ipex-llm 2.3.0b20250506__py3-none-win_amd64.whl → 2.3.0b20250507__py3-none-win_amd64.whl

ipex_llm/transformers/convert.py

@@ -1078,6 +1078,12 @@ def _optimize_pre(model, qtype=None):
     elif model.config.model_type == "qwen2_5_omni":
         from ipex_llm.transformers.models.qwen2_5_omni import merge_qkv
         model.apply(merge_qkv)
+    elif model.config.model_type == "qwen3":
+        from ipex_llm.transformers.models.qwen3 import merge_qkv
+        model.apply(merge_qkv)
+    elif model.config.model_type == "qwen3_moe":
+        from ipex_llm.transformers.models.qwen3_moe import merge_qkv
+        model.apply(merge_qkv)
     return model
 
 
@@ -2106,7 +2112,28 @@ def _optimize_post(model):
             convert_forward(model.token2wav, module.DiTAttention, dit_attention_forward)
             dit_model = model.token2wav.code2wav_dit_model
             dit_model._create_block_diff = MethodType(_create_block_diff, dit_model)
-
+    elif model.config.model_type == "qwen3":
+        modeling_module_name = model.__class__.__module__
+        module = importlib.import_module(modeling_module_name)
+        from ipex_llm.transformers.models.common import rms_norm_forward
+        from ipex_llm.transformers.models.qwen3 import qwen3_model_forward
+        from ipex_llm.transformers.models.qwen3 import qwen3_attention_forward
+        from ipex_llm.transformers.models.common import mlp_silu_forward
+        convert_forward(model, module.Qwen3RMSNorm, rms_norm_forward)
+        convert_forward(model, module.Qwen3Model, qwen3_model_forward)
+        convert_forward(model, module.Qwen3Attention, qwen3_attention_forward)
+        convert_forward(model, module.Qwen3MLP, mlp_silu_forward)
+    elif model.config.model_type == "qwen3_moe":
+        modeling_module_name = model.__class__.__module__
+        module = importlib.import_module(modeling_module_name)
+        from ipex_llm.transformers.models.common import rms_norm_forward
+        from ipex_llm.transformers.models.qwen3_moe import qwen3_moe_model_forward
+        from ipex_llm.transformers.models.qwen3 import qwen3_attention_forward
+        from ipex_llm.transformers.models.qwen3_moe import qwen3_moe_moe_forward
+        convert_forward(model, module.Qwen3MoeRMSNorm, rms_norm_forward)
+        convert_forward(model, module.Qwen3MoeModel, qwen3_moe_model_forward)
+        convert_forward(model, module.Qwen3MoeAttention, qwen3_attention_forward)
+        convert_forward(model, module.Qwen3MoeSparseMoeBlock, qwen3_moe_moe_forward)
     return model
 
 

ipex_llm/transformers/models/qwen3.py

@@ -0,0 +1,115 @@
+#
+# 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.
+#
+
+import torch
+from typing import Optional, List, Tuple
+from transformers.processing_utils import Unpack
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
+from transformers.models.qwen3.modeling_qwen3 import apply_rotary_pos_emb
+from transformers.models.qwen3.modeling_qwen3 import Qwen3Model, Qwen3Attention
+
+from ipex_llm.transformers.kv import DynamicNormalCache
+from ipex_llm.transformers.models.common import merge_qkv_base
+from ipex_llm.transformers.models.common import scaled_dot_product_attention
+from ipex_llm.transformers.models.utils import make_cache_contiguous_inplaced
+
+
+def merge_qkv(module: torch.nn.Module):
+    merge_qkv_base(module, Qwen3Attention)
+
+
+def qwen3_model_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[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,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+) -> MoeModelOutputWithPast:
+    device = input_ids.device if input_ids is not None else inputs_embeds.device
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    use_cache = True if device.type == "xpu" else use_cache
+    if use_cache and not isinstance(past_key_values, DynamicNormalCache):
+        past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
+
+    return Qwen3Model.forward(
+        self=self,
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **flash_attn_kwargs,
+    )
+
+
+def qwen3_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    position_embeddings: Tuple[torch.Tensor, torch.Tensor],
+    attention_mask: Optional[torch.Tensor],
+    past_key_value: Optional[Cache] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    **kwargs: Unpack[FlashAttentionKwargs],
+):
+    bsz, q_len, _ = hidden_states.size()
+    device = hidden_states.device
+
+    qkv = self.qkv_proj(hidden_states)
+    qkv = qkv.view(bsz, q_len, -1, self.head_dim)
+    qkv = qkv.transpose(1, 2)
+    query_states, key_states, value_states = qkv.split([self.config.num_attention_heads,
+                                                        self.config.num_key_value_heads,
+                                                        self.config.num_key_value_heads], dim=1)
+    query_states = self.q_norm(query_states)
+    key_states = self.k_norm(key_states)
+
+    cos, sin = position_embeddings
+    if device.type == "xpu":
+        import xe_addons
+        make_cache_contiguous_inplaced(cos, sin)
+        xe_addons.rotary_half_with_cache_inplaced(query_states, key_states, cos, sin)
+    else:
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin)
+
+    if past_key_value is not None:
+        cache_kwargs = {"sin": sin, "cos": cos, "cache_position": cache_position}
+        key_states, value_states = past_key_value.update(key_states, value_states,
+                                                         self.layer_idx, cache_kwargs)
+    attn_weights = None
+    attn_output = scaled_dot_product_attention(
+        query_states, key_states, value_states,
+        attention_mask, q_len == key_states.size(2), self.scaling
+    )
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    attn_output = attn_output.reshape(bsz, q_len, -1)
+    attn_output = self.o_proj(attn_output)
+    return attn_output, attn_weights

ipex_llm/transformers/models/qwen3_moe.py

@@ -0,0 +1,142 @@
+#
+# 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.
+#
+
+import torch
+from typing import Optional, List
+from transformers.processing_utils import Unpack
+from transformers.modeling_outputs import MoeModelOutputWithPast
+from transformers.modeling_flash_attention_utils import FlashAttentionKwargs
+
+from transformers.models.qwen3_moe.modeling_qwen3_moe import Qwen3MoeModel, Qwen3MoeAttention
+
+from ipex_llm.transformers.kv import DynamicNormalCache
+from ipex_llm.transformers.models.common import merge_qkv_base
+from ipex_llm.transformers.models.utils import use_fuse_moe
+
+
+def merge_qkv(module: torch.nn.Module):
+    merge_qkv_base(module, Qwen3MoeAttention)
+
+
+def qwen3_moe_model_forward(
+    self,
+    input_ids: Optional[torch.LongTensor] = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[List[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,
+    output_router_logits: Optional[bool] = None,
+    cache_position: Optional[torch.LongTensor] = None,
+    **flash_attn_kwargs: Unpack[FlashAttentionKwargs],
+) -> MoeModelOutputWithPast:
+    device = input_ids.device if input_ids is not None else inputs_embeds.device
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    use_cache = True if device.type == "xpu" else use_cache
+    if use_cache and not isinstance(past_key_values, DynamicNormalCache):
+        past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
+
+    return Qwen3MoeModel.forward(
+        self=self,
+        input_ids=input_ids,
+        attention_mask=attention_mask,
+        position_ids=position_ids,
+        past_key_values=past_key_values,
+        inputs_embeds=inputs_embeds,
+        use_cache=use_cache,
+        output_attentions=output_attentions,
+        output_hidden_states=output_hidden_states,
+        output_router_logits=output_router_logits,
+        cache_position=cache_position,
+        **flash_attn_kwargs,
+    )
+
+
