ipex-llm 2.2.0b20250210__py3-none-win_amd64.whl → 2.2.0b20250212__py3-none-win_amd64.whl

ipex_llm/transformers/convert.py

@@ -667,7 +667,6 @@ def _replace_with_low_bit_linear(model, qtype, modules_to_not_convert=None,
                             out_features,
                             mp_group,
                             None,
-                            None,
                             optimize_lm_head,
                             None
                         )
@@ -1062,6 +1061,11 @@ def _optimize_pre(model, qtype=None):
         from ipex_llm.transformers.models.glm import merge_qkv, split_mlp
         model.apply(merge_qkv)
         model.apply(split_mlp)
+    elif model.config.model_type == "baichuan_m1":
+        from ipex_llm.transformers.models.baichuan_m1 import pre_register_inv_freq
+        model.apply(pre_register_inv_freq)
+    elif model.config.model_type == "multi_modality":
+        _optimize_pre(model.language_model)
 
     return model
 
@@ -1994,5 +1998,23 @@ def _optimize_post(model):
         model.llm.config.rope_scaling = {"rope_type": "default"}
         _optimize_post(model.llm)
         model.llm.config.model_type = "megrezo"
+    elif model.config.model_type == "baichuan_m1":
+        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.baichuan_m1 import model_forward
+        from ipex_llm.transformers.models.baichuan_m1 import eager_attention_forward
+        convert_forward(model, module.BaichuanModel, model_forward)
+        convert_forward(model, module.BaichuanRMSNorm, rms_norm_forward)
+        convert_forward(model, module.BaichuanAttention, eager_attention_forward)
+    elif model.config.model_type == "multi_modality":
+        # vision
+        vpm_modeling_module_name = model.vision_model.vision_tower.__class__.__module__
+        vpm_module = importlib.import_module(vpm_modeling_module_name)
+        from ipex_llm.transformers.models.janus import vision_attention_forward
+        convert_forward(model.vision_model, vpm_module.Attention, vision_attention_forward)
+
+        # llm
+        _optimize_post(model.language_model)
 
     return model

ipex_llm/transformers/low_bit_linear.py

@@ -699,7 +699,7 @@ class LowBitLinear(nn.Linear):
                 if is_server() and (not is_spr()) and \
                         self.qtype == SYM_INT4 and x_2d.shape[0] >= TORCH_LINEAR_THRESHOLD:
                     x0_fp32 = ggml_int4_convert_fp32(x0, self.weight_shape, self.weight_length)
-                    result = F.linear(x, x0_fp32)
+                    result = F.linear(x.to(dtype=x0_fp32.dtype), x0_fp32)
                 else:
                     # Weight does not need a convert
                     result = ggml_matmul_src1_x_src0_t(x0, x_2d, self.weight_shape, self.qtype)

ipex_llm/transformers/models/baichuan_m1.py

@@ -0,0 +1,240 @@
+#
+# 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.
+
+# This file is adapted from
+# https://huggingface.co/baichuan-inc/Baichuan-M1-14B-Instruct/blob/main/modeling_baichuan.py
+
+
+import math
+import torch
+import torch.nn.functional as F
+
+from typing import Optional, Tuple, Union
+from transformers.cache_utils import Cache
+from transformers.modeling_outputs import BaseModelOutputWithPast
+from ipex_llm.utils.common import invalidInputError
+from ipex_llm.transformers.models.utils import should_use_fuse_rope, repeat_kv
+from ipex_llm.transformers.models.common import attention_softmax
+from ipex_llm.transformers.models.common import scaled_dot_product_attention
+from ipex_llm.transformers.kv import DynamicNormalCache
+
+
+def pre_register_inv_freq(module: torch.nn.Module):
+    if module.__class__.__name__ == "RotaryEmbedding":
+        inv_freq = module.inv_freq
+        del module.inv_freq
+        module.register_buffer("inv_freq", inv_freq, persistent=False)
+
+
+# copied from Baichuan M1
+def custom_convolution(U, K):
+    """
+    U: Input matrix, shape (bs, seq, h, d)
+    K: Convolution kernel, shape (w, h)
+    Returns: Output matrix V, shape (bs, seq, h, d)
+    """
+    # h, w = K.shape
+    w = K.size(-1)
+    padding = (w - 1, 0)
+    U_padded = F.pad(U, (0, 0, 0, 0, *padding))  # Shape becomes (bs, seq+w-1, h, d)
+    U_unfolded = U_padded.unfold(1, w, 1)  # Shape becomes (bs, seq+w-1, h, d, w)
+    V_unfolded = U_unfolded * K  # Shape remains (bs, seq, h, d, w)
+    V = V_unfolded.sum(dim=-1)  # Shape becomes (bs, seq, h, d)
+    return V
+
+
+def model_forward(
+    self,
+    input_ids: torch.LongTensor = None,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    seqlens: Optional[torch.LongTensor] = None,
+    past_key_values: Optional[Cache] = 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,
+    cache_position: Optional[torch.LongTensor] = 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
+    )
+    return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+    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)
+
+    use_cache = use_cache if use_cache is not None else self.config.use_cache
+    use_cache = True if inputs_embeds.device.type == "xpu" else use_cache
+
+    # IPEX-LLM changes start: remove batch multi-pack and use ipex-llm's kv cache
+    # kept for BC (non `Cache` `past_key_values` inputs)
+    if use_cache and not isinstance(past_key_values, DynamicNormalCache):
+        past_key_values = DynamicNormalCache.from_legacy_cache(past_key_values)
+    # IPEX-LLM changes end
+
+    if cache_position is None:
+        past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+        cache_position = torch.arange(
+            past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1],
+            device=inputs_embeds.device
+        )
+    if position_ids is None:
+        position_ids = cache_position.unsqueeze(0)
+
+    causal_mask = self._update_causal_mask(
+        attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+    )
+
+    hidden_states = inputs_embeds
+
+    # create position embeddings to be shared across the decoder layers
+    # position_embeddings = self.rotary_emb(hidden_states, position_ids)
+    position_embeddings = None
+
+    # 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=causal_mask,
+            position_ids=position_ids,
+            seqlens=None,
+            past_key_value=past_key_values,
+            output_attentions=output_attentions,
+            use_cache=use_cache,
+            cache_position=cache_position,
+            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 use_cache else None
+    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 eager_attention_forward(
+    self,
+    hidden_states: torch.Tensor,
+    attention_mask: Optional[torch.Tensor] = None,
+    position_ids: Optional[torch.LongTensor] = None,
+    seqlens: Optional[torch.LongTensor] = None,
+    past_key_value: Optional[Cache] = None,
+    output_attentions: bool = False,
+    use_cache: bool = False,
+    cache_position: Optional[torch.LongTensor] = None,
+    position_embeddings: Optional[Tuple[torch.Tensor, torch.Tensor]]=None,
+):
+    invalidInputError(seqlens is None, "`seq_lens` must be None")
+
+    bsz, q_len, _ = hidden_states.size()
+    qkv = self.W_pack(hidden_states)
+    qkv = qkv.view(bsz, q_len, self.num_heads + 2 * self.num_key_value_heads, self.head_dim)
+    query_states, key_states, value_states = qkv.split([self.num_heads,
+                                                        self.num_key_value_heads,
+                                                        self.num_key_value_heads], dim=2)
+    # q, k, v: [bsz, seq_len, num_heads, head_dim]
+
+    if past_key_value is None or past_key_value.get_seq_length(self.layer_idx) == 0:    # prefill
+        self.last_k = key_states[:, -1:]
+        self.last_v = value_states[:, -1:]
+
+        key_states = custom_convolution(key_states, self.conv_k)
+        value_states = custom_convolution(value_states, self.conv_v)
+    else:
+        new_key_states = (self.conv_k[0, 0, :, 0, :1] * self.last_k +
+                          self.conv_k[0, 0, :, 0, 1:] * key_states)
+        self.last_k = key_states
+        key_states = new_key_states
+
+        new_value_states = (self.conv_v[0, 0, :, 0, : 1] * self.last_v +
+                            self.conv_v[0, 0, :, 0, 1:] * value_states)
+        self.last_v = value_states
+        value_states = new_value_states
+
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+    # q, k, v: [bsz, num_heads, seq_len, head_dim]
+
+    invalidInputError(should_use_fuse_rope(hidden_states, position_ids, self.training),
+                      "fuse rope must be used")
+    import xe_addons
+    xe_addons.rotary_half_inplaced(self.rotary_emb.inv_freq, position_ids,
+                                   query_states, key_states)
+
+    # ignore sliding window
+    key_states, value_states = past_key_value.update(key_states, value_states,
+                                                     self.layer_idx, None)
+    if self.head_dim <= 128:
+        attn_weights = None
+        attn_output = scaled_dot_product_attention(
+            query_states, key_states, value_states,
+            attention_mask, q_len == key_states.size(2)
+        )
+    else:
+        n_rep = self.num_heads // self.num_key_value_heads
+        key_states = repeat_kv(key_states, n_rep)
+        value_states = repeat_kv(value_states, n_rep)
+        attn_weights = torch.matmul(query_states,
+                                    key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+        if attention_mask is not None:
+            attn_weights = attn_weights + attention_mask
+        attn_weights = attention_softmax(attn_weights)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+    attn_output = attn_output.transpose(1, 2).contiguous()
+    attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
+
+    attn_output = self.o_proj(attn_output)
+
+    if not output_attentions:
+        attn_weights = None
+    return attn_output, attn_weights, past_key_value

ipex_llm/transformers/models/janus.py

@@ -0,0 +1,49 @@
+#
+# 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.
+
+# This file is adapted from
+# https://github.com/deepseek-ai/Janus/blob/main/janus/models/siglip_vit.py
+
+import torch
+
+from ipex_llm.transformers.models.common import scaled_dot_product_attention
+
+
+def vision_attention_forward(self, x: torch.Tensor) -> torch.Tensor:
+    B, N, C = x.shape
+    qkv = (
+        self.qkv(x)
+        .reshape(B, N, 3, self.num_heads, self.head_dim)
+        .permute(2, 0, 3, 1, 4)
+    )
+    q, k, v = qkv.unbind(0)
+    q, k = self.q_norm(q), self.k_norm(k)
+
+    if self.fused_attn:
+        # ipex-llm opt: sdpa
+        x = scaled_dot_product_attention(
+            q, k.contiguous(), v.contiguous(), None, False
+        )
+    else:
+        q = q * self.scale
+        attn = q @ k.transpose(-2, -1)
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+        x = attn @ v
+
+    x = x.transpose(1, 2).reshape(B, N, C)
+    x = self.proj(x)
+    x = self.proj_drop(x)
+    return x

ipex_llm/transformers/models/utils.py

@@ -86,7 +86,7 @@ def use_quantize_kv_cache(linear: torch.nn.Module, x: torch.Tensor,
         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"]]:
+    elif linear.weight.dtype != torch.uint8:    # unquantized
         return False
     else:
         device_name = get_xpu_device_name(x.device)

ipex_llm/vllm/xpu/engine/engine.py

@@ -13,18 +13,28 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 #
-from typing import Dict, Optional
+from vllm.logger import init_logger
+from typing import Dict, Optional, Any, Union, Type
 from vllm.engine.llm_engine import LLMEngine
 from vllm.engine.async_llm_engine import AsyncLLMEngine
 from vllm.engine.arg_utils import AsyncEngineArgs, EngineArgs
 from vllm.entrypoints.llm import LLM
 from vllm.utils import Counter
-from vllm.config import EngineConfig
+from vllm.config import VllmConfig
 from ipex_llm.vllm.xpu.model_convert import _ipex_llm_convert
 from vllm.usage.usage_lib import UsageContext
 from vllm.engine.metrics import StatLoggerBase
 from vllm.engine.multiprocessing.engine import MQLLMEngine
 import signal
+from vllm.engine.arg_utils import (EngineArgs, HfOverrides, PoolerConfig,
+                                   TaskOption)
+from vllm.config import CompilationConfig
+from vllm.v1.engine.llm_engine import LLMEngine as V1LLMEngine
+from vllm import envs
+from vllm.v1.engine.async_llm import AsyncLLM
+import os
+
+logger = init_logger(__name__)
 
 
 class IPEXLLMAsyncLLMEngine(AsyncLLMEngine):
@@ -35,7 +45,7 @@ class IPEXLLMAsyncLLMEngine(AsyncLLMEngine):
     def from_engine_args(
         cls,
         engine_args: AsyncEngineArgs,
-        engine_config: Optional[EngineConfig] = None,
+        engine_config: Optional[VllmConfig] = None,
         start_engine_loop: bool = True,
         usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
         load_in_low_bit: str = "sym_int4",
@@ -49,6 +59,27 @@ class IPEXLLMAsyncLLMEngine(AsyncLLMEngine):
                                         usage_context=usage_context, stat_loggers=stat_loggers)
 
 
+class IPEXLLMAsyncV1Engine(AsyncLLM):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    @classmethod
+    def from_engine_args(
+        cls,
+        engine_args: AsyncEngineArgs,
+        engine_config: Optional[VllmConfig] = None,
+        start_engine_loop: bool = True,
+        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
+        load_in_low_bit: str = "sym_int4",
+        stat_loggers: Optional[Dict[str, StatLoggerBase]]=None,  # noqa
+    ) -> "AsyncLLM":
+        _ipex_llm_convert(load_in_low_bit)
+        return super().from_engine_args(engine_args=engine_args, engine_config=engine_config,
+                                        start_engine_loop=start_engine_loop,
+                                        usage_context=usage_context, stat_loggers=stat_loggers)
+
+
 class IPEXLLMClass(LLM):
     def __init__(
         self,
@@ -57,6 +88,7 @@ class IPEXLLMClass(LLM):
         tokenizer_mode: str = "auto",
         skip_tokenizer_init: bool = False,
         trust_remote_code: bool = False,
+        allowed_local_media_path: str = "",
         tensor_parallel_size: int = 1,
         dtype: str = "auto",
         quantization: Optional[str] = None,
@@ -64,28 +96,48 @@ class IPEXLLMClass(LLM):
         tokenizer_revision: Optional[str] = None,
         seed: int = 0,
         gpu_memory_utilization: float = 0.9,
-        swap_space: int = 4,
+        swap_space: float = 4,
         cpu_offload_gb: float = 0,
-        enforce_eager: bool = False,
-        max_context_len_to_capture: Optional[int] = None,
+        enforce_eager: Optional[bool] = None,
         max_seq_len_to_capture: int = 8192,
         disable_custom_all_reduce: bool = False,
+        disable_async_output_proc: bool = True,
+        hf_overrides: Optional[HfOverrides] = None,
+        mm_processor_kwargs: Optional[Dict[str, Any]]=None,
+        # After positional args are removed, move this right below `model`
+        task: TaskOption = "auto",
+        override_pooler_config: Optional[PoolerConfig] = None,
+        compilation_config: Optional[Union[int, Dict[str, Any]]]=None,
         load_in_low_bit: str = "sym_int4",
         **kwargs,
     ) -> None:
+        '''
+        LLM constructor.
+
+        Note: if enforce_eager is unset (enforce_eager is None)
+        it defaults to False.
+        '''
+
         if "disable_log_stats" not in kwargs:
             kwargs["disable_log_stats"] = True
-        removed_vision_keys = ("image_token_id", "image_feature_size",
-                               "image_input_shape", "image_input_type")
-        if any(k in kwargs for k in removed_vision_keys):
-            raise TypeError(  # noqa
-                "There is no need to pass vision-related arguments anymore.")
+
+        if compilation_config is not None:
+            if isinstance(compilation_config, (int, dict)):
+                compilation_config_instance = CompilationConfig.from_cli(
+                    str(compilation_config))
+            else:
+                compilation_config_instance = compilation_config
+        else:
+            compilation_config_instance = None
+
         engine_args = EngineArgs(
             model=model,
+            task=task,
             tokenizer=tokenizer,
             tokenizer_mode=tokenizer_mode,
             skip_tokenizer_init=skip_tokenizer_init,
             trust_remote_code=trust_remote_code,
+            allowed_local_media_path=allowed_local_media_path,
             tensor_parallel_size=tensor_parallel_size,
             dtype=dtype,
             quantization=quantization,
@@ -96,16 +148,53 @@ class IPEXLLMClass(LLM):
             swap_space=swap_space,
             cpu_offload_gb=cpu_offload_gb,
             enforce_eager=enforce_eager,
-            max_context_len_to_capture=max_context_len_to_capture,
             max_seq_len_to_capture=max_seq_len_to_capture,
             disable_custom_all_reduce=disable_custom_all_reduce,
+            disable_async_output_proc=disable_async_output_proc,
+            hf_overrides=hf_overrides,
+            mm_processor_kwargs=mm_processor_kwargs,
+            override_pooler_config=override_pooler_config,
+            compilation_config=compilation_config_instance,
             **kwargs,
         )
-        self.llm_engine = IPEXLLMLLMEngine.from_engine_args(
+        # Logic to switch between engines is done at runtime instead of import
+        # to avoid import order issues
+        self.engine_class = self.get_engine_class()
+        self.llm_engine = self.engine_class.from_engine_args(
             engine_args, usage_context=UsageContext.LLM_CLASS,
             load_in_low_bit=load_in_low_bit)
+
         self.request_counter = Counter()
 
+    @staticmethod
+    def get_engine_class() -> Type[LLMEngine]:
+        if envs.VLLM_USE_V1:
+            return IPEXLLMLLMV1Engine
+        return IPEXLLMLLMEngine
+
+
+class IPEXLLMLLMV1Engine(V1LLMEngine):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    @classmethod
+    def from_engine_args(
+        cls,
+        engine_args: EngineArgs,
+        usage_context: UsageContext = UsageContext.ENGINE_CONTEXT,
+        stat_loggers: Optional[Dict[str, StatLoggerBase]]=None,
+        enable_multiprocessing: bool = False,
+        load_in_low_bit: str = "sym_int4",
+    ) -> "LLMEngine":
+        """Creates an LLM engine from the engine arguments."""
+        # Create the engine configs.
+
+        _ipex_llm_convert(load_in_low_bit)
+        return super().from_engine_args(engine_args,
+                                        usage_context,
+                                        stat_loggers,
+                                        enable_multiprocessing)
+
 
 class IPEXLLMLLMEngine(LLMEngine):
     def __init__(self, *args, **kwargs):
@@ -134,16 +223,24 @@ class IPEXLLMMQLLMEngine(MQLLMEngine):
 
 
 def run_mp_engine(engine_args: AsyncEngineArgs, usage_context: UsageContext,
-                  ipc_path: str, load_in_low_bit: str):
+                  ipc_path: str, load_in_low_bit: str, engine_alive):
 
     def signal_handler(*_) -> None:
         # Interrupt server on sigterm
         raise KeyboardInterrupt("MQLLMEngine terminated")  # noqa
 
-    signal.signal(signal.SIGTERM, signal_handler)
+    try:
+        signal.signal(signal.SIGTERM, signal_handler)
+
+        engine = IPEXLLMMQLLMEngine.from_engine_args(engine_args=engine_args,
+                                                     usage_context=usage_context,
+                                                     ipc_path=ipc_path,
+                                                     load_in_low_bit=load_in_low_bit)
+        engine.start()
+    except BaseException as e:
+        logger.exception(e)
+        engine_alive.value = False
+        raise e  # noqa
 
-    engine = IPEXLLMMQLLMEngine.from_engine_args(engine_args=engine_args,
-                                                 usage_context=usage_context,
-                                                 ipc_path=ipc_path,
-                                                 load_in_low_bit=load_in_low_bit)
-    engine.start()
+if os.getenv("VLLM_USE_V1"):
+    IPEXLLMAsyncLLMEngine = IPEXLLMAsyncV1Engine