File size: 9,145 Bytes
f128c67 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 | """
statlens — command-line entry point.
Subcommands:
statlens serve spin up local LLM backend + FastAPI; open browser
statlens download fetch the LoRA adapter into the HF cache
(NOT the 64 GB base — that's the user's job)
statlens info show GPU / cache state / paths
statlens classify one-shot CLI run (TSV + context.txt → result folder)
"""
from __future__ import annotations
import os
import signal
import sys
import webbrowser
import click
from . import (
__version__,
DEFAULT_BASE_MODEL_REF,
DEFAULT_LORA_REPO,
DEFAULT_LORA_SUBFOLDER,
DEFAULT_QUANTIZATION,
LOCAL_LLM_PORT,
LOCAL_WEB_PORT,
)
from . import runtime
def _common_serve_opts(f):
"""Decorator that attaches the model/LoRA/quantization options shared by
`serve` and `classify`.
The LLM backend is LLaMA-Factory (`llamafactory-cli api`); options here
map directly to its YAML config.
"""
f = click.option("--quantization", default=DEFAULT_QUANTIZATION,
type=click.Choice(["bitsandbytes", "awq", "gptq", "none"]),
show_default=True,
help="LLM weight-quantization mode (passed to LLaMA-Factory)")(f)
f = click.option("--lora-path", default=None, type=click.Path(),
help="local path to the LoRA folder; "
"if omitted, downloaded from HF")(f)
f = click.option("--base-model", default=None, type=click.Path(),
help="local path to the BF16 base model directory; "
"if omitted, auto-detect from common paths "
"(~/models/qwen3-32b, /root/autodl-tmp/..., HF cache)")(f)
return f
@click.group(context_settings={"help_option_names": ["-h", "--help"]})
@click.version_option(__version__, prog_name="statlens")
def main():
"""statLens — DEA method selector. See `statlens serve --help`."""
# ─────────────────────────────── serve ───────────────────────────────
@main.command()
@_common_serve_opts
@click.option("--web-port", default=LOCAL_WEB_PORT, show_default=True, type=int)
@click.option("--llm-port", default=LOCAL_LLM_PORT, show_default=True, type=int)
@click.option("--no-browser", is_flag=True, help="don't auto-open browser")
@click.option("--no-gpu-check", is_flag=True, help="skip GPU sanity check")
def serve(base_model, lora_path, quantization,
web_port, llm_port, no_browser, no_gpu_check):
"""Start the LLM backend (LLaMA-Factory) + the web app on http://localhost:7860 ."""
if not no_gpu_check:
runtime.check_gpu()
base_path = runtime.resolve_base_model(base_model)
lora_path = runtime.ensure_lora_cached(lora_path)
llm_proc = runtime.start_server(
base_path, lora_path,
port=llm_port,
quantization=quantization,
)
def shutdown(signum=None, frame=None):
runtime.stop_server(llm_proc)
sys.exit(0)
signal.signal(signal.SIGINT, shutdown)
signal.signal(signal.SIGTERM, shutdown)
try:
# show animated progress; abort if subprocess dies
runtime.wait_for_server(port=llm_port, proc=llm_proc)
except (TimeoutError, RuntimeError) as e:
click.secho(f"\n[statlens] {e}", fg="red")
runtime.stop_server(llm_proc)
sys.exit(2)
os.environ["STATLENS_LLM_ENDPOINT"] = f"http://127.0.0.1:{llm_port}/v1"
# LLaMA-Factory's API exposes the model under the name "gpt-3.5-turbo".
os.environ["STATLENS_LLM_MODEL"] = "gpt-3.5-turbo"
web_url = f"http://localhost:{web_port}/"
# Prominent ready banner — only what end users care about.
click.echo("")
click.secho("══════════════════════════════════════════════════════", fg="green")
click.secho(f" ✅ statLens ready", fg="green", bold=True)
click.secho(f" open in browser: {web_url}", fg="green")
click.secho(f" Ctrl+C to stop.", fg="green")
click.secho("══════════════════════════════════════════════════════", fg="green")
click.echo("")
if not no_browser:
import threading
threading.Timer(1.0, lambda: webbrowser.open(web_url)).start()
import uvicorn
from . import server as server_mod
try:
# Quiet uvicorn so it doesn't drown the ready banner; users see HTTP
# requests via the LLM log if they want.
uvicorn.run(server_mod.app, host="127.0.0.1", port=web_port, log_level="warning")
finally:
runtime.stop_server(llm_proc)
# ─────────────────────────── download ───────────────────────────
@main.command()
@click.option("--lora-repo", default=DEFAULT_LORA_REPO, show_default=True)
@click.option("--lora-subfolder", default=DEFAULT_LORA_SUBFOLDER, show_default=True)
def download(lora_repo, lora_subfolder):
"""Pre-fetch the LoRA adapter (~1 GB) into the HF cache.
statLens does NOT auto-download the 64 GB base model. Get it yourself with:
huggingface-cli download Qwen/Qwen3-32B --local-dir ~/models/qwen3-32b
"""
p = runtime.ensure_lora_cached(None, lora_repo, lora_subfolder)
click.secho(f"\n✅ LoRA cached at {p}", fg="green")
click.echo(
"\nReminder: download the BF16 base model separately, e.g.\n"
" huggingface-cli download Qwen/Qwen3-32B --local-dir ~/models/qwen3-32b"
)
# ─────────────────────────── info ───────────────────────────
@main.command()
def info():
"""Show GPU, cache, and default endpoint information."""
click.echo(f"statlens version: {__version__}\n")
try:
import torch
if torch.cuda.is_available():
for i in range(torch.cuda.device_count()):
p = torch.cuda.get_device_properties(i)
click.echo(f" GPU {i}: {p.name} · {p.total_memory/1024**3:.1f} GB")
else:
click.echo(" GPU: (none)")
except Exception as e:
click.echo(f" GPU: (torch not importable: {e})")
click.echo("")
click.echo(f" base model (reference) : {DEFAULT_BASE_MODEL_REF} (NOT auto-downloaded)")
click.echo(f" base model (resolved) : "
f"{os.environ.get('STATLENS_BASE_MODEL', '(unset — pass --base-model)')}")
click.echo(f" default quantization : {DEFAULT_QUANTIZATION}")
click.echo(f" lora repo : {DEFAULT_LORA_REPO} (subfolder {DEFAULT_LORA_SUBFOLDER})")
click.echo("")
from huggingface_hub.constants import HF_HUB_CACHE
click.echo(f" HF cache root : {HF_HUB_CACHE}")
d = runtime.cache_dir_for(DEFAULT_LORA_REPO)
if d.exists():
sz = sum(p.stat().st_size for p in d.rglob("*") if p.is_file())
click.echo(f" · {DEFAULT_LORA_REPO} ✓ ({sz/1024**3:.2f} GB at {d})")
else:
click.echo(f" · {DEFAULT_LORA_REPO} ✗ (run `statlens download` to fetch)")
# ─────────────────────────── classify ───────────────────────────
@main.command()
@_common_serve_opts
@click.option("--tsv", required=True, type=click.Path(exists=True, dir_okay=False))
@click.option("--context", required=True, type=click.Path(exists=True, dir_okay=False))
@click.option("--out", required=True, type=click.Path(file_okay=False))
@click.option("--endpoint", default=None,
help="OpenAI-compatible endpoint (default: spin up a temporary local LLM backend)")
@click.option("--model", default="gpt-3.5-turbo", show_default=True)
@click.option("--keep-llm", is_flag=True, help="don't shut down the LLM backend after classifying")
def classify(base_model, lora_path, quantization,
tsv, context, out, endpoint, model, keep_llm):
"""One-shot CLI: classify a single TSV+context, run pipeline, exit."""
from pathlib import Path
from .statlens_run import run_one
if endpoint is None:
runtime.check_gpu()
base_path = runtime.resolve_base_model(base_model)
lora_resolved = runtime.ensure_lora_cached(lora_path)
proc = runtime.start_server(base_path, lora_resolved,
quantization=quantization)
try:
runtime.wait_for_server()
endpoint = f"http://127.0.0.1:{LOCAL_LLM_PORT}/v1"
run_one(Path(tsv), Path(context), Path(out), endpoint, model)
finally:
if not keep_llm:
runtime.stop_server(proc)
else:
from pathlib import Path
run_one(Path(tsv), Path(context), Path(out), endpoint, model)
if __name__ == "__main__":
main()
|