bigdl-core-npu 2.6.0b2__cp311-cp311-win_amd64.whl → 2.6.0b20241103__cp311-cp311-win_amd64.whl

{bigdl_core_npu-2.6.0b2.dist-info → bigdl_core_npu-2.6.0b20241103.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: bigdl-core-npu
-Version: 2.6.0b2
+Version: 2.6.0b20241103
 Summary: Intel® NPU Acceleration Library
 Home-page: https://github.com/intel/intel-npu-acceleration-library
 Author: Alessandro Palla

{bigdl_core_npu-2.6.0b2.dist-info → bigdl_core_npu-2.6.0b20241103.dist-info}/RECORD

@@ -1,5 +1,5 @@
 intel_npu_acceleration_library/__init__.py,sha256=ZKTIhGMDjF7P6pF-yX8KWcSXbeHWRk24AO_orsa18f8,536
-intel_npu_acceleration_library/_version.py,sha256=o0ui4TyomjRVwLdl9zwG9b1pXUD6KLzj1zUaNOMu9V8,105
+intel_npu_acceleration_library/_version.py,sha256=JfJivKuzXFm3vqyuVhzyZq1BC4oaae5rYiCmb_9n7HI,112
 intel_npu_acceleration_library/compiler.py,sha256=3IdgqjamSC8MLexDBJypIeZRiWIcTFnvQSU1LPXUr7Y,6225
 intel_npu_acceleration_library/device.py,sha256=TbG4cJ197qo7PJQ5zz9zfxbuXB5OTWJlKNaKL4TAlms,7395
 intel_npu_acceleration_library/dtypes.py,sha256=1CV4FIuvlmLsTCS1nCCEwq4EzZmD3thj1_92v5vajpw,3539
@@ -8,10 +8,10 @@ intel_npu_acceleration_library/optimizations.py,sha256=9NY8QoDFbs2LY12jbx6As8g2v
 intel_npu_acceleration_library/quantization.py,sha256=6N_04h1KX6TNbw-ceANV0Pmk4_lQ2Y9C7Pwn5x-zQzo,5566
 intel_npu_acceleration_library/backend/__init__.py,sha256=2NP6Ypr1dGUNXmLGW5GD9xrh0U9KJgqxTd_c7su1RUY,857
 intel_npu_acceleration_library/backend/base.py,sha256=hbHqxSOfWH5BaA5PY6_zaf1Zdg5NrQK6WOfe-hr279k,8605
-intel_npu_acceleration_library/backend/bindings.py,sha256=g802BSzQUqAS-LzgUP4UGfrxYHNG397we31xanDs3B8,8576
+intel_npu_acceleration_library/backend/bindings.py,sha256=cla6JRX7pqUDuRmsXN6K9cAKklHz_mb6butatR2Eu9I,8901
 intel_npu_acceleration_library/backend/compression.py,sha256=Avz_zm2s_ELy5peVQ8zFGn8njBfh9nEGR16mflotBic,630
 intel_npu_acceleration_library/backend/convolution.py,sha256=cN3k78X3Y4Cbf7er-MFq0sJ4OwIvquj8PajpdEDmCo4,2018
-intel_npu_acceleration_library/backend/factory.py,sha256=WqnpZDT3do8213BzSZLr0z8_d3sVyuBhYURZH1rEDAA,37688
+intel_npu_acceleration_library/backend/factory.py,sha256=n63KE8X9eOuv2m2MiQFASjzgnkIM9deGtDC-qSHRMMw,38847
 intel_npu_acceleration_library/backend/linear.py,sha256=RiLUh5FOSxRWHB5kYx7mOPOOrS_vxIeBJ5t3yC6wOiQ,1908
 intel_npu_acceleration_library/backend/matmul.py,sha256=mfGi73-mIbUcXp4kyvCGW0Y9kb4Xp1ppbGNpdJFohuA,1819
 intel_npu_acceleration_library/backend/mlp.py,sha256=BuKVwSI726v3nHQQvtMBbXyWxRTq-WoLZtTxeSeWaaY,2330
@@ -187,7 +187,7 @@ intel_npu_acceleration_library/external/openvino/torch/__init__.py,sha256=RXLzsf
 intel_npu_acceleration_library/functional/__init__.py,sha256=WWKwKOh6Sgovv7mKctA872TbLP98Pg5m5-MREvUmlAA,204
 intel_npu_acceleration_library/functional/scaled_dot_product_attention.py,sha256=yGUcg4tDQOLuUnP1g74cl-ec8TRr2SuAMcNLlN6qLvE,1620
 intel_npu_acceleration_library/lib/Release/cache.json,sha256=CyrSqZUWo0Ec4_7ydOiuKIC0Gm8AybrGdozUqUuHxBw,8840377
-intel_npu_acceleration_library/lib/Release/intel_npu_acceleration_library.dll,sha256=ZxYdf6i90mDDNCd4jgtDDfFXFqGwsopGA6u5I9nWdRo,301056
+intel_npu_acceleration_library/lib/Release/intel_npu_acceleration_library.dll,sha256=DQZEP3lcwX8HIuiWbpbE3-WyMRUCKuxUFtabIHejJXY,304640
 intel_npu_acceleration_library/lib/Release/openvino.dll,sha256=_ifEwHwM-7LuKMhAnlqNuJ2GxsLXbG47easxl5E4shU,12624904
 intel_npu_acceleration_library/lib/Release/openvino_auto_batch_plugin.dll,sha256=hXFvu4oLvfNhCODn5eNYOmkxBb0LEKYXHA0sZLccOXc,195080
 intel_npu_acceleration_library/lib/Release/openvino_auto_plugin.dll,sha256=nh_iDxejjHlkes-KT0IwBzEd4Ec0L3bXQFCl0Dqerf8,472072
@@ -217,7 +217,7 @@ intel_npu_acceleration_library/nn/functional.py,sha256=UfAKBc0u6RtyaMo14ldH2GpEn
 intel_npu_acceleration_library/nn/linear.py,sha256=Q06SoGQeLaI86nA_ky2GnFC6H2Fw1zyMDILKnpYC2eo,5739
 intel_npu_acceleration_library/nn/llm.py,sha256=P6dz36Yf6BHtzWcftaghC6QaMI_WeRfQwrCbO7fD6hk,15002
 intel_npu_acceleration_library/nn/module.py,sha256=klVK4A0O-7fLzEIhGhE6_eVgvyVK_NakAqpDq08Ju1Y,12637
-bigdl_core_npu-2.6.0b2.dist-info/METADATA,sha256=9h9wb4lWrHS8XM0QCcw9PPeZevup5swuMihy_tBBqZE,1536
-bigdl_core_npu-2.6.0b2.dist-info/WHEEL,sha256=WutsMqxRjo8PALJe8NWxuOYrO2lUIIHDIxhZ8tjc8BY,101
-bigdl_core_npu-2.6.0b2.dist-info/top_level.txt,sha256=CH3qQoleRBC1eThu8mCEMxYNKdzJuXCtmeCXRKskt7A,31
-bigdl_core_npu-2.6.0b2.dist-info/RECORD,,
+bigdl_core_npu-2.6.0b20241103.dist-info/METADATA,sha256=7rc3G9Nh-kun6Mc_hRDdMt_2HgqTTMvKVUEemmfAXX8,1543
+bigdl_core_npu-2.6.0b20241103.dist-info/WHEEL,sha256=WutsMqxRjo8PALJe8NWxuOYrO2lUIIHDIxhZ8tjc8BY,101
+bigdl_core_npu-2.6.0b20241103.dist-info/top_level.txt,sha256=CH3qQoleRBC1eThu8mCEMxYNKdzJuXCtmeCXRKskt7A,31
+bigdl_core_npu-2.6.0b20241103.dist-info/RECORD,,

intel_npu_acceleration_library/_version.py

@@ -3,4 +3,4 @@
 # SPDX-License-Identifier: Apache 2.0
 #
 
-__version__ = "2.6.0b2"
+__version__ = "2.6.0b20241103"

intel_npu_acceleration_library/backend/bindings.py

@@ -143,7 +143,7 @@ def init_network_factory(lib: ctypes.CDLL):
     ]
     lib.slice.restype = handler
 
