[fleet_executor] Add entrance of FleetExecutor in AnalysisPredictor for distributed inference

[FeixLiu]

PR types

Others

PR changes

Others

Describe

Add entrance of fleet executor into AnalysisPredictor. Add some helper methods to init NCCL environment for distributed inference.

To use the fleet executor for inference, these configures should be set in DistConfig

bool use_dist_model_ // whether use DistModel or not
std::vector<std::string> trainer_endpoints_ // all trainers' endpoints
std::string current_endpoint_ // current trainer's endpoint
int64_t nranks_ // total ranks (number of trainers)
int64_t local_rank_ // local rank
std::string comm_init_config_ // converter config path (used to init the comm)

Note that, use_dist_model_ muse be set true by calling

EnableDistModel(true);

nranks and rank are set simultaneously by calling

SetRanks(int64_t nranks, int64_t, ranks);

trainer_endpoints and current_endpoint are also set simultaneously by calling

SetEndpoints(std::vector<std::string> trainer_endpoints, std::string current_endpoint);

DistConfig should be set to AnalysisConfig by calling

SetDistConfig(dConfig);

The converter config should some sections like this:

[ring_id -> ranks]
0,0,1,2,3,4,5,6,7
1,0,1,2,3
2,4,5,6,7
21,0,1
22,1,2
23,2,3
24,3,4
25,4,5
26,5,6
27,6,7
[rank -> ring_ids]
0,0,1,21
1,0,1,21,22
2,0,1,22,23
3,0,1,23,24
4,0,2,24,25
5,0,2,25,26
6,0,2,26,27
7,0,2,27

[paddle-bot-old]

Thanks for your contribution!
Please wait for the result of CI firstly. See Paddle CI Manual for details.

[wangxicoding]

LGTM

[gongweibao]

Don't use VLOG(3) for all output logs while you will see redundancy logs you don't need and
var's name should express its clear meaning.

[gongweibao]

Add a concise description of PR like #37725

[gongweibao]

LGTM

[qingqing01]

LGTM

[shangzhizhou]

LGTM

[Superjomn]

LGTM