ipex-llm 2.2.0b20250107__py3-none-win_amd64.whl → 2.2.0b20250109__py3-none-win_amd64.whl

ipex_llm/transformers/models/internlm.py

@@ -87,7 +87,8 @@ def internlm_attention_forward(
         )
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.qkv_proj, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.qkv_proj, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, hidden_states.device
@@ -171,7 +172,8 @@ def internlm2_attention_forward(
         )
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states,
+                                            self.num_heads, self.num_key_value_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, hidden_states.device
@@ -346,7 +348,8 @@ def internlm_xcomposser2_attention_forward(
             query_states, key_states, cos, sin, position_ids, "internlm")
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.wqkv, hidden_states,
+                                            self.num_heads, self.num_key_value_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, device

ipex_llm/transformers/models/llama.py

@@ -72,7 +72,7 @@ def llama_model_forward(
     use_cache = True if inputs.device.type == "xpu" else use_cache
     use_quantize_kv = use_quantize_kv_cache(
         self.layers[0].mlp.down_proj, inputs,
-        self.config.num_attention_heads // self.config.num_key_value_heads
+        self.config.num_attention_heads, self.config.num_key_value_heads
     )
     use_compresskv = should_use_compresskv(inputs, inputs.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)
@@ -116,7 +116,7 @@ def llama_model_forward(
 
 
 def merge_qkv(module: torch.nn.Module):
-    return merge_qkv_base(module, LlamaAttention)
+    merge_qkv_base(module, LlamaAttention)
 
 
 def llama_attention_forward(

ipex_llm/transformers/models/minicpm.py

@@ -51,7 +51,8 @@ from transformers.cache_utils import Cache
 
 
 def merge_qkv(module: torch.nn.Module):
-    return merge_qkv_base(module, "MiniCPMAttention")
+    merge_qkv_base(module, "MiniCPMAttention")
+    merge_qkv_base(module, "MiniCPMSdpaAttention")
 
 
 def apply_residual_scale(module: torch.nn.Module):
@@ -158,7 +159,7 @@ def minicpm_model_forward_wrapper(origin_forward):
         # IPEX-LLM OPT: kv cache and quantize kv cache
         inputs = input_ids if input_ids is not None else inputs_embeds
         use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs,
-                                                self.config.num_attention_heads //
+                                                self.config.num_attention_heads,
                                                 self.config.num_key_value_heads)
         use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
             isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/minicpm3.py

@@ -66,7 +66,9 @@ def minicpm3_model_forward_wrapper(origin_forward):
         inputs = input_ids if input_ids is not None else inputs_embeds
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         use_cache = True if inputs.device.type == "xpu" else use_cache
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs)
+        num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
+                                                num_heads, num_kv_heads)
         if use_cache:
             if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
                 past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

ipex_llm/transformers/models/minicpmv.py

@@ -36,6 +36,7 @@ from transformers.generation.logits_process import RepetitionPenaltyLogitsProces
 # MiniCPM-V-2_5 and MiniCPM-V-2_6
 def merge_qkv(module: torch.nn.Module):
     merge_qkv_base(module, "SiglipAttention")
+    merge_qkv_base(module, "SiglipSdpaAttention")
     merge_qkv_base(module, "Idefics2VisionAttention")
 
 

ipex_llm/transformers/models/mistral.py

@@ -71,7 +71,7 @@ def mistral_model_forward(
     use_cache = use_cache if use_cache is not None else self.config.use_cache
     use_cache = use_cache or inputs.device.type == 'xpu'
     use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
-                                            self.config.num_attention_heads //
+                                            self.config.num_attention_heads,
                                             self.config.num_key_value_heads)
     use_compress_kv = should_use_compresskv(inputs, inputs.size(1)) or \
         isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/mllama.py

@@ -113,7 +113,7 @@ def mllama_text_model_forward(
     use_cache = True if inputs.device.type == "xpu" else use_cache
     use_quantize_kv = use_quantize_kv_cache(
         self.layers[0].mlp.down_proj, inputs,
-        self.config.num_attention_heads // self.config.num_key_value_heads
+        self.config.num_attention_heads, self.config.num_key_value_heads
     )
     if use_cache:
         if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):

