[Bugfix][V1] Sync mamba_block_size via EngineCoreReadyResponse (#50809)
Signed-off-by: “新颐” <lxy537258@antgroup.com> Co-authored-by: “新颐” <lxy537258@antgroup.com>
This commit is contained in:
@@ -90,6 +90,7 @@ class EngineCoreReadyResponse:
|
||||
instance_id: str
|
||||
supports_lora: bool
|
||||
max_loras: int
|
||||
mamba_block_size: int | None = None
|
||||
# KV cache capacity (None for encoder-only/attention-free models).
|
||||
kv_cache_size_tokens: int | None = None
|
||||
kv_cache_max_concurrency: float | None = None
|
||||
|
||||
@@ -1636,6 +1636,7 @@ class EngineCoreProc(EngineCore):
|
||||
max_model_len=self.vllm_config.model_config.max_model_len,
|
||||
num_gpu_blocks=self.vllm_config.cache_config.num_gpu_blocks or 0,
|
||||
block_size=self.vllm_config.cache_config.block_size,
|
||||
mamba_block_size=self.vllm_config.cache_config.mamba_block_size,
|
||||
dp_stats_address=self.frontend_stats_publish_address,
|
||||
dtype=str(self.vllm_config.model_config.dtype).removeprefix("torch."),
|
||||
vllm_version=VLLM_VERSION,
|
||||
|
||||
@@ -758,6 +758,7 @@ class MPClient(EngineCoreClient):
|
||||
# worker for hybrid Mamba models.
|
||||
cache_config = vllm_config.cache_config
|
||||
cache_config.block_size = response.block_size
|
||||
cache_config.mamba_block_size = response.mamba_block_size
|
||||
# Keep these as per-engine cache_config_info values; do not sum across DP.
|
||||
cache_config.kv_cache_size_tokens = (
|
||||
getattr(cache_config, "kv_cache_size_tokens", None)
|
||||
|
||||
Reference in New Issue
Block a user