Merge branch 'v0.7.2-dev-wm' into 'v0.7.2-dev'

[feat]添加VLLM_SPEC_DECODE_EAGER环境变量，用于选择draft model是否强制使用eager模式，在hygon cpu上ds3 mtp提升较大

See merge request dcutoolkit/deeplearing/vllm!91

vllm/config.py

@@ -1854,7 +1854,7 @@ class SpeculativeConfig:
                 max_model_len=None,
                 spec_target_max_model_len=target_model_config.max_model_len,
                 quantization=draft_quantization,
-                enforce_eager=target_model_config.enforce_eager,
+                enforce_eager=True if envs.VLLM_SPEC_DECODE_EAGER else target_model_config.enforce_eager,
                 max_seq_len_to_capture=target_model_config.
                 max_seq_len_to_capture,
                 max_logprobs=target_model_config.max_logprobs,

vllm/envs.py

@@ -94,6 +94,7 @@ if TYPE_CHECKING:
     VLLM_ENABLE_MOE_ALIGN_BLOCK_SIZE_TRITON: bool = False
     VLLM_RAY_PER_WORKER_GPUS: float = 1.0
     VLLM_RAY_BUNDLE_INDICES: str = ""
+    VLLM_SPEC_DECODE_EAGER: bool = False
 
 
 def get_default_cache_root():
@@ -613,6 +614,10 @@ environment_variables: Dict[str, Callable[[], Any]] = {
     # models the alignment is already naturally aligned to 256 bytes.
     "VLLM_CUDA_MEM_ALIGN_KV_CACHE":
     lambda: bool(int(os.getenv("VLLM_CUDA_MEM_ALIGN_KV_CACHE", "1"))),
+
+    # If set, vLLM will disable the draft model in cudagraph mode.
+    "VLLM_SPEC_DECODE_EAGER":
+    lambda: bool(int(os.getenv("VLLM_SPEC_DECODE_EAGER", "0"))),
 }
 
 # end-env-vars-definition