ipex_llm/transformers/models/phi3.py

@@ -249,7 +249,9 @@ def phi3_model_forward_wrapper(origin_model_forward):
         # IPEX-LLM OPT: kv cache and quantize kv cache and sdp
         use_cache = use_cache if use_cache is not None else self.config.use_cache
         inputs = input_ids if input_ids is not None else inputs_embeds
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs)
+        num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
+                                                num_heads, num_kv_heads)
         use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
             isinstance(past_key_values, DynamicCompressCache)
         if use_cache:
@@ -305,7 +307,9 @@ def phi3v_model_forward_wrapper(origin_model_forward):
     ):
         # IPEX-LLM OPT: kv cache and quantize kv cache and sdp
         use_cache = use_cache if use_cache is not None else self.config.use_cache
-        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, input_ids)
+        num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+        use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, input_ids,
+                                                num_heads, num_kv_heads)
         if use_cache:
             if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
                 past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

ipex_llm/transformers/models/qwen.py

@@ -107,7 +107,8 @@ def qwen_attention_forward(
         query_states = query_states * logn_tensor.type_as(query_states).expand_as(query_states)
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.c_attn, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.c_attn, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, device
@@ -205,7 +206,8 @@ def qwen_attention_forward_registered(
         query_states = query_states * logn_tensor.type_as(query_states).expand_as(query_states)
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.c_attn, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.c_attn, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         past_key_value, key_states, value_states,
         kv_seq_len, use_quantize_kv, device

ipex_llm/transformers/models/qwen2.py

@@ -113,10 +113,10 @@ def qwen2_model_forward(
     # ipex-llm changes start
     # IPEX-LLM OPT: kv cache and quantize kv cache
     inputs = input_ids if input_ids is not None else inputs_embeds
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
     use_quantize_kv = (
         self.config.hidden_size != 3584     # disable quantize kv in specific model
-        and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs,
-                                  self.config.num_attention_heads//self.config.num_key_value_heads)
+        and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs, num_heads, num_kv_heads)
     )
     use_compress_kv = should_use_compresskv(inputs, inputs.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)
@@ -305,10 +305,11 @@ def qwen2_model_forward_4_42(
 
     # ipex-llm changes start
     # IPEX-LLM OPT: kv cache and quantize kv cache
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
     use_quantize_kv = (
         self.config.hidden_size != 3584     # disable quantize kv in specific model
         and use_quantize_kv_cache(self.layers[0].mlp.up_proj, inputs_embeds,
-                                  self.config.num_attention_heads//self.config.num_key_value_heads)
+                                  num_heads, num_kv_heads)
     )
     use_compress_kv = should_use_compresskv(inputs_embeds, inputs_embeds.shape[1]) or \
         isinstance(past_key_values, DynamicCompressCache)

ipex_llm/transformers/models/qwen2_moe.py

@@ -73,8 +73,10 @@ def qwen2moe_model_forward(
     return_dict: Optional[bool] = None,
 ):
     use_cache = use_cache if use_cache is not None else self.config.use_cache
-    input = input_ids if input_ids is not None else inputs_embeds
-    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.shared_expert.up_proj, input)
+    inputs = input_ids if input_ids is not None else inputs_embeds
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.shared_expert.up_proj, inputs,
+                                            num_heads, num_kv_heads)
     if use_cache:
         if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
             past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

ipex_llm/transformers/models/qwen2_vl.py

@@ -88,7 +88,9 @@ def qwen2_vl_model_forward(
     # IPEX-LLM OPT start: kv cache and quantize kv cache
     inputs = input_ids if input_ids is not None else inputs_embeds
     use_cache = True if inputs.device.type == "xpu" else use_cache
-    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs)
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.down_proj, inputs,
+                                            num_heads, num_kv_heads)
     if use_cache:
         if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
             past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

ipex_llm/transformers/models/stablelm.py

@@ -69,8 +69,10 @@ def stablelm_model_forward(
 ):
     # IPEX-LLM OPT: kv cache and quantize kv cache
     use_cache = use_cache if use_cache is not None else self.config.use_cache
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
     use_quantize_kv = (self.layers[0].self_attn.head_dim in [64, 80, 96, 128]
-                       and use_quantize_kv_cache(self.layers[0].mlp.up_proj, input_ids))
+                       and use_quantize_kv_cache(self.layers[0].mlp.up_proj, input_ids,
+                                                 num_heads, num_kv_heads))
     if use_cache:
         if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
             past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

ipex_llm/transformers/models/starcoder2.py

@@ -132,7 +132,9 @@ def model_forward(
     return_dict: Optional[bool] = None,
 ):
     use_cache = use_cache if use_cache is not None else self.config.use_cache
-    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.c_fc, input_ids)
+    num_heads, num_kv_heads = self.config.num_attention_heads, self.config.num_key_value_heads
+    use_quantize_kv = use_quantize_kv_cache(self.layers[0].mlp.c_fc, input_ids,
+                                            num_heads, num_kv_heads)
     if use_cache:
         if use_quantize_kv and not isinstance(past_key_values, DynamicFp8Cache):
             past_key_values = DynamicFp8Cache.from_legacy_cache(past_key_values)

