ipex-llm 2.2.0b20250223__py3-none-manylinux2010_x86_64.whl → 2.2.0b20250225__py3-none-manylinux2010_x86_64.whl

ipex_llm/transformers/convert.py

@@ -1070,7 +1070,9 @@ def _optimize_pre(model, qtype=None):
         model.apply(pre_register_inv_freq)
     elif model.config.model_type == "multi_modality":
         _optimize_pre(model.language_model)
-
+    elif model.config.model_type == "deepseek_v3" and model.config.hidden_size == 2048:
+        from ipex_llm.transformers.models.deepseek import padding_mla_v_hd
+        model.apply(padding_mla_v_hd)
     return model
 
 
@@ -2023,6 +2025,17 @@ def _optimize_post(model):
 
         # llm
         _optimize_post(model.language_model)
+    elif model.config.model_type == "deepseek_v3" and model.config.hidden_size == 2048:
+        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.deepseek import deepseek_model_forward
+        from ipex_llm.transformers.models.deepseek import deepseek_attention_forward
+        from ipex_llm.transformers.models.deepseek import deepseek_moe_forward
+        convert_forward(model, module.DeepseekV3RMSNorm, rms_norm_forward)
+        convert_forward(model, module.DeepseekV3Model, deepseek_model_forward)
+        convert_forward(model, module.DeepseekV3Attention, deepseek_attention_forward)
+        convert_forward(model, module.DeepseekV3MoE, deepseek_moe_forward)
 
     return model
 

ipex_llm/transformers/models/common.py

@@ -95,6 +95,33 @@ def padding_attention_hd_base(module: torch.nn.Module, attention_class,
         module.old_head_dim = old_head_dim
 
 
+def padding_mla_v_hd_base(module: torch.nn.Module, attention_class):
+    if (
+        isinstance(attention_class, str) and module.__class__.__name__ == attention_class
+        or not isinstance(attention_class, str) and isinstance(module, attention_class)
+    ):
+        k_head_dim = module.q_head_dim
+        v_head_dim = module.v_head_dim
+        if v_head_dim < k_head_dim:
+            kv_b_proj = module.kv_b_proj
+            w = kv_b_proj.weight.data.view(module.num_heads,
+                                           module.qk_nope_head_dim + module.v_head_dim,
+                                           module.kv_lora_rank)
+            k_w, v_w = w.split([module.qk_nope_head_dim, module.v_head_dim], dim=1)
+            new_v_w = torch.zeros([module.num_heads, k_head_dim, module.kv_lora_rank],
+                                  dtype=v_w.dtype, device=v_w.device)
+            new_v_w[:, :v_head_dim, :] = v_w
+            new_w = torch.cat([k_w, new_v_w], dim=1).view(-1, module.kv_lora_rank)
+
+            new_kv_b_proj = torch.nn.Linear(0, 0, bias=False,
+                                            dtype=new_w.dtype, device=new_w.device)
+            new_kv_b_proj.in_features = new_w.size(1)
+            new_kv_b_proj.out_features = new_w.size(0)
+            new_kv_b_proj.weight = torch.nn.Parameter(new_w, False)
+
+            module.kv_b_proj = new_kv_b_proj
+
+
 def padding_states_hd(states: torch.Tensor, old_head_dim: int, new_head_dim: int):
     bsz, num_heads, seq_len, head_dim = states.size()
     if head_dim == old_head_dim and old_head_dim < new_head_dim:

