deployxa-lm-1.0 / sample_inference.py
mosetireagan's picture
Initial private staging of Deployxa LM 1.0 (GA)
b948218 verified
Raw History Blame Contribute Delete
5.37 kB
#!/usr/bin/env python3
"""
Deployxa LM 1.0 - Sample Inference Script.
Demonstrates loading the fine-tuned Deployxa LM 1.0 adapter on top of
Qwen/Qwen2.5-Coder-1.5B-Instruct using PyTorch, Transformers, and PEFT.
Brand: Deployxa Cloud Inc. (https://deployxa.com)
Product: Deployxa LM 1.0
Tagline: Build. Deploy. Scale.
"""
import sys
import json
import argparse
from pathlib import Path
try:
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel
except ImportError:
print("Error: PyTorch, Transformers, or PEFT not installed.")
print("Install dependencies with: pip install torch transformers peft accelerate")
sys.exit(1)
BASE_MODEL_ID = "Qwen/Qwen2.5-Coder-1.5B-Instruct"
# Default to current folder if adapter_model.safetensors exists here, else Hugging Face hub
DEFAULT_ADAPTER = str(Path(__file__).resolve().parent) if (Path(__file__).resolve().parent / "adapter_model.safetensors").exists() else "deployxa/deployxa-lm-1.0"
APPROVED_CANONICAL_CATEGORIES = [
"BUILD_COMPILATION_FAILED",
"CONTAINER_STARTUP_CRASH",
"HEALTH_CHECK_FAILURE",
"DEPENDENCY_INSTALLATION_FAILED",
"RESOURCE_EXHAUSTION",
"NETWORKING_ROUTING_ERROR",
"DATABASE_MIGRATION_FAILURE",
"PORT_BINDING_CONFLICT",
"INVALID_APPLICATION_CONFIG"
]
def format_diagnostic_prompt(template_id: str, telemetry_logs: str) -> str:
"""Constructs the official Deployxa LM 1.0 diagnostic prompt with pinned taxonomy."""
categories_str = ", ".join(APPROVED_CANONICAL_CATEGORIES)
return (
"<|im_start|>system\n"
"You are DXA-LM, Deployxa's specialist deployment diagnostic and repair advisor.\n"
"Analyze the provided deployment logs and telemetry. Determine the deployment status, "
"exact root cause category, and whether evidence is sufficient to make a diagnosis.\n"
f"Allowed root cause categories: {categories_str}.\n"
"Format response strictly as JSON with keys: deployment_status, root_cause_category, "
"confidence, evidence, reasoning_summary, abstain, abstention_reason.<|im_end|>\n"
"<|im_start|>user\n"
f"Deployment Logs ({template_id}):\n"
f"{telemetry_logs}\n\n"
"Diagnose deployment:<|im_end|>\n"
"<|im_start|>assistant\n"
)
def main():
parser = argparse.ArgumentParser(description="Run Deployxa LM 1.0 Diagnostic Inference")
parser.add_argument("--adapter", default=DEFAULT_ADAPTER, help="Path or Hugging Face repo ID for adapter")
parser.add_argument("--base-model", default=BASE_MODEL_ID, help="Base model identifier")
parser.add_argument("--template", default="solid-start", help="Service or framework template name")
parser.add_argument("--log", default=None, help="Path to error log file (or pass raw text)")
args = parser.parse_args()
sample_log = """
2026-10-18T10:15:20Z [server] Starting HTTP listener on port 3000...
2026-10-18T10:15:20Z [server] uncaughtException: Error: listen EADDRINUSE: address already in use :::3000
at Server.setupListenHandle [as _listen2] (node:net:1855:16)
at listenInCluster (node:net:1903:12)
at Server.listen (node:net:2001:7)
2026-10-18T10:15:21Z [systemd] app.service: Main process exited, code=exited, status=1/FAILURE
"""
if args.log:
p = Path(args.log)
if p.exists() and p.is_file():
sample_log = p.read_text(encoding="utf-8", errors="ignore")
else:
sample_log = args.log
device = "cuda" if torch.cuda.is_available() else ("mps" if torch.backends.mps.is_available() else "cpu")
print(f"Loading base model '{args.base_model}' on {device}...")
tokenizer = AutoTokenizer.from_pretrained(args.base_model, trust_remote_code=True)
base_model = AutoModelForCausalLM.from_pretrained(
args.base_model,
torch_dtype=torch.float16 if device != "cpu" else torch.float32,
device_map="auto" if device == "cuda" else None,
trust_remote_code=True
)
if device != "cuda":
base_model.to(device)
print(f"Attaching Deployxa LM 1.0 adapter '{args.adapter}'...")
model = PeftModel.from_pretrained(base_model, args.adapter)
model.eval()
prompt = format_diagnostic_prompt(args.template, sample_log)
inputs = tokenizer(prompt, return_tensors="pt").to(device)
print("Running neural diagnostic inference...")
with torch.no_grad():
output_ids = model.generate(
**inputs,
max_new_tokens=400,
temperature=0.01,
do_sample=False,
pad_token_id=tokenizer.eos_token_id
)
response_text = tokenizer.decode(output_ids[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True).strip()
# Strip markdown code blocks if wrapped
if response_text.startswith("```json"):
response_text = response_text[7:]
elif response_text.startswith("```"):
response_text = response_text[3:]
if response_text.endswith("```"):
response_text = response_text[:-3]
print("\n" + "=" * 60)
print("DEPLOYXA LM 1.0 PREDICTION OUTPUT:")
print("=" * 60)
try:
data = json.loads(response_text)
print(json.dumps(data, indent=2))
except Exception:
print(response_text)
if __name__ == "__main__":
main()