Skip to content

Commit

Permalink
Merge pull request redhat-et#50 from JudeNiroshan/add_mandatory_params
Browse files Browse the repository at this point in the history 
add missing mandatory parameters for generate_data
  • Loading branch information
@cooktheryan @sallyom
cooktheryan authored and sallyom committed Oct 8, 2024
2 parents a89aa5d + 9c21c7c commit b473e9d
Show file tree
Hide file tree
Showing 3 changed files with 64 additions and 61 deletions.
2 changes: 1 addition & 1 deletion eval/mt_bench/components.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -109,7 +109,7 @@ def stop_vllm_server_by_name():
import os

import torch
from instructlab.eval import mt_bench_answers, mt_bench_judgment
from instructlab.eval.mt_bench import MTBenchEvaluator

os.environ["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
vllm_server = "http://localhost:8000/v1"
Expand Down
121 changes: 61 additions & 60 deletions pipeline.yaml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -592,33 +592,34 @@ deploymentSpec:
\ *\n\ndef data_processing_op(\n sdg: dsl.Input[dsl.Dataset],\n processed_data:\
\ dsl.Output[dsl.Dataset],\n model: dsl.Input[dsl.Artifact],\n max_seq_len:\
\ Optional[int] = 4096,\n max_batch_len: Optional[int] = 20000,\n):\n\
\ import instructlab.training.data_process as dp\n import os\n \
\ from instructlab.training import (\n TrainingArgs,\n DataProcessArgs,\n\
\ )\n\n # define training-specific arguments\n training_args =\
\ TrainingArgs(\n # define data-specific arguments\n model_path=model.path,\n\
\ data_path=f\"{sdg.path}/*_train_msgs*.jsonl\",\n data_output_dir=processed_data.path,\n\
\ # define model-trianing parameters\n max_seq_len=max_seq_len,\n\
\ max_batch_len=max_batch_len,\n # XXX(shanand): We don't\
\ need the following arguments\n # for data processing. Added them\
\ for now to avoid\n # Pydantic validation errors for TrainingArgs\n\
\ ckpt_output_dir=\"data/saved_checkpoints\",\n num_epochs=2,\n\
\ effective_batch_size=3840,\n save_samples=0,\n learning_rate=2e-6,\n\
\ warmup_steps=800,\n is_padding_free=True,\n )\n\n \
\ def data_processing(train_args: TrainingArgs) -> None:\n # early\
\ validation logic here\n if train_args.max_batch_len < train_args.max_seq_len:\n\
\ raise ValueError(\n f\"the `max_batch_len` cannot\
\ be less than `max_seq_len`: {train_args.max_batch_len=} < {train_args.max_seq_len=}\"\
\n )\n\n # process the training data\n if not\
\ os.path.exists(train_args.data_output_dir):\n os.makedirs(train_args.data_output_dir,\
\ exist_ok=True)\n dp.main(\n DataProcessArgs(\n \
\ # XXX(osilkin): make a decision here, either:\n \
\ # 1. the CLI is fully responsible for managing where the data is\
\ written\n # 2. we never cache it and simply write it\
\ to a tmp file every time.\n #\n # An important\
\ reason for why #1 would be preferable is in the case of OpenShift/SELinux\n\
\ # where the user has a defined place for new temporary\
\ data to be written.\n data_output_path=train_args.data_output_dir,\n\
\ model_path=train_args.model_path,\n data_path=train_args.data_path,\n\
\ import os\n\n import instructlab.training.data_process as dp\n \
\ from instructlab.training import (\n DataProcessArgs,\n \
\ TrainingArgs,\n )\n\n # define training-specific arguments\n \
\ training_args = TrainingArgs(\n # define data-specific arguments\n\
\ model_path=model.path,\n data_path=f\"{sdg.path}/*_train_msgs*.jsonl\"\
,\n data_output_dir=processed_data.path,\n # define model-trianing\
\ parameters\n max_seq_len=max_seq_len,\n max_batch_len=max_batch_len,\n\
\ # XXX(shanand): We don't need the following arguments\n \
\ # for data processing. Added them for now to avoid\n # Pydantic\
\ validation errors for TrainingArgs\n ckpt_output_dir=\"data/saved_checkpoints\"\
,\n num_epochs=2,\n effective_batch_size=3840,\n save_samples=0,\n\
\ learning_rate=2e-6,\n warmup_steps=800,\n is_padding_free=True,\n\
\ )\n\n def data_processing(train_args: TrainingArgs) -> None:\n \
\ # early validation logic here\n if train_args.max_batch_len\
\ < train_args.max_seq_len:\n raise ValueError(\n \
\ f\"the `max_batch_len` cannot be less than `max_seq_len`: {train_args.max_batch_len=}\
\ < {train_args.max_seq_len=}\"\n )\n\n # process\
\ the training data\n if not os.path.exists(train_args.data_output_dir):\n\
\ os.makedirs(train_args.data_output_dir, exist_ok=True)\n \
\ dp.main(\n DataProcessArgs(\n # XXX(osilkin):\
\ make a decision here, either:\n # 1. the CLI is fully\
\ responsible for managing where the data is written\n #\
\ 2. we never cache it and simply write it to a tmp file every time.\n\
\ #\n # An important reason for why #1 would\
\ be preferable is in the case of OpenShift/SELinux\n # where\
\ the user has a defined place for new temporary data to be written.\n \
\ data_output_path=train_args.data_output_dir,\n \
\ model_path=train_args.model_path,\n data_path=train_args.data_path,\n\
\ max_seq_len=train_args.max_seq_len,\n chat_tmpl_path=train_args.chat_tmpl_path,\n\
\ )\n )\n\n data_processing(train_args=training_args)\n\
\n"
Expand Down Expand Up @@ -1044,8 +1045,8 @@ deploymentSpec:
\ int,\n batch_size: int,\n device: str = None,\n models_list:\
\ List[str] = None,\n models_folder: Optional[str] = None,\n) -> NamedTuple(\"\
outputs\", best_model=str, best_score=float):\n import json\n import\
\ os\n import torch\n from instructlab.eval.mmlu import MMLUEvaluator,\
\ MMLU_TASKS\n\n mmlu_tasks = mmlu_tasks_list.split(\",\") if mmlu_tasks_list\
\ os\n\n import torch\n from instructlab.eval.mmlu import MMLU_TASKS,\
\ MMLUEvaluator\n\n mmlu_tasks = mmlu_tasks_list.split(\",\") if mmlu_tasks_list\
\ else MMLU_TASKS\n\n if models_list is None and models_folder:\n \
\ models_list = os.listdir(models_folder)\n\n # Device setup and\
\ debug\n gpu_available = torch.cuda.is_available()\n gpu_name = (\n\
Expand Down Expand Up @@ -1114,28 +1115,27 @@ deploymentSpec:
\ # https://github.com/instructlab/eval/blob/main/src/instructlab/eval/mt_bench.py#L36\n\
\ max_workers: str,\n models_list: List[str] = None,\n models_folder:\
\ Optional[str] = None,\n device: str = None,\n) -> NamedTuple(\"outputs\"\
, best_model=str, best_score=float):\n def launch_vllm(\n model_path:\
\ str, gpu_count: int, retries: int = 60, delay: int = 5\n ):\n \
\ import subprocess\n import sys\n import time\n \
\ import requests\n\n if gpu_count > 0:\n command = [\n\
\ sys.executable,\n \"-m\",\n \
\ \"vllm.entrypoints.openai.api_server\",\n \"--model\"\
,\n model_path,\n \"--tensor-parallel-size\"\
,\n str(gpu_count),\n ]\n else:\n \
\ command = [\n sys.executable,\n \"\
-m\",\n \"vllm.entrypoints.openai.api_server\",\n \
\ \"--model\",\n model_path,\n ]\n\n \
\ subprocess.Popen(args=command)\n\n server_url = \"http://localhost:8000/v1\"\
\n print(f\"Waiting for vLLM server to start at {server_url}...\"\
)\n\n for attempt in range(retries):\n try:\n \
\ response = requests.get(f\"{server_url}/models\")\n \
\ if response.status_code == 200:\n print(f\"vLLM\
\ server is up and running at {server_url}.\")\n return\n\
\ except requests.ConnectionError:\n pass\n\n\
\ print(\n f\"Server not available yet, retrying\
\ in {delay} seconds (Attempt {attempt + 1}/{retries})...\"\n \
\ )\n time.sleep(delay)\n\n raise RuntimeError(\n \
\ f\"Failed to start vLLM server at {server_url} after {retries}\
, best_model=str, best_score=float):\n def launch_vllm(model_path: str,\
\ gpu_count: int, retries: int = 60, delay: int = 5):\n import subprocess\n\
\ import sys\n import time\n\n import requests\n\n\
\ if gpu_count > 0:\n command = [\n sys.executable,\n\
\ \"-m\",\n \"vllm.entrypoints.openai.api_server\"\
,\n \"--model\",\n model_path,\n \
\ \"--tensor-parallel-size\",\n str(gpu_count),\n \
\ ]\n else:\n command = [\n sys.executable,\n\
\ \"-m\",\n \"vllm.entrypoints.openai.api_server\"\
