ipex-llm 2.3.0b20250415__py3-none-manylinux2010_x86_64.whl → 2.3.0b20250418__py3-none-manylinux2010_x86_64.whl

ipex_llm/transformers/models/common.py

@@ -17,6 +17,7 @@
 import math
 import torch
 from typing import List
+from ipex_llm.utils.common import invalidInputError
 
 
 def merge_linear(linears: List[torch.nn.Linear]) -> torch.nn.Linear:
@@ -303,3 +304,56 @@ def scaled_dot_product_attention(query: torch.Tensor, key: torch.Tensor,
             )
         attn_output = attn_output.to(dtype)    # workaround ipex 2.1's bug
         return attn_output
+
+
+def linear_forward(x: torch.Tensor, weight: torch.Tensor, qtype: int, out_features: int):
+    if weight.device.type == "xpu":
+        new_shape = x.shape[:-1] + (out_features,)
+        x = x.to(weight.device, dtype=torch.float16)
+        x_2d = x.contiguous().view(-1, x.shape[-1])
+        import xe_linear
+        x = xe_linear.forward_new(x_2d, weight, qtype, out_features)
+        x = x.view(new_shape)
+        return x
+    else:
+        invalidInputError(False,
+                          "Unsupported device type: only support weight on xpu device.")
+
+
+def quantize_linear(weight: torch.Tensor, in_features: int, precision: str):
+    from ipex_llm.transformers.low_bit_linear import FP4Params
+    from ipex_llm.ggml.quantize import ggml_tensor_qtype
+
+    invalidInputError(precision in ggml_tensor_qtype.keys(),
+                      f"{precision} is not supported, "
+                      f"only {ggml_tensor_qtype.keys()} are supported now.")
+    qtype = ggml_tensor_qtype[precision]
+    paramsLowBit = FP4Params(data=weight.data,
+                             requires_grad=False,
+                             quantized=False,
+                             _shape=None,
+                             convert_shape_only=False,
+                             qtype=qtype,
+                             in_features=in_features,
+                             enable_scale_search=False).to("cpu")
+    return paramsLowBit, qtype
+
+
+def moe_group_topk(scores: torch.Tensor, e_score_correction_bias: torch.Tensor,
+                   n_group: int, topk_group: int, top_k: int, norm_topk_prob: float,
+                   routed_scaling_factor: float):
+    import xe_addons
+    topk_idx, topk_weight = xe_addons.moe_group_topk(
+        scores, e_score_correction_bias,
+        n_group, 2, topk_group, top_k,
+        top_k > 1 and norm_topk_prob, 1e-20, routed_scaling_factor
+    )
+    return topk_idx, topk_weight
+
+
+def rotary_two_with_cache_inplaced(query_states: torch.Tensor, key_states: torch.Tensor,
+                                   cos: torch.Tensor, sin: torch.Tensor,
+                                   half_layout: bool):
+    import xe_addons
+    xe_addons.rotary_two_with_cache_inplaced(query_states, key_states,
+                                             cos, sin, half_layout)

ipex_llm/transformers/models/deepseek.py

@@ -228,11 +228,11 @@ def deepseek_attention_forward(
             [k_nope, k_pe.expand([-1, self.num_heads, -1, -1])],
             dim=-1
         )
-        import xe_addons
         cos, sin = position_embeddings
-        xe_addons.rotary_two_with_cache_inplaced(query_states[:, :, :, self.qk_nope_head_dim:],
-                                                 key_states[:, :, :, self.qk_nope_head_dim:],
-                                                 cos, sin, True)
+        from ipex_llm.transformers.models.common import rotary_two_with_cache_inplaced
+        rotary_two_with_cache_inplaced(query_states[:, :, :, self.qk_nope_head_dim:],
+                                       key_states[:, :, :, self.qk_nope_head_dim:],
+                                       cos, sin, True)
     else:
         q_nope, q_pe = torch.split(
             q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
@@ -279,11 +279,11 @@ def fuse_gate_forward(self, x: torch.Tensor):
         )
         scores = logits.sigmoid()
 
-        import xe_addons
-        topk_idx, topk_weight = xe_addons.moe_group_topk(
+        from ipex_llm.transformers.models.common import moe_group_topk
+        topk_idx, topk_weight = moe_group_topk(
             scores, self.e_score_correction_bias,
-            self.n_group, 2, self.topk_group, self.top_k,
-            self.top_k > 1 and self.norm_topk_prob, 1e-20, self.routed_scaling_factor
+            self.n_group, self.topk_group, self.top_k,
+            self.norm_topk_prob, self.routed_scaling_factor
         )
     else:
         topk_idx, topk_weight = self(x)

ipex_llm/transformers/models/glm.py

@@ -98,9 +98,9 @@ def glm_attention_forward(
 
     cos, sin = position_embeddings
     if query_states.device.type == "xpu":
-        import xe_addons
         make_cache_contiguous_inplaced(cos, sin)
-        xe_addons.rotary_two_with_cache_inplaced(query_states, key_states, cos, sin, True)
+        from ipex_llm.transformers.models.common import rotary_two_with_cache_inplaced
+        rotary_two_with_cache_inplaced(query_states, key_states, cos, sin, True)
     else:
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
 

{ipex_llm-2.3.0b20250415.dist-info → ipex_llm-2.3.0b20250418.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.3.0b20250415
+Version: 2.3.0b20250418
 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 ==2.7.0b20250415 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.7.0b20250418 ; 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 ==2.7.0b20250415 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.7.0b20250418 ; (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 ==2.7.0b20250415 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250415 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250415 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.7.0b20250418 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250418 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250418 ; 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 ==2.7.0b20250415 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250415 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250415 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.7.0b20250418 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250418 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250418 ; 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 ==2.7.0b20250415 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.7.0b20250418 ; 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 ==2.7.0b20250415 ; extra == 'xpu-2-6-arl'
+Requires-Dist: bigdl-core-xe-all ==2.7.0b20250418 ; 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'
@@ -155,9 +155,9 @@ 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 ==2.7.0b20250415 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.7.0b20250415 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.7.0b20250415 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.7.0b20250418 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.7.0b20250418 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.7.0b20250418 ; 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'
@@ -178,9 +178,9 @@ 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 ==2.7.0b20250415 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.7.0b20250415 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.7.0b20250415 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.7.0b20250418 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.7.0b20250418 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.7.0b20250418 ; 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'
@@ -201,9 +201,9 @@ 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 ==2.7.0b20250415 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.7.0b20250415 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.7.0b20250415 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.7.0b20250418 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.7.0b20250418 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.7.0b20250418 ; 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-2.3.0b20250415.dist-info → ipex_llm-2.3.0b20250418.dist-info}/RECORD

@@ -153,13 +153,13 @@ ipex_llm/transformers/models/chatglm.py,sha256=DQM63oPIVMMTBQN4O4hPF4WY1aSiTWq4B
 ipex_llm/transformers/models/chatglm2.py,sha256=KyAIX7zGVQDQuwwM3QMBNWZbTeMHEzKUIgAryT0voHc,14933
 ipex_llm/transformers/models/chatglm4.py,sha256=QvUehdaCePB3MNHyWg3dneDxmjtBdxYeKUyQUVcsgfM,16886
 ipex_llm/transformers/models/chatglm4v.py,sha256=Ba9Xtzwtzk_rzg5khGqDrlHfJsDwc5YcM5_yPoord7o,13324
-ipex_llm/transformers/models/common.py,sha256=LVA9nL_qJ61NEkEn9T985PjrrWPGpDTCALknH4Qv5aw,13040
+ipex_llm/transformers/models/common.py,sha256=ueLGko8May2qWdjI-lSH30LXY4NYrqtBDXZekfq9rfQ,15374
 ipex_llm/transformers/models/decilm.py,sha256=P-PBuDPf07GvKggLwJx_wPwIn6esN3rX8ai2JxRuZmE,5246
-ipex_llm/transformers/models/deepseek.py,sha256=w6tGeyJ9joD7lQBiZ6A01Z00g8hAXC1N2yGtJh8kyuk,13096
+ipex_llm/transformers/models/deepseek.py,sha256=LFFA3tOtS3-WVoNmQQsfAymMCwNZiVQyBRZ4ZN_-IaE,13135
 ipex_llm/transformers/models/deepseek_v3.py,sha256=CTgwIKQlUPlUCbOxc9Id5GapWkXOP6pMtkguYrWpCio,10003
 ipex_llm/transformers/models/gemma.py,sha256=_E3Yw8Y45xyNVeLqyVKcpr8kjuICtETeL82cJ-bWJuU,9424
 ipex_llm/transformers/models/gemma2.py,sha256=2WZuv-FLzJyTJFaYxOuzJt47QE64M0lHnzAiO5T6ozI,8049
-ipex_llm/transformers/models/glm.py,sha256=lmeEWd_W2O638VzVW4Gm6cJre5XZcg_QBmPs8NWqXsM,7202
+ipex_llm/transformers/models/glm.py,sha256=-yNcl9Ci42AgAmW8OZ-WCafzJ-B1UUXLiQtI84VUSxA,7254
 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/gptneox.py,sha256=loRh1x_5S6BCeOr_s5xr-N_1SQHL3Y5IiUBAEyoMUqQ,6172
@@ -263,11 +263,11 @@ ipex_llm/vllm/xpu/engine/__init__.py,sha256=pY_CpyuZd72fr6s32ejeKHKFW0K4vUU2rzZj
 ipex_llm/vllm/xpu/engine/engine.py,sha256=NvCMbp0X8NVrOqbwm4FTvXOptTRLzu9jQsy37ZHnTk8,9493
 ipex_llm/vllm/xpu/entrypoints/openai/api_server.py,sha256=IjiSze9vzBCAkLu_VwIcJwuO1jyFna7DLrj6aSL7RaQ,35220
 ipex_llm/vllm/xpu/entrypoints/openai/cli_args.py,sha256=hB398yYtKauASRzevctScdbFIjiiSGMAe1bwEuIHrhY,10893
-ipex_llm-2.3.0b20250415.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
-ipex_llm-2.3.0b20250415.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
-ipex_llm-2.3.0b20250415.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
-ipex_llm-2.3.0b20250415.dist-info/METADATA,sha256=AK6V5GUxadLFvaeVQf6Ym6bSCFwx4ChknK-6c21T2lo,13917
-ipex_llm-2.3.0b20250415.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
-ipex_llm-2.3.0b20250415.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.3.0b20250415.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.3.0b20250415.dist-info/RECORD,,
+ipex_llm-2.3.0b20250418.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
+ipex_llm-2.3.0b20250418.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
+ipex_llm-2.3.0b20250418.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
+ipex_llm-2.3.0b20250418.dist-info/METADATA,sha256=J7dBYJDdMypI4RYNXQlwFb6FKdOpz84r_H1ocBA3ZGc,13917
+ipex_llm-2.3.0b20250418.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
+ipex_llm-2.3.0b20250418.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.3.0b20250418.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.3.0b20250418.dist-info/RECORD,,