Repository navigation
[TRTLLM-17267][refactor] Move the embeddings, mm_encoder and tokenize handlers into serving classes - #21575
Draft
JunyiXu-nv wants to merge 31 commits into
Draft
[TRTLLM-17267][refactor] Move the embeddings, mm_encoder and tokenize handlers into serving classes#21575JunyiXu-nv wants to merge 31 commits into
JunyiXu-nv wants to merge 31 commits into
Conversation
…ss lint for the serving frontends The HTTP server, the SMG gRPC frontend and the OpenEngine gRPC frontend learned engine health by reading private members (_check_health, _executor, _fatal_error, _engine_dead). Add the public health API on the classes that own those members and switch all three frontends to it, with no behavior change: - BaseLLM.check_health() (public rename of _check_health; the private name stays for one release as an alias for out-of-tree callers that look it up), BaseLLM.fatal_error, BaseLLM.shutdown_started, BaseLLM.engine_dead. - GenerationExecutor.fatal_error and engine_dead; GenerationExecutorProxy.engine_dead also reads the sticky worker-death flag set before the fatal error is recorded. Add a private-access lint (tests/unittest/serve/test_private_access_lint.py) that counts cross-package private accesses in tensorrt_llm/serve, tensorrt_llm/commands and tensorrt_llm/grpc against a checked-in baseline (tests/unittest/serve/private_access_baseline.yaml) that only shrinks, with companion test files under tests/unittest/serve/api/visual_gen/ and tests/unittest/grpc/openengine/ so VisualGen-only and OpenEngine-only pull requests keep running the check. Replace the three frontend health mirror tests with tests of the production code, rename the _check_health test doubles to the public name, and register the new CPU test files in l0_cpu.yml. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…erving-frontend private accesses The serving frontends (tensorrt_llm/serve, tensorrt_llm/commands, the OpenEngine gRPC frontend) read private members of LLM, the executors, CompletionOutput, SamplingParams, MultimodalDataTracker, OpenAIDisaggServer and PyExecutor. Give each remaining read a thin public entry point on the class that owns the member, switch the frontends to it with no behavior change, and shrink the private-access baseline from 50 recorded accesses to 19. The reads of the two private LlmArgs attributes stay (exposing them is an LLM-args change for its own PR); the remaining baseline entries belong to other in-flight work. New public members (status prototype where api-stability tracks them): LLM.record_fatal_error / multi_frontend_attach_info / multi_frontend_ipc_dir / engine_in_worker_process / hf_model_dir / hf_model_config / resource_governor_queue / encode_limits / get_kv_cache_load / get_openengine_discovery; GenerationExecutor.record_fatal_error and multi_frontend_ipc_dir; CompletionOutput.spec_dec_counters and text_diff_offset (and the SpecDecCounters rename); SamplingParams.set_request_provided_fields and stop_word_ids; MultimodalDataTracker.multimodal_server_config; OpenAIDisaggServer.address; PyExecutor.set_openengine_discovery and set_openengine_node_server; registered_model_type() in inputs/registry. Also adds tensorrt_llm/llmapi/engine_client.py: the EngineClient protocol naming the engine operations every frontend shares (generate, health, load), checked by tests rather than at runtime. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…a bare OpenAIServer The shut-down-LLM case of the launcher-lost watchdog test now calls the production method unbound on a stand-in object whose generator double defines the public members the method reads, the form the file's other stand-in tests use. New tests must not build servers with object.__new__. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…apshots and a CPU test seam for trtllm-serve Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
… baselines at the merge base The route reference is generated from the live apps; the object.__new__ baseline and the gRPC import allowance record what exists on main today and only shrink. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…n the snapshot matrix An attached frontend of a multi-frontend server registers ten launcher forwarding routes before the local table; seven share a method and path with local handlers and three use a path parameter that matches every batch id, so registration order decides which handler serves them. The new configuration records all of it, a live invariant pins the forwarding-before-local order independently of the reference, and the route-class failure message now says why the msgpack route class matters. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
