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 3, 14:04 UTC
[Bugfix] Ensure embeding model compilation on CPU (#42709)
Signed-off-by: jiang1.li <jiang1.li@intel.com>
Li, Jiang committed
May 15, 2026 at 18:58 UTC
95cfe102a5ffd6bc10c2e897a2d3f3fd3fb250db
1 file changed
+4
-1
vllm/v1/worker/cpu_worker.py
+4
-1
@@ -213,10 +213,13 @@ class CPUWorker(Worker):
213
return kv_cache_size
214
215
def compile_or_warm_up_model(self) -> CompilationTimes:
216
+ # Note: the model has been compiled in determine_available_memory(),
217
+ # Only compile here for models without kv cache
218
+ if len(self.model_runner.kv_caches) == 0:
219
+ self.model_runner.warming_up_model()
220
# Reset the seed to ensure that the random state is not affected by
221
# the model initialization and profiling.
222
set_random_seed(self.model_config.seed)
219
- # Note: the model has been compiled in determine_available_memory()
223
return CompilationTimes(
224
language_model=self.compilation_config.compilation_time,
225
encoder=self.compilation_config.encoder_compilation_time,