moved containers dir and created docker_files dir

This commit is contained in:
2026-04-28 22:36:13 -04:00
parent 3056c19f69
commit 09f7f0187f
8 changed files with 286 additions and 27 deletions
-25
View File
@@ -1,25 +0,0 @@
# Unsloth fine-tuning container for Qwen 3.5 4B on RTX 3090.
#
# Build:
# docker build -f python/prompt_bench/Dockerfile.finetune -t bill-finetune .
#
# Run:
# docker run --rm --device=nvidia.com/gpu=all --ipc=host \
# -v $(pwd)/output:/workspace/output \
# -v $(pwd)/output/finetune_dataset.jsonl:/workspace/dataset.jsonl:ro \
# -v /zfs/models/hf:/models \
# bill-finetune \
# --dataset /workspace/dataset.jsonl \
# --output-dir /workspace/output/qwen-bill-summarizer
FROM ghcr.io/unslothai/unsloth:latest
RUN pip install --no-cache-dir typer
WORKDIR /workspace
COPY python/prompt_bench/finetune.py python/prompt_bench/finetune.py
COPY config/prompts/summarization_prompts.toml config/prompts/summarization_prompts.toml
COPY python/prompt_bench/__init__.py python/prompt_bench/__init__.py
COPY python/__init__.py python/__init__.py
ENTRYPOINT ["python", "-m", "pipelines.prompt_bench.finetune"]
-1
View File
@@ -1 +0,0 @@
"""Prompt benchmarking system for evaluating LLMs via vLLM."""
-179
View File
@@ -1,179 +0,0 @@
"""Docker container lifecycle management for Unsloth fine-tuning."""
from __future__ import annotations
import logging
import subprocess
from pathlib import Path
from typing import Annotated
import typer
from pipelines.tools.containers.lib import check_gpu_free
logger = logging.getLogger(__name__)
CONTAINER_NAME = "bill-finetune"
FINETUNE_IMAGE = "bill-finetune:latest"
REPO_DIR = Path(__file__).resolve().parents[4]
DEFAULT_HF_CACHE = Path("/zfs/models/hf")
def build_image() -> None:
"""Build the fine-tuning Docker image."""
logger.info("Building fine-tuning image: %s", FINETUNE_IMAGE)
result = subprocess.run(
[
"docker",
"build",
"-f",
str(REPO_DIR / "python/prompt_bench/Dockerfile.finetune"),
"-t",
FINETUNE_IMAGE,
".",
],
text=True,
check=False,
)
if result.returncode != 0:
message = "Failed to build fine-tuning image"
raise RuntimeError(message)
logger.info("Image built: %s", FINETUNE_IMAGE)
def start_finetune(
*,
dataset_path: Path,
output_dir: Path,
hf_cache: Path = DEFAULT_HF_CACHE,
) -> None:
"""Run the fine-tuning container.
Args:
dataset_path: Host path to the fine-tuning JSONL dataset.
output_dir: Host path where the trained model will be saved.
hf_cache: Host path to HuggingFace model cache (bind-mounted to avoid re-downloading).
validation_split: Fraction of data held out for validation.
"""
dataset_path = dataset_path.resolve()
output_dir = output_dir.resolve()
if not dataset_path.is_file():
message = f"Dataset not found: {dataset_path}"
raise FileNotFoundError(message)
output_dir.mkdir(parents=True, exist_ok=True)
stop_finetune()
hf_cache = hf_cache.resolve()
hf_cache.mkdir(parents=True, exist_ok=True)
command = [
"docker",
"run",
"--name",
CONTAINER_NAME,
"--device=nvidia.com/gpu=all",
"--ipc=host",
"-v",
f"{hf_cache}:/root/.cache/huggingface",
"-v",
f"{output_dir}:/workspace/output/qwen-bill-summarizer",
"-v",
f"{dataset_path}:/workspace/dataset.jsonl:ro",
FINETUNE_IMAGE,
"--dataset",
"/workspace/dataset.jsonl",
"--output-dir",
"/workspace/output/qwen-bill-summarizer",
]
logger.info("Starting fine-tuning container")
logger.info(" Dataset: %s", dataset_path)
logger.info(" Output: %s", output_dir)
result = subprocess.run(command, text=True, check=False)
if result.returncode != 0:
message = f"Fine-tuning container exited with code {result.returncode}"
raise RuntimeError(message)
logger.info("Fine-tuning complete. Model saved to %s", output_dir)
def stop_finetune() -> None:
"""Stop and remove the fine-tuning container."""
logger.info("Stopping fine-tuning container")
subprocess.run(["docker", "stop", CONTAINER_NAME], capture_output=True, check=False)
subprocess.run(
["docker", "rm", "-f", CONTAINER_NAME], capture_output=True, check=False
)
def logs_finetune() -> str | None:
"""Return recent logs from the fine-tuning container, or None if not running."""
result = subprocess.run(
["docker", "logs", "--tail", "50", CONTAINER_NAME],
capture_output=True,
text=True,
check=False,
)
if result.returncode != 0:
return None
return result.stdout + result.stderr
app = typer.Typer(help="Fine-tuning container management.")
@app.command()
def build() -> None:
"""Build the fine-tuning Docker image."""
build_image()
@app.command()
def run(
dataset: Annotated[Path, typer.Option(help="Fine-tuning JSONL")] = REPO_DIR
/ "data/finetune_dataset.jsonl",
output_dir: Annotated[
Path, typer.Option(help="Where to save the trained model")
] = REPO_DIR / "data/output/qwen-bill-summarizer",
hf_cache: Annotated[
Path, typer.Option(help="Host path to HuggingFace model cache")
] = DEFAULT_HF_CACHE,
log_level: Annotated[str, typer.Option(help="Log level")] = "INFO",
) -> None:
"""Run fine-tuning inside a Docker container."""
logging.basicConfig(
level=log_level, format="%(asctime)s %(levelname)s %(name)s: %(message)s"
)
check_gpu_free()
start_finetune(
dataset_path=dataset,
output_dir=output_dir,
hf_cache=hf_cache,
)
@app.command()
def stop() -> None:
"""Stop and remove the fine-tuning container."""
stop_finetune()
@app.command()
def logs() -> None:
"""Show recent logs from the fine-tuning container."""
output = logs_finetune()
if output is None:
typer.echo("No running fine-tuning container found.")
raise typer.Exit(code=1)
typer.echo(output)
def cli() -> None:
"""Typer entry point."""
app()
if __name__ == "__main__":
cli()
-23
View File
@@ -1,23 +0,0 @@
from __future__ import annotations
import logging
import subprocess
logger = logging.getLogger(__name__)
def check_gpu_free() -> None:
"""Warn if GPU-heavy processes (e.g. Ollama) are running."""
result = subprocess.run(
["nvidia-smi", "--query-compute-apps=pid,process_name", "--format=csv,noheader"],
capture_output=True,
text=True,
check=False,
)
if result.returncode != 0:
logger.warning("Could not query GPU processes: %s", result.stderr.strip())
return
processes = result.stdout.strip()
if processes:
logger.warning("GPU processes detected:\n%s", processes)
logger.warning("Consider stopping Ollama (sudo systemctl stop ollama) before benchmarking")
-70
View File
@@ -1,70 +0,0 @@
"""Docker container lifecycle management for vLLM."""
from __future__ import annotations
import logging
import subprocess
logger = logging.getLogger(__name__)
CONTAINER_NAME = "vllm-bench"
VLLM_IMAGE = "vllm/vllm-openai:v0.19.0"
def start_vllm(
*,
model: str,
port: int,
model_dir: str,
gpu_memory_utilization: float,
) -> None:
"""Start a vLLM container serving the given model.
Args:
model: HuggingFace model directory name (relative to model_dir).
port: Host port to bind.
model_dir: Host path containing HuggingFace model directories.
gpu_memory_utilization: Fraction of GPU memory to use (0-1).
"""
command = [
"docker",
"run",
"-d",
"--name",
CONTAINER_NAME,
"--device=nvidia.com/gpu=all",
"--ipc=host",
"-v",
f"{model_dir}:/models",
"-p",
f"{port}:8000",
VLLM_IMAGE,
"--model",
f"/models/{model}",
"--served-model-name",
model,
"--gpu-memory-utilization",
str(gpu_memory_utilization),
"--max-model-len",
"4096",
]
logger.info("Starting vLLM container with model: %s", model)
stop_vllm()
result = subprocess.run(command, capture_output=True, text=True, check=False)
if result.returncode != 0:
msg = f"Failed to start vLLM container: {result.stderr.strip()}"
raise RuntimeError(msg)
logger.info("vLLM container started: %s", result.stdout.strip()[:12])
def stop_vllm() -> None:
"""Stop and remove the vLLM benchmark container."""
logger.info("Stopping vLLM container")
subprocess.run(["docker", "stop", CONTAINER_NAME], capture_output=True, check=False)
subprocess.run(["docker", "rm", "-f", CONTAINER_NAME], capture_output=True, check=False)
subprocess.run(
["docker", "network", "disconnect", "-f", "bridge", CONTAINER_NAME],
capture_output=True,
check=False,
)
logger.info("vLLM container stopped and removed")