+def qwen3_moe_moe_forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+    batch_size, sequence_length, hidden_dim = hidden_states.shape
+    hidden_states = hidden_states.view(-1, hidden_dim)
+    router_logits = self.gate(hidden_states)
+
+    if router_logits.device == "xpu":
+        import xe_addons
+        selected_experts, routing_weights = xe_addons.moe_softmax_topk(
+            router_logits, self.top_k, self.norm_topk_prob
+        )
+    else:
+        routing_weights = torch.nn.functional.softmax(router_logits, dim=1, dtype=torch.float)
+        routing_weights, selected_experts = torch.topk(routing_weights, self.top_k, dim=-1)
+        if self.norm_topk_prob:
+            routing_weights /= routing_weights.sum(dim=-1, keepdim=True)
+        routing_weights = routing_weights.to(hidden_states.dtype)
+
+    if selected_experts.size(0) == 1:
+        if use_fuse_moe(hidden_states, self.experts[0].down_proj.qtype):
+            if getattr(self, "gates", None) is None:
+                gate_addrs = [expert.gate_proj.weight.data_ptr() for expert in self.experts]
+                up_addrs = [expert.up_proj.weight.data_ptr() for expert in self.experts]
+                down_addrs = [expert.down_proj.weight.data_ptr() for expert in self.experts]
+                gates = torch.tensor(gate_addrs, dtype=torch.uint64, device=hidden_states.device)
+                ups = torch.tensor(up_addrs, dtype=torch.uint64, device=hidden_states.device)
+                downs = torch.tensor(down_addrs, dtype=torch.uint64, device=hidden_states.device)
+                self.register_buffer("gates", gates, persistent=False)
+                self.register_buffer("ups", ups, persistent=False)
+                self.register_buffer("downs", downs, persistent=False)
+
+            import xe_linear
+            final_hidden_states = xe_linear.moe_forward_vec(
+                hidden_states, selected_experts, routing_weights, self.gates, self.ups, self.downs,
+                hidden_states.size(-1), self.experts[0].intermediate_size,
+                self.experts[0].down_proj.qtype
+            )
+        else:
+            idxs = selected_experts.flatten().tolist()
+            outputs = []
+            for i in idxs:
+                expert = self.experts[i]
+                expert_out = expert(hidden_states)
+                outputs.append(expert_out)
+            outs = torch.cat(outputs, dim=0)
+            reshaped_topk_weight = routing_weights.squeeze(0).unsqueeze(-1)
+            final_hidden_states = (outs * reshaped_topk_weight).sum(dim=0, keepdim=True)
+    else:
+        final_hidden_states = torch.zeros(
+            (batch_size * sequence_length, hidden_dim),
+            dtype=hidden_states.dtype, device=hidden_states.device
+        )
+
+        # One hot encode the selected experts to create an expert mask
+        # this will be used to easily index which expert is going to be sollicitated
+        expert_mask = torch.nn.functional.one_hot(selected_experts,
+                                                  num_classes=self.num_experts).permute(2, 1, 0)
+
+        # Loop over all available experts in the model and perform the computation on each expert
+        for expert_idx in range(self.num_experts):
+            expert_layer = self.experts[expert_idx]
+            idx, top_x = torch.where(expert_mask[expert_idx])
+
+            # Index the correct hidden states and compute the expert hidden state for
+            # the current expert. We need to make sure to multiply the output hidden
+            # states by `routing_weights` on the corresponding tokens (top-1 and top-2)
+            current_state = hidden_states[None, top_x].reshape(-1, hidden_dim)
+            current_hidden_states = expert_layer(current_state) * routing_weights[top_x, idx, None]
+
+            # However `index_add_` only support torch tensors for indexing so we'll use
+            # the `top_x` tensor here.
