[Feature][Perf] Support Selective CPU Weight Offloading (#34535)

Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
This commit is contained in:
Wei Zhao
2026-02-13 20:02:24 -08:00
committed by GitHub
parent a0638d052d
commit b37b679770
4 changed files with 44 additions and 2 deletions
+5
View File
@@ -434,6 +434,7 @@ class EngineArgs:
disable_cascade_attn: bool = ModelConfig.disable_cascade_attn
swap_space: float = CacheConfig.swap_space
cpu_offload_gb: float = CacheConfig.cpu_offload_gb
cpu_offload_params: set[str] = get_field(CacheConfig, "cpu_offload_params")
gpu_memory_utilization: float = CacheConfig.gpu_memory_utilization
kv_cache_memory_bytes: int | None = CacheConfig.kv_cache_memory_bytes
max_num_batched_tokens: int | None = None
@@ -942,6 +943,9 @@ class EngineArgs:
"--prefix-caching-hash-algo", **cache_kwargs["prefix_caching_hash_algo"]
)
cache_group.add_argument("--cpu-offload-gb", **cache_kwargs["cpu_offload_gb"])
cache_group.add_argument(
"--cpu-offload-params", **cache_kwargs["cpu_offload_params"]
)
cache_group.add_argument(
"--calculate-kv-scales", **cache_kwargs["calculate_kv_scales"]
)
@@ -1453,6 +1457,7 @@ class EngineArgs:
enable_prefix_caching=self.enable_prefix_caching,
prefix_caching_hash_algo=self.prefix_caching_hash_algo,
cpu_offload_gb=self.cpu_offload_gb,
cpu_offload_params=self.cpu_offload_params,
calculate_kv_scales=self.calculate_kv_scales,
kv_sharing_fast_prefill=self.kv_sharing_fast_prefill,
mamba_cache_dtype=self.mamba_cache_dtype,