src/utils/model_smoke_test.py

from __future__ import annotations

import argparse
import copy
import gc
import importlib.util
import json
import os
import time
import traceback
from pathlib import Path
from typing import Any

import yaml

from src.data.io_utils import write_csv, write_json


REQUIRED_PACKAGES = ["torch", "transformers", "accelerate"]
PASS_STATUSES = {"PASS", "PASS_BACKUP"}
VALID_GROUPS = {"all", "encoders", "retrieval", "llm_small", "llm_large"}
MINIMUM_REQUIRED_MODULES = {
    "Dense retrieval",
    "Reranker",
    "ViFactCheck verifier",
    "AVeriTeC verifier",
    "HealthVer verifier",
    "Vietnamese NLI",
    "English NLI",
}
WIKIKG_REQUIRED_MODULES = {"Extraction"}
LLM_BASELINE_REQUIRED_MODULES = {"LLM judge"}


def package_available(name: str) -> bool:
    return importlib.util.find_spec(name) is not None


def dependency_status() -> dict[str, bool]:
    packages = {name: package_available(name) for name in REQUIRED_PACKAGES}
    packages["bitsandbytes"] = package_available("bitsandbytes")
    return packages


def cuda_status() -> dict[str, Any]:
    if not package_available("torch"):
        return {"torch_available": False, "cuda_available": False, "device_count": 0, "devices": []}
    import torch

    devices = []
    if torch.cuda.is_available():
        for idx in range(torch.cuda.device_count()):
            props = torch.cuda.get_device_properties(idx)
            devices.append(
                {
                    "index": idx,
                    "name": props.name,
                    "total_memory_gb": round(props.total_memory / (1024**3), 3),
                }
            )
    return {
        "torch_available": True,
        "torch_version": torch.__version__,
        "cuda_available": torch.cuda.is_available(),
        "device_count": torch.cuda.device_count() if torch.cuda.is_available() else 0,
        "devices": devices,
    }


def reset_peak_memory() -> None:
    import torch

    if torch.cuda.is_available():
        torch.cuda.empty_cache()
        torch.cuda.reset_peak_memory_stats()


def peak_vram_gb() -> float | None:
    import torch

    if not torch.cuda.is_available():
        return None
    return round(torch.cuda.max_memory_allocated() / (1024**3), 4)


def cleanup_torch() -> None:
    gc.collect()
    if package_available("torch"):
        import torch

        if torch.cuda.is_available():
            torch.cuda.empty_cache()


def dtype_for_device():
    import torch

    if torch.cuda.is_available():
        return torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
    return torch.float32


def build_quantization_config(spec: dict[str, Any]):
    quantization = str(spec.get("quantization", "")).casefold()
    if "4bit" not in quantization:
        return None
    if not package_available("bitsandbytes"):
        raise RuntimeError("bitsandbytes is required for bnb_4bit quantization")
    from transformers import BitsAndBytesConfig

    return BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_quant_type="nf4",
        bnb_4bit_compute_dtype=dtype_for_device(),
        bnb_4bit_use_double_quant=True,
    )


def load_tokenizer(model_id: str):
    from transformers import AutoTokenizer

    return AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)


def smoke_embedding(spec: dict[str, Any]) -> dict[str, Any]:
    import torch
    from transformers import AutoModel

    model_id = spec["model_id"]
    tokenizer = load_tokenizer(model_id)
    model = AutoModel.from_pretrained(
        model_id,
        torch_dtype=dtype_for_device(),
        device_map="auto" if torch.cuda.is_available() else None,
        trust_remote_code=True,
    )
    sentences = [
        "A claim needs evidence with clear provenance.",
        "Kiểm chứng thông tin cần bằng chứng có nguồn.",
        "Hydroxychloroquine does not improve COVID-19 outcomes in this trial.",
        "The statement is not supported by the cited article.",
        "Entity overlap can help retrieve relevant context chunks.",
        "The dataset split must avoid claim-level leakage.",
        "Biomedical verification needs careful terminology.",
        "A reranker scores claim and evidence pairs.",
        "Knowledge graph triples must keep source sentence IDs.",
        "Not enough evidence should not be forced into support or refute.",
    ]
    inputs = tokenizer(sentences, padding=True, truncation=True, max_length=128, return_tensors="pt")
    if torch.cuda.is_available():
        inputs = {key: value.to(model.device) for key, value in inputs.items()}
    with torch.no_grad():
        output = model(**inputs)
        hidden = output.last_hidden_state
        embedding = hidden.mean(dim=1)
    return {"output_shape": list(embedding.shape), "batch": len(sentences)}


def smoke_reranker(spec: dict[str, Any]) -> dict[str, Any]:
    import torch

    model_id = spec["model_id"]
    tokenizer = load_tokenizer(model_id)
    pairs = [
        ("Does HCQ treat COVID-19?", "The trial found no significant difference in outcomes."),
        ("Did the police salary change?", "The salary was 24000 in 2010 and 23000 in 2018."),
    ] * 5
    inputs = tokenizer(
        [claim for claim, _ in pairs],
        [evidence for _, evidence in pairs],
        padding=True,
        truncation=True,
        max_length=256,
        return_tensors="pt",
    )
    quantization_config = build_quantization_config(spec)
    try:
        from transformers import AutoModelForSequenceClassification

        model = AutoModelForSequenceClassification.from_pretrained(
            model_id,
            torch_dtype=dtype_for_device(),
            device_map="auto" if torch.cuda.is_available() else None,
            trust_remote_code=True,
            quantization_config=quantization_config,
        )
        if torch.cuda.is_available():
            inputs = {key: value.to(model.device) for key, value in inputs.items()}
        with torch.no_grad():
            logits = model(**inputs).logits
        return {"output_shape": list(logits.shape), "batch": len(pairs), "loader": "AutoModelForSequenceClassification"}
    except Exception:
        from transformers import AutoModelForCausalLM

        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            torch_dtype=dtype_for_device(),
            device_map="auto" if torch.cuda.is_available() else None,
            trust_remote_code=True,
            quantization_config=quantization_config,
        )
        prompt = "Given a query and a passage, output yes if the passage is relevant.\nQuery: Does HCQ treat COVID-19?\nPassage: The trial found no significant difference.\nAnswer:"
        inputs = tokenizer(prompt, return_tensors="pt")
        if torch.cuda.is_available():
            inputs = {key: value.to(model.device) for key, value in inputs.items()}
        with torch.no_grad():
            output = model.generate(**inputs, max_new_tokens=4)
        return {"output_shape": list(output.shape), "batch": 1, "loader": "AutoModelForCausalLM"}


def smoke_encoder_classifier(spec: dict[str, Any]) -> dict[str, Any]:
    import torch
    from transformers import AutoModelForSequenceClassification

    model_id = spec["model_id"]
    tokenizer = load_tokenizer(model_id)
    texts = [
        "Claim: The statement is supported. Evidence: The source directly states it.",
        "Claim: The claim is false. Evidence: The cited source contradicts it.",
    ]
    inputs = tokenizer(texts, padding=True, truncation=True, max_length=256, return_tensors="pt")
    model = AutoModelForSequenceClassification.from_pretrained(
        model_id,
        num_labels=3,
        torch_dtype=dtype_for_device(),
        device_map="auto" if torch.cuda.is_available() else None,
        trust_remote_code=True,
    )
    if torch.cuda.is_available():
        inputs = {key: value.to(model.device) for key, value in inputs.items()}
    with torch.no_grad():
        logits = model(**inputs).logits
    return {"output_shape": list(logits.shape), "batch": len(texts)}


def smoke_nli(spec: dict[str, Any]) -> dict[str, Any]:
    import torch
    from transformers import AutoModelForSequenceClassification

    model_id = spec["model_id"]
    tokenizer = load_tokenizer(model_id)
    premises = [
        "The clinical trial found no statistically significant improvement.",
        "Bài báo nói sự kiện được tổ chức tại Hải Phòng.",
    ]
    hypotheses = [
        "The treatment improved patient outcomes.",
        "Sự kiện diễn ra tại Hải Phòng.",
    ]
    inputs = tokenizer(premises, hypotheses, padding=True, truncation=True, max_length=256, return_tensors="pt")
    model = AutoModelForSequenceClassification.from_pretrained(
        model_id,
        torch_dtype=dtype_for_device(),
        device_map="auto" if torch.cuda.is_available() else None,
        trust_remote_code=True,
    )
    if torch.cuda.is_available():
        inputs = {key: value.to(model.device) for key, value in inputs.items()}
    with torch.no_grad():
        logits = model(**inputs).logits
    return {"output_shape": list(logits.shape), "batch": len(premises)}


def smoke_causal_json(spec: dict[str, Any]) -> dict[str, Any]:
    import torch
    from transformers import AutoModelForCausalLM, AutoProcessor

    model_id = spec["model_id"]
    quantization_config = build_quantization_config(spec)
    prompt = (
        "Extract one source-grounded fact as strict JSON with keys fact, subject, relation, object.\n"
        "Claim: Masks reduce COVID-19 transmission.\n"
        "Evidence: Broad adoption of face masks may meaningfully reduce community transmission.\n"
        "JSON:"
    )
    processor_causal_error = ""
    tokenizer_causal_error = ""
    try:
        processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
        model = AutoModelForCausalLM.from_pretrained(
            model_id,
            torch_dtype=dtype_for_device(),
            device_map="auto" if torch.cuda.is_available() else None,
            trust_remote_code=True,
            quantization_config=quantization_config,
        )
        inputs = processor(text=prompt, return_tensors="pt")
        if torch.cuda.is_available():
            inputs = {key: value.to(model.device) for key, value in inputs.items()}
        with torch.no_grad():
            output = model.generate(**inputs, max_new_tokens=64, do_sample=False)
        try:
            decoded = processor.decode(output[0], skip_special_tokens=True)
        except Exception:
            decoded = processor.batch_decode(output, skip_special_tokens=True)[0]
        return {"output_text_tail": decoded[-500:], "batch": 1, "loader": "AutoProcessor+AutoModelForCausalLM"}
    except Exception as exc:
        processor_causal_error = repr(exc)
        try:
            tokenizer = load_tokenizer(model_id)
            model = AutoModelForCausalLM.from_pretrained(
                model_id,
                torch_dtype=dtype_for_device(),
                device_map="auto" if torch.cuda.is_available() else None,
                trust_remote_code=True,
                quantization_config=quantization_config,
            )
            inputs = tokenizer(prompt, return_tensors="pt")
            if torch.cuda.is_available():
                inputs = {key: value.to(model.device) for key, value in inputs.items()}
            with torch.no_grad():
                output = model.generate(**inputs, max_new_tokens=64, do_sample=False)
            decoded = tokenizer.decode(output[0], skip_special_tokens=True)
            return {
                "output_text_tail": decoded[-500:],
                "batch": 1,
                "loader": "AutoTokenizer+AutoModelForCausalLM",
                "processor_causal_loader_error": processor_causal_error,
            }
        except Exception as exc:
            tokenizer_causal_error = repr(exc)
            pass

    try:
        from transformers import AutoModelForImageTextToText, AutoProcessor

        processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
        model = AutoModelForImageTextToText.from_pretrained(
            model_id,
            torch_dtype=dtype_for_device(),
            device_map="auto" if torch.cuda.is_available() else None,
            trust_remote_code=True,
            quantization_config=quantization_config,
        )
        inputs = processor(text=prompt, return_tensors="pt")
        if torch.cuda.is_available():
            inputs = {key: value.to(model.device) for key, value in inputs.items()}
        with torch.no_grad():
            output = model.generate(**inputs, max_new_tokens=64, do_sample=False)
        try:
            decoded = processor.batch_decode(output, skip_special_tokens=True)[0]
        except Exception:
            decoded = str(output)
        return {
            "output_text_tail": decoded[-500:],
            "batch": 1,
            "loader": "AutoModelForImageTextToText",
            "processor_causal_loader_error": processor_causal_error,
            "tokenizer_causal_loader_error": tokenizer_causal_error,
        }
    except Exception as image_text_exc:
        raise RuntimeError(
            "All causal/multimodal loaders failed: "
            f"processor_causal={processor_causal_error}; "
            f"tokenizer_causal={tokenizer_causal_error}; "
            f"image_text={image_text_exc!r}"
        ) from image_text_exc


TASK_RUNNERS = {
    "embedding": smoke_embedding,
    "reranker": smoke_reranker,
    "encoder_classifier": smoke_encoder_classifier,
    "nli": smoke_nli,
    "causal_json": smoke_causal_json,
}


def result_key(result: dict[str, Any]) -> str:
    return f"{result['module']}::{result['dataset']}"


def spec_key(spec: dict[str, Any]) -> str:
    return f"{spec['module']}::{spec['dataset']}"


def run_runner(spec: dict[str, Any]) -> tuple[dict[str, Any], float | None]:
    reset_peak_memory()
    runner = TASK_RUNNERS[spec["task"]]
    details = runner(spec)
    peak = peak_vram_gb()
    return details, peak


