File size: 12,186 Bytes
e890160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
381091a
 
 
 
 
 
e890160
 
 
 
 
 
 
 
 
 
1c2f24f
e890160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
c2815cb
 
 
 
 
 
 
 
e890160
c2815cb
 
 
 
 
 
 
e890160
c2815cb
8425a53
c2815cb
 
 
67810ba
c2815cb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e890160
c2815cb
 
e890160
 
c2815cb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e890160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1c2f24f
e890160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
746de52
 
 
 
 
 
 
 
 
 
 
 
 
 
 
e890160
 
 
 
 
 
 
746de52
e890160
 
 
746de52
e890160
746de52
 
 
 
 
 
e890160
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
model_utils.py β€” Model loading, LoRA setup, GPU detection, checkpoint resume.

All persistent state lives on Hugging Face Hub.
GPU is ephemeral β€” we only ever write to Hub, never assume local disk survives.
"""

from __future__ import annotations

import os
from pathlib import Path
from typing import Any, Dict, Optional, Tuple

import torch
from huggingface_hub import HfApi, snapshot_download

# Reduce CUDA allocator fragmentation β€” critical on A10G where free memory
# is non-contiguous after generation KV-cache eviction.
# Must be set before any CUDA allocation.
if "PYTORCH_ALLOC_CONF" not in os.environ:
    os.environ["PYTORCH_ALLOC_CONF"] = "expandable_segments:True"


# ────────────────────────────────────────────────
# GPU Detection
# ────────────────────────────────────────────────

def detect_gpu_tier() -> str:
    """Return 'a100', 'a10g', or 't4' based on GPU name and VRAM."""
    if not torch.cuda.is_available():
        print("[model_utils] No CUDA detected β€” will be extremely slow")
        return "t4"
    vram_gb = torch.cuda.get_device_properties(0).total_memory / 1e9
    name = torch.cuda.get_device_name(0).lower()
    if "a100" in name or vram_gb >= 70:
        return "a100"
    elif "a10" in name or vram_gb >= 20:
        return "a10g"
    else:
        return "t4"


def gpu_scaled_config(cfg: Dict[str, Any]) -> Dict[str, Any]:
    """Auto-scale LoRA rank and seq_length based on GPU tier.

    SRE actions are short (~40 tokens output). The bottleneck is
    env API latency, not context length. So we keep seq_len tight
    (1024 is plenty) and spend VRAM on rank + episodes instead.
    """
    tier = detect_gpu_tier()

    # GPU-tier safe DEFAULTS β€” applied only if the config value is BELOW these floors.
    # These are not caps: if the user deliberately sets a higher value (e.g. lora_rank=64),
    # it is preserved. The override only kicks in to prevent unsafe low values.
    #
    # A10G headroom (verified empirically):
    #   base model: 3.3 GiB  |  one forward (bf16, seq=512, batch=1): 8.9 GiB
    #   peak with loss_batch_size=2: ~17.8 GiB  |  22.3 GiB total  β†’  ~4.5 GiB headroom
    #   lora_rank=64 vs 32: +83 MiB difference β€” negligible
    if tier == "a100":
        safe_defaults = {
            "max_seq_length": 2048,
            "lora_rank": 48,
            "lora_alpha": 48,
            "per_device_train_batch_size": 4,
            "loss_batch_size": 4,
        }
    elif tier == "a10g":
        safe_defaults = {
            "max_seq_length": 768,    # 512β†’768: longer system prompt needs more room; safe with loss_batch_size=1
            "lora_rank": 64,          # verified safe: +83 MiB vs rank-32
            "lora_alpha": 64,         # keep == lora_rank
            "per_device_train_batch_size": 2,
            "loss_batch_size": 1,     # proven safe. New fused-CE path may allow 2 β€” test after validating.
        }
        # Hard floors β€” values below these will always OOM on A10G
        hard_floors = {
            "max_seq_length": 256,
            "loss_batch_size": 1,
        }
    else:  # t4 / unknown
        safe_defaults = {
            "max_seq_length": 512,
            "lora_rank": 16,
            "lora_alpha": 16,
            "per_device_train_batch_size": 1,
            "loss_batch_size": 1,
        }
        hard_floors = {}

    # Apply: only override if config is missing the key entirely, or below the hard floor.
    # Env vars (ANTIATROPOS_<KEY>=<value>) always win over everything.
    hard_floors = locals().get("hard_floors", {})
    for key, default_val in safe_defaults.items():
        env_key = f"ANTIATROPOS_{key.upper()}"

        # Env var always wins
        if env_key in os.environ:
            val = os.environ[env_key]
            cfg[key] = int(val) if isinstance(default_val, int) else float(val)
            print(f"[model_utils] env override: {key} = {cfg[key]}")
            continue

        current = cfg.get(key)
        floor = hard_floors.get(key)

        if current is None:
            # Key not set at all β€” apply safe default
            cfg[key] = default_val
            print(f"[model_utils] GPU {tier}: setting {key} = {default_val} (not in config)")
        elif floor is not None and current < floor:
            # Below hard floor β€” unsafe, must override
            print(f"[model_utils] GPU {tier}: {key}={current} is below safe floor {floor}, "
                  f"raising to {floor}")
            cfg[key] = floor
        else:
            # User set a value >= floor β€” respect it
            if current != default_val:
                print(f"[model_utils] GPU {tier}: keeping user config {key}={current} "
                      f"(default was {default_val})")

    return cfg


# ────────────────────────────────────────────────
# Model Loading
# ────────────────────────────────────────────────

def load_base_model(cfg: Dict[str, Any]):
    """Load base model with Unsloth QLoRA. Returns (model, tokenizer)."""
    from unsloth import FastLanguageModel

    model_name = cfg["base_model"]
    max_seq_length = cfg.get("max_seq_length", 1024)
    load_in_4bit = cfg.get("load_in_4bit", True)

    print(f"[model_utils] Loading {model_name} "
          f"(seq_len={max_seq_length}, 4bit={load_in_4bit})")

    model, tokenizer = FastLanguageModel.from_pretrained(
        model_name=model_name,
        max_seq_length=max_seq_length,
        load_in_4bit=load_in_4bit,
        dtype=None,  # auto-detect bf16/fp16
        trust_remote_code=True,
    )

    return model, tokenizer


def attach_lora(model, cfg: Dict[str, Any], seed: int = 42):
    """Attach LoRA adapters to the base model."""
    from unsloth import FastLanguageModel

    rank = cfg.get("lora_rank", 32)
    alpha = cfg.get("lora_alpha", 32)
    dropout = cfg.get("lora_dropout", 0.0)
    target_modules = cfg.get("lora_target_modules", [
        "q_proj", "k_proj", "v_proj", "o_proj",
        "gate_proj", "up_proj", "down_proj",
    ])

    print(f"[model_utils] Attaching LoRA: rank={rank}, alpha={alpha}, "
          f"dropout={dropout}, targets={len(target_modules)} modules")

    model = FastLanguageModel.get_peft_model(
        model,
        r=rank,
        lora_alpha=alpha,
        lora_dropout=dropout,
        target_modules=target_modules,
        bias="none",
        use_gradient_checkpointing="unsloth",
        random_state=seed,
    )

    if torch.cuda.is_available():
        vram_used = torch.cuda.memory_allocated() / 1e9
        vram_total = torch.cuda.get_device_properties(0).total_memory / 1e9
        print(f"[model_utils] VRAM: {vram_used:.2f} / {vram_total:.2f} GiB")

    trainable = sum(p.numel() for p in model.parameters() if p.requires_grad)
    total = sum(p.numel() for p in model.parameters())
    print(f"[model_utils] Trainable: {trainable:,} / {total:,} "
          f"({100 * trainable / total:.2f}%)")

    return model


# ────────────────────────────────────────────────
# Checkpoint Resume
# ────────────────────────────────────────────────

def find_latest_checkpoint(hub_repo: str) -> Optional[str]:
    """Check Hub for the latest checkpoint subfolder.

