Enable ipex-llm optimization for lm head (intel-analytics#11589)

* basic

* Modify convert.py

* fix

python/llm/src/ipex_llm/transformers/convert.py

@@ -149,9 +149,11 @@ def is_linear_module(module):
         from vllm.model_executor.layers.linear import (
             ColumnParallelLinear, RowParallelLinear, QKVParallelLinear, MergedColumnParallelLinear
         )
-
+        from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead
         VLLM_LINEAR_LIST = [
-            ColumnParallelLinear, RowParallelLinear, QKVParallelLinear, MergedColumnParallelLinear
+            ColumnParallelLinear, RowParallelLinear, QKVParallelLinear,
+            MergedColumnParallelLinear,
+            ParallelLMHead
         ]
         if is_module_in_classes(module, VLLM_LINEAR_LIST):
             if 'xpu' in _VLLM_VERSION:
@@ -167,6 +169,12 @@ def is_linear_module(module):
             else:
                 # For vllm cpu
                 tp_size = 1
+            if isinstance(module, ParallelLMHead) and 'xpu' in _VLLM_VERSION:
+                in_features = module.embedding_dim
+                out_features = module.num_embeddings_per_partition
+                result = True
+                mp_group = None
+                return result, (in_features, out_features, mp_group)
             in_features = module.input_size
             out_features = module.output_size
             result = True

python/llm/src/ipex_llm/vllm/xpu/model_convert.py

@@ -15,18 +15,71 @@
 #
 import torch
 from vllm.logger import init_logger
-from vllm.model_executor.models.llama import LlamaMLP, LlamaAttention
-from vllm.model_executor.models.qwen2 import Qwen2MLP, Qwen2Attention
-from vllm.model_executor.models.qwen import QWenMLP, QWenAttention
+from vllm.model_executor.models.llama import LlamaMLP, LlamaAttention, LlamaForCausalLM
+from vllm.model_executor.models.qwen2 import Qwen2MLP, Qwen2Attention, Qwen2ForCausalLM
+from vllm.model_executor.models.qwen import QWenMLP, QWenAttention, QWenLMHeadModel
 from vllm.model_executor.models.baichuan import BaiChuanMLP, BaiChuanAttention
-from vllm.model_executor.models.chatglm import GLMMLP, GLMAttention
+from vllm.model_executor.models.baichuan import BaiChuanBaseForCausalLM
+from vllm.model_executor.models.chatglm import GLMMLP, GLMAttention, ChatGLMForCausalLM
 from vllm.model_executor.model_loader import get_model
+from vllm.model_executor.layers.sampler import Sampler
 
 from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
 from vllm.model_executor.input_metadata import InputMetadata
 from vllm.config import DeviceConfig
-from typing import Tuple
+from vllm.model_executor.sampling_metadata import SamplingMetadata
+from vllm.model_executor.parallel_utils.communication_op import tensor_model_parallel_gather
+
+from typing import Tuple, Optional
 from ipex_llm.utils.common import invalidInputError
+from vllm.sequence import SamplerOutput
+
+
+def _Llama_sample(
+    self,
+    hidden_states: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+) -> Optional[SamplerOutput]:
+    next_tokens = self.sampler(self.lm_head, hidden_states,
+                               sampling_metadata)
+    return next_tokens
+
+
+def _Qwen2_sample(
+    self,
+    hidden_states: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+) -> Optional[SamplerOutput]:
+    if self.config.tie_word_embeddings:
+        lm_head_weight = self.model.embed_tokens
+    else:
+        lm_head_weight = self.lm_head
+    next_tokens = self.sampler(lm_head_weight, hidden_states,
+                               sampling_metadata)
+    return next_tokens
+
+
+def _Chatglm_sample(
+    self,
+    hidden_states: torch.Tensor,
+    sampling_metadata: SamplingMetadata,
+) -> Optional[SamplerOutput]:
+    next_tokens = self.sampler(self.transformer.output_layer, hidden_states,
+                               sampling_metadata)
+
+    return next_tokens
+
+
+def _sample_get_logits(self, hidden_states: torch.Tensor, embedding: torch.nn.Module,
+                       embedding_bias: Optional[torch.Tensor]) -> torch.Tensor:
+    logits = embedding(hidden_states)
+    if embedding_bias is not None:
+        logits += embedding_bias
+    logits = tensor_model_parallel_gather(logits)
+    # Remove paddings in vocab (if any).
+    if logits is not None:
+        logits = logits[:, :self.org_vocab_size]
+    return logits
 
 
 def _MLP_forward(self, x):
@@ -139,12 +192,26 @@ def _ChatGLM_Attention_forward(
     GLMAttention: _ChatGLM_Attention_forward
 }
 
+_REPLACED_SAMPLER_LAYERS = {
+    LlamaForCausalLM: _Llama_sample,
+    QWenLMHeadModel: _Llama_sample,
+    ChatGLMForCausalLM: _Chatglm_sample,
+    Qwen2ForCausalLM: _Qwen2_sample,
+    BaiChuanBaseForCausalLM: _Llama_sample,
+}
+
 
 def _model_mlp_convert():
     for module, replaced_func in _REPLACED_MLP_LAYERS.items():
         setattr(module, "forward", replaced_func)
 
 
+def _model_sample_convert():
+    setattr(Sampler, "_get_logits", _sample_get_logits)
+    for module, replaced_func in _REPLACED_SAMPLER_LAYERS.items():
+        setattr(module, "sample", replaced_func)
+
+
 def _model_attention_convert():
     for module, replaced_func in _REPLACED_ATTENTION_LAYERS.items():
         setattr(module, "forward", replaced_func)
@@ -160,6 +227,7 @@ def get_load_function(low_bit):
     def _ipex_llm_load_model(self) -> None:
         _model_mlp_convert()
         _model_attention_convert()
+        _model_sample_convert()
 
         from vllm.utils import measure_device_memory
         with measure_device_memory() as m: