File size: 18,971 Bytes
66153d5 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 | """
Phase 5 β Inference Optimization & MLOps
==========================================
β’ vLLM high-throughput inference server
β’ MLflow / W&B experiment tracking
β’ HF Hub model deployment with model cards
β’ Benchmark comparison and reporting
Usage:
# Start inference server
python -m phase5_mlops.serve --model ./merged --port 8000
# Track experiment
python -m phase5_mlops.serve track --run-name my-merge --rouge 0.42 --judge 7.1
# Deploy to HF Hub
python -m phase5_mlops.serve deploy --model ./merged --repo my-org/my-model
"""
from __future__ import annotations
import json
import time
from dataclasses import dataclass, asdict
from pathlib import Path
from typing import Optional
import typer
from rich.console import Console
from rich.table import Table
from rich.panel import Panel
from configs.settings import (
WANDB_PROJECT, MLFLOW_URI, HF_ORG, HF_TOKEN,
VLLM_TENSOR_PARALLEL, VLLM_GPU_MEMORY_UTIL, VLLM_MAX_MODEL_LEN,
EVAL_DIR
)
from utils.logger import logger
app = typer.Typer(help="Phase 5: Inference & MLOps")
console = Console()
# βββββββββββββββββββββββββββββββββββββββββββββ
# 1. vLLM Inference Server
# βββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class GenerationRequest:
prompt: str
max_new_tokens: int = 256
temperature: float = 0.7
top_p: float = 0.95
stop: list[str] = None
@dataclass
class GenerationResponse:
text: str
tokens_used: int
latency_ms: float
model_id: str
class VLLMServer:
"""
Wraps vLLM's AsyncLLMEngine for high-throughput offline + online inference.
Supports: continuous batching, PagedAttention, tensor parallelism.
"""
def __init__(
self,
model_id: str,
tensor_parallel: int = VLLM_TENSOR_PARALLEL,
gpu_memory_util: float = VLLM_GPU_MEMORY_UTIL,
max_model_len: int = VLLM_MAX_MODEL_LEN,
quantization: Optional[str] = None, # "awq", "gptq", "squeezellm", None
dtype: str = "bfloat16",
):
self.model_id = model_id
self._engine = None
self._engine_args = dict(
model = model_id,
tensor_parallel_size = tensor_parallel,
gpu_memory_utilization = gpu_memory_util,
max_model_len = max_model_len,
quantization = quantization,
dtype = dtype,
trust_remote_code = True,
tokenizer = model_id,
)
def _load(self):
if self._engine is None:
from vllm import LLM, SamplingParams
logger.info(f"[vLLM] Loading: {self.model_id}")
self._llm = LLM(**self._engine_args)
self._SamplingParams = SamplingParams
logger.success("[vLLM] Engine ready")
def generate(self, requests: list[GenerationRequest]) -> list[GenerationResponse]:
"""Batch inference β all requests processed together with continuous batching."""
self._load()
from vllm import SamplingParams
prompts = [r.prompt for r in requests]
params = [
SamplingParams(
temperature = r.temperature,
top_p = r.top_p,
max_tokens = r.max_new_tokens,
stop = r.stop or [],
)
for r in requests
]
t0 = time.perf_counter()
outputs = self._llm.generate(prompts, params)
elapsed = (time.perf_counter() - t0) * 1000
responses = []
for out in outputs:
text = out.outputs[0].text
toks = len(out.outputs[0].token_ids)
responses.append(GenerationResponse(
text = text,
tokens_used = toks,
latency_ms = elapsed / len(outputs),
model_id = self.model_id,
))
return responses
def generate_one(self, prompt: str, **kwargs) -> GenerationResponse:
return self.generate([GenerationRequest(prompt=prompt, **kwargs)])[0]
def benchmark_throughput(self, n_requests: int = 50) -> dict:
"""Measure tokens/sec across n_requests synthetic requests."""
prompts = [
f"Explain the concept of {topic} in detail."
for topic in (["machine learning", "quantum computing", "climate change",
"evolution", "blockchain"] * 10)[:n_requests]
]
reqs = [GenerationRequest(p, max_new_tokens=128) for p in prompts]
t0 = time.perf_counter()
responses = self.generate(reqs)
elapsed = time.perf_counter() - t0
total_toks = sum(r.tokens_used for r in responses)
tps = total_toks / elapsed
result = {
"n_requests": n_requests,
"total_tokens": total_toks,
"elapsed_s": round(elapsed, 2),
"tokens_per_sec": round(tps, 1),
"avg_latency_ms": round(elapsed * 1000 / n_requests, 1),
}
logger.info(f"[vLLM] Throughput: {tps:.0f} tokens/s, avg latency: {result['avg_latency_ms']}ms")
return result
def start_api_server(self, host: str = "0.0.0.0", port: int = 8000):
"""
Launch vLLM's built-in OpenAI-compatible API server as a subprocess.
Compatible with any OpenAI SDK client.
"""
import subprocess, sys
cmd = [
sys.executable, "-m", "vllm.entrypoints.openai.api_server",
"--model", self.model_id,
"--host", host,
"--port", str(port),
"--tensor-parallel-size", str(self._engine_args["tensor_parallel_size"]),
"--gpu-memory-utilization", str(self._engine_args["gpu_memory_utilization"]),
"--max-model-len", str(self._engine_args["max_model_len"]),
"--dtype", self._engine_args["dtype"],
"--trust-remote-code",
]
if HF_TOKEN:
cmd += ["--token", HF_TOKEN]
logger.info(f"[vLLM] Starting API server at http://{host}:{port}")
logger.info(f" β OpenAI-compatible: POST http://{host}:{port}/v1/chat/completions")
subprocess.run(cmd)
# βββββββββββββββββββββββββββββββββββββββββββββ
# 2. Experiment Tracking
# βββββββββββββββββββββββββββββββββββββββββββββ
@dataclass
class ExperimentMetrics:
run_name: str
model_id: str
merge_strategy: str = ""
base_models: list[str] = None
avg_rouge1: float = 0.0
avg_rouge2: float = 0.0
avg_rougeL: float = 0.0
avg_bertscore: float = 0.0
avg_faithfulness: float = 0.0
hallucination_rate: float = 0.0
avg_judge_score: float = 0.0
tokens_per_sec: float = 0.0
gap_categories: list[str] = None
notes: str = ""
class ExperimentTracker:
"""Unified tracker: logs to W&B + MLflow simultaneously."""
def __init__(self, use_wandb: bool = True, use_mlflow: bool = True):
self.use_wandb = use_wandb
self.use_mlflow = use_mlflow
def log(self, metrics: ExperimentMetrics, artifacts: Optional[list[Path]] = None):
if self.use_wandb:
self._log_wandb(metrics, artifacts)
if self.use_mlflow:
self._log_mlflow(metrics, artifacts)
def _log_wandb(self, m: ExperimentMetrics, artifacts: Optional[list[Path]]):
try:
import wandb
run = wandb.init(
project = WANDB_PROJECT,
name = m.run_name,
config = {
"model_id": m.model_id,
"merge_strategy": m.merge_strategy,
"base_models": m.base_models or [],
},
)
run.log({
"eval/rouge1": m.avg_rouge1,
"eval/rouge2": m.avg_rouge2,
"eval/rougeL": m.avg_rougeL,
"eval/bertscore": m.avg_bertscore,
"eval/faithfulness": m.avg_faithfulness,
"eval/hallucination_rate": m.hallucination_rate,
"eval/judge_score": m.avg_judge_score,
"perf/tokens_per_sec": m.tokens_per_sec,
})
if artifacts:
for path in artifacts:
run.save(str(path))
run.finish()
logger.info(f"[W&B] Logged run: {m.run_name}")
except Exception as e:
logger.warning(f"W&B logging failed: {e}")
def _log_mlflow(self, m: ExperimentMetrics, artifacts: Optional[list[Path]]):
try:
import mlflow
mlflow.set_tracking_uri(MLFLOW_URI)
mlflow.set_experiment(WANDB_PROJECT)
with mlflow.start_run(run_name=m.run_name):
mlflow.set_tags({
"model_id": m.model_id,
"merge_strategy": m.merge_strategy,
"base_models": str(m.base_models),
"gaps": str(m.gap_categories),
})
mlflow.log_metrics({
"rouge1": m.avg_rouge1,
"rouge2": m.avg_rouge2,
"rougeL": m.avg_rougeL,
"bertscore": m.avg_bertscore,
"faithfulness": m.avg_faithfulness,
"hallucination_rate": m.hallucination_rate,
"judge_score": m.avg_judge_score,
"tokens_per_sec": m.tokens_per_sec,
})
if artifacts:
for path in artifacts:
mlflow.log_artifact(str(path))
logger.info(f"[MLflow] Logged run: {m.run_name}")
except Exception as e:
logger.warning(f"MLflow logging failed: {e}")
# βββββββββββββββββββββββββββββββββββββββββββββ
# 3. Model Card Generator
# βββββββββββββββββββββββββββββββββββββββββββββ
MODEL_CARD_TEMPLATE = """\
---
language:
- en
tags:
- llm-pipeline
- merged-model
- peft
license: apache-2.0
pipeline_tag: text-generation
---
# {model_name}
Produced by the **LLM Pipeline** β automated discovery, merging, evaluation, and fine-tuning.