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_name
+from ipex_llm.transformers.utils import get_xpu_device_name
 from ipex_llm.transformers.low_bit_linear import SYM_INT4, SYM_INT8, FP8E5, IQ2_XXS, FP4, FP8E4,\
     FP6, ASYM_INT4
 
@@ -74,7 +74,8 @@ def append_kv_cache(cache_k, cache_v, key_states, value_states):
     return new_cache_k, new_cache_v
 
 
-def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor, kv_group: int = 1) -> bool:
+def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor,
+                          num_heads: int, num_kv_heads: int) -> bool:
     if os.environ.get("BIGDL_QUANTIZE_KV_CACHE", None) is not None:
         warnings.warn(
             "`BIGDL_QUANTIZE_KV_CACHE` is deprecated and will be removed in future releases. "
@@ -90,8 +91,11 @@ def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor, kv_group: in
     else:
         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
+            num_kv_heads >= 4
+            and (
+                device_name in ["mtl", "lnl", "arl"] and num_heads // num_kv_heads <= 4
+                or device_name in ["arc", "bmg"] and x.size(0) > 1
+            )
         )
 
 
@@ -168,7 +172,7 @@ 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]
@@ -183,7 +187,7 @@ 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
@@ -192,19 +196,6 @@ 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

ipex_llm/transformers/models/yuan.py

@@ -158,7 +158,8 @@ def yuan_attention_forward(
                                                         "yuan")
 
     # IPEX-LLM OPT: kv cache and quantzie kv cache
-    use_quantize_kv = use_quantize_kv_cache(self.qk_proj, hidden_states)
+    use_quantize_kv = use_quantize_kv_cache(self.qk_proj, hidden_states,
+                                            self.num_heads, self.num_heads)
     key_states, value_states = update_past_key_value(
         None if past_key_value is None else (past_key_value[0], past_key_value[1]),
         key_states, value_states,

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

@@ -154,24 +154,12 @@ def get_autocast_dtype(x):
                           f"Device {x.device} is not supported.")
 
 
-_ipex_version = None
-
-
-def get_ipex_version():
-
-    global _ipex_version
-    if _ipex_version is not None:
-        return _ipex_version
-
-    import intel_extension_for_pytorch as ipex
-    _ipex_version = ipex.__version__
-    return _ipex_version
-
-
 def get_xpu_device_name(device: torch.device):
     if device.type != "xpu":
         return device.type
     else:
+        # possiable device name:
+        # ["arc", "pvc", "mtl", "lnl", "bmg", "arl", "legacy", "unknown"]
         import xe_linear
         return xe_linear.get_xpu_device_name(device)
 

ipex_llm/transformers/xpu_ops.py

@@ -20,9 +20,9 @@ import xe_batch
 import xe_addons
 
 
-@torch.library.register_fake("ipex_llm::forward_new")
-def _(x, weight, qtype, input_size):
-    return torch.empty_like(x)
+# @torch.library.register_fake("ipex_llm::forward_new")
+# def _(x, weight, qtype, input_size):
+#     return ???
 
 
 # @torch.library.register_fake("ipex_llm::dequant")
@@ -32,32 +32,38 @@ def _(x, weight, qtype, input_size):
 
 @torch.library.register_fake("ipex_llm::mlp_forward_xpu")
 def _(x, weight1, weight2, batch_size, state_size, output_size, act_type, qtype):
-    return torch.empty_like(x)
+    return torch.empty([batch_size, output_size],
+                       dtype=x.dtype, device=x.device)
 
 
-# @torch.library.register_fake("ipex_llm::rwkv_linear_attention_v4")
-# def _(time_decay, time_first, key, value, num_state, den_state, max_state)
-    # return ???
+@torch.library.register_fake("ipex_llm::rwkv_linear_attention_v4")
+def _(time_decay, time_first, key, value, num_state, den_state, max_state):
+    return torch.empty_like(key)
 
 
-# @torch.library.register_fake("ipex_llm::rwkv_linear_attention_v5")
-# def _(time_decay, time_first, receptance, key, value, state)
-    # return ???
+@torch.library.register_fake("ipex_llm::rwkv_linear_attention_v5")
+def _(time_decay, time_first, receptance, key, value, state):
+    bsz, n_heads, seq_len, head_dim = key.shape
+    return torch.empty([bsz, seq_len, n_heads, head_dim],
+                       dtype=key.dtype, device=key.device)
 
 
-# @torch.library.register_fake("ipex_llm::rwkv_time_shift")
-# def _(hidden, shifted, mix):
-    # return ???
+@torch.library.register_fake("ipex_llm::rwkv_time_shift")
+def _(hidden, shifted, mix):
+    bsz, seq_len, hidden_size = hidden.shape
+    return torch.empty([mix.size(0), bsz, seq_len, hidden_size],
+                       dtype=hidden.dtype, device=hidden.device)
 
 
-# @torch.library.register_fake("ipex_llm::dequantize_rows")
-# def _(x, weight, qtype, state_size, output_size):
-    # return ???
+@torch.library.register_fake("ipex_llm::dequantize_rows")
+def _(x, weight, qtype, state_size, output_size):
+    return torch.empty([x.size(0), x.size(1), state_size],
+                       dtype=torch.float, device=weight.device)
 
 
-@torch.library.register_fake("ipex_llm::batch_forward")
-def _(x, weight, qtype):
-    return torch.empty_like(x)
+# @torch.library.register_fake("ipex_llm::batch_forward")
+# def _(x, weight, qtype):
+#     return ???
 
 
 @torch.library.register_fake("ipex_llm::sdp")

{ipex_llm-2.2.0b20250107.dist-info → ipex_llm-2.2.0b20250109.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: ipex-llm
-Version: 2.2.0b20250107
+Version: 2.2.0b20250109
 Summary: Large Language Model Develop Toolkit
 Home-page: https://github.com/intel-analytics/ipex-llm
 Author: BigDL Authors
@@ -27,10 +27,10 @@ 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 ==2.6.0b20250107 ; extra == 'cpp'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250109 ; extra == 'cpp'
 Requires-Dist: setuptools ; extra == 'cpp'
 Provides-Extra: cpp-arl
-Requires-Dist: bigdl-core-cpp ==2.6.0b20250107 ; extra == 'cpp-arl'
+Requires-Dist: bigdl-core-cpp ==2.6.0b20250109 ; extra == 'cpp-arl'
 Requires-Dist: setuptools ; extra == 'cpp-arl'
 Requires-Dist: onednn-devel ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
 Requires-Dist: onednn ==2024.1.1 ; (platform_system == "Windows") and extra == 'cpp-arl'
@@ -67,7 +67,7 @@ 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 ==2.6.0b20250107 ; (platform_system == "Windows") and extra == 'npu'
+Requires-Dist: bigdl-core-npu ==2.6.0b20250109 ; (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'
@@ -87,9 +87,9 @@ 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 ==2.6.0b20250107 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250107 ; extra == 'xpu'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250107 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250109 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250109 ; extra == 'xpu'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250109 ; extra == 'xpu'
 Provides-Extra: xpu-2-1
 Requires-Dist: py-cpuinfo ; extra == 'xpu-2-1'
 Requires-Dist: protobuf ; extra == 'xpu-2-1'
@@ -104,9 +104,9 @@ 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 ==2.6.0b20250107 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250107 ; extra == 'xpu-2-1'
-Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250107 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-21 ==2.6.0b20250109 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-batch-21 ==2.6.0b20250109 ; extra == 'xpu-2-1'
+Requires-Dist: bigdl-core-xe-addons-21 ==2.6.0b20250109 ; 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'
@@ -124,7 +124,7 @@ 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 ==2.6.0b20250107 ; extra == 'xpu-2-6'
+Requires-Dist: bigdl-core-xe-all ==2.6.0b20250109 ; extra == 'xpu-2-6'
 Requires-Dist: intel-openmp ; (platform_machine == "x86_64" or platform_machine == "AMD64") and extra == 'xpu-2-6'
 Provides-Extra: xpu-arc
 Requires-Dist: py-cpuinfo ; extra == 'xpu-arc'
@@ -137,9 +137,9 @@ 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 ==2.6.0b20250107 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250107 ; extra == 'xpu-arc'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250107 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250109 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250109 ; extra == 'xpu-arc'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250109 ; 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'
@@ -160,9 +160,9 @@ 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 ==2.6.0b20250107 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250107 ; extra == 'xpu-arl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250107 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250109 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250109 ; extra == 'xpu-arl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250109 ; 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'
@@ -183,9 +183,9 @@ 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 ==2.6.0b20250107 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250107 ; extra == 'xpu-lnl'
-Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250107 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-23 ==2.6.0b20250109 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-batch-23 ==2.6.0b20250109 ; extra == 'xpu-lnl'
+Requires-Dist: bigdl-core-xe-addons-23 ==2.6.0b20250109 ; 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'