ipex-llm 2.2.0b20250106__py3-none-win_amd64.whl → 2.2.0b20250106.post1__py3-none-win_amd64.whl

ipex_llm/transformers/models/chatglm2.py

@@ -269,7 +269,7 @@ def chatglm2_attention_forward(
     # IPEX-LLM OPT: fuse rope
     inv_freq, position_ids = rotary_pos_emb
     rot_dim = inv_freq.size(-1) * 2
-    if should_use_fuse_rope(hidden_states, rotary_pos_emb[1], self.training):
+    if should_use_fuse_rope(hidden_states, position_ids, self.training):
         import xe_addons
         xe_addons.rotary_two_inplaced(inv_freq, position_ids,
                                       query_states[..., :rot_dim], key_states[..., :rot_dim])
@@ -321,197 +321,6 @@ def chatglm2_attention_forward(
     return output, past_key_value
 
 
-@torch.jit.script
-def apply_rotary_pos_emb_original(x: torch.Tensor, rope_cache: torch.Tensor) -> torch.Tensor:
-    # x: [sq, b, np, hn]
-    sq, b, np, hn = x.size(0), x.size(1), x.size(2), x.size(3)
-    rot_dim = rope_cache.shape[-2] * 2
-    x, x_pass = x[..., :rot_dim], x[..., rot_dim:]
-    # truncate to support variable sizes
-    rope_cache = rope_cache[:sq]
-    xshaped = x.reshape(sq, -1, np, rot_dim // 2, 2)
-    rope_cache = rope_cache.view(sq, -1, 1, xshaped.size(3), 2)
-    x_out2 = torch.stack(
-        [
-            xshaped[..., 0] * rope_cache[..., 0] - xshaped[..., 1] * rope_cache[..., 1],
-            xshaped[..., 1] * rope_cache[..., 0] + xshaped[..., 0] * rope_cache[..., 1],
-        ],
-        -1,
-    )
-    x_out2 = x_out2.flatten(3)
-    return torch.cat((x_out2, x_pass), dim=-1)
-
-
-def codegeex_model_forward(
-    self,
-    input_ids,
-    position_ids: Optional[torch.Tensor]=None,
-    attention_mask: Optional[torch.BoolTensor]=None,
-    full_attention_mask: Optional[torch.BoolTensor]=None,
-    past_key_values: Optional[Tuple[Tuple[torch.Tensor, torch.Tensor], ...]]=None,
-    inputs_embeds: Optional[torch.Tensor]=None,
-    use_cache: Optional[bool]=None,
-    output_hidden_states: Optional[bool]=None,
-    return_dict: Optional[bool]=None,
-):
-    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
-
-    if inputs_embeds is None:
-        batch_size, seq_length = input_ids.shape
-        inputs_embeds = self.embedding(input_ids)
-    else:
-        inputs_embeds = inputs_embeds.transpose(0, 1).contiguous()
-        seq_length, batch_size, _ = inputs_embeds.shape
-        input_ids = torch.empty((batch_size, seq_length),
-                                dtype=inputs_embeds.dtype, device=inputs_embeds.device)
-
-    if full_attention_mask is None:
-        if (attention_mask is not None and not attention_mask.all()) or (
-                past_key_values and seq_length != 1):
-            full_attention_mask = self.get_masks(input_ids,
-                                                 past_key_values,
-                                                 padding_mask=attention_mask)
-
-    # ipex-llm changes begin
-    # 1. replace `rotary_pos_emb` with `inv_freq` and `position_ids`
-    # 2. generate `causal_mask` and replace `full_attention_mask` with it
-    if position_ids is None:
-        if past_key_values is None:
-            position_ids = torch.arange(seq_length, dtype=torch.int64, device=inputs_embeds.device)
-        else:
-            if isinstance(past_key_values, DynamicCompressCache):
-                kv_length = past_key_values.get_seq_length()
-            else:
-                kv_length = past_key_values[0][0].size(0)
-            position_ids = torch.arange(kv_length, kv_length + seq_length,
-                                        dtype=torch.int64, device=inputs_embeds.device)
-        position_ids = position_ids.repeat(batch_size, 1)
-    use_fuse_rope = input_ids.device.type == "xpu" and not self.training
-
-    # Rotary positional embeddings
-    rotary_pos_emb = self.rotary_pos_emb(self.seq_length)
-    if position_ids is not None:
-        rotary_pos_emb = rotary_pos_emb[position_ids]
-    else:
-        rotary_pos_emb = rotary_pos_emb[None, :seq_length]
-    if use_fuse_rope:
-        # Repeat cos sin here, call only once for each token.
-        # Chatglm2's rotary embedding is similar to gptj's, is rotate_every_two.
-        # If put this to attension forward, it will generate too many times.
-        cos, sin = rotary_pos_emb.split(rotary_pos_emb.shape[-1] // 2, dim=-1)
-        cos = cos.squeeze(-1)
-        sin = sin.squeeze(-1)
-        cos = torch.repeat_interleave(cos[:, :, None, :], 2, 3)
-        sin = torch.repeat_interleave(sin[:, :, None, :], 2, 3)
-        rotary_pos_emb = (cos, sin)
-    else:
-        rotary_pos_emb = rotary_pos_emb.transpose(0, 1).contiguous()
-
-    # `full_attention_mask` is not None only when
-    #  `past_key_values` is not None and `seq_length` > 1
-    if full_attention_mask is not None:
-        causal_mask = torch.zeros([batch_size, 1, seq_length, full_attention_mask.size(-1)],
-                                  dtype=inputs_embeds.dtype, device=inputs_embeds.device)
-        mask_value = torch.finfo(inputs_embeds.dtype).min
-        causal_mask.masked_fill_(full_attention_mask, mask_value)
-    elif self.training or (inputs_embeds.device.type != "xpu" and past_key_values is None):
-        full_attention_mask = self.get_masks(input_ids,
-                                             past_key_values,
-                                             padding_mask=attention_mask)
-        causal_mask = torch.zeros([batch_size, 1, seq_length, full_attention_mask.size(-1)],
-                                  dtype=inputs_embeds.dtype, device=inputs_embeds.device)
-        mask_value = torch.finfo(inputs_embeds.dtype).min
-        causal_mask.masked_fill_(full_attention_mask, mask_value)
-    else:
-        causal_mask = None
-
-    # Run encoder.
-    hidden_states, presents, all_hidden_states, all_self_attentions = self.encoder(
-        inputs_embeds, causal_mask,
-        rotary_pos_emb=rotary_pos_emb,
-        kv_caches=past_key_values, use_cache=use_cache, output_hidden_states=output_hidden_states
-    )
-    # ipex-llm changes end
-
-    if not return_dict:
-        return tuple(v for v in [hidden_states, presents, all_hidden_states, all_self_attentions]
-                     if v is not None)
-
-    return BaseModelOutputWithPast(
-        last_hidden_state=hidden_states,
-        past_key_values=presents,
-        hidden_states=all_hidden_states,
-        attentions=all_self_attentions,
-    )
-
-
-def codegeex_attention_forward(
-    self, hidden_states, attention_mask, rotary_pos_emb, kv_cache=None, use_cache=True
-):
-    q_len, bsz, _ = hidden_states.size()
-    n_head = self.num_attention_heads_per_partition
-    n_kv_head = self.num_multi_query_groups_per_partition if self.multi_query_attention else n_head
-    head_dim = self.hidden_size_per_attention_head
-
-    past_key_value = None if kv_cache is None else (kv_cache[0].permute(1, 2, 0, 3),
-                                                    kv_cache[1].permute(1, 2, 0, 3))
-    qkv = self.query_key_value(hidden_states)
-    qkv = qkv.view(q_len, bsz, n_head + 2 * n_kv_head, head_dim)
-    # [seq_len, bsz, n_head, head_dim] -> [bsz, n_head, seq_len, head_dim]
-    qkv = qkv.permute(1, 2, 0, 3)
-    query_layer, key_layer, value_layer = qkv.split([n_head,
-                                                     n_kv_head,
-                                                     n_kv_head], dim=1)
-    kv_seq_len = key_layer.shape[2]
-    if past_key_value is not None:
-        kv_seq_len += past_key_value[0].shape[2]
-
-    # apply relative positional encoding (rotary embedding)
-    if len(rotary_pos_emb) == 2 and isinstance(rotary_pos_emb, tuple):
-        cos, sin = rotary_pos_emb
-        rot_dim = cos.shape[-1]
-        query_layer = query_layer.transpose(1, 2)
-        key_layer = key_layer.transpose(1, 2)
-        query_layer_cur = query_layer[..., :rot_dim]
-        key_layer_cur = key_layer[..., :rot_dim]
-        # ipex_llm's apply_rotary_embedding can change the origin storage,
-        # so query_layer will get the result directly.
-        torch.ops.torch_ipex.apply_rotary_embedding(query_layer_cur, sin, cos, query_layer_cur)
-        torch.ops.torch_ipex.apply_rotary_embedding(key_layer_cur, sin, cos, key_layer_cur)
-        query_layer = query_layer.transpose(1, 2)
-        key_layer = key_layer.transpose(1, 2)
-    else:
-        query_layer = apply_rotary_pos_emb_original(query_layer, rotary_pos_emb)
-        key_layer = apply_rotary_pos_emb_original(key_layer, rotary_pos_emb)
-
-    key_layer, value_layer = update_past_key_value(
-        past_key_value, key_layer, value_layer,
-        kv_seq_len, False, hidden_states.device
-    )
-    # past_key_value: [bsz, n_kv_head, seq_len, head_dim] -> [seq_len, bsz, n_kv_head, head_dim]
-    past_key_value = (key_layer.permute(2, 0, 1, 3),
-                      value_layer.permute(2, 0, 1, 3)) if use_cache else None
-
-    # =================
-    # Output. [sq, b, h]
-    # =================
-    context_layer = scaled_dot_product_attention(
-        query_layer, key_layer, value_layer,
-        attention_mask, q_len == kv_seq_len
-    )
-
-    context_layer = context_layer.permute(2, 0, 1, 3).contiguous().view(q_len,
-                                                                        bsz,
-                                                                        n_head * head_dim)
-    output = self.dense(context_layer)
-
-    return output, past_key_value
-
 import torch.nn.functional as F
 
 

ipex_llm/transformers/models/minicpmv.py

@@ -53,10 +53,10 @@ def siglip_attention_forward(
     qkv = qkv.transpose(1, 2)
     query_states, key_states, value_states = qkv.chunk(3, dim=1)
 
-    from ipex_llm.transformers.utils import get_xpu_device_type
+    from ipex_llm.transformers.utils import get_xpu_device_name
     if (
         self.head_dim == 72
-        and get_xpu_device_type(query_states) in ["arc", "flex"] and
+        and get_xpu_device_name(query_states.device) == "arc" and
         query_states.dtype in [torch.float, torch.half]
     ):
         n_heads, kv_length = query_states.size(1), key_states.size(2)

ipex_llm/transformers/models/sd.py

@@ -36,7 +36,7 @@ import math
 import torch
 from typing import Optional
 
-from ipex_llm.transformers.utils import get_xpu_device_type
+from ipex_llm.transformers.utils import get_xpu_device_name
 from ipex_llm.transformers.models.common import padding_qkv_hd
 from ipex_llm.transformers.models.common import scaled_dot_product_attention
 from diffusers.models.attention_processor import Attention
@@ -144,7 +144,7 @@ class AttnProcessor2_0:
 
 def upcast_vae(self):
     # workaround overflow and ipex's bugs
-    if get_xpu_device_type(self.vae.post_quant_conv.weight) in ["arc", "flex", "pvc"]:
+    if get_xpu_device_name(self.vae.post_quant_conv.weight.device) == "arc":
         self.vae.to(torch.bfloat16)
     else:
         self.vae.decoder.up_blocks.to(torch.bfloat16)

ipex_llm/transformers/models/utils.py

@@ -19,7 +19,7 @@ import torch
 import warnings
 from ipex_llm.utils.common import invalidInputError
 from ipex_llm.ggml.quantize import ggml_tensor_qtype
-from ipex_llm.transformers.utils import get_ipex_version, get_xpu_device_type
+from ipex_llm.transformers.utils import get_ipex_version, get_xpu_device_name
 from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4,\
     FP6, ASYM_INT4
 
