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 15, 21:00 UTC
[V1][Minor] Remove obsolete FIXME comment (#14304)
Signed-off-by: Nick Hill <nhill@redhat.com>
Nick Hill committed
Mar 5, 2025 at 11:59 UTC
a32c8669cacaab56b753c3d737f3a7a099c30fba
1 file changed
-5
vllm/v1/worker/gpu_input_batch.py
-5
@@ -298,11 +298,6 @@ class InputBatch:
298
if sampling_params.logit_bias is not None:
299
self.logit_bias[req_index] = sampling_params.logit_bias
300
301
- # FIXME: this implementation is incorrect. We create this mask
302
- # then apply -inf to these specific tokens, which means we never
303
- # select the allowed tokens! We cannot do the reverse, since
304
- # this will impact the requests that do not have allowed_token_ids.
305
- # This feature is currently disabled on V1 (we reject in Processor).
301
if sampling_params.allowed_token_ids:
302
self.has_allowed_token_ids.add(req_id)
303
if self.allowed_token_ids_mask_cpu_tensor is None: