LLM: Fix vLLM CPU version error (#11206)

Fix vLLM CPU version error

docker/llm/serving/cpu/docker/start-vllm-service.sh

@@ -11,7 +11,7 @@ python -m ipex_llm.vllm.cpu.entrypoints.openai.api_server \
   --device cpu \
   --dtype bfloat16 \
   --enforce-eager \
-  --load-in-low-bit sym_int4 \
+  --load-in-low-bit bf16 \
   --max-model-len 4096 \
   --max-num-batched-tokens 10240 \
   --max-num-seqs 12 \

docker/llm/serving/cpu/docker/vllm_offline_inference.py

@@ -49,7 +49,7 @@ llm = LLM(model="YOUR_MODEL",
           device="cpu",
           dtype="bfloat16",
           enforce_eager=True,
-          load_in_low_bit="sym_int4",
+          load_in_low_bit="bf16",
           tensor_parallel_size=1)
 # Generate texts from the prompts. The output is a list of RequestOutput objects
 # that contain the prompt, generated text, and other information.

python/llm/example/CPU/vLLM-Serving/offline_inference.py

@@ -46,7 +46,7 @@ sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
 
 # Create an LLM.
 # llm = LLM(model="facebook/opt-125m")
-llm = LLM(model="YOUR_MODEL_PATH", device="cpu", load_in_low_bit="sym_int4")
+llm = LLM(model="YOUR_MODEL_PATH", device="cpu", load_in_low_bit="bf16")
 # Generate texts from the prompts. The output is a list of RequestOutput objects
 # that contain the prompt, generated text, and other information.
 outputs = llm.generate(prompts, sampling_params)

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

@@ -54,6 +54,7 @@ import sys
 
 _IS_VLLM_AVAILABLE = None
 _USE_VLLM = False
+_VLLM_VERSION = None
 
 
 def is_auto_gptq_available():
@@ -77,6 +78,14 @@ def is_vllm_available():
     return _IS_VLLM_AVAILABLE
 
 
+def get_package_version(package_name):
+    result = subprocess.run(['pip', 'list'], capture_output=True, text=True)
+    for line in result.stdout.splitlines():
+        if line.startswith(package_name):
+            return line.split()[1]
+    return None
+
+
 def get_use_vllm():
     return _USE_VLLM
 
@@ -133,13 +142,24 @@ def is_linear_module(module):
     is_awq = is_auto_awq_available() and isinstance(module, WQLinear_GEMM)
     if is_vllm_available():
         # Only convert vllm modules
-        from vllm.model_executor.layers.linear import (
-            ColumnParallelLinear, RowParallelLinear, QKVParallelLinear, MergedColumnParallelLinear
-        )
+        global _VLLM_VERSION
+        if _VLLM_VERSION is None:
+            _VLLM_VERSION = get_package_version('vllm')
+        if 'xpu' in _VLLM_VERSION:
+            # For vllm xpu
             from vllm.model_executor.parallel_utils.parallel_state import (
                 get_tensor_model_parallel_group,
                 get_tensor_model_parallel_world_size
             )
+            tp_size = get_tensor_model_parallel_world_size()
+        else:
+            # For vllm cpu
+            tp_size = 1
+
+        from vllm.model_executor.layers.linear import (
+            ColumnParallelLinear, RowParallelLinear, QKVParallelLinear, MergedColumnParallelLinear
+        )
+
         VLLM_LINEAR_LIST = [
             ColumnParallelLinear, RowParallelLinear, QKVParallelLinear, MergedColumnParallelLinear
         ]
@@ -148,7 +168,6 @@ def is_linear_module(module):
             out_features = module.output_size
             result = True
             mp_group = None
-            tp_size = get_tensor_model_parallel_world_size()
             if isinstance(module, RowParallelLinear) and tp_size >= 2:
                 mp_group = get_tensor_model_parallel_group()
                 in_features = module.input_size_per_partition