- AGENTS.md: the CPU test command sets LLM_MODELS_ROOT, which the runner requires, and the stage-mapping note says the mapper matches list-entry text, so a file listed through its directory maps by the directory. - check_contracts matches each from-import as <module>.<name>, so 'from pkg import sub' and 'from .. import sub' count as imports of the submodule; new test for both forms and for the allowance entry. - The VisualGen /v1/models test asserts the 200 status. - The AGENTS.md command check also recognizes commands that start with LLM_MODELS_ROOT=. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…on the current main The disaggregated frontend now registers its routes with KvHintsRoute (NVIDIA#20269), which strips kv_hints from request bodies unless the server trusts its ingress. Regenerate the route reference: the six disaggregated configurations record KvHintsRoute where they recorded APIRoute; their route sets and order are unchanged. The route-class invariant, which a regenerated reference cannot hide, now requires KvHintsRoute on every API route of the disaggregated frontend and APIRoute on the coordinator app. Dropping the KV-hint policy from the frontend fails it even after a regeneration, and both apps still reject the aggregated server's msgpack route class. The object.__new__ baseline records the sites on main at the merge base: test_kv_hints.py builds one server that way (NVIDIA#20269), and test_harmony_parsing.py no longer builds any (NVIDIA#21176). The harmony parsing tests are now registered as a whole file (NVIDIA#21411), so the node-id entry that the CPU-selection test used as its example is gone. The node-id check now runs select_entries on a one-entry list of its own, so it no longer depends on which node-id entries l0_cpu.yml holds. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…ail on silent regressions The route snapshot now requires every configuration other than the three invalid ones to build, so a configuration that starts failing at construction cannot pass through a regenerated reference. The object.__new__ ratchet resolves 'from ... import <server class> as <name>' aliases and records each baseline site as its enclosing function or class plus the server class, so swapping one site for another in the same file no longer passes. The baseline keeps the same 52 sites in the same 26 files. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…lm/serve/_base/msgspec_route.py Generated move: _msgpack_decoder, _MsgspecRequest and _MsgspecRoute leave tensorrt_llm/serve/openai_server.py byte for byte; the module imports _MsgspecRoute back for its app router and keeps no re-export (the three names are private). The two disaggregated test importers now import the new module. OpenAIServer.health, .version and .get_model are removed here and replaced by the system router added in the following commits. tensorrt_llm/serve/_base/msgspec_route.py joins legacy-files.txt (its code comes from a legacy-formatted file), and the generated lint configs are regenerated with scripts/legacy_utils.py gen-configs. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…ve the system endpoints through it New structure: tensorrt_llm/serve/_base/ (shared serving infrastructure) and tensorrt_llm/serve/api/ (endpoint packages, one router plus one serving class each; the package __init__ imports nothing so routers load lazily, by role). serve/_base/routing.py provides make_router(), which always uses the msgpack-decoding route class, and attach_routers(), which imports and attaches only the routers whose RouterSpec matches the server role. The first endpoint package, serve/api/ops/system/, serves /health, /version and /v1/models for every role. OpenAIServer builds a SystemServing with the injected _health_probe (today's probe body, SIGINT side effect included) and attaches the routers before the role dispatch; the 12 add_api_route lines for the three paths leave the four role tables. Same routes, methods, route classes, bodies and conflict-relevant order: the recorded route reference is unchanged. The layering tests gain the new packages' contracts, the route-snapshot failure message points new endpoints at the router pattern, and the serve agent docs describe where a new endpoint goes. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…w serve packages tensorrt_llm/serve/_base/ and tensorrt_llm/serve/api/ops/ join the multi-GPU trigger list: the msgpack route class decodes the disaggregated frontend's worker requests, and /health is the disaggregated readiness probe, so changes under these directories keep running the multi-GPU stages that exercise that path. The three new modules join the strict mypy file set, and the type-check pre-commit hook now also triggers on changes under tensorrt_llm/serve/ and tensorrt_llm/commands/serve_cli/, so edits to typed serving modules get local type checking; the hook still checks only the declared file set. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