## Model Details
| Property | Value |
|----------------|-------|
| Base models | {base_models} |
| Merge strategy | `{merge_strategy}` |
| Fine-tuning | LoRA/QLoRA (PEFT) |
| Scale | 7B parameters |
| Precision | bfloat16 |
## Evaluation Results
| Metric | Score |
|--------------------|-------|
| ROUGE-1 | {rouge1:.3f} |
| ROUGE-2 | {rouge2:.3f} |
| ROUGE-L | {rougeL:.3f} |
| BERTScore F1 | {bertscore:.3f} |
| Faithfulness | {faithfulness:.3f} |
| Hallucination rate | {halluc:.1%} |
| Judge Score | {judge:.1f}/10 |
## Training Details
### Knowledge Gaps Addressed
{gap_categories}
### Merge Configuration
- **Strategy**: {merge_strategy}
- **Density** (TIES/Breadcrumbs): 0.7
- **Alpha**: 0.5
## Usage
```python
from transformers import AutoTokenizer, AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("{repo_id}", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("{repo_id}")
inputs = tok("Your prompt here", return_tensors="pt")
output = model.generate(**inputs, max_new_tokens=256)
print(tok.decode(output[0], skip_special_tokens=True))
```
## Limitations
- Evaluated primarily on English text
- May hallucinate on out-of-training-distribution topics
- 4-bit quantization used during evaluation (not deployment)
## Citation
```
@misc{{llm-pipeline-2024,
title={{Automated LLM Discovery, Merging, and Fine-Tuning Pipeline}},
year={{2024}}
}}
```
"""
def generate_model_card(metrics: ExperimentMetrics, repo_id: str) -> str:
model_name = repo_id.split("/")[-1]
base_models = ", ".join(f"`{m}`" for m in (metrics.base_models or []))
gap_cats = "\n".join(f"- {g}" for g in (metrics.gap_categories or [])) or "- None detected"
return MODEL_CARD_TEMPLATE.format(
model_name = model_name,
base_models = base_models,
merge_strategy = metrics.merge_strategy,
rouge1 = metrics.avg_rouge1,
rouge2 = metrics.avg_rouge2,
rougeL = metrics.avg_rougeL,
bertscore = metrics.avg_bertscore,
faithfulness = metrics.avg_faithfulness,
halluc = metrics.hallucination_rate,
judge = metrics.avg_judge_score,
gap_categories = gap_cats,
repo_id = repo_id,
)
# βββββββββββββββββββββββββββββββββββββββββββββ
# 4. HF Hub Deployment
# βββββββββββββββββββββββββββββββββββββββββββββ
def deploy_to_hub(
model_path: str,
repo_id: str,
metrics: Optional[ExperimentMetrics] = None,
private: bool = False,
commit_msg: str = "Upload merged model",
) -> str:
"""
Push model + tokenizer + model card to HF Hub.
Returns the repo URL.
