Merge branch 'main' into cuda-graph

predibase · Jan 4, 2024 · 08061d3 · 08061d3
2 parents e00a09f + 6bfc3a2
commit 08061d3
Showing 1 changed file with 0 additions and 2 deletions.
diff --git a/server/lorax_server/models/flash_causal_lm.py b/server/lorax_server/models/flash_causal_lm.py
@@ -941,8 +941,6 @@ def decode(self, generated_ids: Union[torch.Tensor, List[int]]) -> str:
         )
 
     def forward(self, batch: FlashCausalLMBatch, adapter_data: AdapterBatchData) -> Tuple[torch.Tensor, torch.Tensor]:
-        global CACHE_MANAGER
-
         prefill = batch.cu_seqlen_prefill is not None
         model = self.model
         if (