-    lib.compile.argtypes = [handler]
+    lib.compile.argtypes = [handler, ctypes.c_int]
     lib.compile.restype = handler
 
     lib.get_output_tensor_shape_size.argtypes = [handler, ctypes.c_int]
@@ -237,6 +237,19 @@ def init_network_factory(lib: ctypes.CDLL):
     ]
     lib.dq_split_linear.restype = handler
 
+    lib.dq_split_linear_prefill.argtypes = [
+        handler,
+        handler,
+        ctypes.c_int,
+        ctypes.c_int,
+        ctypes.c_int,
+        ctypes.c_bool,
+        ctypes.c_char_p,
+        ctypes.c_char_p,
+        ctypes.c_bool,
+    ]
+    lib.dq_split_linear_prefill.restype = handler
+
     for op in get_supported_ops():
         fn = getattr(lib, op.name)
         fn.argtypes = [handler] * (op.inputs + 1) + list(op.parameters)

intel_npu_acceleration_library/backend/factory.py

@@ -421,6 +421,7 @@ class NNFactory(BaseNPUBackendWithPrefetch):
         act_dtype: npt.DTypeLike = np.float16,
         wt_dtype: npt.DTypeLike = np.float16,
         scale_factor: bool = True,
+        is_prefill: bool = False,
     ) -> ctypes._Pointer:
         """Generate a linear layer for dynamic quantization linear layer.
 
@@ -432,15 +433,18 @@ class NNFactory(BaseNPUBackendWithPrefetch):
             bias (bool, optional): enable/disable bias. Defaults to False.
             act_dtype (npt.DTypeLike, optional): activation dtype. Defaults to np.float16.
             wt_dtype (npt.DTypeLike, optional): weight dtype. Defaults to np.float16.
+            scale_factor (bool, optional): enable/disable mul scale factor. Default to True,
+            is_prefill (bool, optional): enable/disable prefill linear optimization. Default to True.
 
         Returns:
             ctypes._Pointer: output node
         """
-        return backend_lib.dq_split_linear(self._mm, input_node, n_splits,
-                                           input_channels, outout_channels, bias,
-                                           self.get_backend_dtype(act_dtype),
-                                           self.get_backend_dtype(wt_dtype),
-                                           scale_factor)
+        func = backend_lib.dq_split_linear_prefill if is_prefill else backend_lib.dq_split_linear
+        return func(self._mm, input_node, n_splits,
+                    input_channels, outout_channels, bias,
+                    self.get_backend_dtype(act_dtype),
+                    self.get_backend_dtype(wt_dtype),
+                    scale_factor)
 
     @return_tensor
     def reshape(
@@ -897,6 +901,27 @@ class NNFactory(BaseNPUBackendWithPrefetch):
             auto_pad,  # auto_pad
         )
 
+    @return_tensor
+    def scaled_dot_product_attention(
+        self, query: ctypes._Pointer, key: ctypes._Pointer,
+        value: ctypes._Pointer, attn_mask: ctypes._Pointer,
+        is_causal: bool
+    ) -> ctypes._Pointer:
+        """Constructs a ScaledDotProductAttention operation.
+        Args:
+            query (ctypes._Pointer): query
+            key (ctypes._Pointer): key
+            value (ctypes._Pointer): value
+            attn_mask (ctypes._Pointer): attention mask
+            is_causal (ctypes._Pointer): causal/not causal
+        Returns:
+            ctypes._Pointer: output node
+        """
+        return backend_lib.scaled_dot_product_attention(self._mm,
+                                                        query, key,
+                                                        value, attn_mask,
+                                                        is_causal)
+
     def get_tensor_shape(self, node):
         """Get tensor shape.
 
@@ -946,7 +971,7 @@ class NNFactory(BaseNPUBackendWithPrefetch):
         else:
             raise RuntimeError("Unsupported dtype")
 
-    def compile(self):
+    def compile(self, npu_dpu_groups=4):
         """Finalize and compile a model."""
         self.out = []
         self.torch_out = []
@@ -954,7 +979,7 @@ class NNFactory(BaseNPUBackendWithPrefetch):
             backend_lib.result(self._mm, node)
 
         # Compile the model
-        backend_lib.compile(self._mm)
+        backend_lib.compile(self._mm, npu_dpu_groups)
 
         for idx, node in enumerate(self.output_nodes):
             output_shape = self.get_tensor_shape(node)