+            final_hidden_states.index_add_(0, top_x, current_hidden_states.to(hidden_states.dtype))
+    final_hidden_states = final_hidden_states.reshape(batch_size, sequence_length, hidden_dim)
+    return final_hidden_states, router_logits

ipex_llm/vllm/xpu/model_convert.py

@@ -65,13 +65,14 @@ def _model_sample_convert():
 def _ipex_llm_convert(load_in_low_bit):
     # import pdb
     # pdb.set_trace()
-    from vllm.worker.xpu_model_runner import XPUModelRunner
+    from vllm.worker.xpu_model_runner import XPUModelRunner, XPUModelRunnerBase
     from ipex_llm.vllm.xpu.ipex_llm_wrapper import get_ipex_llm_wrapper
     from ipex_llm.vllm.xpu.ipex_llm_v1_wrapper import get_ipex_llm_v1_wrapper
     import vllm.executor.ray_utils as ray_utils_v0
     import vllm.v1.executor.ray_utils as ray_utils_v1
     from vllm.v1.worker.gpu_model_runner import GPUModelRunner
     setattr(XPUModelRunner, "load_model", get_load_function(load_in_low_bit))
+    setattr(XPUModelRunnerBase, "load_model", get_load_function(load_in_low_bit))
     setattr(GPUModelRunner, "load_model", get_load_function(load_in_low_bit))
     setattr(ray_utils_v0, "RayWorkerWrapper", get_ipex_llm_wrapper(load_in_low_bit))
     setattr(ray_utils_v1, "RayWorkerWrapper", get_ipex_llm_v1_wrapper(load_in_low_bit))

{ipex_llm-2.3.0b20250506.dist-info → ipex_llm-2.3.0b20250507.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.3.0b20250506
+Version: 2.3.0b20250507
 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 ==2.7.0b20250506 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.7.0b20250507 ; 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 ==2.7.0b20250506 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.7.0b20250507 ; (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 ==2.7.0b20250506 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250506 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250506 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.7.0b20250507 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250507 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250507 ; 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 ==2.7.0b20250506 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250506 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250506 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.7.0b20250507 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.7.0b20250507 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.7.0b20250507 ; 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 ==2.7.0b20250506 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.7.0b20250507 ; 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 ==2.7.0b20250506 ; extra == 'xpu-2-6-arl'