def run_one(spec: dict[str, Any], mode: str, allow_cpu: bool) -> dict[str, Any]:
    deps = dependency_status()
    cuda = cuda_status()
    missing = [name for name, ok in deps.items() if name in REQUIRED_PACKAGES and not ok]
    if "4bit" in str(spec.get("quantization", "")).casefold() and not deps.get("bitsandbytes"):
        missing.append("bitsandbytes")
    result: dict[str, Any] = {
        "module": spec["module"],
        "group": spec.get("group", ""),
        "dataset": spec["dataset"],
        "main_model": spec["main_model"],
        "model_id": spec["model_id"],
        "backup": spec.get("backup", ""),
        "backup_model_id": spec.get("backup_model_id", ""),
        "task": spec["task"],
        "quantization": spec.get("quantization", ""),
        "max_test_batch": spec.get("max_test_batch", ""),
        "load_status": "PENDING",
        "peak_vram_gb": "",
        "decision": "pending",
        "error": "",
        "details": {},
        "tested_model_id": "",
        "used_backup": False,
    }
    if missing:
        result["load_status"] = "BLOCKED_MISSING_DEPENDENCY"
        result["decision"] = "install_dependencies"
        result["error"] = "Missing packages: " + ", ".join(sorted(set(missing)))
        return result
    if spec.get("requires_cuda", True) and not cuda.get("cuda_available") and not allow_cpu:
        result["load_status"] = "BLOCKED_NO_CUDA"
        result["decision"] = "wait_for_cuda_or_run_with_allow_cpu"
        result["error"] = "CUDA is required by config"
        return result
    if mode == "metadata":
        result["load_status"] = "PENDING_METADATA_ONLY"
        result["decision"] = "run_full_smoke"
        return result

    start = time.time()
    try:
        details, peak = run_runner(copy.deepcopy(spec))
        result["details"] = details
        result["peak_vram_gb"] = peak
        result["load_status"] = "PASS"
        result["decision"] = "keep"
        result["tested_model_id"] = spec["model_id"]
    except Exception as exc:  # noqa: BLE001 - report must capture model load failures.
        main_error = repr(exc)
        main_traceback = traceback.format_exc(limit=8)
        backup_model_id = spec.get("backup_model_id")
        if backup_model_id:
            try:
                cleanup_torch()
                backup_spec = copy.deepcopy(spec)
                backup_spec["model_id"] = backup_model_id
                backup_spec["main_model"] = spec.get("backup", backup_model_id)
                details, peak = run_runner(backup_spec)
                result["details"] = details
                result["details"]["main_model_error"] = main_error
                result["peak_vram_gb"] = peak
                result["load_status"] = "PASS_BACKUP"
                result["decision"] = "fallback"
                result["tested_model_id"] = backup_model_id
                result["used_backup"] = True
            except Exception as backup_exc:  # noqa: BLE001
                result["load_status"] = "FAIL"
                result["decision"] = "fallback_or_fix"
                result["error"] = f"main={main_error}; backup={backup_exc!r}"
                result["traceback"] = main_traceback + "\n--- BACKUP TRACEBACK ---\n" + traceback.format_exc(limit=8)
                try:
                    result["peak_vram_gb"] = peak_vram_gb()
                except Exception:  # noqa: BLE001
                    result["peak_vram_gb"] = ""
        else:
            result["load_status"] = "FAIL"
            result["decision"] = "fallback_or_fix"
            result["error"] = main_error
            result["traceback"] = main_traceback
            try:
                result["peak_vram_gb"] = peak_vram_gb()
            except Exception:  # noqa: BLE001
                result["peak_vram_gb"] = ""
    finally:
        cleanup_torch()
        result["elapsed_sec"] = round(time.time() - start, 3)
    return result


def status_is_pass(status: str) -> bool:
    return status in PASS_STATUSES


def filter_specs(specs: list[dict[str, Any]], group: str) -> list[dict[str, Any]]:
    if group == "all":
        return specs
    return [spec for spec in specs if spec.get("group") == group]


def placeholder_result(spec: dict[str, Any]) -> dict[str, Any]:
    return {
        "module": spec["module"],
        "group": spec.get("group", ""),
        "dataset": spec["dataset"],
        "main_model": spec["main_model"],
        "model_id": spec["model_id"],
        "backup": spec.get("backup", ""),
        "backup_model_id": spec.get("backup_model_id", ""),
        "task": spec["task"],
        "quantization": spec.get("quantization", ""),
        "max_test_batch": spec.get("max_test_batch", ""),
        "load_status": "NOT_RUN",
        "peak_vram_gb": "",
        "decision": "run_smoke_test",
        "error": "",
        "details": {},
        "tested_model_id": "",
        "used_backup": False,
    }


