resume: --lora-adapter-repo support in loop.py

physix/training/loop.py

@@ -76,6 +76,15 @@ class TrainingConfig(BaseModel):
     #: before GRPO begins. Without this the cold base model rarely produces
     #: any reward signal in early steps.
     sft_checkpoint: Optional[str] = None
+    #: Optional Hub repo id (or local path) of an existing LoRA adapter to
+    #: warm-start GRPO from — e.g. a previous GRPO run that was interrupted
+    #: and pushed checkpoints to ``hub_checkpoint_repo_id``. When set, the
+    #: base ``model_name`` is loaded and this adapter is applied as the
+    #: starting trainable LoRA (skipping the fresh ``get_peft_model`` call).
+    #: SFT is unnecessary in this case (the adapter is already downstream
+    #: of an SFT warm-start), so leave ``sft_checkpoint`` unset when using
+    #: this flag.
+    lora_adapter_repo: Optional[str] = None
     output_dir: str = "runs/physix-1.5b-rl"
     max_seq_length: int = 2048
     lora_r: int = 16
@@ -219,6 +228,20 @@ def _load_model_and_tokenizer(
     policy that already knows the JSON format and equation grammar, so early
     rollouts produce meaningful reward signal instead of all scoring zero.
     """
+    if config.lora_adapter_repo:
+        _log.info(
+            "Resuming from existing LoRA adapter %s (warm-start, skipping fresh PEFT init)",
+            config.lora_adapter_repo,
+        )
+        model, tokenizer = FastLanguageModel.from_pretrained(
+            model_name=config.lora_adapter_repo,
+            max_seq_length=config.max_seq_length,
+            load_in_4bit=True,
+            dtype=None,
+        )
+        FastLanguageModel.for_training(model)
+        return model, tokenizer
+
     if config.sft_checkpoint:
         _log.info(
             "Loading SFT-warmed model from %s (GRPO will refine from here)",
@@ -485,7 +508,16 @@ def _parse_args() -> TrainingConfig:
         help="How to persist the final adapter (merged_16bit is deployable).",
     )
     parser.add_argument("--sft-checkpoint", default=None,
-                        help="Path to a LoRA adapter from sft.py to warm-start from.")
+                        help="Path to a merged SFT model from sft.py to warm-start from.")
+    parser.add_argument(
+        "--lora-adapter-repo",
+        default=None,
+        help=(
+            "Hub repo id (or local path) of an existing LoRA adapter to warm-start "
+            "GRPO from — e.g. a previous run's checkpoint at "
+            "user/physix-1.5b-rl-ckpt. Mutually exclusive with --sft-checkpoint."
+        ),
+    )
     parser.add_argument("--wandb-project", default="physix-live")
     parser.add_argument("--wandb-run-name", default=None)
     parser.add_argument("--push-to-hub", action="store_true")
@@ -504,9 +536,17 @@ def _parse_args() -> TrainingConfig:
 
     args = parser.parse_args()
 
+    if args.sft_checkpoint and args.lora_adapter_repo:
+        parser.error(
+            "--sft-checkpoint and --lora-adapter-repo are mutually exclusive. "
+            "Use --lora-adapter-repo to resume from a prior GRPO run, or "
+            "--sft-checkpoint for a fresh GRPO from a merged SFT model."
+        )
+
     return TrainingConfig(
         model_name=args.model,
         sft_checkpoint=args.sft_checkpoint,
+        lora_adapter_repo=args.lora_adapter_repo,
         output_dir=args.output_dir,
         num_steps=args.num_steps,
         learning_rate=args.learning_rate,