A high-throughput and memory-efficient inference and serving engine for LLMs
[Bug] Migrate Reset cache for both v2 and v1 model runner (#42759)
Signed-off-by: yewentao256 <zhyanwentao@126.com> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Wentao Ye committed
Jun 12, 2026 at 09:38 UTC
fbc3a1907aeb6beff59461e535045f17ac14306e
2 files changed
+3
-2
vllm/v1/worker/gpu/model_runner.py
-2
@@ -367,8 +367,6 @@ class GPUModelRunner(LoRAModelRunnerMixin):
367
from vllm.v1.worker.gpu_model_runner import GPUModelRunner as GPUModelRunnerV1
368
369
GPUModelRunnerV1.reload_weights(self, *args, **kwargs) # type: ignore[arg-type]
370
- self.reset_encoder_cache()
371
- self.reset_mm_cache()
370
371
def apply_sparse_weight_patches(self, *args, **kwargs) -> None:
372
# TODO: Use full version instead of import when fully migrated to v2
vllm/v1/worker/gpu_model_runner.py
+3
@@ -5391,6 +5391,9 @@ class GPUModelRunner(
5391
weights_not_loaded,
5392
)
5393
5394
+ self.reset_encoder_cache()
5395
+ self.reset_mm_cache()
5396
+
5397
def _get_prompt_logprobs_dict(
5398
self,
5399
hidden_states: torch.Tensor,