[recipe, training] fix: Correct adam_eps default and add non-default config logging

examples/resiliency/fault_tolerance/fault_tolerance_example.py

@@ -150,7 +150,7 @@ def create_config(
         fp16=False,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         use_distributed_optimizer=True,
         clip_grad=1.0,
         lr=1e-4,

examples/resiliency/straggler_detection/straggler_detection_example.py

@@ -118,7 +118,7 @@ def create_config(
         fp16=False,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         use_distributed_optimizer=True,
         clip_grad=1.0,
         lr=1e-4,

src/megatron/bridge/recipes/deepseek/deepseek_v2.py

@@ -196,7 +196,7 @@ def _deepseek_common(
         lr_decay_iters=lr_decay_iters,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         weight_decay=0.1,
         max_lr=lr,
         min_lr=min_lr,

src/megatron/bridge/recipes/deepseek/deepseek_v3.py

@@ -271,7 +271,7 @@ def _deepseek_v3_common(
         lr_decay_iters=lr_decay_iters,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         weight_decay=0.1,
         max_lr=lr,
         min_lr=min_lr,

src/megatron/bridge/recipes/llama/llama2.py

@@ -194,7 +194,7 @@ def _llama2_common(
         lr_decay_iters=lr_decay_iters,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         weight_decay=0.1,
         max_lr=lr,
         min_lr=min_lr,

src/megatron/bridge/recipes/moonlight/moonlight_16b.py

@@ -277,7 +277,7 @@ def _moonlight_common(
             lr_decay_iters=train_iters,
             adam_beta1=0.9,
             adam_beta2=0.95,
-            adam_eps=1e-5,
+            adam_eps=1e-8,
             weight_decay=0.1,
             max_lr=lr,
             min_lr=min_lr,
@@ -616,7 +616,7 @@ def _moonlight_finetune_common(
         min_lr=min_lr,
         adam_beta1=0.9,
         adam_beta2=0.98,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         weight_decay=0.1,
     )
 

src/megatron/bridge/recipes/nemotronh/nemotron_3_nano.py

@@ -496,7 +496,6 @@ def _nemotron_3_nano_finetune_common(
         lr_decay_iters=lr_decay_iters,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-8,
         weight_decay=0.1,
         max_lr=finetune_lr,
         min_lr=min_lr,

src/megatron/bridge/recipes/nemotronh/nemotronh.py

@@ -282,7 +282,7 @@ def _nemotronh_common(
         lr_decay_iters=lr_decay_iters,
         adam_beta1=0.9,
         adam_beta2=0.95,
-        adam_eps=1e-5,
+        adam_eps=1e-8,
         weight_decay=0.1,
         max_lr=lr,
         min_lr=min_lr,

src/megatron/bridge/recipes/utils/optimizer_utils.py

@@ -88,7 +88,7 @@ def distributed_fused_adam_with_cosine_annealing(
     lr_decay_iters: Optional[int] = None,
     adam_beta1: float = 0.9,
     adam_beta2: float = 0.95,
-    adam_eps: float = 1e-5,
+    adam_eps: float = 1e-8,
     weight_decay: float = 0.1,
     max_lr: float = 1e-4,
     min_lr: Optional[float] = None,
@@ -156,7 +156,7 @@ def distributed_fused_adam_with_cosine_annealing_samples(
     lr_decay_samples: Optional[int] = None,
     adam_beta1: float = 0.9,
     adam_beta2: float = 0.95,
-    adam_eps: float = 1e-5,
+    adam_eps: float = 1e-8,
     weight_decay: float = 0.1,
     max_lr: float = 1e-4,
     min_lr: Optional[float] = None,

src/megatron/bridge/training/config.py

@@ -16,7 +16,7 @@
 import os
 import signal
 from abc import ABC, abstractmethod
-from dataclasses import dataclass, field, fields
+from dataclasses import MISSING, dataclass, field, fields
 from pathlib import Path
 from typing import Any, Dict, Literal, Optional, Tuple, Union
 
@@ -30,6 +30,8 @@
 )
 from megatron.core.transformer.enums import AttnBackend, CudaGraphScope
 from megatron.core.transformer.module import MegatronModule
+from megatron.core.transformer.transformer_config import MLATransformerConfig as MCoreMLATransformerConfig
+from megatron.core.transformer.transformer_config import TransformerConfig as MCoreTransformerConfig
 
 from megatron.bridge.data.datasets.packed_sequence import PackedSequenceSpecs
 from megatron.bridge.models import GPTModelProvider, T5ModelProvider
@@ -1658,6 +1660,161 @@ def _calculate_scheduler_steps(self) -> None:
             else:
                 self.scheduler.lr_warmup_steps = self.scheduler.lr_warmup_iters * self.train.global_batch_size
 
+    def log_non_default_values(self) -> None:
+        """Log configuration values that differ from Megatron Core defaults.
+
+        For configs that inherit from Megatron Core (e.g., OptimizerConfig, DDPConfig,
+        TransformerConfig), this method logs only the values that differ from the Mcore
+        defaults. This makes it easier to spot unintended deviations from baseline settings.
+
+        For configs that don't inherit from Mcore, key values are logged via
+        `_get_key_config_values`, which excludes None values and callables.
+        """
+        # Determine the correct Mcore parent class for the model config
+        # Some models (e.g., DeepSeek) use MLATransformerConfig instead of TransformerConfig
+        model_mcore_class = _get_mcore_transformer_parent(self.model)
+
+        # Map of config names to their (config object, Mcore parent class or None)
+        mcore_configs = [
+            ("optimizer", self.optimizer, MCoreOptimizerConfig),
+            ("ddp", self.ddp, MCoreDistributedDataParallelConfig),
+            ("model", self.model, model_mcore_class),
+        ]
+
+        # Non-Mcore configs - log all values
+        non_mcore_configs = [
+            ("train", self.train),
+            ("scheduler", self.scheduler),
+            ("checkpoint", self.checkpoint),
+            ("logger", self.logger),
+            ("tokenizer", self.tokenizer),
+            ("rng", self.rng),
+        ]
+
+        log_lines = [""]
+        log_lines.append("=" * 70)
+        log_lines.append("Configuration Summary (Non-Default Values vs Megatron Core)")
+        log_lines.append("=" * 70)
+
+        # Log non-default values for Mcore configs
+        for config_name, config_obj, mcore_class in mcore_configs:
+            non_defaults = _get_non_default_values(config_obj, mcore_class)
+            if non_defaults:
+                log_lines.append(f"\n[{config_name}] Non-default values (vs Mcore {mcore_class.__name__}):")
+                for field_name, (current_val, default_val) in sorted(non_defaults.items()):
+                    log_lines.append(f"  {field_name}: {current_val!r}  (Mcore default: {default_val!r})")
+
+        # Log key values for non-Mcore configs
+        log_lines.append("\n" + "-" * 70)
+        log_lines.append("Other Configuration Values:")
+        log_lines.append("-" * 70)
+
+        for config_name, config_obj in non_mcore_configs:
+            if config_obj is None:
+                continue
+            key_values = _get_key_config_values(config_obj)
+            if key_values:
+                log_lines.append(f"\n[{config_name}]:")
+                for field_name, value in sorted(key_values.items()):
+                    log_lines.append(f"  {field_name}: {value!r}")
+
+        log_lines.append("\n" + "=" * 70)
+
+        print_rank_0("\n".join(log_lines))
+
+
+def _get_mcore_transformer_parent(model_config: Any) -> type:
+    """Determine the correct Mcore TransformerConfig parent class for a model.
+
+    Some models (e.g., DeepSeek v2/v3) inherit from MLATransformerConfig instead of
+    the base TransformerConfig. This function checks the inheritance chain to find
+    the appropriate Mcore class to use as the baseline for comparison.
+
+    Args:
+        model_config: The model configuration object.
+
+    Returns:
+        The appropriate Mcore TransformerConfig class (MCoreMLATransformerConfig or
+        MCoreTransformerConfig).
+    """
+    # Check if the model inherits from MLATransformerConfig
+    if isinstance(model_config, MCoreMLATransformerConfig):
+        return MCoreMLATransformerConfig
+    return MCoreTransformerConfig
+
+
+def _get_non_default_values(config_obj: Any, mcore_class: type) -> Dict[str, Tuple[Any, Any]]:
+    """Get values that differ from Mcore parent class defaults.
+
+    Args:
+        config_obj: The config object to compare.
+        mcore_class: The Megatron Core parent class to compare against.
+
+    Returns:
+        Dictionary mapping field name to (current_value, default_value) for non-default fields.
+    """
+    non_defaults = {}
+
+    # Get default values from Mcore class
+    mcore_defaults = {}
+    for f in fields(mcore_class):
+        if f.name.startswith("_"):
+            continue
+        if f.default is not MISSING:
+            mcore_defaults[f.name] = f.default
+        elif f.default_factory is not MISSING:
+            mcore_defaults[f.name] = f.default_factory()
+
+    # Compare current values against Mcore defaults
+    for f in fields(config_obj):
+        if f.name.startswith("_"):
+            continue
+        field_name = f.name
+        current_value = getattr(config_obj, field_name, None)
+
+        if field_name in mcore_defaults:
+            default_value = mcore_defaults[field_name]
+            # Skip callable values (like functions) and complex objects
+            if callable(current_value) or callable(default_value):
+                continue
+            # Compare values
+            try:
+                if current_value != default_value:
+                    non_defaults[field_name] = (current_value, default_value)
+            except (TypeError, ValueError):
+                # Some types may not be directly comparable (e.g., torch.dtype)
+                if str(current_value) != str(default_value):
+                    non_defaults[field_name] = (current_value, default_value)
+
+    return non_defaults
+
+
+def _get_key_config_values(config_obj: Any) -> Dict[str, Any]:
+    """Get key configuration values for non-Mcore configs.
+
+    Args:
+        config_obj: The config object to extract values from.
+
+    Returns:
+        Dictionary mapping field name to value for key fields.
+    """
+    values = {}
+    if not hasattr(config_obj, "__dataclass_fields__"):
+        return values
+
+    for f in fields(config_obj):
+        if f.name.startswith("_"):
+            continue
+        value = getattr(config_obj, f.name, None)
+        # Skip None values and complex objects
+        if value is None:
+            continue
+        if callable(value):
+            continue
+        values[f.name] = value
+
+    return values
+
 
 def runtime_config_update(cfg: ConfigContainer) -> None:
     """Apply runtime configuration updates prior to initialization.

src/megatron/bridge/training/setup.py

@@ -110,6 +110,7 @@ def setup(
         SetupOutput containing the populated state, model, optimizer, scheduler, dataloaders, and ckpt context.
     """
     cfg = state.cfg
+    maybe_log_and_save_config(cfg)
 
     # Conditionally enable experimental features for Megatron Core
     set_experimental_flag(cfg.dist.enable_megatron_core_experimental)
@@ -305,7 +306,6 @@ def modelopt_pre_wrap_hook(model):
     # Print setup timing.
     print_rank_0("done with setup ...")
     timers.log(["model-and-optimizer-setup", "train/valid/test-data-iterators-setup"], barrier=True)
-    maybe_log_and_save_config(cfg)
 
     return SetupOutput(
         state,
@@ -484,7 +484,14 @@ def _validate_and_set_vocab_size(model_vocab_size: Optional[int], tokenizer_voca
 
 
 def maybe_log_and_save_config(cfg: ConfigContainer) -> None:
-    """Save configuration to disk and log it on rank 0."""
+    """Save configuration to disk and log non-default values on rank 0.
+
+    Instead of printing the full config YAML, this now logs only the values
+    that differ from Megatron Core defaults, making it easier to spot
+    unintended configuration deviations.
+
+    The full config can still be saved to a file via logger.save_config_filepath.
+    """
 
     if get_rank_safe() != 0:
         return
@@ -495,6 +502,4 @@ def maybe_log_and_save_config(cfg: ConfigContainer) -> None:
         except Exception as e:
             print_rank_0(f"Error saving config to file {cfg.logger.save_config_filepath}: {e}")
 
-    print("------- Task Configuration -------")
-    cfg.print_yaml()
-    print("----------------------------------")
+    cfg.log_non_default_values()

tests/functional_tests/training/test_callbacks.py

@@ -182,7 +182,7 @@ def test_callbacks(self):
                 fp16=False,
                 adam_beta1=0.9,
                 adam_beta2=0.95,
-                adam_eps=1e-5,
+                adam_eps=1e-8,
                 use_distributed_optimizer=True,
                 clip_grad=1.0,
                 lr=3e-3,

tests/functional_tests/training/test_decentralized_pg.py

@@ -132,7 +132,7 @@ def test_pretrain_with_decentralized_pg(self, tmp_path):
                     fp16=False,
                     adam_beta1=0.9,
                     adam_beta2=0.95,
-                    adam_eps=1e-5,
+                    adam_eps=1e-8,
                     use_distributed_optimizer=True,
                     clip_grad=1.0,
                     lr=3e-3,
@@ -257,7 +257,7 @@ def test_pretrain_with_decentralized_pg_disabled(self, tmp_path):
                     fp16=False,
                     adam_beta1=0.9,
                     adam_beta2=0.95,
-                    adam_eps=1e-5,
+                    adam_eps=1e-8,
                     use_distributed_optimizer=True,
                     clip_grad=1.0,
                     lr=3e-3,
@@ -389,7 +389,7 @@ def test_pretrain_with_decentralized_pg_and_pp(self, tmp_path):
                     fp16=False,
                     adam_beta1=0.9,
                     adam_beta2=0.95,
-                    adam_eps=1e-5,
+                    adam_eps=1e-8,
                     use_distributed_optimizer=True,
                     clip_grad=1.0,
                     lr=3e-3,
@@ -521,7 +521,7 @@ def test_pretrain_with_decentralized_pg_and_cp(self, tmp_path):
                     fp16=False,
                     adam_beta1=0.9,
                     adam_beta2=0.95,
-                    adam_eps=1e-5,
+                    adam_eps=1e-8,
                     use_distributed_optimizer=True,
                     clip_grad=1.0,
                     lr=3e-3,
@@ -653,7 +653,7 @@ def test_pretrain_with_decentralized_pg_combined_parallelism(self, tmp_path):
                     fp16=False,
                     adam_beta1=0.9,
                     adam_beta2=0.95,
-                    adam_eps=1e-5,
+                    adam_eps=1e-8,
                     use_distributed_optimizer=True,
                     clip_grad=1.0,
                     lr=3e-3,
@@ -785,7 +785,7 @@ def test_pretrain_with_decentralized_pg_and_tp(self, tmp_path):
                     fp16=False,
                     adam_beta1=0.9,
                     adam_beta2=0.95,
-                    adam_eps=1e-5,
+                    adam_eps=1e-8,
                     use_distributed_optimizer=True,
                     clip_grad=1.0,
                     lr=3e-3,

tests/functional_tests/training/test_finetune_dora.py

@@ -126,7 +126,7 @@ def _create_optimizer_config(self, lr=3e-3):
             fp16=False,
             adam_beta1=0.9,
             adam_beta2=0.95,
-            adam_eps=1e-5,
+            adam_eps=1e-8,
             use_distributed_optimizer=True,
             clip_grad=1.0,
             lr=lr,

tests/functional_tests/training/test_finetune_lora.py

@@ -244,7 +244,7 @@ def _create_optimizer_config(self, lr=3e-3):
             optimizer="adam",
             adam_beta1=0.9,
             adam_beta2=0.95,
-            adam_eps=1e-5,
+            adam_eps=1e-8,
             use_distributed_optimizer=True,
             clip_grad=1.0,
             lr=lr,