… tests that build routes by hand Two multi-frontend tests build an OpenAIServer without its constructor and call register_routes() themselves. Since /health, /version and /v1/models are attached through the system router by the constructor, those tests no longer saw the three routes: one asserted that /version stays local on an attached frontend (404), the other reported the three per-frontend-safe entries as removed. Both helpers now set the serving object and attach the routers before the hand registration, in the constructor's order, and the route enumeration walks included routers, which FastAPI 0.139 keeps as objects without a path in app.routes. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…n the OpenEngine metrics test The HTTP-enablement case builds a real OpenAIServer through its constructor and looked for /prometheus/metrics by reading .path off every entry of app.routes. Since the constructor attaches the system router, FastAPI 0.139 keeps an included-router object there that has no .path, and the child process failed with AttributeError. The test now matches a GET /prometheus/metrics request against the routes with the public route-matching API, which works on both ends of the pinned FastAPI range. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…d name The check that endpoint packages under tensorrt_llm/serve/api do not import each other read the package off the from-module alone, so `from tensorrt_llm.serve import api` inside an endpoint package passed and `from tensorrt_llm.serve.api.ops import system` inside ops/system failed. It now reads the package off <module>.<name> per imported name, as the layering contracts do, and a synthetic-edge test pins both forms. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…e/api/visual_gen Move the eight VisualGen request and response models (ImageGenerationRequest, ImageEditRequest, ImageObject, ImageGenerationResponse, MediaReferenceItem, VideoGenerationRequest, VideoJob, VideoJobList) out of tensorrt_llm/serve/openai_protocol.py into the new package tensorrt_llm/serve/api/visual_gen/ (protocol.py), unchanged. openai_protocol keeps the eight names importable through a lazy module __getattr__: importing it no longer loads the VisualGen models or tensorrt_llm.visual_gen, and the re-export is part of its own interface, so it never warns and has no removal release. The imports only the moved block used (UploadFile, AliasChoices, MediaRole) leave openai_protocol with it. Every importer in the repository now imports the new module; the benchmarking guide links the new file. protocol.py joins legacy-files.txt (its code comes from a legacy-formatted file) and the derived lint configs are regenerated. No behavior change. OpenAPI component names may change with the move; they are not a contract. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…isualGen team in CODEOWNERS VisualGen serving code now lives under tensorrt_llm/serve/api/visual_gen/. The new rule, placed after the /tensorrt_llm/serve rule so that it wins, makes @NVIDIA/trt-llm-torch-visual-gen-devs the owner of that directory: changes there need the VisualGen team's approval and no longer need runtime-devs. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…y re-export under a visual_gen path Add tests/unittest/api_stability/visual_gen/test_visual_gen_serve_api.py: it runs the checks of test_serve_api.py on the models of references/trtllm_serve_api.yaml that are defined in tensorrt_llm.serve.api.visual_gen.protocol, and checks that tensorrt_llm.serve.openai_protocol re-exports the eight VisualGen names lazily (importing it does not load the VisualGen module; unknown names still raise AttributeError). The file lives under a visual_gen/ path and has its own l0_cpu.yml entry so that CI selection for VisualGen-only pull requests keeps running it; the reference YAML is unchanged. Add the eight identity rows of the VisualGen names to the VisualGen shim identity table: identity only, no warning and no ban on the old path, because the re-export is openai_protocol's permanent interface. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…its no-warning contract Resolve every pinned model with the full gate's own loader, so a VisualGen model pinned in the reference YAML that is later renamed or removed fails this gate too instead of being dropped silently; the four known models stay a secondary check. Make the lazy-load probe's diagnostic best-effort and identity-based, so a failure prints which modules hold the VisualGen classes instead of crashing on module-level __getattr__ hooks while walking sys.modules. In the same fresh-interpreter probe, assert that reading each of the eight names from openai_protocol emits no FutureWarning or DeprecationWarning. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
