File size: 5,968 Bytes
23ce76a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# /// script
# requires-python = ">=3.11"
# dependencies = [
#   "torch>=2.1,<2.7",
#   "transformers>=4.46,<4.50",
#   "datasets",
#   "hqq>=0.2.8",
#   "accelerate",
#   "tqdm",
#   "huggingface_hub",
# ]
# ///
"""
HSAQ-quantize the LoRA-merged guardian model for swarm serving.

Source: mxguru1/master-chief-guardian-8b-v1 (bf16 merged, 16 GB)
Output: HSAQ-quantized state_dict + tokenizer + config in
        mxguru1/master-chief-guardian-8b-v1-hsaq (HF dataset)

This is Phase 1A of the wargame thesis-test. The artifact is then loaded
by a local FastAPI shim (Phase 1B) that mimics Ollama's endpoints, so the
wargame can route to it via LOCAL_HF_SHIMS in adversarial_wargame_swarm.py.

Cost: ~$1.50, ~30 min on A100 80GB.
"""
from __future__ import annotations
import json, logging, os, subprocess, sys, time
from datetime import UTC, datetime
from pathlib import Path
import torch

if not torch.cuda.is_available():
    subprocess.check_call([sys.executable, "-m", "pip", "install", "torch", "--force-reinstall",
                            "--index-url", "https://download.pytorch.org/whl/cu124"])
    import importlib; importlib.reload(torch)
    if not torch.cuda.is_available(): sys.exit(1)

logging.basicConfig(level=logging.INFO, format="%(asctime)s %(levelname)s %(name)s | %(message)s")
logger = logging.getLogger("GuardianHSAQ")

sys.path.insert(0, "/opt/hsaq")
from quantization.hsaq.config import HSAQConfig
from quantization.hsaq.pipeline import HSAQPipeline

MODEL_ID = "mxguru1/master-chief-guardian-8b-v1"
HF_TOKEN = os.environ.get("HF_TOKEN")
RUN_TAG = datetime.now(UTC).strftime("%Y%m%d_%H%M%S")
OUT = Path("/tmp/guardian_hsaq"); OUT.mkdir(parents=True, exist_ok=True)


def main():
    start = time.time()
    report = {"run_tag": RUN_TAG, "approach": "hsaq_quantize_for_serving", "model_id": MODEL_ID}
    try:
        cfg = HSAQConfig(
            model_id=MODEL_ID,
            output_dir=str(OUT.parent),
            hf_token=HF_TOKEN,
            gpu_budget_gb=11.2,
            enable_2bit=False,
            enable_pruning=False,
            train_lora=False,  # already LoRA-merged
            calibration_samples=128,
        )
        pipe = HSAQPipeline(cfg)

        logger.info("Stage 1: load + profile %s", MODEL_ID)
        model, tokenizer = pipe._load_model()
        model = model.to("cuda:0")
        if tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token

        sensitivity = pipe.profiler.profile(model)
        candidates = pipe._build_layer_candidates(sensitivity, model)
        from quantization.hsaq.assignment import assign_bit_widths
        weight_budget_gb = pipe._compute_weight_budget()
        assignment = assign_bit_widths(candidates, weight_budget_gb)
        name_to_bits = {a.component: a.chosen.bits for a in assignment.assignments}

        logger.info("Stage 2: HQQ apply to all 281 Linears")
        n_hqq = pipe._apply_per_module_hqq(model, name_to_bits, device="cuda:0")
        logger.info("HQQ applied to %d Linears", n_hqq)

        # Save model artifact
        artifact_dir = OUT / "guardian-hsaq"
        artifact_dir.mkdir(parents=True, exist_ok=True)
        try:
            from hqq.models.hf.base import AutoHQQHFModel
            AutoHQQHFModel.save_quantized(model, str(artifact_dir))
            save_method = "AutoHQQHFModel.save_quantized"
        except Exception as exc:
            logger.warning("AutoHQQHFModel.save_quantized failed (%s); falling back to state_dict", exc)
            torch.save(model.state_dict(), artifact_dir / "pytorch_model.bin")
            save_method = "state_dict"
        tokenizer.save_pretrained(str(artifact_dir))
        model.config.save_pretrained(str(artifact_dir))

        # Per-layer bits manifest for the shim
        manifest = {
            "model_id": MODEL_ID,
            "save_method": save_method,
            "n_linears_quantized": n_hqq,
            "name_to_bits": name_to_bits,
            "group_size": 64,
            "axis": 0,
            "compute_dtype": "bfloat16",
            "total_weight_gb": round(assignment.total_weights_gb, 3),
        }
        (artifact_dir / "hsaq_manifest.json").write_text(json.dumps(manifest, indent=2))
        report["artifact"] = manifest
        report["artifact_size_bytes"] = sum(p.stat().st_size for p in artifact_dir.rglob("*") if p.is_file())
        logger.info("Artifact saved: %s (%.2f GB)", artifact_dir, report["artifact_size_bytes"] / 1e9)

        # Upload as dataset (HQQ artifacts aren't standard HF model format; dataset is fine)
        from huggingface_hub import HfApi
        repo_id = "mxguru1/master-chief-guardian-8b-v1-hsaq"
        api = HfApi(token=HF_TOKEN)
        api.create_repo(repo_id=repo_id, repo_type="dataset", exist_ok=True)
        for p in artifact_dir.rglob("*"):
            if p.is_file():
                api.upload_file(
                    path_or_fileobj=str(p),
                    path_in_repo=p.name,
                    repo_id=repo_id,
                    repo_type="dataset",
                )
        logger.info("Uploaded artifact to https://huggingface.co/datasets/%s", repo_id)
        report["status"] = "success"
        report["artifact_repo"] = repo_id

    except Exception as e:
        logger.exception("Run failed")
        report["status"] = "failed"
        report["error"] = repr(e)
    finally:
        report["elapsed_s"] = round(time.time() - start, 1)
        # Always upload the report
        from huggingface_hub import HfApi
        api = HfApi(token=HF_TOKEN)
        rid = f"mxguru1/guardian-hsaq-job-{RUN_TAG}"
        try:
            api.create_repo(repo_id=rid, repo_type="dataset", exist_ok=True)
            p = OUT / "report.json"; p.write_text(json.dumps(report, indent=2))
            api.upload_file(path_or_fileobj=str(p), path_in_repo="report.json", repo_id=rid, repo_type="dataset")
        except Exception:
            pass


if __name__ == "__main__":
    main()