def merge_results(existing_report: dict[str, Any], specs: list[dict[str, Any]], new_results: list[dict[str, Any]]) -> list[dict[str, Any]]:
    merged = {spec_key(spec): placeholder_result(spec) for spec in specs}
    for result in existing_report.get("results", []) or []:
        key = result_key(result)
        if key in merged:
            if result.get("load_status") == "BLOCKED_MISSING_DEPENDENCY":
                continue
            merged[key].update(result)
    for result in new_results:
        merged[result_key(result)] = result
    return [merged[spec_key(spec)] for spec in specs]


def modules_pass(results: list[dict[str, Any]], required_modules: set[str]) -> bool:
    status_by_module = {result["module"]: result["load_status"] for result in results}
    return all(status_is_pass(status_by_module.get(module, "")) for module in required_modules)


def main() -> None:
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", type=Path, default=Path("configs/model/model_stack.yaml"))
    parser.add_argument("--report", type=Path, default=Path("outputs/stats/model_smoke_test_report.json"))
    parser.add_argument("--memory-csv", type=Path, default=Path("outputs/stats/model_memory_report.csv"))
    parser.add_argument("--table", type=Path, default=Path("outputs/tables/T4_model_stack.csv"))
    parser.add_argument("--mode", choices=["full", "metadata"], default=os.environ.get("SMOKE_TEST_MODE", "full"))
    parser.add_argument("--group", choices=sorted(VALID_GROUPS), default="all")
    parser.add_argument("--merge-existing", action="store_true")
    parser.add_argument("--allow-cpu", action="store_true")
    parser.add_argument("--strict", action="store_true")
    args = parser.parse_args()

    config = yaml.safe_load(args.config.read_text(encoding="utf-8"))
    specs = config["models"]
    selected_specs = filter_specs(specs, args.group)
    if not selected_specs and args.group != "all":
        print(f"No model specs selected for group={args.group}")
    new_results = [run_one(spec, mode=args.mode, allow_cpu=args.allow_cpu) for spec in selected_specs]
    existing_report = {}
    if args.merge_existing and args.report.exists():
        try:
            existing_report = json.loads(args.report.read_text(encoding="utf-8"))
        except json.JSONDecodeError:
            existing_report = {}
    results = merge_results(existing_report, specs, new_results)
    overall_pass = all(status_is_pass(result["load_status"]) for result in results)
    minimum_pass = modules_pass(results, MINIMUM_REQUIRED_MODULES)
    wikikg_pass = modules_pass(results, WIKIKG_REQUIRED_MODULES)
    llm_baseline_pass = modules_pass(results, LLM_BASELINE_REQUIRED_MODULES)
    report = {
        "mode": args.mode,
        "requested_group": args.group,
        "overall_pass": overall_pass,
        "minimum_pass": minimum_pass,
        "wikikg_pass": wikikg_pass,
        "llm_baseline_pass": llm_baseline_pass,
        "dependencies": dependency_status(),
        "cuda": cuda_status(),
        "results": results,
    }
    write_json(args.report, report)

    memory_rows = [
        {
            "module": result["module"],
            "group": result.get("group", ""),
            "dataset": result["dataset"],
            "model_id": result["model_id"],
            "tested_model_id": result.get("tested_model_id", ""),
            "task": result["task"],
            "load_status": result["load_status"],
            "peak_vram_gb": result["peak_vram_gb"],
            "elapsed_sec": result.get("elapsed_sec", ""),
            "max_test_batch": result["max_test_batch"],
            "error": result["error"],
        }
        for result in results
    ]
    write_csv(args.memory_csv, memory_rows)

    t4_rows = [
        {
            "Module": result["module"],
            "Dataset": result["dataset"],
            "Main model": result["main_model"],
            "Backup": result["backup"],
            "Load status": result["load_status"],
            "Peak VRAM GB": result["peak_vram_gb"],
            "Max test batch": result["max_test_batch"],
            "Decision": result["decision"],
        }
        for result in results
    ]
    write_csv(args.table, t4_rows)
    print(
        f"Wrote model smoke report to {args.report}. "
        f"PASS={overall_pass} MINIMUM_PASS={minimum_pass} WIKIKG_PASS={wikikg_pass}"
    )
    if args.strict and not overall_pass:
        raise SystemExit(1)


if __name__ == "__main__":
    main()