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 14, 05:30 UTC
[Core] Remove unused `prev_sampled_token_ids_invalid_indices` input batch field (#26514)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Oct 9, 2025 at 13:22 UTC
2e54db4d2b468f1dd1580238b74d80da0695995f
2 files changed
-4
vllm/v1/worker/gpu_input_batch.py
-1
@@ -251,7 +251,6 @@ class InputBatch:
251
252
# Cached reference to the GPU tensor of previously sampled tokens
253
self.prev_sampled_token_ids: Optional[torch.Tensor] = None
254
- self.prev_sampled_token_ids_invalid_indices: Optional[set[int]] = None
254
self.prev_req_id_to_index: Optional[dict[str, int]] = None
255
256
@property
vllm/v1/worker/gpu_model_runner.py
-3
@@ -2305,9 +2305,6 @@ class GPUModelRunner(LoRAModelRunnerMixin, KVConnectorModelRunnerMixin):
2305
# These will be copied into input_ids in the next step
2306
# when preparing inputs.
2307
self.input_batch.prev_sampled_token_ids = sampled_token_ids
2308
- self.input_batch.prev_sampled_token_ids_invalid_indices = (
2309
- invalid_req_indices_set
2310
- )
2308
self.input_batch.prev_req_id_to_index = {
2309
req_id: i
2310
for i, req_id in enumerate(self.input_batch.req_ids)