CommuodTS_N2xOt4lnyS-8ZP0vWAMSEBYQOElEuSjk,8865
+ipex_llm-2.3.0b20250605.dist-info/WHEEL,sha256=6iYPr8vTHsyDK75jr9X0V3I9wPSVmtwr_8fdATBciGk,98
+ipex_llm-2.3.0b20250605.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.3.0b20250605.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.3.0b20250605.dist-info/RECORD,,