|
| 1 | +import time |
| 2 | +import torch |
| 3 | +import torch.multiprocessing as mp |
| 4 | +import torch.distributed as dist |
| 5 | +from typing import List |
| 6 | +from lightllm.server.router.model_infer.infer_batch import g_infer_context, InferReq |
| 7 | +from lightllm.server.core.objs.req import PDChunkedPrefillReq |
| 8 | +from lightllm.utils.log_utils import init_logger |
| 9 | +from lightllm.server.router.model_infer.mode_backend.generic_post_process import sample |
| 10 | +from lightllm.utils.envs_utils import get_env_start_args |
| 11 | +from lightllm.server.router.model_infer.mode_backend.dp_backend.pre_process import padded_prepare_decode_inputs |
| 12 | + |
| 13 | +from .impl_for_pd_decode import PDNIXLBackendForDecodeNode |
| 14 | + |
| 15 | +logger = init_logger(__name__) |
| 16 | + |
| 17 | + |
| 18 | +class PDNIXLDPBackendForDecodeNode(PDNIXLBackendForDecodeNode): |
| 19 | + def __init__(self, prefill_task_queue: mp.Queue, prefill_done_queue: mp.Queue, nix_meta_queue: mp.Queue) -> None: |
| 20 | + super().__init__(prefill_task_queue, prefill_done_queue, nix_meta_queue) |
| 21 | + self.enable_decode_microbatch_overlap = get_env_start_args().enable_decode_microbatch_overlap |
| 22 | + |
| 23 | + def init_custom(self): |
| 24 | + super().init_custom() |
| 25 | + |
| 26 | + self.reduce_tensor = torch.tensor([0], dtype=torch.int32, device="cuda", requires_grad=False) |
| 27 | + from lightllm.server.router.model_infer.mode_backend.dp_backend.pre_process import padded_prepare_prefill_inputs |
| 28 | + kwargs, run_reqs, padded_req_num = padded_prepare_prefill_inputs([], 1, is_multimodal=self.is_multimodal) |
| 29 | + self.model.forward(**kwargs) |
| 30 | + assert len(run_reqs) == 0 and padded_req_num == 1 |
| 31 | + |
| 32 | + return |
| 33 | + |
| 34 | + def decode(self): |
| 35 | + |
| 36 | + uninit_reqs, aborted_reqs, ok_finished_reqs, prefill_reqs, decode_reqs = self._get_classed_reqs( |
| 37 | + g_infer_context.infer_req_ids, |
| 38 | + no_decode=False, |
| 39 | + ) |
| 40 | + # filter out remote prefilling reqs |
| 41 | + prefill_reqs, aborted_reqs, decode_reqs, _ = self._decode_filter_reqs(prefill_reqs, aborted_reqs, decode_reqs) |
| 42 | + |
| 43 | + self._filter_reqs(aborted_reqs) |
| 44 | + |
| 45 | + # allocate kv cache, do remote prefill |
| 46 | + if prefill_reqs: |
| 47 | + # TODO: we could allocate cache later after remote prefill done and get a signal from remote |
| 48 | + # but it will have a risk to not have enough cache for this request. |
| 49 | + kwargs, run_reqs = self._prepare_remote_prefill_inputs(prefill_reqs) |
| 50 | + for idx, run_req in enumerate(run_reqs): |
| 51 | + run_req: InferReq = run_req |
| 52 | + shm_req: PDChunkedPrefillReq = run_req.shm_req |
| 53 | + # forward each req to remote prefill |
| 54 | + # since the token index are the same across TPs, we only need to trigger prefill on master |
| 55 | + if self.is_master_in_dp: |
| 56 | + run_req.remote_prefill_start = time.time() |
| 57 | + self.to_remote_queue.put(self._build_remote_prefill_task(idx, kwargs, run_req)) |
| 58 | + |
| 59 | + shm_req.set_pd_req_rank_state(self.rank_in_dp, 0) # set in progress state |
| 60 | + run_req.in_prefill_or_transfer = True |
| 61 | + self.remote_prefilled_reqs[shm_req.group_req_id] = run_req |
| 62 | + |
| 63 | + self.reduce_tensor.fill_(len(decode_reqs)) |
| 64 | + dist.all_reduce(self.reduce_tensor, op=dist.ReduceOp.MAX) |
| 65 | + max_decode_num = self.reduce_tensor.item() |
| 66 | + if max_decode_num != 0: |
| 67 | + if not self.enable_decode_microbatch_overlap: |
| 68 | + self.normal_decode(decode_reqs, max_decode_num, uninit_reqs, ok_finished_reqs) |
| 69 | + else: |
| 70 | + self.overlap_decode(decode_reqs, max_decode_num, uninit_reqs, ok_finished_reqs) |
| 71 | + self._overlap_req_init_and_filter(uninit_reqs=uninit_reqs, ok_finished_reqs=ok_finished_reqs, clear_list=True) |
| 72 | + return |
| 73 | + |
| 74 | + def normal_decode(self, decode_reqs: List[InferReq], max_decode_num: int, uninit_reqs, ok_finished_reqs): |
| 75 | + |
| 76 | + kwargs, run_reqs, padded_req_num = padded_prepare_decode_inputs( |
| 77 | + decode_reqs, max_decode_num, is_multimodal=self.is_multimodal |
| 78 | + ) |
| 79 | + logits = self.model.forward(**kwargs) |
| 80 | + self._overlap_req_init_and_filter(uninit_reqs=uninit_reqs, ok_finished_reqs=ok_finished_reqs, clear_list=True) |
| 81 | + if len(run_reqs) != 0: |
| 82 | + logits = logits[0 : len(run_reqs), :] |
| 83 | + next_token_ids, next_token_probs = sample(logits, run_reqs, self.eos_id) |
| 84 | + next_token_ids = next_token_ids.detach().cpu().numpy() |
| 85 | + next_token_logprobs = torch.log(next_token_probs).detach().cpu().numpy() |
| 86 | + self._post_handle( |
| 87 | + run_reqs, next_token_ids, next_token_logprobs, is_chuncked_mode=False, do_filter_finished_reqs=False |
| 88 | + ) |
| 89 | + return |
| 90 | + |
| 91 | + def overlap_decode(self, decode_reqs: List[InferReq], max_decode_num: int, uninit_reqs, ok_finished_reqs): |
| 92 | + from lightllm.server.router.model_infer.mode_backend.dp_backend.pre_process import ( |
| 93 | + padded_overlap_prepare_decode_inputs, |
| 94 | + ) |
| 95 | + |
| 96 | + ( |
| 97 | + micro_batch, |
| 98 | + run_reqs, |
| 99 | + padded_req_num, |
| 100 | + micro_batch1, |
| 101 | + run_reqs1, |
| 102 | + padded_req_num1, |
| 103 | + ) = padded_overlap_prepare_decode_inputs(decode_reqs, max_decode_num, is_multimodal=self.is_multimodal) |
| 104 | + |
| 105 | + logits, logits1 = self.model.microbatch_overlap_decode(micro_batch, micro_batch1) |
| 106 | + self._overlap_req_init_and_filter(uninit_reqs=uninit_reqs, ok_finished_reqs=ok_finished_reqs, clear_list=True) |
| 107 | + req_num, req_num1 = len(run_reqs), len(run_reqs1) |
| 108 | + all_logits = torch.empty((req_num + req_num1, logits.shape[1]), dtype=logits.dtype, device=logits.device) |
| 109 | + |
| 110 | + all_logits[0:req_num, :].copy_(logits[0:req_num, :], non_blocking=True) |
| 111 | + all_logits[req_num : (req_num + req_num1), :].copy_(logits1[0:req_num1, :], non_blocking=True) |
| 112 | + |
| 113 | + all_run_reqs = run_reqs + run_reqs1 |
| 114 | + if all_run_reqs: |
| 115 | + next_token_ids, next_token_probs = sample(all_logits, all_run_reqs, self.eos_id) |
| 116 | + next_token_ids = next_token_ids.detach().cpu().numpy() |
| 117 | + next_token_logprobs = torch.log(next_token_probs).detach().cpu().numpy() |
| 118 | + self._post_handle( |
| 119 | + all_run_reqs, next_token_ids, next_token_logprobs, is_chuncked_mode=False, do_filter_finished_reqs=False |
| 120 | + ) |
| 121 | + return |
0 commit comments