xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2025-12-13 14:35:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

[FEATURE] support custom vllm tuned config path for fused moe triton kernels (#22791)

Browse Source 
Signed-off-by: Chi Zhang <zhangchi.usc1992@bytedance.com>
This commit is contained in:
Chi Zhang 2025-08-13 20:27:25 +08:00 committed by GitHub
parent 653124bd46
commit 98deac3879
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 26 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/envs.py
View File

@ -158,6 +158,7 @@ if TYPE_CHECKING:
VLLM_USE_TRTLLM_ATTENTION: Optional[str] = None VLLM_USE_TRTLLM_ATTENTION: Optional[str] = None
VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False VLLM_USE_FLASHINFER_MOE_MXFP4_MXFP8: bool = False
VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False VLLM_USE_FLASHINFER_MOE_MXFP4_BF16: bool = False
VLLM_TUNED_CONFIG_FOLDER: Optional[str] = None
def get_default_cache_root(): def get_default_cache_root():
@ -1120,6 +1121,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
# never removed from memory until the server terminates. # never removed from memory until the server terminates.
"VLLM_ENABLE_RESPONSES_API_STORE": "VLLM_ENABLE_RESPONSES_API_STORE":
lambda: bool(int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))), lambda: bool(int(os.getenv("VLLM_ENABLE_RESPONSES_API_STORE", "0"))),
# Allows vllm to find tuned config under customized folder
"VLLM_TUNED_CONFIG_FOLDER":
lambda: os.getenv("VLLM_TUNED_CONFIG_FOLDER", None),
} }
# --8<-- [end:env-vars-definition] # --8<-- [end:env-vars-definition]

28
vllm/model_executor/layers/fused_moe/fused_moe.py
View File

@ -701,20 +701,32 @@ def get_moe_configs(
block_shape = [block_n, block_k] if block_n and block_k else None block_shape = [block_n, block_k] if block_n and block_k else None
json_file_name = get_config_file_name(E, N, dtype, block_shape) json_file_name = get_config_file_name(E, N, dtype, block_shape)
config_file_path = os.path.join( config_file_paths = []
# note that we prioritize user defined config
user_defined_config_folder = envs.VLLM_TUNED_CONFIG_FOLDER
if user_defined_config_folder is not None:
user_defined_config_file_path = os.path.join(
user_defined_config_folder, json_file_name)
config_file_paths.append(user_defined_config_file_path)
default_config_file_path = os.path.join(
os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name) os.path.dirname(os.path.realpath(__file__)), "configs", json_file_name)
if os.path.exists(config_file_path): config_file_paths.append(default_config_file_path)
with open(config_file_path) as f:
logger.info("Using configuration from %s for MoE layer.", for config_file_path in config_file_paths:
config_file_path) if os.path.exists(config_file_path):
# If a configuration has been found, return it with open(config_file_path) as f:
return {int(key): val for key, val in json.load(f).items()} logger.info("Using configuration from %s for MoE layer.",
config_file_path)
# If a configuration has been found, return it
return {int(key): val for key, val in json.load(f).items()}
# If no optimized configuration is available, we will use the default # If no optimized configuration is available, we will use the default
# configuration # configuration
logger.warning( logger.warning(
("Using default MoE config. Performance might be sub-optimal! " ("Using default MoE config. Performance might be sub-optimal! "
"Config file not found at %s"), config_file_path) "Config file not found at %s"), config_file_paths)
return None return None