Add mistral small3 24B config and recipe

Signed-off-by: Joosung Yoon <joosungy@nvidia.com>

Author: eagle705

nemo/collections/llm/__init__.py

@@ -141,6 +141,7 @@
     MistralConfig7B,
     MistralModel,
     MistralNeMoConfig12B,
+    MistralSmall3Config24B,
     MixtralConfig,
     MixtralConfig8x3B,
     MixtralConfig8x7B,
@@ -251,6 +252,7 @@
     "MaskedTokenLossReduction",
     "MistralConfig7B",
     "MistralNeMoConfig12B",
+    "MistralSmall3Config24B",
     "MistralModel",
     "MixtralConfig",
     "MixtralConfig8x3B",

nemo/collections/llm/gpt/model/__init__.py

@@ -110,7 +110,12 @@
     Llama33NemotronSuper49BConfig,
     LlamaNemotronModel,
 )
-from nemo.collections.llm.gpt.model.mistral import MistralConfig7B, MistralModel, MistralNeMoConfig12B
+from nemo.collections.llm.gpt.model.mistral import (
+    MistralConfig7B,
+    MistralModel,
+    MistralNeMoConfig12B,
+    MistralSmall3Config24B,
+)
 from nemo.collections.llm.gpt.model.mixtral import (
     MixtralConfig,
     MixtralConfig8x3B,
@@ -194,6 +199,7 @@
     "MistralConfig7B",
     "MistralModel",
     "MistralNeMoConfig12B",
+    "MistralSmall3Config24B",
     "MixtralConfig8x3B",
     "MixtralConfig8x7B",
     "MixtralConfig8x22B",

nemo/collections/llm/gpt/model/mistral.py

@@ -102,6 +102,26 @@ class MistralNeMoConfig123B(MistralConfig7B):
     params_dtype: torch.dtype = torch.bfloat16
 
 
+@dataclass
+class MistralSmall3Config24B(MistralConfig7B):
+    """
+    https://mistral.ai/news/mistral-small-3/
+    """
+
+    num_layers: int = 40
+    hidden_size: int = 5120
+    ffn_hidden_size: int = 32768
+    num_attention_heads: int = 32
+    kv_channels: int = 128
+    seq_length: int = 32768
+
+    window_size: List[int] = None
+    cp_comm_type: str = None
+    rotary_percent: float = 1.0
+    rotary_base: float = 100000000.0
+    params_dtype: torch.dtype = torch.bfloat16
+
+
 class MistralModel(GPTModel):
     """ """
 

nemo/collections/llm/recipes/__init__.py

@@ -68,6 +68,7 @@
     mamba2_hybrid_8b,
     mistral_7b,
     mistral_nemo_12b,
+    mistral_small3_24b,
     mixtral_8x7b,
     mixtral_8x7b_16k,
     mixtral_8x7b_64k,
@@ -171,6 +172,7 @@
     "nemotron_nano_12b_v2",
     "mistral_7b",
     "mistral_nemo_12b",
+    "mistral_small3_24b",
     "hyena_base",
     "hyena_1b",
     "hyena_7b",