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

@@ -24,8 +24,10 @@ from vllm.model_executor.models.chatglm import GLMMLP, GLMAttention
 from vllm.attention import Attention, AttentionMetadata
 from vllm.lora.worker_manager import LRUCacheWorkerLoRAManager
 from vllm.config import DeviceConfig
-from typing import Tuple
+
+from vllm._C import ops
 from ipex_llm.utils.common import invalidInputError
+from typing import List, Optional, Tuple, Union
 
 
 def _MLP_forward(self, x):
@@ -42,7 +44,7 @@ def _Attention_forward(
     kv_cache: torch.Tensor,
     attn_metadata: AttentionMetadata,
 ) -> torch.Tensor:
-    qkv = self.qkv_proj(hidden_states)
+    qkv = self.qkv_proj(hidden_states).to(dtype=kv_cache.dtype)
     q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1)
     q, k = self.rotary_emb(positions, q, k)
     attn_output = self.attn(q, k, v, kv_cache, attn_metadata, self.kv_scale)
@@ -145,9 +147,61 @@ def _model_attention_convert():
 
 
 def _ipex_llm_convert(load_in_low_bit):
-    from vllm.worker.model_runner import ModelRunner
+    from vllm.worker.cpu_model_runner import CPUModelRunner
     import vllm.model_executor.model_loader as model_loader
-    setattr(ModelRunner, "load_model", get_load_function(load_in_low_bit))
+    setattr(CPUModelRunner, "load_model", get_load_function(load_in_low_bit))
+
+    from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
+    setattr(RotaryEmbedding, "forward", _ipex_llm_rotary_embedding_forward)
+    from vllm.model_executor.layers.layernorm import RMSNorm
+    setattr(RMSNorm, "forward", _ipex_llm_rmsnorm_forward)
+
+
+def _ipex_llm_rotary_embedding_forward(
+    self,
+    positions: torch.Tensor,
+    query: torch.Tensor,
+    key: torch.Tensor,
+    offsets: Optional[torch.Tensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    self.cos_sin_cache = self.cos_sin_cache.to(positions.device, dtype=query.dtype)
+
+    # ops.rotary_embedding()/batched_rotary_embedding()
+    # are in-place operations that update the query and key tensors.
+    if offsets is not None:
+        ops.batched_rotary_embedding(positions, query, key, self.head_size,
+                                     self.cos_sin_cache,
+                                     self.is_neox_style, self.rotary_dim,
+                                     offsets)
+    else:
+        ops.rotary_embedding(positions, query, key, self.head_size,
+                             self.cos_sin_cache, self.is_neox_style)
+    return query, key
+
+
+def _ipex_llm_rmsnorm_forward(
+    self,
+    x: torch.Tensor,
+    residual: Optional[torch.Tensor] = None,
+) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
+    x = x.to(dtype=self.weight.data.dtype)
+    if residual is not None:
+        residual = residual.to(dtype=self.weight.data.dtype)
+        ops.fused_add_rms_norm(
+            x,
+            residual,
+            self.weight.data,
+            self.variance_epsilon,
+        )
+        return x, residual
+    out = torch.empty_like(x)
+    ops.rms_norm(
+        out,
+        x,
+        self.weight.data,
+        self.variance_epsilon,
+    )
+    return out
 
 
 def get_load_function(low_bit):
@@ -155,11 +209,15 @@ def get_load_function(low_bit):
         _model_mlp_convert()
         _model_attention_convert()
 
-        self.model = get_model(self.model_config,
-                               self.device_config,
+        self.model = get_model(
+            model_config=self.model_config,
+            load_config=self.load_config,
+            device_config=self.device_config,
+            vision_language_config=self.vision_language_config,
             lora_config=self.lora_config,
             parallel_config=self.parallel_config,
             scheduler_config=self.scheduler_config)
+
         from ipex_llm import optimize_model
         optimize_model(self.model, low_bit=low_bit, torch_dtype=self.model_config.dtype)