Fix dynamic_prompt_cache for chunked prefill

Author: sufubao

lightllm/common/basemodel/basemodel.py

@@ -251,7 +251,6 @@ def _create_inferstate(self, model_input: ModelInput, microbatch_index: int = 0)
         infer_state.is_prefill = model_input.is_prefill
         infer_state.is_token_healing = self.is_token_healing
         infer_state.return_all_prompt_logics = self.return_all_prompt_logics
-        infer_state.use_dynamic_prompt_cache = self.use_dynamic_prompt_cache
         infer_state.batch_size = model_input.batch_size
         infer_state.total_token_num = model_input.total_token_num
         infer_state.max_len_in_batch = model_input.max_len_in_batch

lightllm/models/deepseek2/layer_infer/transformer_layer_infer.py

@@ -255,7 +255,7 @@ def _decompress_kv(
         b_kv_start_loc,
         skip_sample=False,
     ):
-        if infer_state.use_dynamic_prompt_cache and not skip_sample:
+        if not skip_sample:
             if is_fp8:
                 kv = infer_state.mem_manager.kv_buffer[self.layer_num_][:, :, :-2].view(torch.float8_e4m3fn)
                 kv_scale = infer_state.mem_manager.kv_buffer[self.layer_num_][:, :, -2:].view(torch.bfloat16)

lightllm/server/api_start.py

@@ -95,10 +95,6 @@ def normal_or_p_d_start(args):
     if args.graph_max_len_in_batch == 0:
         args.graph_max_len_in_batch = args.max_req_total_len
 
-    # mode setting check.
-    if not args.disable_chunked_prefill:
-        assert args.disable_dynamic_prompt_cache is False
-        assert args.disable_chunked_prefill is False
     if args.output_constraint_mode != "none":
         assert args.disable_dynamic_prompt_cache is False
         assert args.disable_chunked_prefill is False