forked from pytorch/FBGEMM
-
Notifications
You must be signed in to change notification settings - Fork 0
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
Implement PS KV DB for FBGEMM TBE operator (pytorch#2664)
Summary: Pull Request resolved: pytorch#2664 Implemented kv_db::EmbeddingKVDB interface to connect remote PS service instead of using local SSD based rocksDB embedding. Initialize PS KV DB when ps_hosts is not None. Reviewed By: sryap Differential Revision: D56715840 fbshipit-source-id: 53ebb514bceb21ee4c124afed46907875f9e1750
- Loading branch information
1 parent
eb7b841
commit 4449cbc
Showing
3 changed files
with
238 additions
and
41 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
89 changes: 89 additions & 0 deletions
89
fbgemm_gpu/src/ps_split_embeddings_cache/ps_split_table_batched_embeddings.cpp
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,89 @@ | ||
| /* | ||
| * Copyright (c) Meta Platforms, Inc. and affiliates. | ||
| * All rights reserved. | ||
| * | ||
| * This source code is licensed under the BSD-style license found in the | ||
| * LICENSE file in the root directory of this source tree. | ||
| */ | ||
|
|
||
| #include "./ps_table_batched_embeddings.h" | ||
|
|
||
| #include <torch/custom_class.h> | ||
| #include "fbgemm_gpu/sparse_ops_utils.h" | ||
|
|
||
| using namespace at; | ||
| using namespace ps; | ||
|
|
||
| namespace { | ||
| class EmbeddingParameterServerWrapper : public torch::jit::CustomClassHolder { | ||
| public: | ||
| EmbeddingParameterServerWrapper( | ||
| const std::vector<std::string>& tps_ips, | ||
| const std::vector<int64_t>& tps_ports, | ||
| int64_t tbe_id, | ||
| int64_t maxLocalIndexLength = 54, | ||
| int64_t num_threads = 32) { | ||
| TORCH_CHECK( | ||
| tps_ips.size() == tps_ports.size(), | ||
| "tps_ips and tps_ports must have the same size"); | ||
| std::vector<std::pair<std::string, int>> tpsHosts = {}; | ||
| for (int i = 0; i < tps_ips.size(); i++) { | ||
| tpsHosts.push_back(std::make_pair(tps_ips[i], tps_ports[i])); | ||
| } | ||
|
|
||
| impl_ = std::make_shared<ps::EmbeddingParameterServer>( | ||
| std::move(tpsHosts), tbe_id, maxLocalIndexLength, num_threads); | ||
| } | ||
|
|
||
| void | ||
| set_cuda(Tensor indices, Tensor weights, Tensor count, int64_t timestep) { | ||
| return impl_->set_cuda(indices, weights, count, timestep); | ||
| } | ||
|
|
||
| void get_cuda(Tensor indices, Tensor weights, Tensor count) { | ||
| return impl_->get_cuda(indices, weights, count); | ||
| } | ||
|
|
||
| void set(Tensor indices, Tensor weights, Tensor count) { | ||
| return impl_->set(indices, weights, count); | ||
| } | ||
|
|
||
| void get(Tensor indices, Tensor weights, Tensor count) { | ||
| return impl_->get(indices, weights, count); | ||
| } | ||
|
|
||
| void compact() { | ||
| return impl_->compact(); | ||
| } | ||
|
|
||
| void flush() { | ||
| return impl_->flush(); | ||
| } | ||
|
|
||
| void cleanup() { | ||
| return impl_->cleanup(); | ||
| } | ||
|
|
||
| private: | ||
| // shared pointer since we use shared_from_this() in callbacks. | ||
| std::shared_ptr<EmbeddingParameterServer> impl_; | ||
| }; | ||
|
|
||
| static auto embedding_parameter_server_wrapper = | ||
| torch::class_<EmbeddingParameterServerWrapper>( | ||
| "fbgemm", | ||
| "EmbeddingParameterServerWrapper") | ||
| .def(torch::init< | ||
| const std::vector<std::string>, | ||
| const std::vector<int64_t>, | ||
| int64_t, | ||
| int64_t, | ||
| int64_t>()) | ||
| .def("set_cuda", &EmbeddingParameterServerWrapper::set_cuda) | ||
| .def("get_cuda", &EmbeddingParameterServerWrapper::get_cuda) | ||
| .def("compact", &EmbeddingParameterServerWrapper::compact) | ||
| .def("flush", &EmbeddingParameterServerWrapper::flush) | ||
| .def("set", &EmbeddingParameterServerWrapper::set) | ||
| .def("get", &EmbeddingParameterServerWrapper::get) | ||
| .def("cleanup", &EmbeddingParameterServerWrapper::cleanup); | ||
| } // namespace |
64 changes: 64 additions & 0 deletions
64
fbgemm_gpu/src/ps_split_embeddings_cache/ps_table_batched_embeddings.h
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,64 @@ | ||
| /* | ||
| * Copyright (c) Meta Platforms, Inc. and affiliates. | ||
| * All rights reserved. | ||
| * | ||
| * This source code is licensed under the BSD-style license found in the | ||
| * LICENSE file in the root directory of this source tree. | ||
| */ | ||
|
|
||
| #pragma once | ||
| #include "../ssd_split_embeddings_cache/kv_db_table_batched_embeddings.h" | ||
|
|
||
| #include <folly/experimental/coro/BlockingWait.h> | ||
| #include "mvai_infra/experimental/ps_training/tps_client/TrainingParameterServiceClient.h" | ||
|
|
||
| namespace ps { | ||
|
|
||
| class EmbeddingParameterServer : public kv_db::EmbeddingKVDB { | ||
| public: | ||
| explicit EmbeddingParameterServer( | ||
| std::vector<std::pair<std::string, int>>&& tps_hosts, | ||
| int64_t tbe_id, | ||
| int64_t maxLocalIndexLength = 54, | ||
| int64_t num_threads = 32) | ||
| : tps_client_( | ||
| std::make_shared<mvai_infra::experimental::ps_training::tps_client:: | ||
| TrainingParameterServiceClient>( | ||
| std::move(tps_hosts), | ||
| tbe_id, | ||
| maxLocalIndexLength, | ||
| num_threads)) {} | ||
|
|
||
| void set( | ||
| const at::Tensor& indices, | ||
| const at::Tensor& weights, | ||
| const at::Tensor& count) override { | ||
| RECORD_USER_SCOPE("EmbeddingParameterServer::set"); | ||
| folly::coro::blockingWait( | ||
| tps_client_->set(indices, weights, count.item().toLong())); | ||
| } | ||
| void get( | ||
| const at::Tensor& indices, | ||
| const at::Tensor& weights, | ||
| const at::Tensor& count) override { | ||
| RECORD_USER_SCOPE("EmbeddingParameterServer::get"); | ||
| folly::coro::blockingWait( | ||
| tps_client_->get(indices, weights, count.item().toLong())); | ||
| } | ||
| void flush() override {} | ||
| void compact() override {} | ||
| // cleanup cached results in server side | ||
| // This is a test helper, please do not use it in production | ||
| void cleanup() { | ||
| folly::coro::blockingWait(tps_client_->cleanup()); | ||
| } | ||
|
|
||
| private: | ||
| void flush_or_compact(const int64_t /*timestep*/) override {} | ||
|
|
||
| std::shared_ptr<mvai_infra::experimental::ps_training::tps_client:: | ||
| TrainingParameterServiceClient> | ||
| tps_client_; | ||
| }; // class EmbeddingKVDB | ||
|
|
||
| } // namespace ps |