API Updates: fleshing out RAG APIs, introduce "llama stack" CLI command (#51)

* add tools to chat completion request * use templates for generating system prompts * Moved ToolPromptFormat and jinja templates to llama_models.llama3.api * <WIP> memory changes - inlined AgenticSystemInstanceConfig so API feels more ergonomic - renamed it to AgentConfig, AgentInstance -> Agent - added a MemoryConfig and `memory` parameter - added `attachments` to input and `output_attachments` to the response - some naming changes * InterleavedTextAttachment -> InterleavedTextMedia, introduce memory tool * flesh out memory banks API * agentic loop has a RAG implementation * faiss provider implementation * memory client works * re-work tool definitions, fix FastAPI issues, fix tool regressions * fix agentic_system utils * basic RAG seems to work * small bug fixes for inline attachments * Refactor custom tool execution utilities * Bug fix, show memory retrieval steps in EventLogger * No need for api_key for Remote providers * add special unicode character ↵ to showcase newlines in model prompt templates * remove api.endpoints imports * combine datatypes.py and endpoints.py into api.py * Attachment / add TTL api * split batch_inference from inference * minor import fixes * use a single impl for ChatFormat.decode_assistant_mesage * use interleaved_text_media_as_str() utilityt * Fix api.datatypes imports * Add blobfile for tiktoken * Add ToolPromptFormat to ChatFormat.encode_message so that tools are encoded properly * templates take optional --format={json,function_tag} * Rag Updates * Add `api build` subcommand -- WIP * fix * build + run image seems to work * <WIP> adapters * bunch more work to make adapters work * api build works for conda now * ollama remote adapter works * Several smaller fixes to make adapters work Also, reorganized the pattern of __init__ inside providers so configuration can stay lightweight * llama distribution -> llama stack + containers (WIP) * All the new CLI for api + stack work * Make Fireworks and Together into the Adapter format * Some quick fixes to the CLI behavior to make it consistent * Updated README phew * Update cli_reference.md * llama_toolchain/distribution -> llama_toolchain/core * Add termcolor * update paths * Add a log just for consistency * chmod +x scripts * Fix api dependencies not getting added to configuration * missing import lol * Delete utils.py; move to agentic system * Support downloading of URLs for attachments for code interpreter * Simplify and generalize `llama api build` yay * Update `llama stack configure` to be very simple also * Fix stack start * Allow building an "adhoc" distribution * Remote `llama api []` subcommands * Fixes to llama stack commands and update docs * Update documentation again and add error messages to llama stack start * llama stack start -> llama stack run * Change name of build for less confusion * Add pyopenapi fork to the repository, update RFC assets * Remove conflicting annotation * Added a "--raw" option for model template printing --------- Co-authored-by: Hardik Shah <hjshah@fb.com> Co-authored-by: Ashwin Bharambe <ashwin@meta.com> Co-authored-by: Dalton Flanagan <6599399+dltn@users.noreply.github.com>
2024-09-03 22:39:39 -07:00 · 2024-09-03 22:39:39 -07:00 · 7bc7785b0d
commit 7bc7785b0d
parent 35093c0b6f
141 changed files with 8252 additions and 4032 deletions
--- a/llama_toolchain/post_training/api/api.py
+++ b/llama_toolchain/post_training/api/api.py
@ -0,0 +1,209 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the terms described in the LICENSE file in
+# the root directory of this source tree.
+
+from datetime import datetime
+from enum import Enum
+
+from typing import Any, Dict, List, Optional, Protocol
+
+from llama_models.schema_utils import json_schema_type, webmethod
+
+from pydantic import BaseModel, Field
+
+from llama_models.llama3.api.datatypes import *  # noqa: F403
+from llama_toolchain.dataset.api import *  # noqa: F403
+from llama_toolchain.common.training_types import *  # noqa: F403
+
+
+class OptimizerType(Enum):
+    adam = "adam"
+    adamw = "adamw"
+    sgd = "sgd"
+
+
+@json_schema_type
+class OptimizerConfig(BaseModel):
+    optimizer_type: OptimizerType
+    lr: float
+    lr_min: float
+    weight_decay: float
+
+
+@json_schema_type
+class TrainingConfig(BaseModel):
+    n_epochs: int
+    batch_size: int
+    shuffle: bool
+    n_iters: int
+
+    enable_activation_checkpointing: bool
+    memory_efficient_fsdp_wrap: bool
+    fsdp_cpu_offload: bool
+
+
+@json_schema_type
+class FinetuningAlgorithm(Enum):
+    full = "full"
+    lora = "lora"
+    qlora = "qlora"
+    dora = "dora"
+
+
+@json_schema_type
+class LoraFinetuningConfig(BaseModel):
+    lora_attn_modules: List[str]
+    apply_lora_to_mlp: bool
+    apply_lora_to_output: bool
+    rank: int
+    alpha: int
+
+
+@json_schema_type
+class QLoraFinetuningConfig(LoraFinetuningConfig):
+    pass
+
+
+@json_schema_type
+class DoraFinetuningConfig(LoraFinetuningConfig):
+    pass
+
+
+@json_schema_type
+class PostTrainingJobLogStream(BaseModel):
+    """Stream of logs from a finetuning job."""
+
+    job_uuid: str
+    log_lines: List[str]
+
+
+@json_schema_type
+class PostTrainingJobStatus(Enum):
+    running = "running"
+    completed = "completed"
+    failed = "failed"
+    scheduled = "scheduled"
+
+
+@json_schema_type
+class RLHFAlgorithm(Enum):
+    dpo = "dpo"
+
+
+@json_schema_type
+class DPOAlignmentConfig(BaseModel):
+    reward_scale: float
+    reward_clip: float
+    epsilon: float
+    gamma: float
+
+
+@json_schema_type
+class PostTrainingSFTRequest(BaseModel):
+    """Request to finetune a model."""
+
+    job_uuid: str
+
+    model: str
+    dataset: TrainEvalDataset
+    validation_dataset: TrainEvalDataset
+
+    algorithm: FinetuningAlgorithm
+    algorithm_config: Union[
+        LoraFinetuningConfig, QLoraFinetuningConfig, DoraFinetuningConfig
+    ]
+
+    optimizer_config: OptimizerConfig
+    training_config: TrainingConfig
+
+    # TODO: define these
+    hyperparam_search_config: Dict[str, Any]
+    logger_config: Dict[str, Any]
+
+
+@json_schema_type
+class PostTrainingRLHFRequest(BaseModel):
+    """Request to finetune a model."""
+
+    job_uuid: str
+
+    finetuned_model: URL
+
+    dataset: TrainEvalDataset
+    validation_dataset: TrainEvalDataset
+
+    algorithm: RLHFAlgorithm
+    algorithm_config: Union[DPOAlignmentConfig]
+
+    optimizer_config: OptimizerConfig
+    training_config: TrainingConfig
+
+    # TODO: define these
+    hyperparam_search_config: Dict[str, Any]
+    logger_config: Dict[str, Any]
+
+
+class PostTrainingJob(BaseModel):
+    job_uuid: str
+
+
+@json_schema_type
+class PostTrainingJobStatusResponse(BaseModel):
+    """Status of a finetuning job."""
+
+    job_uuid: str
+    status: PostTrainingJobStatus
+
+    scheduled_at: Optional[datetime] = None
+    started_at: Optional[datetime] = None
+    completed_at: Optional[datetime] = None
+
+    resources_allocated: Optional[Dict[str, Any]] = None
+
+    checkpoints: List[Checkpoint] = Field(default_factory=list)
+
+
+@json_schema_type
+class PostTrainingJobArtifactsResponse(BaseModel):
+    """Artifacts of a finetuning job."""
+
+    job_uuid: str
+    checkpoints: List[Checkpoint] = Field(default_factory=list)
+
+    # TODO(ashwin): metrics, evals
+
+
+class PostTraining(Protocol):
+    @webmethod(route="/post_training/supervised_fine_tune")
+    def supervised_fine_tune(
+        self,
+        request: PostTrainingSFTRequest,
+    ) -> PostTrainingJob: ...
+
+    @webmethod(route="/post_training/preference_optimize")
+    def preference_optimize(
+        self,
+        request: PostTrainingRLHFRequest,
+    ) -> PostTrainingJob: ...
+
+    @webmethod(route="/post_training/jobs")
+    def get_training_jobs(self) -> List[PostTrainingJob]: ...
+
+    # sends SSE stream of logs
+    @webmethod(route="/post_training/job/logs")
+    def get_training_job_logstream(self, job_uuid: str) -> PostTrainingJobLogStream: ...
+
+    @webmethod(route="/post_training/job/status")
+    def get_training_job_status(
+        self, job_uuid: str
+    ) -> PostTrainingJobStatusResponse: ...
+
+    @webmethod(route="/post_training/job/cancel")
+    def cancel_training_job(self, job_uuid: str) -> None: ...
+
+    @webmethod(route="/post_training/job/artifacts")
+    def get_training_job_artifacts(
+        self, job_uuid: str
+    ) -> PostTrainingJobArtifactsResponse: ...