    Checkpoints are stored as: <hub_repo>/checkpoint-<step>/
    Returns the path to download, or None if no checkpoint exists.
    """
    if not hub_repo:
        return None

    try:
        api = HfApi()
        # List all files in the repo, find checkpoint dirs
        files = api.list_repo_files(hub_repo, repo_type="model")
        checkpoint_dirs = set()
        for f in files:
            # checkpoint-123/adapter_model.safetensors
            parts = f.split("/")
            if len(parts) >= 2 and parts[0].startswith("checkpoint-"):
                try:
                    step = int(parts[0].split("-")[1])
                    checkpoint_dirs.add(step)
                except (ValueError, IndexError):
                    continue

        if not checkpoint_dirs:
            return None

        latest_step = max(checkpoint_dirs)
        ckpt_path = f"checkpoint-{latest_step}"
        print(f"[model_utils] Found Hub checkpoint: {hub_repo}/{ckpt_path}")
        return ckpt_path

    except Exception as e:
        print(f"[model_utils] Could not check Hub for checkpoints: {e}")
        return None


def download_checkpoint(hub_repo: str, checkpoint_path: str,
                        local_dir: str = "/tmp/antiatropos_ckpt") -> str:
    """Download a checkpoint from Hub to local disk.

    Returns the local path containing adapter files.
    """
    print(f"[model_utils] Downloading checkpoint {hub_repo}/{checkpoint_path}...")
    snapshot_download(
        repo_id=hub_repo,
        repo_type="model",
        local_dir=local_dir,
        allow_patterns=[f"{checkpoint_path}/*"],
    )
    return str(Path(local_dir) / checkpoint_path)


def load_from_checkpoint(model, tokenizer, ckpt_local_path: str):
    """Load LoRA weights from a local checkpoint directory."""
    from peft import PeftModel

    print(f"[model_utils] Loading adapter from {ckpt_local_path}")
    # For Unsloth models, we reload the adapter
    model.load_adapter(ckpt_local_path)
    return model


# ────────────────────────────────────────────────
# Save & Push
# ────────────────────────────────────────────────

def save_checkpoint(model, tokenizer, output_dir: str, step: int) -> str:
    """Save adapter + tokenizer locally. Returns the checkpoint path."""
    ckpt_dir = str(Path(output_dir) / f"checkpoint-{step}")
    Path(ckpt_dir).mkdir(parents=True, exist_ok=True)

    model.save_pretrained(ckpt_dir)
    tokenizer.save_pretrained(ckpt_dir)
    print(f"[model_utils] Checkpoint saved: {ckpt_dir}")
    return ckpt_dir


def push_to_hub(
    local_dir: str,
    hub_repo: str,
    commit_message: str = "",
    path_in_repo: str = "",
) -> None:
    """Push a local directory to a Hub model repo.

    Args:
        local_dir:     Local folder to upload.
        hub_repo:      HF model repo (e.g. 'Keshav051/antiatropos-qlora').
        commit_message: Commit message.
        path_in_repo:  Subfolder INSIDE the Hub repo (e.g. 'run_001/checkpoint-0005').
                       If empty, uploads to the repo root (legacy behaviour).
    """
    if not hub_repo:
        print("[model_utils] No hub_model_repo configured, skipping push")
        return

    try:
        from huggingface_hub import upload_folder

        kwargs = dict(
            folder_path=local_dir,
            repo_id=hub_repo,
            repo_type="model",
            commit_message=commit_message or "Upload from AntiAtropos training",
        )
        if path_in_repo:
            kwargs["path_in_repo"] = path_in_repo

        upload_folder(**kwargs)
        dest = f"{hub_repo}/{path_in_repo}" if path_in_repo else hub_repo
        print(f"[model_utils] Pushed β†’ {dest}")
    except Exception as e:
        print(f"[model_utils] Push failed: {e}")


def push_adapter_to_hub(model, tokenizer, hub_repo: str,
                        step: int, output_dir: str = "/tmp/antiatropos_final") -> None:
    """Save final adapter and push to Hub."""
    if not hub_repo:
        print("[model_utils] No hub_model_repo configured, skipping final push")
        return

    Path(output_dir).mkdir(parents=True, exist_ok=True)
    model.save_pretrained(output_dir)
    tokenizer.save_pretrained(output_dir)
    push_to_hub(output_dir, hub_repo, f"AntiAtropos QLoRA step {step}")