A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 13:30 UTC
[caching] Add enable_prompt_embeds and cpu_offload_gb to compile hashes. (#29435)
Signed-off-by: zhxchen17 <zhxchen17@fb.com>
Zhengxu Chen committed
Nov 25, 2025 at 16:46 UTC
0abc79482a6d476f58530907443c46134ba6e2e1
2 files changed
+1
-4
vllm/config/cache.py
+1
-3
@@ -144,7 +144,7 @@ class CacheConfig:
144
145
kv_offloading_backend: KVOffloadingBackend | None = None
146
"""The backend to use for KV cache offloading. Supported backends include
147
- 'native' (vLLM native CPU offloading), 'lmcache' This option must be used
147
+ 'native' (vLLM native CPU offloading), 'lmcache' This option must be used
148
together with kv_offloading_size."""
149
150
def compute_hash(self) -> str:
@@ -167,8 +167,6 @@ class CacheConfig:
167
"num_gpu_blocks_override",
168
"enable_prefix_caching",
169
"prefix_caching_hash_algo",
170
- # `cpu_offload_gb` does not use `torch.compile` yet.
171
- "cpu_offload_gb",
170
"cpu_kvcache_space_bytes",
171
"mamba_page_size_padded",
172
# Post-init/derived counters
vllm/config/model.py
-1
@@ -345,7 +345,6 @@ class ModelConfig:
345
"logprobs_mode",
346
"disable_cascade_attn",
347
"skip_tokenizer_init",
348
- "enable_prompt_embeds",
348
"served_model_name",
349
"config_format",
350
"hf_token",