feat: move init_req_to_token_indexes and copy_kv_index_to_req to alloc fun

Author: niushengxiao

lightllm/common/basemodel/basemodel.py

@@ -11,9 +11,7 @@
 from lightllm.common.basemodel.infer_struct import InferStateInfo
 from lightllm.common.mem_manager import MemoryManager
 from lightllm.common.req_manager import ReqManager
-from lightllm.common.infer_utils import init_req_to_token_indexes
 from lightllm.common.build_utils import repair_config
-from lightllm.common.basemodel.triton_kernel.copy_kv_index_to_req import copy_kv_index_to_req
 from lightllm.common.basemodel.layer_infer.cache_tensor_manager import g_cache_manager
 from lightllm.common.basemodel.cuda_graph import CudaGraph
 from lightllm.common.quantization import Quantcfg
@@ -332,14 +330,6 @@ def _prefill(
         model_input: ModelInput,
     ):
         infer_state = self._create_inferstate(model_input)
-        init_req_to_token_indexes(
-            self.req_manager.req_to_token_indexs,
-            model_input.b_req_idx,
-            model_input.b_seq_len,
-            infer_state.b_ready_cache_len,
-            model_input.max_len_in_batch,
-            infer_state.mem_index,
-        )
 
         infer_state.init_some_extra_state(self, model_input.input_ids)
         return self._context_forward(model_input.input_ids, infer_state)
@@ -360,12 +350,6 @@ def _decode(
             find_graph_batch_size = self.graph.find_closest_graph_batch_size(model_input.batch_size)
             padded_model_input = self._create_padded_decode_model_input(model_input, find_graph_batch_size)
             infer_state = self._create_inferstate(padded_model_input)
-            copy_kv_index_to_req(
-                self.req_manager.req_to_token_indexs,
-                infer_state.b_req_idx,
-                infer_state.b_seq_len,
-                infer_state.mem_index,
-            )
             infer_state.init_some_extra_state(self, padded_model_input.input_ids)
 
             if self.graph.need_capture(find_graph_batch_size):
@@ -381,12 +365,6 @@ def _decode(
             )
         else:
             infer_state = self._create_inferstate(model_input)
-            copy_kv_index_to_req(
-                self.req_manager.req_to_token_indexs,
-                infer_state.b_req_idx,
-                infer_state.b_seq_len,
-                infer_state.mem_index,
-            )
             infer_state.init_some_extra_state(self, model_input.input_ids)
             model_output = self._token_forward(model_input.input_ids, infer_state)
 
@@ -471,25 +449,9 @@ def microbatch_overlap_prefill(self, model_input0: ModelInput, model_input1: Mod
         input_ids0, input_ids1 = model_input0.input_ids, model_input1.input_ids
 
         infer_state0 = self._create_inferstate(model_input0, 0)
-        init_req_to_token_indexes(
-            self.req_manager.req_to_token_indexs,
-            model_input0.b_req_idx,
-            model_input0.b_seq_len,
-            infer_state0.b_ready_cache_len,
-            model_input0.max_len_in_batch,
-            infer_state0.mem_index,
-        )
         infer_state0.init_some_extra_state(self, input_ids0)
 
         infer_state1 = self._create_inferstate(model_input1, 1)
-        init_req_to_token_indexes(
-            self.req_manager.req_to_token_indexs,
-            model_input1.b_req_idx,
-            model_input1.b_seq_len,
-            infer_state1.b_ready_cache_len,
-            model_input1.max_len_in_batch,
-            infer_state1.mem_index,
-        )
         infer_state1.init_some_extra_state(self, input_ids1)
 
         model_output0, model_output1 = self._overlap_tpsp_context_forward(
@@ -531,20 +493,8 @@ def microbatch_overlap_decode(self, model_input0: ModelInput, model_input1: Mode
             padded_model_input0 = self._create_padded_decode_model_input(model_input0, find_graph_batch_size)
             padded_model_input1 = self._create_padded_decode_model_input(model_input1, find_graph_batch_size)
             infer_state0 = self._create_inferstate(padded_model_input0, 0)
-            copy_kv_index_to_req(
-                self.req_manager.req_to_token_indexs,
-                infer_state0.b_req_idx,
-                infer_state0.b_seq_len,
-                infer_state0.mem_index,
-            )
             infer_state0.init_some_extra_state(self, padded_model_input0.input_ids)
             infer_state1 = self._create_inferstate(padded_model_input1, 1)
-            copy_kv_index_to_req(
-                self.req_manager.req_to_token_indexs,
-                infer_state1.b_req_idx,
-                infer_state1.b_seq_len,
-                infer_state1.mem_index,
-            )
             infer_state1.init_some_extra_state(self, padded_model_input1.input_ids)
 
             if self.graph.need_capture(find_graph_batch_size):
@@ -569,20 +519,8 @@ def microbatch_overlap_decode(self, model_input0: ModelInput, model_input1: Mode
             model_output1 = self._create_unpad_decode_model_output(model_output1, origin_batch_size=origin_batch_size)
         else:
             infer_state0 = self._create_inferstate(model_input0, 0)
-            copy_kv_index_to_req(
-                self.req_manager.req_to_token_indexs,
-                infer_state0.b_req_idx,
-                infer_state0.b_seq_len,
-                infer_state0.mem_index,
-            )
             infer_state0.init_some_extra_state(self, model_input0.input_ids)
             infer_state1 = self._create_inferstate(model_input1, 1)
-            copy_kv_index_to_req(
-                self.req_manager.req_to_token_indexs,
-                infer_state1.b_req_idx,
-                infer_state1.b_seq_len,
-                infer_state1.mem_index,
-            )
             infer_state1.init_some_extra_state(self, model_input1.input_ids)
 
             model_output0, model_output1 = self._overlap_tpsp_token_forward(
@@ -683,10 +621,12 @@ def _check_max_len_infer(self):
             logger.info("begin check max_len infer")
             dummy_input_ids = torch.ones(self.batch_max_tokens, dtype=torch.int32, device="cuda")
             b_req_idx = torch.tensor([self.req_manager.alloc()], dtype=torch.int32, device="cuda")
-            mem_indexes = self.mem_manager.alloc(len(dummy_input_ids)).cuda()
             b_seq_len = torch.ones(1, dtype=torch.int32, device="cuda")
             b_seq_len[:] = self.batch_max_tokens
             b_ready_cache_len = torch.zeros(1, dtype=torch.int32, device="cuda")
+            mem_indexes = self.mem_manager.alloc(
+                len(dummy_input_ids), b_req_idx, b_seq_len, b_ready_cache_len, True
+            ).cuda()
             total_token_num = self.batch_max_tokens
             b_mtp_index = torch.zeros(1, dtype=torch.int32, device="cuda")
             model_input = ModelInput(

lightllm/common/basemodel/cuda_graph.py

@@ -196,13 +196,13 @@ def warmup(self, model):
             total_token_num = batch_size * seq_len
             max_len_in_batch = self.graph_max_len_in_batch
             input_ids = torch.tensor([1 for _ in range(batch_size)], dtype=torch.int32, device="cuda")
-            mem_indexes = model.mem_manager.alloc(len(input_ids)).cuda()
             b_req_idx = torch.tensor(
                 [model.req_manager.HOLD_REQUEST_ID for _ in range(batch_size)], dtype=torch.int32, device="cuda"
             )
             b_seq_len = torch.empty(batch_size, dtype=torch.int32, device="cuda")
             b_seq_len.fill_(seq_len)
             b_mtp_index = torch.zeros(batch_size, dtype=torch.int32, device="cuda")
+            mem_indexes = model.mem_manager.alloc(len(input_ids), b_req_idx, b_seq_len).cuda()
 
             model_input = ModelInput(
                 batch_size=batch_size,
@@ -252,13 +252,13 @@ def warmup_overlap(self, model):
                 total_token_num = batch_size * seq_len
                 max_len_in_batch = self.graph_max_len_in_batch
                 input_ids = torch.tensor([1 for _ in range(batch_size)], dtype=torch.int32, device="cuda")
-                mem_indexes = model.mem_manager.alloc(len(input_ids)).cuda()
                 b_req_idx = torch.tensor(
                     [model.req_manager.HOLD_REQUEST_ID for _ in range(batch_size)], dtype=torch.int32, device="cuda"
                 )
                 b_seq_len = torch.empty(batch_size, dtype=torch.int32, device="cuda")
                 b_seq_len.fill_(seq_len)
                 b_mtp_index = torch.zeros(batch_size, dtype=torch.int32, device="cuda")
+                mem_indexes = model.mem_manager.alloc(len(input_ids), b_req_idx, b_seq_len).cuda()
 
                 micro_batch = ModelInput(
                     is_prefill=False,

lightllm/common/infer_utils.py

@@ -1,6 +1,4 @@
-def init_req_to_token_indexes(
-    req_to_token_indexs, b_req_idx, b_seq_len, b_ready_cache_len, max_len_in_batch, alloc_mem_index
-):
+def init_req_to_token_indexes(req_to_token_indexs, b_req_idx, b_seq_len, b_ready_cache_len, alloc_mem_index):
     start_index = 0
     b_seq_len_numpy = b_seq_len.cpu().numpy()
     b_ready_cache_len_numpy = b_ready_cache_len.cpu().numpy()

lightllm/common/mem_manager.py

@@ -12,6 +12,8 @@
 from lightllm.utils.envs_utils import get_unique_server_name, get_env_start_args
 from lightllm.distributed.pynccl import PyNcclCommunicator
 from lightllm.utils.dist_utils import get_current_device_id
+from lightllm.common.infer_utils import init_req_to_token_indexes
+from lightllm.common.basemodel.triton_kernel.copy_kv_index_to_req import copy_kv_index_to_req
 
 logger = init_logger(__name__)
 
@@ -52,6 +54,7 @@ def __init__(self, size, dtype, head_num, head_dim, layer_num, always_copy=False
             layer_num,
         )
         self.HOLD_TOKEN_MEMINDEX = self.size
+        self.req_to_token_indexs = None
 
     def get_cell_size(self):
         return 2 * self.head_num * self.head_dim * self.layer_num * torch._utils._element_size(self.dtype)
@@ -243,7 +246,9 @@ def _write_kv_move_data_p2p(self, token_indexes: torch.Tensor, buffer_tensor: to
     def _free_buffers(self):
         self.kv_buffer = None
 
-    def alloc(self, need_size) -> torch.Tensor:
+    def alloc(
+        self, need_size, b_req_idx=None, b_seq_len=None, b_ready_cache_len=None, is_prefill=False
+    ) -> torch.Tensor:
         if need_size > self.mark_end - self.mark_start:
             logger.error(f"warn no enough cache need_size {need_size} left_size {self.can_use_mem_size}")
             assert False, "error alloc state"
@@ -255,8 +260,29 @@ def alloc(self, need_size) -> torch.Tensor:
 
         self.can_use_mem_size -= need_size
         self.shared_can_use_token_num.set_value(self.can_use_mem_size)
+
+        if self.req_to_token_indexs is not None:
+            assert b_req_idx is not None and b_seq_len is not None, "b_req_idx and b_seq_len must be provided"
+            if is_prefill:
+                init_req_to_token_indexes(
+                    self.req_to_token_indexs,
+                    b_req_idx,
+                    b_seq_len,
+                    b_ready_cache_len,
+                    ans,
+                )
+            else:
+                copy_kv_index_to_req(
+                    self.req_to_token_indexs,
+                    b_req_idx.cuda(),
+                    b_seq_len.cuda(),
+                    ans.cuda(),
+                )
         return ans
 
+    def set_prefix_cache_to_req(self, req_idx: int, start: int, end: int, values: torch.Tensor):
+        self.req_to_token_indexs[req_idx, start:end] = values
+
     def free(self, free_index: Union[torch.Tensor, List[int]]):
         """_summary_
 

lightllm/common/req_manager.py

@@ -62,6 +62,7 @@ def __init__(self, max_request_num, max_sequence_length, mem_manager: MemoryMana
         self.req_to_token_indexs = torch.zeros(
             (max_request_num + 1, max_sequence_length), dtype=torch.int32, device="cuda"
         )
+        mem_manager.req_to_token_indexs = self.req_to_token_indexs
         self.mem_manager = mem_manager
         self.req_sampling_params_manager = ReqSamplingParamsManager(max_request_num)
         self.max_request_num = max_request_num

lightllm/server/router/model_infer/infer_batch.py

@@ -340,7 +340,9 @@ def _match_radix_cache(self):
                 self.shared_kv_node = share_node
                 ready_cache_len = share_node.node_prefix_total_len
                 # 从 cpu 到 gpu 是流内阻塞操作
-                g_infer_context.req_manager.req_to_token_indexs[self.req_idx, 0:ready_cache_len] = value_tensor
+                g_infer_context.req_manager.mem_manager.set_prefix_cache_to_req(
+                    self.req_idx, 0, ready_cache_len, value_tensor
+                )
                 self.cur_kv_len = int(ready_cache_len)  # 序列化问题, 该对象可能为numpy.int64，用 int(*)转换
                 self.shm_req.prompt_cache_len = self.cur_kv_len  # 记录 prompt cache 的命中长度
 
@@ -461,7 +463,7 @@ def diverse_copy(self, req_manager, is_prefill):
             req = g_infer_context.requests_mapping[req_id]
             req.finish_status.set_status(FinishStatus.NO_FINISH)
             input_len = req.get_chuncked_input_token_len()
-            req_manager.req_to_token_indexs[req.req_idx][prefix_len:input_len] = cache_token_id
+            req_manager.mem_manager.set_prefix_cache_to_req(req.req_idx, prefix_len, input_len, cache_token_id)
             assert input_len == pre_input_len
 
 

lightllm/server/router/model_infer/mode_backend/generic_padded_pre_process.py

@@ -78,7 +78,9 @@ def padded_prepare_prefill_inputs(
     g_infer_state_lock.acquire()
     if g_infer_context.radix_cache is not None:
         g_infer_context.radix_cache.free_radix_cache_to_get_enough_token(input_ids.shape[0] - padded_req_num)
-    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(input_ids.shape[0] - padded_req_num)
+    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(
+        input_ids.shape[0] - padded_req_num, b_req_idx, b_seq_len, b_ready_cache_len, True
+    )
     g_infer_state_lock.release()
 
     if padded_req_num > 0:
@@ -163,7 +165,9 @@ def padded_prepare_decode_inputs(
     g_infer_state_lock.acquire()
     if g_infer_context.radix_cache is not None:
         g_infer_context.radix_cache.free_radix_cache_to_get_enough_token(b_seq_len.shape[0] - padded_req_num)
-    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(b_seq_len.shape[0] - padded_req_num)
+    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(
+        b_seq_len.shape[0] - padded_req_num, b_req_idx, b_seq_len
+    )
     g_infer_state_lock.release()
 
     if padded_req_num > 0:

lightllm/server/router/model_infer/mode_backend/generic_pre_process.py

@@ -56,7 +56,9 @@ def prepare_prefill_inputs(
     g_infer_state_lock.acquire()
     if g_infer_context.radix_cache is not None:
         g_infer_context.radix_cache.free_radix_cache_to_get_enough_token(input_ids.shape[0])
-    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(input_ids.shape[0])
+    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(
+        input_ids.shape[0], b_req_idx, b_seq_len, b_ready_cache_len, True
+    )
     g_infer_state_lock.release()
 
     model_input = ModelInput(
@@ -112,7 +114,7 @@ def prepare_decode_inputs(req_objs: List[InferReq]) -> Tuple[ModelInput, List[In
     g_infer_state_lock.acquire()
     if g_infer_context.radix_cache is not None:
         g_infer_context.radix_cache.free_radix_cache_to_get_enough_token(b_seq_len.shape[0])
-    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(b_seq_len.shape[0])
+    mem_indexes = g_infer_context.req_manager.mem_manager.alloc(b_seq_len.shape[0], b_req_idx, b_seq_len)
     g_infer_state_lock.release()
 
     model_input = ModelInput(

test/benchmark/static_inference/model_infer.py

@@ -242,7 +242,9 @@ def run_forward_once(
         b_seq_len[i] = input_len
 
     total_token_num = batch_size * input_len
-    mem_indexes = model_part.req_manager.mem_manager.alloc(test_data.shape[0]).cuda()
+    mem_indexes = model_part.req_manager.mem_manager.alloc(
+        test_data.shape[0], b_req_idx, b_seq_len, b_ready_cache_len, True
+    ).cuda()
 
     rank_id = model_kvargs["rank_id"]
 
@@ -303,7 +305,7 @@ def run_forward_once(
         step_start = time.time()
         total_token_num += batch_size
         b_seq_len += 1
-        mem_indexes = model_part.req_manager.mem_manager.alloc(predict_ids.shape[0]).cuda()
+        mem_indexes = model_part.req_manager.mem_manager.alloc(predict_ids.shape[0], b_req_idx, b_seq_len).cuda()
         max_len_in_batch = input_len + i + 1
         logits = decode_fn(
             model_part,

test/benchmark/static_inference/model_infer_mtp.py

@@ -126,7 +126,9 @@ def run_forward_once(args, input_len, output_len, batch_size, main_model, draft_
         b_seq_len[i] = input_len
 
     total_token_num = input_len * batch_size
-    mem_indexes = main_model.req_manager.mem_manager.alloc(test_data.shape[0]).cuda()
+    mem_indexes = main_model.req_manager.mem_manager.alloc(
+        test_data.shape[0], b_req_idx, b_seq_len, b_ready_cache_len, True
+    ).cuda()
     # Main model Prefill
     model_input = ModelInput(
         batch_size=batch_size,
@@ -193,7 +195,9 @@ def run_forward_once(args, input_len, output_len, batch_size, main_model, draft_
 
     nopad_b_seq_idx = torch.tensor(nopad_b_seq_idx, dtype=torch.int32, device="cuda")
     nopad_b_seq_len = torch.tensor(nopad_b_seq_len, dtype=torch.int32, device="cuda")
-    mem_indexes = main_model.req_manager.mem_manager.alloc(batch_size * (len(draft_models) + 1)).cuda()
+    mem_indexes = main_model.req_manager.mem_manager.alloc(
+        batch_size * (len(draft_models) + 1), nopad_b_seq_idx, nopad_b_seq_len
+    ).cuda()
 
     model_input = ModelInput(
         batch_size=batch_size * (len(draft_models) + 1),