text-generation-inference/backends/trtllm/csrc/backend.cpp
Funtowicz Morgan 17367438f3
Give TensorRT-LLMa proper CI/CD 😍 (#2886)
* test(ctest) enable address sanitizer

* feat(trtllm): expose finish reason to Rust

* feat(trtllm): fix logits retrieval

* misc(ci): enabe building tensorrt-llm

* misc(ci): update Rust action toolchain

* misc(ci): let's try to build the Dockerfile for trtllm

# Conflicts:
#	Dockerfile_trtllm

* misc(ci): provide mecanism to cache inside container

* misc(ci): export aws creds as output of step

* misc(ci): let's try this way

* misc(ci): again

* misc(ci): again

* misc(ci): add debug profile

* misc(ci): add debug profile

* misc(ci): lets actually use sccache ...

* misc(ci): do not build with ssl enabled

* misc(ci): WAT

* misc(ci): WAT

* misc(ci): WAT

* misc(ci): WAT

* misc(ci): WAT

* misc(backend): test with TGI S3 conf

* misc(backend): test with TGI S3 conf

* misc(backend): once more?

* misc(backend): let's try with GHA

* misc(backend): missing env directive

* misc(backend): make sure to correctly set IS_GHA_BUILD=true in wf

* misc(backend): ok let's debug smtg

* misc(backend): WWWWWWWWWWWWWAAAAAAAA

* misc(backend): kthxbye retry s3

* misc(backend): use session token

* misc(backend): add more info

* misc(backend): lets try 1h30

* misc(backend): lets try 1h30

* misc(backend): increase to 2h

* misc(backend): lets try...

* misc(backend): lets try...

* misc(backend): let's build for ci-runtime

* misc(backend): let's add some more tooling

* misc(backend): add some tags

* misc(backend): disable Werror for now

* misc(backend): added automatic gha detection

* misc(backend): remove leak sanitizer which is included in asan

* misc(backend): forward env

* misc(backend): forward env

* misc(backend): let's try

* misc(backend): let's try

* misc(backend): again

* misc(backend): again

* misc(backend): again

* misc(backend): again

* misc(backend): again

* misc(backend): fix sscache -> sccache

* misc(backend): fix sscache -> sccache

* misc(backend): fix sscache -> sccache

* misc(backend): let's actually cache things now

* misc(backend): let's actually cache things now

* misc(backend): attempt to run the testS?

* misc(backend): attempt to run the tests?

* misc(backend): attempt to run the tests?

* change runner size

* fix: Correctly tag docker images (#2878)

* fix: Correctly tag docker images

* fix: Correctly tag docker images

* misc(llamacpp): maybe?

* misc(llamacpp): maybe?

* misc(llamacpp): maybe?

* misc(ci): gogogo

* misc(ci): gogogo

* misc(ci): gogogo

* misc(ci): gogogo

* misc(ci): gogogo

* misc(ci): gogogo

* misc(ci): go

* misc(ci): go

* misc(ci): go

* misc(ci): use bin folder

* misc(ci): make the wf callable for reuse

* misc(ci): make the wf callable for reuse (bis)

* misc(ci): make the wf callable for reuse (bis)

* misc(ci): give the wf a name

* Create test-trtllm.yml

* Update test-trtllm.yml

* Create build-trtllm2

* Rename build-trtllm2 to 1-build-trtllm2

* Rename test-trtllm.yml to 1-test-trtllm2.yml

* misc(ci): fw secrets

* Update 1-test-trtllm2.yml

* Rename 1-build-trtllm2 to 1-build-trtllm2.yml

* Update 1-test-trtllm2.yml

* misc(ci): use ci-build.yaml as main dispatcher

* Delete .github/workflows/1-test-trtllm2.yml

* Delete .github/workflows/1-build-trtllm2.yml

* misc(ci): rights?

* misc(ci): rights?

* misc(ci): once more?

* misc(ci): once more?

* misc(ci): baby more time?

* misc(ci): baby more time?

* misc(ci): try the permission above again?

* misc(ci): try the permission above again?

* misc(ci): try the permission scoped again?

* misc(ci): install tensorrt_llm_executor_static

* misc(ci): attempt to rebuild with sccache?

* misc(ci):run the tests on GPU instance

* misc(ci): let's actually setup sccache in the build.rs

* misc(ci): reintroduce variables

* misc(ci): enforce sccache

* misc(ci): correct right job name dependency

* misc(ci): detect dev profile for debug

* misc(ci): detect gha build

* misc(ci): detect gha build

* misc(ci): ok debug

* misc(ci): wtf

* misc(ci): wtf2

* misc(ci): wtf3

* misc(ci): use commit HEAD instead of merge commit for image id

* misc(ci): wtfinfini

* misc(ci): wtfinfini

* misc(ci): KAMEHAMEHA

* Merge TRTLLM in standard CI

* misc(ci): remove input machine

* misc(ci): missing id-token for AWS auth

* misc(ci): missing id-token for AWS auth

* misc(ci): missing id-token for AWS auth

* misc(ci): again...

* misc(ci): again...

* misc(ci): again...

* misc(ci): again...

* misc(ci): missing benchmark

* misc(ci): missing backends

* misc(ci): missing launcher

* misc(ci): give everything aws needs

* misc(ci): give everything aws needs

* misc(ci): fix warnings

* misc(ci): attempt to fix sccache not building trtllm

* misc(ci): attempt to fix sccache not building trtllm again

---------

Co-authored-by: Guillaume LEGENDRE <glegendre01@gmail.com>
Co-authored-by: Hugo Larcher <hugo.larcher@huggingface.co>
Co-authored-by: Pauline Bailly-Masson <155966238+paulinebm@users.noreply.github.com>
2025-01-21 10:19:16 +01:00

81 lines
3.5 KiB
C++

#include <ranges>
#include <nlohmann/json.hpp>
#include "backend.hpp"
#include "hardware.hpp"
namespace huggingface::tgi::backends::trtllm {
tle::ParallelConfig backend_workspace_t::parallel_config() const {
// Single engine (TP = PP = 1) -> using leader mode (no MPI involved)
const auto world_size = config_["/pretrained_config/mapping/world_size"_json_pointer].get<size_t>();
auto mode = tle::CommunicationMode::kLEADER;
std::optional<tle::OrchestratorConfig> orchestratorConfig = std::nullopt;
if (world_size > 1) {
SPDLOG_INFO("Detected sharded engine deployment, using orchestrator mode");
mode = tle::CommunicationMode::kORCHESTRATOR;
orchestratorConfig = std::make_optional<tle::OrchestratorConfig>(true, executor_worker_path_, nullptr,
true);
} else {
SPDLOG_INFO("Detected single engine deployment, using leader mode");
}
return tle::ParallelConfig(tle::CommunicationType::kMPI, mode, std::nullopt, std::nullopt, orchestratorConfig);
}
tle::ExecutorConfig backend_workspace_t::executor_config() const {
// Retrieve the compute capabilities to enable some options at runtime
const auto compute_capabilities = hardware::cuda::compute_capabilities_t();
// Allocate the config
tle::ExecutorConfig executor_config(/* maxBeamWidth = */ 1);
// Set the parallel config as inferred
executor_config.setParallelConfig(parallel_config());
// Define some configuration variables
executor_config.setKvCacheConfig(tle::KvCacheConfig(true));
executor_config.setEnableChunkedContext(compute_capabilities.is_at_least_ampere());
executor_config.setSchedulerConfig(tle::SchedulerConfig(tle::CapacitySchedulerPolicy::kMAX_UTILIZATION));
return executor_config;
}
backend_t::backend_t(std::filesystem::path &engines_folder, std::filesystem::path &executor_worker_path)
: workspace(engines_folder, executor_worker_path), executor_(executor_factory_initializer(workspace)) {}
size_t backend_t::num_tokens_ready() const noexcept {
return executor_.getNumResponsesReady();
}
std::expected<request_id_t, backend_error_t>
backend_t::submit(std::span<const token_id_t> token_ids, const generation_params_t g_params,
const sampling_params_t s_params) noexcept {
SPDLOG_DEBUG("Submit {:d} tokens for scheduling ({}, {})", token_ids.size(), g_params, s_params);
return executor_.enqueueRequest(tle::Request{
{token_ids.begin(), token_ids.end()}, // Making actual copy of the tokens
static_cast<tle::SizeType32>(g_params.max_new_tokens),
true,
(tle::SamplingConfig) s_params,
tle::OutputConfig{ /* returnLogProbs= */ true},
std::nullopt,
std::nullopt,
std::nullopt,
std::nullopt,
workspace.generation_config().stop_words
});
}
std::vector<tle::Response> backend_t::pull_tokens() noexcept {
SPDLOG_TRACE(FMT_STRING("Pulling out tokens ({:d} available)"), num_tokens_ready());
return executor_.awaitResponses();
}
void backend_t::cancel(request_id_t request_id) noexcept {
SPDLOG_TRACE(FMT_STRING("Cancelling request: {:d}"), request_id);
executor_.cancelRequest(request_id);
}
}