update version

README.md

@@ -82,7 +82,7 @@ VLLM_INSTALL_PUNICA_KERNELS=1 python3 setup.py install
 + 若使用 pip install 下载安装过慢，可添加源：-i https://pypi.tuna.tsinghua.edu.cn/simple/
 
 ## 验证
-- python -c "import vllm; print(vllm.\_\_version__)"，版本号与官方版本同步，查询该软件的版本号，例如0.5.5；
+- python -c "import vllm; print(vllm.\_\_version__)"，版本号与官方版本同步，查询该软件的版本号，例如0.6.0；
 
 ## Known Issue
 - 无

benchmarks/kernels/benchmark_paged_attention.py

@@ -7,6 +7,7 @@ import torch
 from vllm import _custom_ops as ops
 from vllm.utils import (STR_DTYPE_TO_TORCH_DTYPE, FlexibleArgumentParser,
                         create_kv_caches_with_random)
+import vllm.envs as envs
 
 NUM_BLOCKS = 1024
 PARTITION_SIZE = 512
@@ -104,6 +105,24 @@ def main(
 
         for _ in range(num_iters):
             if version == "v1":
+                if envs.VLLM_USE_OPT_OP:
+                    ops.paged_attention_v1_opt(
+                        output,
+                        query,
+                        key_cache,
+                        value_cache,
+                        num_kv_heads,
+                        scale,
+                        block_tables,
+                        seq_lens,
+                        block_size,
+                        max_seq_len,
+                        alibi_slopes,
+                        kv_cache_dtype,
+                        k_scale,
+                        v_scale,
+                    )
+                else:
                     ops.paged_attention_v1(
                     output,
                     query,
@@ -121,6 +140,7 @@ def main(
                     v_scale,
                 )
             elif version == "v2":
+                if envs.VLLM_USE_OPT_OP:
                     ops.paged_attention_v2(
                         output,
                         exp_sums,
@@ -140,6 +160,26 @@ def main(
                         k_scale,
                         v_scale,
                     )
+                else:
+                    ops.paged_attention_v2_opt(
+                    output,
+                    exp_sums,
+                    max_logits,
+                    tmp_output,
+                    query,
+                    key_cache,
+                    value_cache,
+                    num_kv_heads,
+                    scale,
+                    block_tables,
+                    seq_lens,
+                    block_size,
+                    max_seq_len,
+                    alibi_slopes,
+                    kv_cache_dtype,
+                    k_scale,
+                    v_scale,
+                )
             else:
                 raise ValueError(f"Invalid version: {version}")
         torch.cuda.synchronize()

setup.py

@@ -404,8 +404,8 @@ except Exception as e:
                   stacklevel=2)
     __commit__ = "COMMIT_HASH_PLACEHOLDER"
 
-__version__ = "0.5.5"
-__dcu_version__ = f'0.5.5+{version}' 
+__version__ = "0.6.0"
+__dcu_version__ = f'0.6.0+{version}' 
 
 """
     

vllm/engine/llm_engine.py

@@ -166,6 +166,24 @@ class LLMEngine:
 
     @classmethod
     def validate_outputs(
+        cls,
+        outputs: GenericSequence[object],
+        output_type: Type[_O],
+    ) -> List[_O]:
+        do_validate = cls.DO_VALIDATE_OUTPUT
+
+        outputs_: List[_O]
+        if TYPE_CHECKING or do_validate:
+            outputs_ = []
+            for output in outputs:
+                if not isinstance(output, output_type):
+                    raise TypeError(f"Expected output of type {output_type}, "
+                                    f"but found type {type(output)}")
+
+                outputs_.append(output)
+        else:
+            outputs_ = outputs
+
         return outputs_
 
     tokenizer: Optional[BaseTokenizerGroup]