Option 1 for dynamic batching OOM safety

src/dataloaders/build_dataloader.py

@@ -26,6 +26,7 @@ def build_dataloader(
         return torch_data_loader
 
     def build_torch_dataloader(self, torch_dataset):
+        self.dataset = torch_dataset
         sampler = self.get_torch_dataloader_sampler(torch_dataset)
         if "train" in self.args.mode:
             torch_data_loader = DataLoader(
@@ -64,6 +65,8 @@ def collate_fn(self, batch):
         batch = [item for item in batch if item is not None]
         if len(batch) == 0:
             return None
+        max_len = max(item["elm_input_ids"].shape[0] for item in batch)
+        batch = [self.dataset.pad_to_batch(item, max_len) for item in batch]
         self._assert_same_structure_and_shapes(batch)
         return torch.utils.data.dataloader.default_collate(batch)
 

src/dataloaders/data_representation/base.py

@@ -158,6 +158,19 @@ def pad_input(self, tokens: list) -> list:
         padding_len = self.args.llm_input_len - len(tokens)
         return [self.llm_tokenizer.pad_token_id] * padding_len + tokens  # left side padding
 
+    def pad_to_batch(self, item: dict, target_len: int) -> dict:
+        pad_len = target_len - item["elm_input_ids"].shape[0]
+        if pad_len == 0:
+            return item
+        pad_values = {"elm_input_ids": self.llm_tokenizer.pad_token_id, "elm_labels": -100, "elm_attention_mask": 0}
+        for key, value in pad_values.items():
+            if key in item:
+                item[key] = torch.nn.functional.pad(item[key], (pad_len, 0), value=value)  # left side padding
+        if "signal_id_indices" in item:
+            idx = item["signal_id_indices"]
+            item["signal_id_indices"] = torch.where(idx >= 0, idx + pad_len, idx)  # left-pad shifts every real position by pad_len
+        return item
+
     def make_prompt(
         self,
         text: str,

src/dataloaders/data_representation/rgb.py

@@ -72,8 +72,8 @@ def prepare_training_set(
         signal_id_indices = self.find_signal_token_indices(truncated_padded_input)
         attention_mask = self.create_attention_mask(truncated_padded_input)
         labels = self.create_labels(truncated_padded_input)
-        assert len(truncated_padded_input) == len(attention_mask) == len(labels) == self.args.llm_input_len, (
-            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)} != {self.args.llm_input_len}"
+        assert len(truncated_padded_input) == len(attention_mask) == len(labels), (
+            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)}"
         )
         elm = {
             "elm_input_ids": torch.tensor(truncated_padded_input, dtype=torch.int64),
@@ -113,12 +113,6 @@ def augment_image(self, image: np.array):
 
     def trunc_pad_input(self, prompt: str):
         prompt_tokens = self.llm_tokenizer.encode(prompt, add_special_tokens=False)
-        if "train" in self.args.mode:
-            prompt_len = len(prompt_tokens)
-            if prompt_len == self.args.llm_input_len:
-                return prompt_tokens
-            elif prompt_len < self.args.llm_input_len:
-                return self.pad_input(prompt_tokens)
+        if "train" in self.args.mode and len(prompt_tokens) > self.args.llm_input_len:
             return self.truncate_input_preserving_signal_tokens(prompt_tokens)
-        else:
-            return prompt_tokens
+        return prompt_tokens

src/dataloaders/data_representation/signal.py

@@ -47,8 +47,8 @@ def prepare_training_set(
         labels = self.create_labels(truncated_padded_input)
         # print("signal_id_indices", len(signal_id_indices), "\n")
         assert len(signal_id_indices) == self.args.num_encoder_tokens
-        assert len(truncated_padded_input) == len(attention_mask) == len(labels) == self.args.llm_input_len, (
-            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)} != {self.args.llm_input_len}"
+        assert len(truncated_padded_input) == len(attention_mask) == len(labels), (
+            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)}"
         )
         elm = {
             "elm_input_ids": torch.tensor(truncated_padded_input, dtype=torch.int64),
@@ -76,16 +76,9 @@ def prepare_eval_inference_set(
 
     def trunc_pad_input(self, prompt: str):
         prompt_tokens = self.llm_tokenizer.encode(prompt, add_special_tokens=False)
-        if "train" in self.args.mode:
-            prompt_len = len(prompt_tokens)
-            # print("prompt len", prompt_len, "\n")
-            if prompt_len == self.args.llm_input_len:
-                return prompt_tokens
-            elif prompt_len < self.args.llm_input_len:
-                return self.pad_input(prompt_tokens)
+        if "train" in self.args.mode and len(prompt_tokens) > self.args.llm_input_len:
             return self.truncate_input_preserving_signal_tokens(prompt_tokens)
-        else:
-            return prompt_tokens
+        return prompt_tokens
 
     def transform_ecg_signal(self, ecg_signal):
         if self.args.elm == "base_elf":

src/dataloaders/data_representation/stacked_signal.py

@@ -55,8 +55,8 @@ def prepare_training_set(
         signal_id_indices = self.find_signal_token_indices(truncated_padded_input)
         attention_mask = self.create_attention_mask(truncated_padded_input)
         labels = self.create_labels(truncated_padded_input)
-        assert len(truncated_padded_input) == len(attention_mask) == len(labels) == self.args.llm_input_len, (
-            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)} != {self.args.llm_input_len}"
+        assert len(truncated_padded_input) == len(attention_mask) == len(labels), (
+            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)}"
         )
         elm = {
             "elm_input_ids": torch.tensor(truncated_padded_input, dtype=torch.int64),
@@ -90,12 +90,6 @@ def signal_to_stacked_signal(self, signal):
 
     def trunc_pad_input(self, prompt: str):
         prompt_tokens = self.llm_tokenizer.encode(prompt, add_special_tokens=False)
-        if "train" in self.args.mode:
-            prompt_len = len(prompt_tokens)
-            if prompt_len == self.args.llm_input_len:
-                return prompt_tokens
-            elif prompt_len < self.args.llm_input_len:
-                return self.pad_input(prompt_tokens)
+        if "train" in self.args.mode and len(prompt_tokens) > self.args.llm_input_len:
             return self.truncate_input_preserving_signal_tokens(prompt_tokens)
-        else:
-            return prompt_tokens
+        return prompt_tokens

src/dataloaders/data_representation/symbolic.py

@@ -55,8 +55,8 @@ def prepare_training_set(
             self.check_labels(labels)
             self.check_attention_mask(truncated_padded_input, attention_mask)
 
-        assert len(truncated_padded_input) == len(attention_mask) == len(labels) == self.args.llm_input_len, (
-            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)} != {self.args.llm_input_len}"
+        assert len(truncated_padded_input) == len(attention_mask) == len(labels), (
+            f"Length mismatch: {len(truncated_padded_input)} != {len(attention_mask)} != {len(labels)}"
         )
         # print("truncated_padded_ecg_tokens", truncated_padded_ecg_tokens)
         # print("signal_id_indices", signal_id_indices)
@@ -86,12 +86,8 @@ def trunc_pad_input(self, ecg_tokens: np.ndarray, prompt: str):
             min_ecg_token_len = int(self.args.min_ecg_tokens_len)
             before_len, after_len, ecg_token_len = len(before), len(after), len(ecg_tokens)
 
-            if before_len + after_len + ecg_token_len == self.args.llm_input_len:
-                # return before + ecg_tokens + after, self.convert_ecg_tokens(ecg_tokens)
+            if before_len + after_len + ecg_token_len <= self.args.llm_input_len:
                 return before + ecg_tokens + after
-            elif before_len + after_len + ecg_token_len < self.args.llm_input_len:
-                # return self.pad_input(before + ecg_tokens + after), self.convert_ecg_tokens(ecg_tokens)
-                return self.pad_input(before + ecg_tokens + after)
 
             if before_len + min_ecg_token_len > self.args.llm_input_len:
                 raise ValueError("before + min_ecg exceeds llm_input_len; lower min_ecg_tokens_len.")

src/main_trainer.py

@@ -13,7 +13,7 @@
 
 from utils.checkpoint_manager import CheckpointManager
 from utils.seed_manager import set_seed
-from utils.gpu_manager import is_main, init_dist, cleanup, GPUSetup, broadcast_value
+from utils.gpu_manager import is_main, init_dist, cleanup, GPUSetup, broadcast_value, assert_max_batch_fits
 from utils.dir_file_manager import setup_experiment_folders
 from utils.wandb_manager import setup_wandb, cleanup_wandb
 
@@ -67,6 +67,8 @@ def main():
         if args.resume_ckpt and checkpoint_manager:
             start_epoch = checkpoint_manager.resume_checkpoint(args.resume_ckpt, elm, optimizer)
         runner = run_rl_train if getattr(args, "train_phase", "sft") == "rl" else run_train
+        if not args.dev and runner is run_train:
+            assert_max_batch_fits(elm, next(iter(dataloader)), args, optimizer)
         for epoch in range(start_epoch, args.epochs):
             train_result = runner(elm, optimizer, dataloader, epoch, args, checkpoint_manager)
             should_stop = False

src/optimizers/optimizer_setup.py

@@ -143,6 +143,15 @@ def _build_muon_optimizer(self, model):
 
         return MuonAdamW(muon_opt, adamw_opt, adamw_lr_ratio)
 
+    def estimated_state_bytes(self) -> int:
+        """Resident optimizer-state bytes once allocated: Adam/AdamW keep 2 tensors
+        (exp_avg, exp_avg_sq) per param; Muon keeps 1 momentum buffer per param."""
+        def nbytes(param_groups, per_param):
+            return per_param * sum(p.numel() * p.element_size() for g in param_groups for p in g["params"])
+        if self._is_muon:
+            return nbytes(self.optimizer.muon.param_groups, 1) + nbytes(self.optimizer.adamw.param_groups, 2)
+        return nbytes(self.optimizer.param_groups, 2)
+
     def _log_config(self):
         if is_main():
             if self._is_muon:

src/utils/gpu_manager.py

@@ -70,6 +70,46 @@ def train_dev_break(enabled: bool, batch: dict, loss_value: float) -> bool:
     return broadcast_value(should_break, src=0)
 
 
+def assert_max_batch_fits(model, batch: dict, args, optimizer=None, margin: float = 0.10) -> None:
+    if batch is None or not torch.cuda.is_available():
+        return
+    F, ids = torch.nn.functional, batch["elm_input_ids"]
+    pad = args.llm_input_len - ids.shape[1]
+    if pad <= 0:
+        return  # first batch already at the cap; training itself already stresses it
+    probe = dict(batch)
+    probe["elm_input_ids"] = F.pad(ids, (pad, 0), value=int(ids.reshape(-1)[0]))
+    probe["elm_attention_mask"] = F.pad(batch["elm_attention_mask"], (pad, 0), value=0)
+    if "elm_labels" in batch:
+        probe["elm_labels"] = F.pad(batch["elm_labels"], (pad, 0), value=-100)
+    if "signal_id_indices" in batch:
+        probe["signal_id_indices"] = torch.where(batch["signal_id_indices"] >= 0, batch["signal_id_indices"] + pad, batch["signal_id_indices"])
+    device = next(model.parameters()).device
+    probe = {k: batch_to_device(v, device) for k, v in probe.items()}
+    torch.cuda.reset_peak_memory_stats(device)
+    try:
+        model(**probe).loss.backward()
+    except torch.cuda.OutOfMemoryError as e:
+        raise RuntimeError(f"[fit-check] worst-case batch (batch_size={ids.shape[0]} x llm_input_len="
+                           f"{args.llm_input_len}) does not fit; lower --batch_size / --llm_input_len or "
+                           f"add --gradient_checkpointing. ({e})") from e
+    finally:
+        model.zero_grad(set_to_none=True)
+        torch.cuda.empty_cache()
+    # backward peak (params + grads + activations) + optimizer state resident during real steps
+    opt_bytes = optimizer.estimated_state_bytes() if optimizer is not None else 0
+    peak = torch.cuda.max_memory_allocated(device) + opt_bytes
+    free, _ = torch.cuda.mem_get_info(device)
+    budget = free + torch.cuda.memory_allocated(device)  # free accounts for other processes; + our resident params
+    gb = 1024 ** 3
+    if peak * (1 + margin) > budget:
+        raise RuntimeError(f"[fit-check] worst-case peak ~{peak / gb:.1f} GB (incl. optimizer state) leaves "
+                           f"<{margin:.0%} headroom of {budget / gb:.1f} GB available; lower --batch_size / "
+                           f"--llm_input_len or add --gradient_checkpointing.")
+    if is_main():
+        print(f"[fit-check] OK: worst-case ~{peak / gb:.1f} GB / {budget / gb:.1f} GB available ({margin:.0%} margin kept).")
+
+
 class GPUSetup:
     def __init__(self, args: argparse.Namespace):
         self.args = args
@@ -82,7 +122,7 @@ def setup_gpu(self, model: torch.nn.Module, find_unused_parameters) -> torch.nn.
         if is_main():
             print(f"find_unused_parameters: {find_unused_parameters}")
         if self.args.torch_compile:
-            model = torch.compile(model)
+            model = torch.compile(model, dynamic=True)
         return model
 
     def get_device(self) -> torch.device: