【Inference】Fix Bug set_preids_token_penalty_multi_scores (#10492)

* fix

* check model_inputs[pre_ids] update in dynamic insert

Author: ckl117

csrc/gpu/set_preids_token_penalty_multi_scores.cu

@@ -41,11 +41,11 @@ __global__ void set_preids_token_penalty_multi_scores_kernel(const bool *stop_fl
     T *logits_now = logits + bi * length;
     int tid = threadIdx.x;
 
-    if (tid < bs && !stop_flags[tid]) {
-        int64_t *pre_ids_now = pre_ids + tid * length_id;
-        const int64_t *input_ids_now = input_ids + tid * length_input_ids;
-        const int seq_len_dec = seq_lens_decoder[tid];
-        const int seq_len_enc = seq_lens_encoder[tid];
+    if (bi < bs && !stop_flags[bi]) {
+        int64_t *pre_ids_now = pre_ids + bi * length_id;
+        const int64_t *input_ids_now = input_ids + bi * length_input_ids;
+        const int seq_len_dec = seq_lens_decoder[bi];
+        const int seq_len_enc = seq_lens_encoder[bi];
         if (seq_len_dec == 0 && seq_len_enc == 0) return; // stopped
 
         const int step_idx_now = step_idx[bi];

llm/predict/predictor.py

@@ -1311,6 +1311,8 @@ def insert_task(self, pos, task_id, repeat_num):
         self.model_inputs["stop_flags"][pos] = False
         self.model_inputs["result_id"][pos][0] = task_id
         self.model_inputs["step_idx"][pos, 0] = 1
+        self.model_inputs["pre_ids"][pos][0] = self.input_ids[query_id][-1]
+        self.model_inputs["pre_ids"][pos][1:] = -1
         self.model_inputs["not_need_stop"][0] = True
 
         num_prefill_blocks = length // self.block_size