… one submodule per API tensorrt_llm/serve/openai_protocol.py held every OpenAI-dialect request and response model plus the helpers that turn them into engine parameters. It becomes the package tensorrt_llm/serve/openai_protocol/ with one submodule per API: common, sampling, tools, chat, completions, embeddings, responses, tokenize, models, kv_cache, control and disagg. Every definition moves unchanged (93 definitions; generated by an AST tool and checked definition by definition against the former module). The package __init__.py is a lazy facade: a name-to-module table and a module __getattr__ import a submodule only when one of its names is first read, so every import from tensorrt_llm.serve.openai_protocol keeps working for every name the former module had, the underscore helpers that tests import included, and the VisualGen models re-exported from tensorrt_llm.serve.api.visual_gen.protocol. The facade is the package's own interface, not a compatibility shim: no warning, no removal release, and no importer in the repository changes. Its TYPE_CHECKING block imports the public names from their submodules so type checkers resolve them; the VisualGen names are left out of it on purpose and appear only as strings in the table, so that importing the facade never loads the VisualGen package. DisaggScheduleStyle, which the former module imported and two importers took from it, stays reachable through the facade. Submodules import each other directly, never through the package, along an acyclic graph: common imports no sibling; sampling, tools, models, tokenize, embeddings and control import common; completions and responses import common and sampling; chat imports common, sampling and tools; kv_cache imports common, tools and chat; disagg imports common, completions and chat. torch is imported only by sampling and disagg, openai_harmony only by chat, and the serving-extension registry only by sampling, so code that needs only the tool types no longer loads the chat or Responses models. The 12 submodules are registered in legacy-files.txt (their code comes from a legacy-formatted file) and the deleted module's line is removed; the derived lint configs are regenerated with scripts/legacy_utils.py gen-configs. The facade is new code and stays under the default formatter. No behavior change. OpenAPI component names may change for classes that share a short name with an openai type, because pydantic qualifies them with the module path; they are not a contract. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…ziness and layering tests/unittest/serve/openai_protocol/test_openai_protocol_facade.py keeps the package tensorrt_llm/serve/openai_protocol/ honest: every name the former single-file module defined (plus DisaggScheduleStyle) is reachable through the facade and is the object its submodule defines; the facade table is frozen at those names; __all__, dir() and the TYPE_CHECKING block agree with the table; unknown names raise AttributeError; the submodule import graph stays inside the allowed acyclic graph and no submodule imports the package itself; importing the package loads no submodule, importing tools loads neither chat nor responses nor completions, and reading ChatCompletionRequest does not load the serving extensions; the private helpers that other tests import through the package stay importable. tests/unittest/serve/api/visual_gen/test_openai_protocol_facade_visual_gen.py holds the VisualGen assertions under a visual_gen/ path so that CI selection runs them for VisualGen-only changes: the eight VisualGen models map to tensorrt_llm.serve.api.visual_gen.protocol, nothing but a lookup of one of them loads that module, and the facade names it only as strings in its table, never in an import statement. tests/unittest/serve/test_layering.py gains the contract that the protocol package imports no handler module (servers, handlers and routers import the models, never the reverse), and tensorrt_llm/serve/AGENTS.md tells new code to import the submodule that defines a model rather than the package. Both test files are collected by the existing unittest/serve and unittest/serve/api/visual_gen/ entries of l0_cpu.yml and carry the cpu_only marker. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…s as strings and bind the test's culprit list unconditionally The table comment in tensorrt_llm/serve/openai_protocol/__init__.py said that a TYPE_CHECKING import of the VisualGen models would load their package at run time, which a TYPE_CHECKING block never does. The real reason is the static import-isolation check that lets CI run only the VisualGen tests for VisualGen-only changes: it forbids an import edge from this package to the VisualGen package even under TYPE_CHECKING, so the facade names that module only as a string. In tests/unittest/serve/openai_protocol/test_openai_protocol_facade.py the list of submodules that import the serving extensions was bound only on the failing path and read in the assertion message; it is now bound before the branch, so an unconditional message can never raise UnboundLocalError. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…junyix-feat-serve-embeddings-mm-encoder Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com> # Conflicts: # pyproject.toml # tensorrt_llm/grpc/openengine/streaming.py # tensorrt_llm/serve/openai_protocol.py # tensorrt_llm/serve/openai_server.py # tests/integration/test_lists/test-db/l0_cpu.yml # tests/unittest/serve/__init__.py # tests/unittest/serve/api/__init__.py # tests/unittest/serve/api/visual_gen/__init__.py
…gs, multimodal encoder and tokenize Create tensorrt_llm/serve/_base/serving.py with BaseServing and GenerateBaseServing, and the EmbeddingServing, MMEncoderServing and TokenizeServing skeletons under tensorrt_llm/serve/api/openai/. Every dependency is a keyword-only constructor argument. The new Group B files are registered in legacy-files.txt and the generated lint configs are regenerated. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…tokenize handlers into their serving classes Generated move. OpenAIServer loses openai_embedding, _normalize_embedding_input, _init_embedding_batcher (to EmbeddingServing), openai_mm_encoder (to MMEncoderServing) and tokenize (to TokenizeServing); register_embedding_routes is deleted, its route is served by the embeddings router. create_error_response, _attach_multimodal_data and await_disconnected move to the serving bases in tensorrt_llm/serve/_base/serving.py and stay reachable on OpenAIServer as aliases of the moved objects while chat and completions still call them. _record_generator_termination moves to the same module. tensorrt_llm/serve/encode_batcher.py becomes tensorrt_llm/serve/api/openai/embeddings/batcher.py; the old module is deleted and every import of it is rewritten. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…tokenize routers and serving objects Each new package gets an api_router.py built with make_router() and a RouterSpec in ROUTERS for exactly the roles the route had. The composition root builds RenderResources once after _init_llm and injects it, constructs the serving objects inside the role branches, puts them on app.state, and attaches the routers after the role switch, where every attach_router can check its app.state key. The two route registrations the routers now serve leave register_routes and register_mm_encoder_routes. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…enize serving classes on CPU Construction tests (every dependency is a required keyword-only argument), route tests through the real composition root and routers (msgpack bodies included), the batcher reference the health probe reads, the injected RenderResources carrying the server chat template, the serving bases and the borrowed-helper identity on OpenAIServer. The encode-only fake now encodes fixed embeddings and scripted outputs can carry disaggregated params; the app factory passes the chat-template knobs. The runtime-API tests build EmbeddingServing and patch the termination telemetry where the moved helper looks it up; the lazy-import guard also checks that importing tensorrt_llm.serve loads no endpoint package. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
… and the serve guide The multi-GPU trigger table gains the three package prefixes, so code moved out of tensorrt_llm/serve/openai_server.py keeps its trigger; the three router modules join the mypy file list; the serve guide states the constructor rule of serving classes. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
… stub provide the tokenizer The test runs the real OpenAIServer constructor with _init_llm replaced by a stub. The constructor now builds the tokenize serving object for the LLM role from self.tokenizer, so the stub sets it too; a missing dependency fails at construction on purpose instead of at request time. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
…ition root with the tokenize serving object Both tests build a partial OpenAIServer and register its routes the way the constructor does. The constructor now puts the tokenize serving object on app.state and attaches the role's routers after the local table, so the re-enactments do the same; the tokenize router refuses to attach without its serving object. Signed-off-by: JunyiXu-nv <219237550+JunyiXu-nv@users.noreply.github.com>
JunyiXu-nv
force-pushed
the
dev-junyix-feat-serve-embeddings-mm-encoder
branch
from
November 4, 2026 12:51
d4d2d0f to
600bcd5
Compare
1 task done
This branch has not been deployed
This file contains hidden or 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
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stacked on #21544 ("Split openai_protocol into a package") and #21445 ("Add public runtime accessors for the remaining serving-frontend private accesses"): the first commits below belong to those PRs and disappear from this diff when they merge; this PR's own change is the last 7 commits, on top of the merge of the two (
d2e8afb843).@coderabbitai summary
Description
Part of the trtllm-serve refactor (epic TRTLLM-17056, task TRTLLM-17267: move the remaining handlers off
OpenAIServerinto one serving class per endpoint). This PR moves the EMBEDDING role'sPOST /v1/embeddings, the MM_ENCODER role'sPOST /v1/chat/completionsand the LLM roles'POST /_internal/tokenizeoffOpenAIServerintoEmbeddingServing,MMEncoderServingandTokenizeServingundertensorrt_llm/serve/api/openai/{embeddings,mm_encoder,tokenize}/, each with anapi_router.pycreated bymake_router()and attached throughROUTERSfor exactly the roles the route had. It creates the serving bases intensorrt_llm/serve/_base/serving.py:BaseServing(holdscreate_error_response) andGenerateBaseServing(holdsawait_disconnectedand_attach_multimodal_data; the base of the completions, chat and Responses serving classes that follow), and moves the module helper_record_generator_terminationthere.tensorrt_llm/serve/encode_batcher.pymoves whole totensorrt_llm/serve/api/openai/embeddings/batcher.py; the old module is deleted and every import of it is rewritten. The composition root (OpenAIServer.__init__) buildsRenderResourcesonce after_init_llmand injects it, constructs the serving objects inside the role branches, puts them onapp.state, and attaches the routers after the role switch, where each router can check that its serving object exists.Builds on #21448 (the router skeleton and
ops.system), #21406 (route, middleware-order and shim-identity snapshots and the fixture factory), #21544 (the protocol package) and #21445 (the public runtime accessors that replaced the private engine reads in the moved code).Why the tokenize route moves in this PR: the handler needs only the tokenizer and has no blocker beyond this PR's, and moving it now keeps the completions and chat moves to their large handlers. Why the composition root keeps
self.embedding_batcher: the lifespan starts and stops the batcher and the health probe reports it; the attribute is bound toEmbeddingServing.embedding_batcher, the same object (tested).Behavior
No behavior change. Every route keeps its path, method, msgpack route class and role set (the route snapshot and the middleware-order snapshot are unchanged). Notes a reviewer may want to check:
openai_mm_encoderno longer callsRenderResources.from_server(self)per request; it uses theRenderResourcesthe composition root built once after_init_llm. Equivalent: every attributefrom_serverreads is set before__init__returns and never reassigned, andharmony_adapteris never set on the MM_ENCODER role (only the harmony chat path sets it, on the LLM role).create_mm_embedding_responsereturns the 500 error response, the handler then calls.model_dump()on it, and theAttributeErroris answered by the handler'sexcept Exceptionas a 400. The new test pins today's 400; the fix is a separate change (see "Coordination notes").{"error": ...}) embeds the route function's source location; that text changes with any move and is not a contract.Moves
Pure-move check (head `600bcd503b`, run from a clean checkout)
Pure-move check: base d2e8afb, head 600bcd5, tools 517fc468330a
ROUTERSEncodeBatcherEncodeFnInputTooLongErrorQueueFullError_QueuedRequestOpenAIServer._attach_multimodal_dataOpenAIServer.await_disconnectedOpenAIServer.create_error_response_record_generator_terminationOpenAIServer.register_embedding_routesOpenAIServer._init_embedding_batcherOpenAIServer._normalize_embedding_inputOpenAIServer.openai_embeddingOpenAIServer.openai_mm_encoderOpenAIServer.tokenizeOpenAIServer.__init__OpenAIServer._attach_multimodal_dataOpenAIServer.await_disconnectedOpenAIServer.create_error_responseOpenAIServer.register_mm_encoder_routesOpenAIServer.register_routesBaseServingGenerateBaseServing.__init__GenerateBaseServingtensorrt_llm/serve/api/openai/embeddings/api_router.py::*EmbeddingServing.__init__EmbeddingServingtensorrt_llm/serve/api/openai/mm_encoder/api_router.py::*MMEncoderServing.__init__MMEncoderServingtensorrt_llm/serve/api/openai/tokenize/api_router.py::*TokenizeServing.__init__TokenizeServingOpenAIServer._generate_serving_kwargsReplaced
OpenAIServer.register_embedding_routes->tensorrt_llm/serve/api/openai/embeddings/api_router.py::openai_embedding: its last route is served by the embeddings router; /health, /version and /v1/models are already served by serve/api/ops/systemAdded
OpenAIServer._generate_serving_kwargs: constructor arguments of every GenerateBaseServing, read once from the serverBaseServing: serving base class; holds create_error_responseGenerateBaseServing: base of the generate serving classesGenerateBaseServing.__init__: explicit dependencies instead of server attributes, so missing wiring fails at constructionEmbeddingServing: serving class of /v1/embeddingsEmbeddingServing.__init__: explicit dependencies; builds the batcherMMEncoderServing: serving class of the multimodal-encoder /v1/chat/completionsMMEncoderServing.__init__: explicit dependencies, injected RenderResourcesTokenizeServing: serving class of /_internal/tokenizeTokenizeServing.__init__: explicit dependenciestensorrt_llm/serve/api/openai/embeddings/api_router.py::*: new router moduletensorrt_llm/serve/api/openai/mm_encoder/api_router.py::*: new router moduletensorrt_llm/serve/api/openai/tokenize/api_router.py::*: new router moduleRewritten
OpenAIServer.openai_mm_encoder(tensorrt_llm/serve/api/openai/mm_encoder/serving.py):RenderResources.from_server(self)->self._render_resourcesModified (declared)
OpenAIServer.__init__: RenderResources built once; serving objects built in the role branches; attach_routers after the role switch--- base
+++ head
@@ -94,6 +94,7 @@
self._init_visual_gen()
else:
self._init_llm(chat_template)
@asynccontextmanager
async def lifespan(app: FastAPI):
@@ -166,20 +167,29 @@
detail = detail[:4096] + ' ... [truncated %d bytes]' % (len(detail) - 4096)
return JSONResponse(status_code=400, content={'error': detail})
self.app.state.ops_serving_system = SystemServing(model=self.model, health_probe=self._health_probe)
if self.server_role is ServerRole.VISUAL_GEN:
assert self._is_visual_gen, 'generator must be a VisualGen for VISUAL_GEN server'
self.register_visual_gen_routes()
elif self.server_role is ServerRole.MM_ENCODER:
assert isinstance(self.generator, MultimodalEncoder), 'generator must be a MultimodalEncoder for multimodal encoder'
assert getattr(self.generator.args, 'encode_only', False), 'generator must be an encode_only=True LLM for the embedding server'
if self._collect_perf_metrics:
self.app.add_middleware(PerfMetricsMiddleware, expose_headers=self._expose_perf_metrics, writer=self._perf_metrics_writer, adjusted_clock=self._adjusted_steady_clock)
self.app.add_middleware(ServerArrivalTimeMiddleware, adjusted_clock=self._adjusted_steady_clock)
OpenAIServer.register_routes: /_internal/tokenize comes from the tokenize router--- base
+++ head
@@ -28,7 +28,6 @@
self.app.add_api_route('/v1/responses', self.openai_responses, methods=['POST'])
self.app.add_api_route('/v1/responses/{response_id}', self.openai_responses_get_response, methods=['GET'])
self.app.add_api_route('/v1/responses/{response_id}', self.openai_responses_delete_response, methods=['DELETE'])
self.app.add_api_route('/start_profile', self.start_profile, methods=['POST'])
self.app.add_api_route('/stop_profile', self.stop_profile, methods=['POST'])
self._register_runtime_control_routes()
OpenAIServer.register_mm_encoder_routes: /v1/chat/completions comes from the multimodal-encoder router--- base
+++ head
@@ -1,3 +1,2 @@
def register_mm_encoder_routes(self):
self.app.add_api_route('/metrics', self.get_iteration_stats, methods=['GET'])
ROUTERS: the embeddings, multimodal-encoder and tokenize routers join the routing table for their role groups--- base
+++ head
@@ -1 +1 @@
-ROUTERS: tuple[RouterSpec, ...] = (RouterSpec('tensorrt_llm.serve.api.ops.system.api_router', roles=ALL_ROLE_GROUPS),)
+ROUTERS: tuple[RouterSpec, ...] = (RouterSpec('tensorrt_llm.serve.api.ops.system.api_router', roles=ALL_ROLE_GROUPS), RouterSpec('tensorrt_llm.serve.api.openai.embeddings.api_router', roles=frozenset({RoleGroup.EMBEDDING})), RouterSpec('tensorrt_llm.serve.api.openai.mm_encoder.api_router', roles=frozenset({RoleGroup.MM_ENCODER})), RouterSpec('tensorrt_llm.serve.api.openai.tokenize.api_router', roles=frozenset({RoleGroup.LLM})))
13 definitions moved verbatim; 0 differences; 1 rewritten (declared); 1 files moved; 3 borrowed; 1 replaced; 13 added; 4 modified (declared); shim exports 0 names