ipex_llm/transformers/models/deepseek.py

@@ -0,0 +1,303 @@
+#
+# 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.
+#
+# Some parts of this file is adapted from
+# https://huggingface.co/deepseek-ai/DeepSeek-R1/blob/main/modeling_deepseek.py
+# which is licensed under Apache License 2.0:
+#
+# https://github.com/OpenBMB/MiniCPM/blob/main/LICENSE
+#
+
+import torch
+import warnings
+
+from typing import Optional, Tuple, List, Union
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from transformers.modeling_attn_mask_utils import _prepare_4d_causal_attention_mask
+
+from ipex_llm.utils.common.log4Error import invalidInputError
+from ipex_llm.transformers.kv import DynamicNormalCache
+from ipex_llm.transformers.models.common import padding_mla_v_hd_base
+from ipex_llm.transformers.models.common import scaled_dot_product_attention
+from ipex_llm.transformers.models.utils import rotate_half
+
+
+def padding_mla_v_hd(module: torch.nn.Module):
+    padding_mla_v_hd_base(module, "DeepseekV3Attention")
+
+
+def deepseek_model_forward(
+    self,
+    input_ids: 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,
+    return_dict: Optional[bool] = None,
+) -> Union[Tuple, BaseModelOutputWithPast]:
+    output_attentions = (
+        output_attentions if output_attentions is not None
+        else self.config.output_attentions
+    )
+    output_hidden_states = (
+        output_hidden_states if output_hidden_states is not None
+        else self.config.output_hidden_states
+    )
+
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+    return_dict = (
+        return_dict if return_dict is not None else self.config.use_return_dict
+    )
+
+    # retrieve input_ids and inputs_embeds
+    invalidInputError((input_ids is None) ^ (inputs_embeds is None),
+                      "You cannot specify both input_ids and inputs_embeds at the same time, "
+                      "and must specify either one")
+
+    if inputs_embeds is None:
+        inputs_embeds = self.embed_tokens(input_ids)
+
+    batch_size, seq_length = inputs_embeds.shape[:2]
+
+    # IPEX-LLM OPT start: kv cache
+    past_key_values_length = 0
+    use_cache = True if inputs_embeds.device.type == "xpu" else use_cache
+    if use_cache:
+        if not isinstance(past_key_values, DynamicNormalCache):
+            past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
+        past_key_values_length = past_key_values.get_usable_length(seq_length)
+    # IPEX-LLM OPT end: kv cache
+
+    if position_ids is None:
+        position_ids = torch.arange(
+            past_key_values_length,
+            seq_length + past_key_values_length,
+            dtype=torch.long,
+            device=inputs_embeds.device,
+        )
+        position_ids = position_ids.unsqueeze(0)
+
+    # IPEX-LLM OPT start: fuse rope
+    if inputs_embeds.device.type == "xpu" and position_ids is not None:
+        cos, sin = self.layers[0].self_attn.rotary_emb(inputs_embeds,
+                                                       seq_length + past_key_values_length)
+        cos = cos[position_ids[0]].contiguous()
+        sin = sin[position_ids[0]].contiguous()
+        position_embeddings = (cos, sin)
+    else:
+        position_embeddings = None
+    # IPEX-LLM OPT end: fuse rope
+
+    # 4d mask is passed through the layers
+    attention_mask = _prepare_4d_causal_attention_mask(
+        attention_mask,
+        (batch_size, seq_length),
+        inputs_embeds,
+        past_key_values_length,
+    )
+
+    # embed positions
+    hidden_states = inputs_embeds
+
+    # decoder layers
+    all_hidden_states = () if output_hidden_states else None
+    all_self_attns = () if output_attentions else None
+    next_decoder_cache = None
+
+    for decoder_layer in self.layers:
+        if output_hidden_states:
+            all_hidden_states += (hidden_states,)
+
+        layer_outputs = decoder_layer(
+            hidden_states,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_value=past_key_values,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            position_embeddings=position_embeddings,
+        )
+
+        hidden_states = layer_outputs[0]
+
+        if use_cache:
+            next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+        if output_attentions:
+            all_self_attns += (layer_outputs[1],)
+
+    hidden_states = self.norm(hidden_states)
+
+    # add hidden states from the last decoder layer
+    if output_hidden_states:
+        all_hidden_states += (hidden_states,)
+
+    next_cache = next_decoder_cache
+    if not return_dict:
+        return tuple(
+            v
+            for v in [hidden_states, next_cache, all_hidden_states, all_self_attns]
+            if v is not None
+        )
+    return BaseModelOutputWithPast(
+        last_hidden_state=hidden_states,
