diff --git a/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py b/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py index 3a61d058..c0d0bbb3 100644 --- a/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py +++ b/server/text_generation_server/models/custom_modeling/flash_santacoder_modeling.py @@ -363,6 +363,7 @@ class Block(nn.Module): slots, input_lengths, max_s, + prefill_cache_indices ): hidden_states, residual = self.ln_1(hidden_states, residual) hidden_states = self.attn(