,\n \"--model\",\n model_path,\n \
\ ]\n\n subprocess.Popen(args=command)\n\n server_url =\
\ \"http://localhost:8000/v1\"\n print(f\"Waiting for vLLM server\
\ to start at {server_url}...\")\n\n for attempt in range(retries):\n\
\ try:\n response = requests.get(f\"{server_url}/models\"\
)\n if response.status_code == 200:\n \
\ print(f\"vLLM server is up and running at {server_url}.\")\n \
\ return\n except requests.ConnectionError:\n \
\ pass\n\n print(\n f\"Server not available\
\ yet, retrying in {delay} seconds (Attempt {attempt + 1}/{retries})...\"\
\n )\n time.sleep(delay)\n\n raise RuntimeError(\n\
\ f\"Failed to start vLLM server at {server_url} after {retries}\
\ retries.\"\n )\n\n # This seems like excessive effort to stop\
\ the vllm process, but merely saving & killing the pid doesn't work\n \
\ # Also, the base image does not include `pkill` cmd, so can't pkill\
Expand All @@ -1160,8 +1160,8 @@ deploymentSpec:
\ to terminate process with PID {process.info['pid']}.\"\n \
\ )\n except Exception as e:\n print(\n\
\ f\"Failed to terminate process with PID {process.info['pid']}.\
\ Error: {e}\"\n )\n\n import json\n import torch\n\
\ import os\n\n from instructlab.eval.mt_bench import MTBenchEvaluator\n\
\ Error: {e}\"\n )\n\n import json\n import os\n\
\n import torch\n from instructlab.eval.mt_bench import MTBenchEvaluator\n\
\n os.environ[\"PYTORCH_CUDA_ALLOC_CONF\"] = \"expandable_segments:True\"\
\n vllm_server = \"http://localhost:8000/v1\"\n\n gpu_available =\
\ torch.cuda.is_available()\n gpu_name = (\n torch.cuda.get_device_name(torch.cuda.current_device())\n\
Expand Down Expand Up @@ -1235,19 +1235,20 @@ deploymentSpec:
- "\nimport kfp\nfrom kfp import dsl\nfrom kfp.dsl import *\nfrom typing import\
\ *\n\ndef sdg_op(\n num_instructions_to_generate: int,\n taxonomy:\
\ dsl.Input[dsl.Dataset],\n sdg: dsl.Output[dsl.Dataset],\n repo_branch:\
\ Optional[str],\n repo_pr: Optional[int],\n):\n import openai\n \
\ from instructlab.sdg import generate_data\n from instructlab.sdg.utils.taxonomy\
\ import read_taxonomy\n from os import getenv\n\n api_key = getenv(\"\
api_key\")\n model = getenv(\"model\")\n endpoint = getenv(\"endpoint\"\
)\n client = openai.OpenAI(base_url=endpoint, api_key=api_key)\n\n \
\ taxonomy_base = \"main\" if repo_branch or (repo_pr and int(repo_pr)\
\ Optional[str],\n repo_pr: Optional[int],\n):\n from os import getenv\n\
\n import openai\n from instructlab.sdg import generate_data\n \
\ from instructlab.sdg.utils.taxonomy import read_taxonomy\n\n api_key\
\ = getenv(\"api_key\")\n model = getenv(\"model\")\n endpoint = getenv(\"\
endpoint\")\n client = openai.OpenAI(base_url=endpoint, api_key=api_key)\n\
\n taxonomy_base = \"main\" if repo_branch or (repo_pr and int(repo_pr)\
\ > 0) else \"empty\"\n\n print(\"Generating syntetic dataset for:\"\
)\n print()\n print(read_taxonomy(taxonomy.path, taxonomy_base))\n\
\n # generate_data has a magic word for its taxonomy_base argument -\
\ `empty`\n # it allows generating from the whole repo, see:\n # https://github.com/instructlab/sdg/blob/c6a9e74a1618b1077cd38e713b8aaed8b7c0c8ce/src/instructlab/sdg/utils/taxonomy.py#L230\n\
\ generate_data(\n client=client,\n num_instructions_to_generate=num_instructions_to_generate,\n\
\ output_dir=sdg.path,\n taxonomy=taxonomy.path,\n \
\ taxonomy_base=taxonomy_base,\n model_name=model,\n )\n\n"
\ taxonomy_base=taxonomy_base,\n model_name=model,\n chunk_word_count=1000,\n\
\ server_ctx_size=4096,\n )\n\n"
image: quay.io/tcoufal/ilab-sdg:latest
pipelineInfo:
description: InstructLab pipeline
Expand Down
2 changes: 2 additions & 0 deletions sdg/components.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -62,4 +62,6 @@ def sdg_op(
taxonomy=taxonomy.path,
taxonomy_base=taxonomy_base,
model_name=model,
chunk_word_count=1000,
server_ctx_size=4096,
)

0 comments on commit b473e9d

Please sign in to comment.