"""
from huggingface_hub import HfApi, create_repo
if not HF_TOKEN:
raise EnvironmentError("Set HF_TOKEN env var before deploying")
api = HfApi(token=HF_TOKEN)
# Create repo if needed
url = create_repo(repo_id, token=HF_TOKEN, private=private, exist_ok=True)
logger.info(f"[Hub] Repo: {url}")
# Write model card
if metrics:
card_text = generate_model_card(metrics, repo_id)
card_path = Path(model_path) / "README.md"
with open(card_path, "w") as f:
f.write(card_text)
logger.info("[Hub] Model card written")
# Upload
logger.info(f"[Hub] Uploading {model_path} β {repo_id} ...")
api.upload_folder(
folder_path = model_path,
repo_id = repo_id,
commit_message = commit_msg,
ignore_patterns = ["*.log", "__pycache__"],
)
logger.success(f"[Hub] Deployed: https://huggingface.co/{repo_id}")
return f"https://huggingface.co/{repo_id}"
# βββββββββββββββββββββββββββββββββββββββββββββ
# 5. Pipeline Summary Report
# βββββββββββββββββββββββββββββββββββββββββββββ
def print_pipeline_summary(
eval_dir: Path = EVAL_DIR,
top_n: int = 10,
) -> None:
"""Load all saved eval JSONs and print a ranked leaderboard."""
eval_files = sorted(eval_dir.glob("eval_*.json"))
if not eval_files:
console.print("[yellow]No evaluation results found.[/yellow]")
return
rows = []
for ef in eval_files:
try:
with open(ef) as f:
d = json.load(f)
rows.append(d)
except Exception:
continue
rows.sort(key=lambda x: x.get("avg_rouge1", 0), reverse=True)
table = Table(title="Pipeline Leaderboard", show_lines=True)
table.add_column("Rank", width=5, style="dim")
table.add_column("Model", style="cyan")
table.add_column("Dataset", style="dim")
table.add_column("ROUGE-1", justify="right", style="green")
table.add_column("BERTScore", justify="right")
table.add_column("Faithfulness", justify="right")
table.add_column("Judge", justify="right")
table.add_column("Halluc %", justify="right", style="red")
for i, r in enumerate(rows[:top_n], 1):
table.add_row(
str(i),
r.get("model_id", "")[-40:],
r.get("dataset", ""),
f"{r.get('avg_rouge1',0):.3f}",
f"{r.get('avg_bertscore',0):.3f}",
f"{r.get('avg_faithfulness',0):.3f}",
f"{r.get('avg_judge_score',0):.1f}",
f"{r.get('hallucination_rate',0)*100:.1f}%",
)
console.print(table)
# βββββββββββββββββββββββββββββββββββββββββββββ
# CLI
# βββββββββββββββββββββββββββββββββββββββββββββ
@app.command("serve")
def serve(
model: str = typer.Argument(..., help="Model ID or local path"),
port: int = typer.Option(8000),
host: str = typer.Option("0.0.0.0"),
bench: bool = typer.Option(False, "--bench", help="Run throughput benchmark then exit"),
):
server = VLLMServer(model)
if bench:
result = server.benchmark_throughput()
console.print(Panel(str(result), title="Throughput Benchmark"))
return
server.start_api_server(host=host, port=port)
@app.command("track")
def track_cmd(
run_name: str = typer.Argument(...),
model: str = typer.Option(..., "--model"),
strategy: str = typer.Option(""),
rouge1: float = typer.Option(0.0),
bertscore: float = typer.Option(0.0),
judge: float = typer.Option(0.0),
tps: float = typer.Option(0.0),
wandb_flag: bool = typer.Option(True, "--wandb/--no-wandb"),
mlflow_flag: bool = typer.Option(True, "--mlflow/--no-mlflow"),
):
metrics = ExperimentMetrics(
run_name = run_name,
model_id = model,
merge_strategy = strategy,
avg_rouge1 = rouge1,
avg_bertscore = bertscore,
avg_judge_score = judge,
tokens_per_sec = tps,
)
tracker = ExperimentTracker(use_wandb=wandb_flag, use_mlflow=mlflow_flag)
tracker.log(metrics)
console.print(f"[green]β Logged experiment: {run_name}[/green]")
@app.command("deploy")
def deploy_cmd(
model: str = typer.Argument(..., help="Local model path"),
repo: str = typer.Option(..., "--repo", help="HF repo: username/model-name"),
private: bool = typer.Option(False),
):
deploy_to_hub(model, repo, private=private)
@app.command("leaderboard")
def leaderboard():
print_pipeline_summary()
if __name__ == "__main__":
app()
|