+        past_key_values=next_cache,
+        hidden_states=all_hidden_states,
+        attentions=all_self_attns,
+    )
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, position_ids, unsqueeze_dim=1):
+    cos = cos[position_ids].unsqueeze(unsqueeze_dim)
+    sin = sin[position_ids].unsqueeze(unsqueeze_dim)
+
+    b, h, s, d = q.shape
+    q = q.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    b, h, s, d = k.shape
+    k = k.view(b, h, s, d // 2, 2).transpose(4, 3).reshape(b, h, s, d)
+
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+def deepseek_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Cache] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    **kwargs,
+) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+    if "padding_mask" in kwargs:
+        warnings.warn(
+            "Passing `padding_mask` is deprecated and will be removed in v4.37. "
+            "Please make sure use `attention_mask` instead.`"
+        )
+
+    bsz, q_len, _ = hidden_states.size()
+
+    if self.q_lora_rank is None:
+        q = self.q_proj(hidden_states)
+    else:
+        q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
+    q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
+
+    compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
+    compressed_kv, k_pe = torch.split(
+        compressed_kv, [self.kv_lora_rank, self.qk_rope_head_dim], dim=-1
+    )
+    k_pe = k_pe.view(bsz, q_len, 1, self.qk_rope_head_dim).transpose(1, 2)
+    kv = (
+        self.kv_b_proj(self.kv_a_layernorm(compressed_kv))
+        .view(bsz, q_len, self.num_heads, self.qk_nope_head_dim + self.q_head_dim)
+        .transpose(1, 2)
+    )
+
+    k_nope, value_states = torch.split(
+        kv, [self.qk_nope_head_dim, self.q_head_dim], dim=-1
+    )
+    kv_seq_len = value_states.shape[-2]
+    if past_key_value is not None:
+        kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+    position_embeddings = kwargs.get("position_embeddings", None)
+    if position_embeddings is not None:
+        query_states = q
+        key_states = torch.cat(
+            [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)
+    else:
+        q_nope, q_pe = torch.split(
+            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
+        )
+        cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
+        q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
+
+        query_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
+        query_states[:, :, :, : self.qk_nope_head_dim] = q_nope
+        query_states[:, :, :, self.qk_nope_head_dim:] = q_pe
+
+        key_states = k_pe.new_empty(bsz, self.num_heads, q_len, self.q_head_dim)
+        key_states[:, :, :, : self.qk_nope_head_dim] = k_nope
+        key_states[:, :, :, self.qk_nope_head_dim:] = k_pe
+
+    if past_key_value is not None:
+        key_states, value_states = past_key_value.update(key_states, value_states,
+                                                         self.layer_idx, None)
+
+    attn_weights = None
+    attn_output = scaled_dot_product_attention(
+        query_states, key_states, value_states,
+        attention_mask, q_len == kv_seq_len, self.softmax_scale
+    )
+    attn_output = attn_output[:, :, :, :self.v_head_dim]
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+
+    attn_output = attn_output.reshape(bsz, q_len, self.num_heads * self.v_head_dim)
+
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+
+    return attn_output, attn_weights, past_key_value
+
+
+def moe_infer_decode(self, x: torch.Tensor, topk_ids: torch.Tensor, topk_weight: torch.Tensor):
+    idxs = topk_ids.flatten().tolist()
+    outputs = []
+    for i in idxs:
+        expert = self.experts[i]
+        expert_out = expert(x)
+        outputs.append(expert_out)
+    outs = torch.cat(outputs, dim=0)
+    reshaped_topk_weight = topk_weight.squeeze(0).unsqueeze(-1).to(outs.dtype)
+    final_out = (outs * reshaped_topk_weight).sum(dim=0, keepdim=True)
+    return final_out
+
+
+def deepseek_moe_forward(self, hidden_states: torch.Tensor):
+    identity = hidden_states
+    orig_shape = hidden_states.shape
+    topk_idx, topk_weight = self.gate(hidden_states)
+    hidden_states = hidden_states.view(-1, hidden_states.shape[-1])
+    flat_topk_idx = topk_idx.view(-1)
+    if not self.training:
+        # IPEX-LLM OPT start : add special moe_infer implementation for decoding
+        if topk_idx.size(0) == 1:
+            y = moe_infer_decode(self, hidden_states, topk_idx, topk_weight)
+        else:
+            y = self.moe_infer(hidden_states, topk_idx, topk_weight)
+        y = y.view(*orig_shape)
+        # IPEX-LLM OPT end
+    if self.config.n_shared_experts is not None:
+        y = y + self.shared_experts(identity)
+    return y

ipex_llm/transformers/models/minicpm3.py

@@ -1,3 +1,25 @@
+#
+# 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.
+#
+# Some parts of this file is adapted from
+# https://hf-mirror.com/openbmb/MiniCPM3-4B/blob/main/modeling_minicpm.py
+# which is licensed under Apache License 2.0:
+#
+# https://github.com/OpenBMB/MiniCPM/blob/main/LICENSE
+#
+
 import torch
 import warnings
 
@@ -122,9 +144,6 @@ def minicpm3_attention_forward(
 
     q = self.q_b_proj(self.q_a_layernorm(self.q_a_proj(hidden_states)))
     q = q.view(bsz, q_len, self.num_heads, self.q_head_dim).transpose(1, 2)
-    q_nope, q_pe = torch.split(
-        q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
-    )
 
     compressed_kv = self.kv_a_proj_with_mqa(hidden_states)
     compressed_kv, k_pe = torch.split(
@@ -169,6 +188,9 @@ def minicpm3_attention_forward(
         else:
             invalidInputError(f"unknown rope method: {self.rotary_emb.__class__.__name__}")
     else:
+        q_nope, q_pe = torch.split(
+            q, [self.qk_nope_head_dim, self.qk_rope_head_dim], dim=-1
+        )
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
         q_pe, k_pe = apply_rotary_pos_emb(q_pe, k_pe, cos, sin, position_ids)
 

ipex_llm/transformers/xgrammar.py

@@ -0,0 +1,47 @@
+#
+# 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.
+#
+
+
+from transformers import PreTrainedTokenizer, LogitsProcessor
+from ipex_llm.utils.modules import insert_fake_module
+
+insert_fake_module("xgrammar.kernels.apply_token_bitmask_inplace_cuda")
+insert_fake_module("xgrammar.kernels.apply_token_bitmask_inplace_triton")
+insert_fake_module(
+    "xgrammar.kernels.apply_token_bitmask_inplace_cuda.apply_token_bitmask_inplace_cuda"
+)
+insert_fake_module(
+    "xgrammar.kernels.apply_token_bitmask_inplace_triton.apply_token_bitmask_inplace_triton"
+)
+
+import xgrammar as xgr
+
+
+def create_json_logits_processor(tokenizer: PreTrainedTokenizer, vocab_size: int, schema=None):
+    tokenizer_info = xgr.TokenizerInfo.from_huggingface(tokenizer, vocab_size=vocab_size)
+    grammar_compiler = xgr.GrammarCompiler(tokenizer_info)
+    if schema is None:
+        compiled_grammar = grammar_compiler.compile_builtin_json_grammar()
+    else:
+        compiled_grammar = grammar_compiler.compile_json_schema(schema)
+    processor = xgr.contrib.hf.LogitsProcessor(compiled_grammar)
+    return processor
+
+
+def reset_json_logits_processor(processor: LogitsProcessor) -> LogitsProcessor:
+    compiled_grammar = processor.compiled_grammar
+    new_processor = xgr.contrib.hf.LogitsProcessor(compiled_grammar)
+    return new_processor

{ipex_llm-2.2.0b20250223.dist-info → ipex_llm-2.2.0b20250225.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250223
+Version: 2.2.0b20250225
 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 ==2.6.0b20250223 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250225 ; 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 ==2.6.0b20250223 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250225 ; (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 ==2.6.0b20250223 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250223 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250223 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250225 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250225 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250225 ; 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 ==2.6.0b20250223 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250223 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250223 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250225 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250225 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250225 ; 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 ==2.6.0b20250223 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.6.0b20250225 ; 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'
@@ -133,9 +133,9 @@ 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 ==2.6.0b20250223 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250223 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250223 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250225 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250225 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250225 ; 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'
@@ -156,9 +156,9 @@ 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 ==2.6.0b20250223 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250223 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250223 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250225 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250225 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250225 ; 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'
@@ -179,9 +179,9 @@ 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 ==2.6.0b20250223 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250223 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250223 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250225 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250225 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250225 ; 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-2.2.0b20250223.dist-info → ipex_llm-2.2.0b20250225.dist-info}/RECORD

@@ -94,7 +94,7 @@ 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=--X5moNSvtV8Mtzg2fh9v1Ej1iyYyocQwo5pHlhkPqo,102230
+ipex_llm/transformers/convert.py,sha256=294hk2uMQPN0DrPIpqhfgFvR7klvlHKS53DUUhoWaeU,103273
 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
@@ -113,6 +113,7 @@ ipex_llm/transformers/speculative.py,sha256=0XNLgc9dGswJHVPrXo4iM7pPxkWwfFfJMECc
 ipex_llm/transformers/streamer.py,sha256=RrVlLblzCOtABRUpaMXAyaMnCGgLUtAi_YesLumRbww,4842
 ipex_llm/transformers/training_patch.py,sha256=oxMkUtqyvqJiprw6dE3skkYfD1HOmUlH9N0hBkbn0G0,10799
 ipex_llm/transformers/utils.py,sha256=a-2wbflSd_yYnC5qcMoY5HLR1yT_QpxeX_WpGpaDLrA,17457
+ipex_llm/transformers/xgrammar.py,sha256=dd-e0DO0s-t-idngVzncnPAa_Gxb8YNoRJ3WROkwrs8,1840
 ipex_llm/transformers/xpu_customize_fwd.py,sha256=PUBYLnTbaBXUs3Dnte9Gqln2XFk8iA62SmloWjr7GJI,7668
 ipex_llm/transformers/xpu_ops.py,sha256=z95iTtcDQvNyJOvB4A6B_ECTYjHp4A7x-FsssoETOMs,4914
 ipex_llm/transformers/awq/__init__.py,sha256=Du5gu3-eeAkeDO_dEMBTzrDBA66DSN3uL3-rn8WGXQw,875
@@ -152,8 +153,9 @@ ipex_llm/transformers/models/chatglm.py,sha256=UHai1t2AUtGmF765_eHF8LUMVQzp_oCBx
 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=L6y45M_wjS2_HqchmCUxRlQZUNuSNCGOiynAQrGh918,14124
-ipex_llm/transformers/models/common.py,sha256=VKouwfP3q7nQkYnDeDYfjLwKgr_Qk9uc3ZCISaxbfn4,11646
+ipex_llm/transformers/models/common.py,sha256=0OTRaXekOPApRdQ8UKl5Du8DOtKJ6awnQIStvYvFQOI,13018
 ipex_llm/transformers/models/decilm.py,sha256=P-PBuDPf07GvKggLwJx_wPwIn6esN3rX8ai2JxRuZmE,5246
+ipex_llm/transformers/models/deepseek.py,sha256=2w2bWbbuYi__fPs56vE9Wq5bdiZCF2NkYJNXf-b9LjQ,11130
 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
@@ -166,7 +168,7 @@ ipex_llm/transformers/models/internvl.py,sha256=Vx0vENIEQLX2M6P398mw5TOhpks0U8xf
 ipex_llm/transformers/models/janus.py,sha256=0URo2NC8_2CGaOl3CiVB3IFTVsYyplMFgjBJdPDNBsY,1509
 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/minicpm3.py,sha256=37P_yMjw8RIzy27qL_E7kzbQRNW6f0xYQNK9xtoe5kI,10183
 ipex_llm/transformers/models/minicpmv.py,sha256=PP05b5iTnrMpiseCn8iJcxKJDnfq7WqXp9Mrch0kKZ0,9876
 ipex_llm/transformers/models/mistral.py,sha256=uVhkdXaq15v1P3QY0emVsA7SxUbAWChHEEXYN-drjpQ,7449
 ipex_llm/transformers/models/mllama.py,sha256=ZyRq9DTKsvk1AlRbr-z6ngjS3Sr_7YuGZ6-Yr1MBBAM,10937
@@ -260,11 +262,11 @@ 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=D577nxWlyoWaHXNXIEvS3ViKSSWL3XZq8D8t6izD7x4,33250
 ipex_llm/vllm/xpu/entrypoints/openai/cli_args.py,sha256=hB398yYtKauASRzevctScdbFIjiiSGMAe1bwEuIHrhY,10893
-ipex_llm-2.2.0b20250223.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
-ipex_llm-2.2.0b20250223.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
-ipex_llm-2.2.0b20250223.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
-ipex_llm-2.2.0b20250223.dist-info/METADATA,sha256=cD2rMxvz0NtJa5GHFRIrjjpe2SutmWllkgqGNSqrGOs,12369
-ipex_llm-2.2.0b20250223.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
-ipex_llm-2.2.0b20250223.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
-ipex_llm-2.2.0b20250223.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
-ipex_llm-2.2.0b20250223.dist-info/RECORD,,
+ipex_llm-2.2.0b20250225.data/scripts/ipex-llm-init,sha256=fLQsT2dRL6H5bThb4GuIWotAuqoLsIxFwA-0c2qmaO8,6672
+ipex_llm-2.2.0b20250225.data/scripts/llm-chat,sha256=TdUnUmNapzuoe1c8IzrdVOQwWEg8IqsMSBRlOD3daZM,2249
+ipex_llm-2.2.0b20250225.data/scripts/llm-cli,sha256=RXGPlLElHxcKzoUxljEMBIAXbzCDysXL-Nxw-xF-7LU,2457
+ipex_llm-2.2.0b20250225.dist-info/METADATA,sha256=3_EbaWHFTUZ3JXASEqCh8-KfRdJ-s0TRsdOk6L2-Fyo,12369
+ipex_llm-2.2.0b20250225.dist-info/WHEEL,sha256=PPJcBMAZibF_2GFE9NmOJGqiaSMPiNFbJd6QaJjdA6Y,109
+ipex_llm-2.2.0b20250225.dist-info/entry_points.txt,sha256=TiUyBB2MRmfF3ko-pyAEzqeBCRnyhu27bNOAsWPp3e8,61
+ipex_llm-2.2.0b20250225.dist-info/top_level.txt,sha256=CGCMHM-SyqUabU4h8RqJ2KTYckQUO3LvIWwmUQ6Qbzw,9
+ipex_llm-2.2.0b20250225.dist-info/RECORD,,