@@ -85,16 +85,14 @@ def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor, kv_group: in
         return os.environ["IPEX_LLM_QUANTIZE_KV_CACHE"] == "1"
     elif os.environ.get("IPEX_LLM_LOW_MEM", None) is not None:
         return os.environ["IPEX_LLM_LOW_MEM"] == "1"
+    elif linear.qtype in [ggml_tensor_qtype["fp16"], ggml_tensor_qtype["bf16"]]:
+        return False
     else:
-        return x.device.type == 'xpu' and kv_cache_device_check(x, kv_group) \
-            and hasattr(linear, "qtype") and \
-            linear.qtype != ggml_tensor_qtype["fp16"] and linear.qtype != ggml_tensor_qtype["bf16"]
-
-
-def kv_cache_device_check(x: torch.Tensor, kv_group: int) -> bool:
-    return (get_xpu_device_type(x) in ["mtl", "lnl"] and kv_group <= 1) or \
-        ((get_xpu_device_type(x) == "arc" or get_xpu_device_type(x) == "flex") and
-            1 < x.size(0) and x.size(0) <= 8)
+        device_name = get_xpu_device_name(x.device)
+        return (
+            device_name in ["mtl", "lnl", "arl"] and kv_group == 1
+            or device_name in ["arc", "bmg"] and x.size(0) > 1
+        )
 
 
 def init_fp8_kv_cache(batch_size, num_heads, current_length, head_dim, device):
@@ -170,7 +168,7 @@ def should_use_fuse_rope(hidden_states, position_ids, training):
 
 def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
     if model_family in ["llama", "baichuan", "internlm", "aquila", "gpt_neox", "mistral",
-                        "mixtral", "qwen2", "yuan", "stablelm", "qwen2_moe"]:
+                        "qwen2", "yuan", "stablelm", "qwen2_moe"]:
         # The first two dimensions of cos and sin are always 1, so we can `squeeze` them.
         cos = cos.squeeze(1).squeeze(0)  # [seq_len, dim]
         sin = sin.squeeze(1).squeeze(0)  # [seq_len, dim]
@@ -185,7 +183,7 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
         q_embed = (q * cos) + (rotate_half(q) * sin)
         k_embed = (k * cos) + (rotate_half(k) * sin)
         return q_embed, k_embed
-    elif model_family in ["gptj", "chatglm"]:
+    elif model_family in ["chatglm"]:
         q_embed = (q * cos) + (rotate_every_two(q) * sin)
         k_embed = (k * cos) + (rotate_every_two(k) * sin)
         return q_embed, k_embed
@@ -194,19 +192,6 @@ def apply_rotary_pos_emb(q, k, cos, sin, position_ids, model_family):
                           f"{model_family} is not supported.")
 
 
-def apply_ipex_rotate_every_two(q, k, cos, sin):
-    # ipex's apply_rotary_embedding_two_qk can change the origin storage,
-    # so q/k will get the result directly.
-    from ipex_llm.transformers.utils import get_ipex_version
-    if get_ipex_version() >= "2.1.10+xpu":
-        torch.ops.torch_ipex.apply_rotary_embedding_two_qk(
-            q, k, sin, cos, q, k
-        )
-    else:
-        torch.ops.torch_ipex.apply_rotary_embedding(q, sin, cos, q)
-        torch.ops.torch_ipex.apply_rotary_embedding(k, sin, cos, k)
-
-
 def is_enough_kv_cache_room_4_36(past_key_value, idx, seq_len=1):
     # to determinate if is enough kv cache room in transformers==4.36
     # seq_len for current seq len
@@ -226,57 +211,6 @@ def is_enough_kv_cache_room_4_31(past_key_value, seq_len=1):
         (past_key_value[0].size(2) + seq_len) * past_key_value[0].size(3)
 
 
-def use_flash_attention(query, key, attention_mask=None):
-    # here we support query's shape is always [batch_size, head_num, q_len, head_dim],
-    # key's shape is always [batch_size, head_num, k_len, head_dim]
-    invalidInputError(query.dim() == 4,
-                      "Here query input of use_flash_attention should be [batch_size, "
-                      "head_num, q_len, head_dim]")
-    invalidInputError(key.dim() == 4,
-                      "Here key input of use_flash_attention should be [batch_size, "
-                      "head_num, k_len, head_dim]")
-    bsz, _, q_len, _ = query.size()
-    k_len = key.size()[2]
-    # check whether ipex flash attention can be used
-    if q_len != k_len:
-        # now only use flash attention for first token
-        # as it seems have no performance benifit for rest token now
-        return False
-    if query.device.type != "xpu":
-        # ipex flash attention only support for xpu
-        return False
-    ipex_version = get_ipex_version()
-    if ipex_version <= "2.0.110+xpu":
-        # ipex flash attention is supported from ipex 2.1
-        return False
-    if not torch.xpu.has_xetla():
-        # ipex flash attention is only supported for xetla
-        # may update this later
-        return False
-    elif get_xpu_device_type(query) != "pvc":
-        return False
-    if query.dtype not in [torch.float32, torch.float16]:
-        # only use flash attention for fp32/fp16 input
-        return False
-    if bsz > 1:
-        # as flash attention doesn't support attn_mask in ipex 2.1,
-        # so it will cause output error for padded batch input
-        if attention_mask is None:
-            return True
-        else:
-            # TODO: below logic may change for different model
-            # attention mask shape : [bsz, 1, q_len, k_len]
-            if attention_mask[0].squeeze()[0, 0].item() != 0:
-                # first batch contains padding
-                # otherwise we suppose it should be a upper triangular matrix
-                # at the same time, the diagonal is also 0
-                return False
-            elif not attention_mask.equal(attention_mask[0].repeat(bsz, 1, 1, 1)):
-                # check whether mask of every batch is the same
-                return False
-    return True
-
-
 def use_sdp(q_len, kv_len, head_dim, query_states):
     return (
         query_states.device.type == "xpu"
@@ -315,38 +249,16 @@ def mlp_fusion_check(x, qtype, training):
     if training or x.requires_grad:
         return False
     if qtype == FP6:
-        device = get_xpu_device_type(x)
-        if device in ["mtl", "lnl"]:
+        device = get_xpu_device_name(x.device)
+        if device in ["mtl", "lnl", "arl"]:
             return False
     return True
 
 
-def use_decoding_fast_path(proj,
-                           use_fuse_rope,
-                           enough_kv_room,
-                           bs,
-                           qtype_check=decoding_fast_path_qtype_check):
-    if proj is None:
-        return False
-    device = get_xpu_device_type(proj.weight)
-    if not qtype_check(proj):
-        return False
-    if not use_fuse_rope:
-        return False
-    if not enough_kv_room:
-        return False
-    if bs != 1:
-        return False
-
-    if device in ["uhd"]:
-        return False
-    return True
-
-
 def use_xmx(x: torch.Tensor, qtype: int):
-    device = get_xpu_device_type(x)
+    device = get_xpu_device_name(x.device)
     return (
-        device in ["arc", "flex", "pvc"]
+        device in ["arc", "pvc"]
         and qtype in [SYM_INT4, SYM_INT8, FP8E4, FP8E5]
         and (
             (device == "pvc" and 1 < x.size(0) <= 16)
@@ -370,7 +282,7 @@ def fp16_fusion_check(proj, x, training):
         return False
     if x.requires_grad:
         return False
-    device_type = get_xpu_device_type(x)
+    device_type = get_xpu_device_name(x.device)
     if device_type != "pvc":
         return False
     return True
@@ -439,7 +351,7 @@ def should_use_compresskv(x: torch.Tensor, prompt_len: int):
     else:
         if use_compress_kv is None:
             return (
-                get_xpu_device_type(x) in ["mtl", "lnl"]
+                get_xpu_device_name(x.device) in ["mtl", "lnl", "arl"]
                 and prompt_len >= 1800
                 and prompt_len <= 4500
             )

ipex_llm/transformers/npu_pipeline_model/convert_pipeline.py

@@ -473,10 +473,6 @@ def convert_llm_for_deploy(model: torch.nn.Module,
                        "n_splits_linear": n_splits_linear,
                        "n_splits_down_proj": n_splits_down_proj,
                        "lm_head_low_bit": lm_head_low_bit}
-        model.config.update(update_dict)
-        model.config.save_pretrained(save_directory)
-        if model.can_generate():
-            model.generation_config.save_pretrained(save_directory)
 
         from .qwen import convert_qwen_layer, convert_fused_qwen_layer
         from .qwen import convert_lm_head_and_embedding
@@ -537,8 +533,6 @@ def convert_llm_for_deploy(model: torch.nn.Module,
                        "n_splits_linear": n_splits_linear,
                        "n_splits_down_proj": n_splits_down_proj,
                        "lm_head_low_bit": lm_head_low_bit}
-        model.config.update(update_dict)
-        model.config.save_pretrained(save_directory)
 
         from .llama import convert_llama_layer, convert_fused_llama_layer
         from .llama import convert_lm_head_and_embedding
@@ -577,8 +571,6 @@ def convert_llm_for_deploy(model: torch.nn.Module,
                        "n_splits_linear": n_splits_linear,
                        "n_splits_down_proj": n_splits_down_proj,
                        "lm_head_low_bit": lm_head_low_bit}
-        model.config.update(update_dict)
-        model.config.save_pretrained(save_directory)
 
         from .minicpm import convert_minicpm_layer, convert_fused_minicpm_layer
         from .minicpm import convert_lm_head_and_embedding
@@ -595,3 +587,8 @@ def convert_llm_for_deploy(model: torch.nn.Module,
                                       save_directory, weight_dir,
                                       convert_model=True,
                                       max_prompt_len=max_prompt_len)
+
+    model.config.update(update_dict)
+    model.config.save_pretrained(save_directory)
+    if model.can_generate():
+        model.generation_config.save_pretrained(save_directory)

ipex_llm/transformers/speculative.py

@@ -432,8 +432,7 @@ def _check_and_extend_kv_cache(past_key_values, max_step_draft, kv_alloc_block_l
     from ipex_llm.transformers.models.utils import is_enough_kv_cache_room_4_31, \
         extend_kv_cache
     enough_kv_room = True
-    if model_type not in ["chatglm", "qwen", "baichuan", "llama", "mistral",
-                          "gptj", "opt"]:
+    if model_type not in ["chatglm", "qwen", "baichuan", "llama", "mistral", "opt"]:
         return past_key_values, False
     cache_k = past_key_values[0][0]
     if model_type == "chatglm":
@@ -527,7 +526,7 @@ def _crop_past_key_values(self, past_key_values, new_cache_size, _enable_ipex=Fa
                         v[:-(new_cache_size), :, :, :])
                     for k, v in past_key_values
                 ]
-        elif self.config.model_type in ["baichuan", "gptj"]:
+        elif self.config.model_type in ["baichuan"]:
             past_key_values = [
                 (k[:, :, :-(new_cache_size), :],
                     v[:, :, :-(new_cache_size), :])
@@ -796,13 +795,6 @@ def _non_cpu_ipex_verify(self, verify_input_ids, past_key_values, cur_attention_
                                     device=verify_input_ids.device)
         position_ids = position_ids.unsqueeze(0).repeat(1, 1) + past_key_value_len
         forward_args["position_ids"] = position_ids
-    elif self.config.model_type == "gptj":
-        past_length = past_key_values[0][0].size(2)
-        input_len = verify_input_ids.shape[1]
-        position_ids = torch.arange(past_length, input_len + past_length,
-                                    dtype=torch.long, device=verify_input_ids.device)
-        position_ids = position_ids.unsqueeze(0).view(-1, input_len)
-        forward_args["position_ids"] = position_ids
 
     return self(**forward_args)
 
@@ -971,10 +963,6 @@ def speculative_generate(self,
                         past_key_value_len = past_key_values[0][0].shape[0]
                     position_ids = torch.Tensor([[past_key_value_len + step_draft]]).long()
                     forward_args["position_ids"] = position_ids
-                elif self.config.model_type == "gptj":
-                    past_length = draft_past_key_values[0][0].size(2)
-                    position_ids = torch.Tensor([[past_length]]).long().to(self.device)
-                    forward_args["position_ids"] = position_ids
 
                 if _enable_ipex:
                     if any(keyword in self.config.model_type

ipex_llm/transformers/utils.py

@@ -168,27 +168,14 @@ def get_ipex_version():
     return _ipex_version
 
 
-def get_xpu_device_type(x):
-    if x.device.type != "xpu":
-        return x.device.type
-    name = torch.xpu.get_device_name(x.device.index)
-    if name.startswith("Intel(R) Arc(TM) A"):
-        return "arc"
-    elif name.startswith("Intel(R) Graphics [0xe20b]"):
-        return "bmg"
-    elif name.startswith("Intel(R) Arc(TM)"):
-        if 'V' in name:
-            return "lnl"
-        else:
-            return "mtl"
-    elif name.startswith("Intel(R) Data Center GPU Flex"):
-        return "flex"
-    elif name.startswith("Intel(R) Data Center GPU Max"):
-        return "pvc"
-    elif name.startswith("Intel(R) UHD"):
-        return "uhd"
+def get_xpu_device_name(device: torch.device):
+    if device.type != "xpu":
+        return device.type
     else:
-        return "others"
+        # possiable device name:
+        # ["arc", "pvc", "mtl", "lnl", "bmg", "arl", "legacy", "unknown"]
+        import xe_linear
+        return xe_linear.get_xpu_device_name(device)
 
 
 def load_imatrix_data(imatrix_file):