+Requires-Dist: bigdl-core-xe-all ==2.7.0b20250507 ; 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-2.3.0b20250506.dist-info → ipex_llm-2.3.0b20250507.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=22wrJqv9sWR9HnPLzio3B296jvY6NinrMPnt1c0_MFQ,36352
-ipex_llm/libs/bloom.dll,sha256=9en-Ji91waynOJHs8VXrPAMV2JrMawAeUrx4dTmiJP4,507904
-ipex_llm/libs/gptneox-api.dll,sha256=UKL0fcB3g3JI5njBv4j2UULRS_6jvfl6iSpp0NwtiHc,24576
-ipex_llm/libs/gptneox.dll,sha256=Wt945Qf_5GbY5jfpHg7mG9xn6LoDvoGyJjmx-oH6Yuw,568320
-ipex_llm/libs/libbloom_avx.dll,sha256=3X7c6OiRyh40OVt7x9CKthiNEyEje4cjE8Dc18JUHMo,536576
-ipex_llm/libs/libbloom_vnni.dll,sha256=bTkMhPRpruA2ub8-IOsyqq09UXTWKmrBdDlazlLvDNQ,508416
-ipex_llm/libs/libgptneox_avx.dll,sha256=gsX3W1UFTEZY4rk7Pb2HspiQ1TJKQKcIwIQY_ENDCaQ,596992
-ipex_llm/libs/libgptneox_vnni.dll,sha256=AgPQA4Dt5ZzLIqrRwIpMWh0LjaLYMGmNAbwJJSzGh8A,568832
-ipex_llm/libs/libllama_avx.dll,sha256=c5kbeImKfxR3j1OtVD5ojVDg9bmEHDHFtnJWr3_IApA,591360
-ipex_llm/libs/libllama_vnni.dll,sha256=hFszK131YSqH1JSYqc7V6cpPJR9xoXJK0LhO9fecXQ8,563200
-ipex_llm/libs/libstarcoder_avx.dll,sha256=eYYr_hJjrenb3Jua3-bxrWpXj1WUbryJpxCfdjFIw_I,627712
-ipex_llm/libs/libstarcoder_vnni.dll,sha256=47uszx8WV0cMlN9DbqlPJ2oiSepk1OUDtCNVGGqTh5I,599552
-ipex_llm/libs/llama-api.dll,sha256=mzm_HqP6hSe1dpFKWAN2dl7mFBHQd79NOkPunLxwMCc,25600
-ipex_llm/libs/llama.dll,sha256=69NmJ5PhYO3NM1ZEcWFzTL7dTHIDo-6grG2E84-YqAg,562688
-ipex_llm/libs/main-bloom.exe,sha256=_EpnRlG0qUFrdY8AbMZmKALfv3l3OA4tewupZpKrm0A,103424
-ipex_llm/libs/main-gptneox.exe,sha256=hJvEs3hA6NrIzqlWkFQvikjnNs6maHm1ThPhG1t3Kvw,98816
-ipex_llm/libs/main-llama.exe,sha256=YBt0PyyVpOAxQdZZ8bwnV-gpMoPpg_Lqmsxm9WuzdJ0,99840
-ipex_llm/libs/main-starcoder.exe,sha256=EGAWFWMsvAJBw43r0JwItz-okFge5e6F7yOpitDqnCY,157696
-ipex_llm/libs/pipeline.dll,sha256=kzD55sWU-71D0qcUL-YlOwFK_xAySW_pIQ3H6ev_L_4,73216
-ipex_llm/libs/quantize-bloom.exe,sha256=wCTjZtTUkjhXDMfk78ib5g9no76EL3GfpCmnuiD4lW4,126464
-ipex_llm/libs/quantize-bloom_vnni.exe,sha256=0oIOV9Y6C-FQH-0C0ZOk34ROX616bvxfZFz_vsL4G60,128000
-ipex_llm/libs/quantize-gptneox.exe,sha256=nJ0hBR-kdE9UWWV-iMN_gb13s2pqhixaro35m12OXd8,104448
-ipex_llm/libs/quantize-gptneox_vnni.exe,sha256=ZVjbdMhQT9_5J9Q-dVDdUcFxmZIX-ZVnwzoKtvE85Hg,104960
-ipex_llm/libs/quantize-llama.exe,sha256=F75XdM4fjQ4ot5caKLmixnzTMPSBtBSrYhZ12DPblSU,110080
-ipex_llm/libs/quantize-llama_vnni.exe,sha256=_J3eM4LgcSG7ACb8DTSb5OPFvSv7wMKQ-Kdewxf070M,110592
-ipex_llm/libs/quantize-starcoder.exe,sha256=zvpmhrrbm2pXA5MrhyZj0buhvswwkS_Vuq32Sdm8w3s,127488
-ipex_llm/libs/quantize-starcoder_vnni.exe,sha256=QoT-09g3aP44ln_ULGa-1sQJD3c2IsjA9SnrrdkHjt4,128512
-ipex_llm/libs/starcoder-api.dll,sha256=-h3oVRrPWo168PRcbOFa_n1pIVVjDXoNqP98dPatZKw,21504
-ipex_llm/libs/starcoder.dll,sha256=1igLCjVlgHCs22hdxokewN2affmiXbCC2OeVtIaI-DA,599040
+ipex_llm/libs/bloom-api.dll,sha256=Tt4tHy_XQ-MEiEw6Up7SpsWRlvqYtGwwpH26WdHArOc,36352
+ipex_llm/libs/bloom.dll,sha256=EA5yFTwHb_PGvMvz4Cyb1tnvmAxdzxC5-wuV6p0OMKI,507904
+ipex_llm/libs/gptneox-api.dll,sha256=LIaPSIwg7huhRjoUgyzE9F2O-DODUbywfz2UwLDNOeg,24576
+ipex_llm/libs/gptneox.dll,sha256=WatAJmN3tQVb0rWRptVvF-dtkAPa0E7ysgg2iz7PCN8,568320
+ipex_llm/libs/libbloom_avx.dll,sha256=6NBma3ZePdGuPD80NiLLbwGd2Ja_fHFyz5j7wbYfb9M,536576
+ipex_llm/libs/libbloom_vnni.dll,sha256=0friR2pBOZ7sLw51TabkjtpbUuIOrnNKn5wN6UdBiKo,508416
+ipex_llm/libs/libgptneox_avx.dll,sha256=KtoxICvRPw7QLaBsh2_TtEl3Uwo2Mhoblm2Zvu2Kzrc,596992
+ipex_llm/libs/libgptneox_vnni.dll,sha256=fFnNCW-WzG0Zu2m-nyEiQfefNvmhPM94Bx3AVp3nu8Q,568832
+ipex_llm/libs/libllama_avx.dll,sha256=BWXO_CZT6Qps9hZI9KzJJ5Z4gUl_NflK_TlmyoJe4-I,591360
+ipex_llm/libs/libllama_vnni.dll,sha256=EmfDxxZ5bqjWCsgzjQBj5pstoC8Ky_hXuepz8CWn6Rc,563200
+ipex_llm/libs/libstarcoder_avx.dll,sha256=-MkXNLpWj0HxIomfy4hrv2Q-Bts8beRG06giVAPOlXk,627712
+ipex_llm/libs/libstarcoder_vnni.dll,sha256=C2jUNQ-64b8Teprc6I564lvpkqdPvZADUEuO0ytOb8M,599552
+ipex_llm/libs/llama-api.dll,sha256=oZ2RtujrylsJQdVRuEcsuD2OlX9j8mdaP-1mDTF7vho,25600
+ipex_llm/libs/llama.dll,sha256=3drx1vyctJLIDjG12df8t4ejyo5BmMA9aPtisNn5otQ,562688
+ipex_llm/libs/main-bloom.exe,sha256=fnieKQvKt1ppyEWY9bitNt2iaDkW6CBKHqEPBYCH20I,103424
+ipex_llm/libs/main-gptneox.exe,sha256=xV7yvYUhvnLTdrYNOZqR-FO3dpK2sGknRffCTBnux-g,98816
+ipex_llm/libs/main-llama.exe,sha256=HbI0RIngIauJl_molsqiKarm2jnN9k-XPFTRy72I-rE,99840
+ipex_llm/libs/main-starcoder.exe,sha256=IbO6csHarglFS76k0F6zLqdqdVsqNzx5us_6wvsiNt8,157696
+ipex_llm/libs/pipeline.dll,sha256=bE_ZKwL0jhC5npmY3Rst4bX5SRavoefJSgmmCGmh3ac,73216
+ipex_llm/libs/quantize-bloom.exe,sha256=7_p6sqNt1DflgDNZM4kP4byMmGQ_EtPPrXaBogeKVto,126464
+ipex_llm/libs/quantize-bloom_vnni.exe,sha256=mJ9hYQDSBmmzu2z-iHFKjUA7Gor7PlEPagvjnzXva1Q,128000
+ipex_llm/libs/quantize-gptneox.exe,sha256=uw_e-hgup0bgB3boPo2SM7nvI27IeJDaJE2u1h5Kp0U,104448
+ipex_llm/libs/quantize-gptneox_vnni.exe,sha256=vjJusWjA_7-jVXNz-dmyoh1m-4JjGn_bEkxUTi6YOMU,104960
+ipex_llm/libs/quantize-llama.exe,sha256=zq8cbPhW6raS0-m4S1v2BrHt4A3Z_lDfXLbtm0hy7pE,110080
+ipex_llm/libs/quantize-llama_vnni.exe,sha256=YYBlgR3sj7O6FExih8liLarKbSfdtiGaGUp3kLQOgg0,110592
+ipex_llm/libs/quantize-starcoder.exe,sha256=iWxQHYSXAxCKRUu7uRZKPoFO4AiINjEcziIOwVF57FY,127488
+ipex_llm/libs/quantize-starcoder_vnni.exe,sha256=c22imSa2x6MNcEOTVuWLoXz1hM0C6NEh6ai7kTsowwY,128512
+ipex_llm/libs/starcoder-api.dll,sha256=EMtZBU0dAHFtgRepVnPLl-KU-6yp1PM1j62AyS7IKW8,21504
+ipex_llm/libs/starcoder.dll,sha256=KJxkaBk43ERKcIgTFHJ-vfd0GdWXea25zDAWSJIbTGo,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
@@ -87,7 +87,7 @@ 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=BreA3EY6hkNq0rVixb_sUuTLzMrcWXTt3yRsshCPHHQ,1214
-ipex_llm/transformers/convert.py,sha256=fmunTuZCTgTnZZXfDaCNHlO4TMDBGUgqPRKuOLp7x9Y,107586
+ipex_llm/transformers/convert.py,sha256=mJWejcYwe2gCqPsLFchq_umhkwG_6tchikEHGWyQWy8,109431
 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=src_HcVDKFwQ1V8hdTrFQw5RIwUewM9VOR47GVTPJG4,21187
@@ -174,6 +174,8 @@ ipex_llm/transformers/models/qwen2.py,sha256=zK-FpUaxEhjD4gZa1ZvArodAilz29T_cpeA
 ipex_llm/transformers/models/qwen2_5_omni.py,sha256=gNnWvGiPntn6GlyIUtCVYCSIMPU6FG9TCS7fhWbwerY,18779
 ipex_llm/transformers/models/qwen2_moe.py,sha256=a0gYo-ngf8SxaEnBdZUJDnPS6Mkn_poDd8xqhx50icI,19516
 ipex_llm/transformers/models/qwen2_vl.py,sha256=G-9e2oN4f5p5IWQ-zsBZuONxTura3BjlgyT2meigbHQ,13579
+ipex_llm/transformers/models/qwen3.py,sha256=n7kbImp1i3-6Tki2y2NwMKZ6f6tItyLu_0USmCF7_qY,4756
+ipex_llm/transformers/models/qwen3_moe.py,sha256=4S0nxMRA3gneaAcXNbPowgZq70_rb3L9PObRQtUNsww,6812
 ipex_llm/transformers/models/qwen_vl.py,sha256=lwNwCJNsBvOu1TGNDW_E2IlCjI7XZwHY8qIcZpuNDFc,17187
 ipex_llm/transformers/models/rwkv4.py,sha256=H4KMtxN0JA2ZTXnonHpsUUJ5xULemo-D1Jzl0ri_UY8,6123
 ipex_llm/transformers/models/rwkv5.py,sha256=OkRNj1pCAZg1z2Fw-I0DEnxLEdZyPeRSQ6msrkxLOCs,10710
@@ -251,16 +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=NnD7xj7O9ZHVUedCxqAH8GEZvagokZmBXrjYSLpigkY,9541
+ipex_llm/vllm/xpu/model_convert.py,sha256=vkzH9quwVcjoAviPWz2IdbPec9u2YTdG6KNUEt6j7dU,9643
 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=uWHyyHLw-B8wXBnQw9_MCG81tKK9Jb0dyq1xfYHgoNw,45905
 ipex_llm/vllm/xpu/entrypoints/openai/cli_args.py,sha256=hB398yYtKauASRzevctScdbFIjiiSGMAe1bwEuIHrhY,10893
-ipex_llm-2.3.0b20250506.data/scripts/ipex-llm-init.bat,sha256=HPtCYuDYwEatq7dAwOvdfVcHYCpAVdbj75K1qh0vQek,2578
-ipex_llm-2.3.0b20250506.data/scripts/llm-chat.ps1,sha256=6qrs-hGVAV8IKh7Jx8nq_XrnZcjd7qGU5wndArM7Yag,2769
-ipex_llm-2.3.0b20250506.data/scripts/llm-cli.ps1,sha256=3qBtTLs_EjYDnM8YyCpJhzLnGCKTEGssu9UNqfkjVXs,3009
-ipex_llm-2.3.0b20250506.dist-info/METADATA,sha256=LtfQ-1Y1ougxU2AMzpnbh3XdCE0Gb1xgf7LIAGyB6iQ,8865
-ipex_llm-2.3.0b20250506.dist-info/WHEEL,sha256=6iYPr8vTHsyDK75jr9X0V3I9wPSVmtwr_8fdATBciGk,98
-ipex_llm-2.3.0b20250506.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.3.0b20250506.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.3.0b20250506.dist-info/RECORD,,
+ipex_llm-2.3.0b20250507.data/scripts/ipex-llm-init.bat,sha256=HPtCYuDYwEatq7dAwOvdfVcHYCpAVdbj75K1qh0vQek,2578
+ipex_llm-2.3.0b20250507.data/scripts/llm-chat.ps1,sha256=6qrs-hGVAV8IKh7Jx8nq_XrnZcjd7qGU5wndArM7Yag,2769
+ipex_llm-2.3.0b20250507.data/scripts/llm-cli.ps1,sha256=3qBtTLs_EjYDnM8YyCpJhzLnGCKTEGssu9UNqfkjVXs,3009
+ipex_llm-2.3.0b20250507.dist-info/METADATA,sha256=dOnD7OAqHtDZF8uGb-xz3Xhw_pNyTL3ntanZ-5kehDY,8865
+ipex_llm-2.3.0b20250507.dist-info/WHEEL,sha256=6iYPr8vTHsyDK75jr9X0V3I9wPSVmtwr_8fdATBciGk,98
+ipex_llm-2.3.0b20250507.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.3.0b20250507.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.3.0b20250507.dist-info/RECORD,,