nemo/collections/llm/recipes/mistral_small3_24b.py

@@ -0,0 +1,309 @@
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+
+from typing import Callable, Optional
+
+import lightning.pytorch as pl
+import nemo_run as run
+import torch
+from lightning.pytorch.callbacks.callback import Callback
+from megatron.core.distributed import DistributedDataParallelConfig
+
+from nemo import lightning as nl
+from nemo.collections.llm.api import finetune, pretrain
+from nemo.collections.llm.gpt.data.mock import MockDataModule
+from nemo.collections.llm.gpt.model.mistral import MistralModel, MistralSmall3Config24B
+from nemo.collections.llm.peft import PEFT_STR2CLS
+from nemo.collections.llm.recipes.finetune_default import default_finetune_recipe
+from nemo.collections.llm.recipes.log.default import default_log, default_resume, tensorboard_logger
+from nemo.collections.llm.recipes.optim.adam import distributed_fused_adam_with_cosine_annealing
+from nemo.collections.llm.recipes.precision.mixed_precision import bf16_mixed
+from nemo.lightning.pytorch.callbacks.megatron_comm_overlap import MegatronCommOverlapCallback
+from nemo.utils.exp_manager import TimingCallback
+
+NAME = "mistral_small3_24b"
+
+
+@run.cli.factory(name=NAME)
+def model() -> run.Config[pl.LightningModule]:
+    """
+    Factory function to create a Mistral-Small-3-24B model configuration.
+
+    Returns:
+        run.Config[pl.LightningModule]: Configuration for the Mistral-Small-3-24B model.
+
+    Examples:
+        CLI usage:
+            $ nemo llm pretrain model=mistral_small3_24b ...
+
+        Python API usage:
+            >>> model_config = model()
+            >>> print(model_config)
+    """
+    return run.Config(MistralModel, config=run.Config(MistralSmall3Config24B))
+
+
+def trainer(
+    tensor_parallelism: int = 4,
+    pipeline_parallelism: int = 2,
+    pipeline_parallelism_type: Optional[torch.dtype] = None,
+    virtual_pipeline_parallelism: Optional[int] = None,
+    context_parallelism: int = 1,
+    sequence_parallelism: bool = True,
+    num_nodes: int = 1,
+    num_gpus_per_node: int = 8,
+    max_steps: int = 1168251,
+    callbacks: Optional[list[run.Config[Callback]]] = None,
+) -> run.Config[nl.Trainer]:
+    """
+    Configure the NeMo Lightning Trainer for Mistral-Small-3-24B model.
+
+    This function sets up the distributed training strategy and other training parameters.
+
+    Args:
+        tensor_parallelism (int): Degree of tensor model parallelism.
+        pipeline_parallelism (int): Degree of pipeline model parallelism.
+        pipeline_parallelism_type (Optional[torch.dtype]): Data type for pipeline parallelism.
+        virtual_pipeline_parallelism (Optional[int]): Size of virtual pipeline parallelism.
+        context_parallelism (int): Degree of context parallelism.
+        sequence_parallelism (bool): Whether to use sequence parallelism.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        max_steps (int): Maximum number of training steps.
+        callbacks (Optional[list[run.Config[Callback]]]): List of callback configurations.
+
+    Returns:
+        run.Config[nl.Trainer]: Configuration for the NeMo Lightning Trainer.
+
+    Examples:
+        CLI usage:
+            $ nemo llm pretrain trainer=mistral_small3_24b ...
+
+        Python API usage:
+            >>> trainer_config = trainer(num_nodes=2, num_gpus_per_node=8)
+            >>> print(trainer_config)
+
+    Note:
+        For more information on distributed training strategies, refer to the
+        NeMo documentation on multi-GPU and multi-node training.
+    """
+    strategy = run.Config(
+        nl.MegatronStrategy,
+        tensor_model_parallel_size=tensor_parallelism,
+        pipeline_model_parallel_size=pipeline_parallelism,
+        pipeline_dtype=pipeline_parallelism_type,
+        virtual_pipeline_model_parallel_size=virtual_pipeline_parallelism,
+        context_parallel_size=context_parallelism,
+        sequence_parallel=sequence_parallelism,
+        gradient_as_bucket_view=True,
+        ckpt_async_save=True,
+        ckpt_parallel_load=True,
+        ddp=run.Config(
+            DistributedDataParallelConfig,
+            check_for_nan_in_grad=True,
+            grad_reduce_in_fp32=True,
+            overlap_grad_reduce=True,
+            overlap_param_gather=True,
+        ),
+    )
+
+    trainer = run.Config(
+        nl.Trainer,
+        accelerator="gpu",
+        accumulate_grad_batches=1,
+        callbacks=callbacks,
+        devices=num_gpus_per_node,
+        limit_test_batches=50,
+        limit_val_batches=32,
+        log_every_n_steps=10,
+        max_steps=max_steps,
+        num_nodes=num_nodes,
+        plugins=bf16_mixed(),
+        strategy=strategy,
+        use_distributed_sampler=False,
+        val_check_interval=2000,
+    )
+
+    return trainer
+
+
+@run.cli.factory(target=pretrain, name=NAME)
+def pretrain_recipe(
+    dir: Optional[str] = None, name: str = "default", num_nodes: int = 1, num_gpus_per_node: int = 8, fn=pretrain
+) -> run.Partial:
+    """
+    Create a pre-training recipe for Mistral-Small-3-24B model.
+
+    This function sets up a complete configuration for pre-training, including
+    model, trainer, data, logging, optimization, and resumption settings.
+
+    Args:
+        dir (Optional[str]): Directory for saving logs and checkpoints.
+        name (str): Name of the pre-training run.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        fn (Callable): The pre-training function to use.
+
+    Returns:
+        run.Partial: Partial configuration for pre-training.
+
+    Examples:
+        CLI usage:
+            $ nemo llm pretrain --factory mistral_small3_24b
+            $ nemo llm pretrain --factory "mistral_small3_24b(num_nodes=2, name='my_pretrain')"
+
+        Python API usage:
+            >>> recipe = pretrain_recipe(name="mistral_small3_24b", num_nodes=2)
+            >>> print(recipe)
+    """
+    return run.Partial(
+        fn,
+        model=model(),
+        trainer=trainer(
+            num_nodes=num_nodes,
+            num_gpus_per_node=num_gpus_per_node,
+            callbacks=[run.Config(TimingCallback)],
+        ),
+        data=run.Config(MockDataModule, seq_length=8192, global_batch_size=512, micro_batch_size=1),
+        log=default_log(dir=dir, name=name, tensorboard_logger=tensorboard_logger(name=name)),
+        optim=distributed_fused_adam_with_cosine_annealing(max_lr=3e-4),
+        resume=default_resume(),
+    )
+
+
+@run.cli.factory(target=pretrain, name=NAME + "_optimized")
+def pretrain_recipe_performance(
+    dir: Optional[str] = None,
+    name: str = "default",
+    num_nodes: int = 1,
+    num_gpus_per_node: int = 8,
+    fn: Callable = pretrain,
+) -> run.Partial:
+    """
+    Create a performance-optimized pre-training recipe for Mistral-Small-3-24B model.
+
+    This recipe enables performance optimizations that may not be suitable for all use cases.
+    It builds upon the standard pre-training recipe and adds additional performance enhancements.
+
+    Args:
+        dir (Optional[str]): Directory for saving logs and checkpoints.
+        name (str): Name of the pre-training run.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        fn (Callable): The pre-training function to use.
+
+    Returns:
+        run.Partial: Partial configuration for performance-optimized pre-training.
+
+    Examples:
+            $ nemo llm pretrain --factory mistral_small3_24b_optimized
+
+        Python API usage:
+            >>> recipe = pretrain_recipe_performance(name="mistral_small3_24b_perf", num_nodes=4)
+            >>> print(recipe)
+
+    Note:
+        Use this recipe with caution and only when you need maximum performance.
+        It may not be suitable for all hardware configurations or use cases.
+    """
+    recipe = pretrain_recipe(name=name, dir=dir, num_nodes=num_nodes, num_gpus_per_node=num_gpus_per_node, fn=fn)
+
+    recipe.trainer.callbacks.append(
+        run.Config(
+            MegatronCommOverlapCallback,
+            tp_comm_overlap=True,
+        )
+    )
+    return recipe
+
+
+@run.cli.factory(target=finetune, name=NAME)
+def finetune_recipe(
+    dir: Optional[str] = None,
+    resume_path: str = "mistralai/Mistral-Small-3-24B-Instruct-2501",
+    name: str = "default",
+    num_nodes: int = 1,
+    num_gpus_per_node: int = 8,
+    peft_scheme: Optional[str] = 'lora',
+    seq_length: Optional[int] = None,
+    packed_sequence: bool = False,
+) -> run.Partial:
+    """
+    Create a fine-tuning recipe for Mistral-Small-3-24B model.
+
+    This function sets up a complete configuration for fine-tuning, including
+    model, trainer, data, logging, optimization, and resumption settings.
+    The recipe uses LoRA (Low-Rank Adaptation) for efficient fine-tuning, unless peft_scheme is set to None.
+
+    Args:
+        dir (Optional[str]): Directory for saving logs and checkpoints.
+        resume_path (str): Path to the NeMo checkpoint
+        name (str): Name of the fine-tuning run.
+        num_nodes (int): Number of compute nodes to use.
+        num_gpus_per_node (int): Number of GPUs per node.
+        peft_scheme (Optional[str]): Name of the peft scheme to use for fine-tuning.
+            Allowed values: 'lora'/'dora'/'none'/None.
+        seq_length (int): Maximum number of tokens per microbatch.
+        packed_sequence (Optional[bool]): Packing multiple training sequences into one long sequence for training
+            efficiency. Default sequence length is 2048.
+
+    Returns:
+        run.Partial: Partial configuration for fine-tuning.
+
+    Examples:
+        CLI usage:
+            $ nemo llm finetune --factory mistral_small3_24b
+
+        Python API usage:
+            >>> recipe = finetune_recipe(name="mistral_small3_24b_finetune", num_nodes=2)
+            >>> print(recipe)
+
+    Note:
+        This recipe uses the SQuAD dataset for fine-tuning.
+    """
+
+    # For unpacked sequence, most samples in SQuAD dataset are shorter than 2K
+    if seq_length is None:
+        seq_length = 4096 if packed_sequence else 2048
+
+    recipe = default_finetune_recipe(
+        model(),
+        resume_path,
+        dir,
+        name,
+        num_nodes,
+        num_gpus_per_node,
+        packed_sequence,
+    )
+    if peft_scheme is None or peft_scheme.lower() == 'none':
+        recipe.trainer.strategy.tensor_model_parallel_size = 4
+        recipe.trainer.strategy.pipeline_model_parallel_size = 2
+        recipe.optim.config.lr = 5e-6
+    elif peft_scheme.lower() in ['lora', 'dora']:
+        recipe.peft = run.Config(
+            PEFT_STR2CLS[peft_scheme.lower()], target_modules=['linear_qkv', 'linear_proj'], dim=32
+        )
+        recipe.optim.config.lr = 1e-4
+    else:
+        raise ValueError(f"Unrecognized peft scheme: {peft_scheme}")
+
+    # Sequence length settings in the model and dataset must agree
+    recipe.model.config.seq_length = seq_length
+    recipe.data.seq_length = seq_length
+    if packed_sequence:
+        recipe.data.dataset_kwargs = {'pad_to_max_length': True}
+        recipe.data.packed_sequence_specs = run.Config(PackedSequenceSpecs, packed_sequence_size=seq_length)
+
+    return recipe