diff --git a/src/megatron/bridge/perf_recipes/llama/__init__.py b/src/megatron/bridge/perf_recipes/llama/__init__.py
index 5169592aad..8099e4566c 100644
--- a/src/megatron/bridge/perf_recipes/llama/__init__.py
+++ b/src/megatron/bridge/perf_recipes/llama/__init__.py
@@ -77,6 +77,9 @@
     llama3_70b_sft_32gpu_gb200_fp8mx_config,
 )
 from megatron.bridge.perf_recipes.llama.gb200.llama31 import (
+    llama31_8b_pretrain_8gpu_gb200_nvfp4_config,
+    llama31_8b_pretrain_72gpu_gb200_nvfp4_config,
+    llama31_8b_pretrain_512gpu_gb200_fp8cs_config,
     llama31_405b_pretrain_128gpu_gb200_bf16_config,
     llama31_405b_pretrain_128gpu_gb200_fp8cs_config,
     llama31_405b_pretrain_128gpu_gb200_fp8mx_config,
@@ -115,6 +118,9 @@
     llama3_70b_sft_32gpu_gb300_fp8mx_config,
 )
 from megatron.bridge.perf_recipes.llama.gb300.llama31 import (
+    llama31_8b_pretrain_8gpu_gb300_nvfp4_config,
+    llama31_8b_pretrain_72gpu_gb300_nvfp4_config,
+    llama31_8b_pretrain_512gpu_gb300_fp8cs_config,
     llama31_405b_pretrain_128gpu_gb300_bf16_config,
     llama31_405b_pretrain_128gpu_gb300_fp8cs_config,
     llama31_405b_pretrain_128gpu_gb300_fp8mx_config,
diff --git a/src/megatron/bridge/perf_recipes/llama/common.py b/src/megatron/bridge/perf_recipes/llama/common.py
index 93aa3e9908..59dd834160 100644
--- a/src/megatron/bridge/perf_recipes/llama/common.py
+++ b/src/megatron/bridge/perf_recipes/llama/common.py
@@ -29,6 +29,7 @@
     llama3_70b_peft_config,
     llama3_70b_pretrain_config,
     llama3_70b_sft_config,
+    llama31_8b_pretrain_config,
     llama31_405b_pretrain_config,
 )
 from megatron.bridge.training.comm_overlap import (
@@ -62,6 +63,199 @@ def _llama_benchmark_common(cfg: ConfigContainer) -> None:
         cfg.rng.te_rng_tracker = cfg.model.use_te_rng_tracker = False
 
 
+def _llama31_8b_common(cfg: ConfigContainer) -> None:
+    """Set common configurations for all Llama3.1 8B configs."""
+    cfg.optimizer.adam_beta2 = 0.95
+    cfg.optimizer.adam_eps = 1e-05
+    cfg.optimizer.bf16 = True
+    cfg.optimizer.fp8_recipe = 'tensorwise'
+    cfg.optimizer.overlap_param_gather = True
+    cfg.optimizer.params_dtype = torch.bfloat16
+    cfg.optimizer.use_distributed_optimizer = True
+    cfg.optimizer.weight_decay = 0.1
+    cfg.optimizer.bucket_size = 768000000
+    cfg.ddp.data_parallel_sharding_strategy = 'optim_grads_params'
+    cfg.ddp.delay_wgrad_compute = False
+    cfg.ddp.overlap_grad_reduce = True
+    cfg.ddp.overlap_param_gather = True
+    cfg.ddp.use_distributed_optimizer = True
+    cfg.ddp.bucket_size = 768000000
+    cfg.ddp.grad_reduce_in_fp32 = False
+    cfg.ddp.average_in_collective = False
+    cfg.ddp.check_for_nan_in_grad = False
+    cfg.model.add_bias_linear = False
+    cfg.model.apply_rope_fusion = True
+    cfg.model.attention_dropout = 0.0
+    cfg.model.gradient_accumulation_fusion = True
+    cfg.model.hidden_dropout = 0.0
+    cfg.model.hidden_size = 4096
+    cfg.model.kv_channels = 128
+    cfg.model.masked_softmax_fusion = True
+    cfg.model.microbatch_group_size_per_vp_stage = 1
+    cfg.model.normalization = 'RMSNorm'
+    cfg.model.num_attention_heads = 32
+    cfg.model.num_layers = 32
+    cfg.model.num_layers_at_end_in_bf16 = 0
+    cfg.model.num_layers_at_start_in_bf16 = 0
+    cfg.model.num_query_groups = 8
+    cfg.model.overlap_p2p_comm = True
+    cfg.model.params_dtype = torch.bfloat16
+    cfg.model.persist_layer_norm = True
+    cfg.model.pipeline_dtype = torch.bfloat16
+    cfg.model.recompute_modules = ['core_attn']
+    cfg.model.tp_only_amax_red = True
+    cfg.model.use_te_rng_tracker = True
+    cfg.model.wgrad_deferral_limit = 50
+    cfg.model.batch_p2p_comm = False
+    cfg.model.fused_single_qkv_rope = True
+    cfg.model.cuda_graph_warmup_steps = 1
+    cfg.model.vocab_size = 128256
+    cfg.train.check_optimizer_step_success = False
+    cfg.train.decrease_batch_size_if_needed = False
+    cfg.train.empty_unused_memory_level = 0
+    cfg.train.exit_signal_handler = False
+    cfg.train.exit_signal_handler_for_dataloader = False
+    cfg.train.exit_signal_handler_for_training = False
+    cfg.train.iterations_to_skip = []
+    cfg.train.manual_gc = True
+    cfg.train.manual_gc_eval = False
+    cfg.train.manual_gc_interval = 500
+    cfg.train.skip_sync_grad_norm_across_mp = True
+    cfg.scheduler.end_weight_decay = 0.1
+    cfg.scheduler.lr_decay_style = 'cosine'
+    cfg.scheduler.lr_warmup_init = 0.0
+    cfg.scheduler.lr_warmup_samples = 0
+    cfg.scheduler.lr_wsd_decay_style = 'exponential'
+    cfg.scheduler.override_opt_param_scheduler = False
+    cfg.scheduler.start_weight_decay = 0.1
+    cfg.scheduler.use_checkpoint_opt_param_scheduler = False
+    cfg.scheduler.weight_decay_incr_style = 'constant'
+    cfg.dataset.add_extra_token_to_sequence = True
+    cfg.dataset.allow_ambiguous_pad_tokens = False
+    cfg.dataset.create_attention_mask = False
+    cfg.dataset.data_parallel_size = 1
+    cfg.dataset.data_sharding = True
+    cfg.dataset.dataloader_type = 'single'
+    cfg.dataset.defer_npy_index_mmap = True
+    cfg.dataset.drop_last = True
+    cfg.dataset.drop_last_partial_validation_sequence = True
+    cfg.dataset.eod_mask_loss = False
+    cfg.dataset.fast_cache_load = True
+    cfg.dataset.hybrid_context_parallel = False
+    cfg.dataset.inter_document_masking = False
+    cfg.dataset.mid_level_dataset_surplus = 0.005
+    cfg.dataset.mmap_bin_files = True
+    cfg.dataset.mock = False
+    cfg.dataset.num_dataset_builder_threads = 1
+    cfg.dataset.num_workers = 8
+    cfg.dataset.persistent_workers = True
+    cfg.dataset.pin_memory = True
+    cfg.dataset.reset_attention_mask = False
+    cfg.dataset.reset_position_ids = False
+    cfg.dataset.seq_length = 8192
+    cfg.dataset.sequence_length = 8192
+    cfg.dataset.sequence_parallel_size = 0
+    cfg.checkpoint.also_save_hf_checkpoint = False
+    cfg.checkpoint.async_ckpt_cpu_priority = 10
+    cfg.checkpoint.async_ckpt_io_priority = 3
+    cfg.checkpoint.async_ckpt_use_cpu_shm = False
+    cfg.checkpoint.async_save = False
+    cfg.checkpoint.async_strategy = 'mcore'
+    cfg.checkpoint.async_write_results_mp_mode = 'fork'
+    cfg.checkpoint.auto_detect_ckpt_format = False
+    cfg.checkpoint.ckpt_assume_constant_structure = False
+    cfg.checkpoint.ckpt_convert_update_legacy_dist_opt_format = False
+    cfg.checkpoint.ckpt_format = 'torch_dist'
+    cfg.checkpoint.ckpt_fully_parallel_load_exchange_algo = 'broadcast'
+    cfg.checkpoint.ckpt_fully_parallel_load_process_group = 'dp'
+    cfg.checkpoint.ckpt_fully_parallel_save_process_group = 'dp'
+    cfg.checkpoint.ckpt_load_validate_sharding_integrity = True
+    cfg.checkpoint.dist_ckpt_optim_fully_reshardable = False
+    cfg.checkpoint.dist_ckpt_save_pre_mcore_014 = False
+    cfg.checkpoint.dist_ckpt_strictness = 'log_all'
+    cfg.checkpoint.dist_ckpt_workers = 1
+    cfg.checkpoint.distrib_optim_fully_reshardable_mem_efficient = False
+    cfg.checkpoint.exit_on_missing_checkpoint = False
+    cfg.checkpoint.finetune = False
+    cfg.checkpoint.fully_parallel_load = True
+    cfg.checkpoint.fully_parallel_save = True
+    cfg.checkpoint.hf_distributed_save = False
+    cfg.checkpoint.hf_save_every_n_ranks = 1
+    cfg.checkpoint.hf_trust_remote_code = False
+    cfg.checkpoint.load_optim = False
+    cfg.checkpoint.load_rng = False
+    cfg.checkpoint.most_recent_k = -1
+    cfg.checkpoint.non_persistent_local_ckpt_algo = 'fully_parallel'
+    cfg.checkpoint.replication = False
+    cfg.checkpoint.replication_factor = 2
+    cfg.checkpoint.save_optim = True
+    cfg.checkpoint.save_rng = True
+    cfg.checkpoint.save_tokenizer_assets = True
+    cfg.checkpoint.storage_writers_per_rank = 1
+    cfg.checkpoint.strict_fsdp_dtensor_load = False
+    cfg.checkpoint.use_checkpoint_args = False
+    cfg.checkpoint.use_mp_args_from_checkpoint_args = False
+    cfg.checkpoint.use_persistent_ckpt_worker = True
+    cfg.checkpoint.use_tokenizer_model_from_checkpoint_args = True
+    cfg.checkpoint.verify_integrity = False
+    cfg.logger.barrier_with_L1_time = True
+    cfg.logger.filter_warnings = True
+    cfg.logger.log_device_memory_used = False
+    cfg.logger.log_energy = False
+    cfg.logger.log_interval = 1
+    cfg.logger.log_l2_norm_grad_to_tensorboard = False
+    cfg.logger.log_loss_scale_to_tensorboard = True
+    cfg.logger.log_max_attention_logit = False
+    cfg.logger.log_memory_to_tensorboard = False
+    cfg.logger.log_num_zeros_in_grad = False
+    cfg.logger.log_params_norm = False
+    cfg.logger.log_progress = False
+    cfg.logger.log_runtime_to_tensorboard = False
+    cfg.logger.log_throughput = False
+    cfg.logger.log_throughput_to_tensorboard = False
+    cfg.logger.log_timers_to_tensorboard = False
+    cfg.logger.log_validation_ppl_to_tensorboard = False
+    cfg.logger.log_world_size_to_tensorboard = False
+    cfg.logger.logging_level = 20
+    cfg.logger.mlflow_log_artifacts = True
+    cfg.logger.moe_routing_trace_capture_hidden_states = False
+    cfg.logger.moe_routing_trace_capture_logits = False
+    cfg.logger.moe_routing_trace_dump_weights = False
+    cfg.logger.runtime_time_unit = 'hours'
+    cfg.logger.set_level_for_all_loggers = False
+    cfg.logger.skip_train_metrics_log = False
+    cfg.logger.tensorboard_log_interval = 1
+    cfg.logger.tensorboard_queue_size = 1000
+    cfg.logger.throughput_window_size = 100
+    cfg.logger.timing_log_level = 0
+    cfg.logger.timing_log_option = 'minmax'
+    cfg.tokenizer.force_system_message = False
+    cfg.tokenizer.hf_tokenizer_kwargs = {'use_fast': True}
+    cfg.tokenizer.make_vocab_size_divisible_by = 1
+    cfg.tokenizer.null_tokenizer_pad_id = -1
+    cfg.tokenizer.pad_vocab_size = False
+    cfg.tokenizer.rank = 0
+    cfg.tokenizer.sp_tokenizer_kwargs = {}
+    cfg.tokenizer.tensor_model_parallel_size = 1
+    cfg.tokenizer.tiktoken_num_special_tokens = 1000
+    cfg.tokenizer.tokenizer_hf_no_include_special_tokens = False
+    cfg.tokenizer.tokenizer_hf_no_use_fast = False
+    cfg.tokenizer.tokenizer_sentencepiece_ignore_extra_whitespaces = True
+    cfg.tokenizer.tokenizer_sentencepiece_legacy = False
+    cfg.tokenizer.tokenizer_type = "NullTokenizer"
+    cfg.tokenizer.trust_remote_code = False
+    cfg.tokenizer.vocab_extra_ids = 0
+    cfg.rng.data_parallel_random_init: False
+    cfg.rng.inference_rng_tracker = False
+    cfg.rng.te_rng_tracker = True
+    cfg.ddp.align_param_gather = False
+    cfg.optimizer.overlap_param_gather_with_optimizer_step = False
+    cfg.validation.full_validation = False
+    cfg.validation.multiple_validation_sets = False
+    cfg.validation.skip_train = False
+    cfg.validation.test_mode = False
+
+    
 def _llama2_70b_precision_config(compute_dtype: str):
     """Get the precision configs for the given compute dtype and FP8 recipe."""
     precision_config = _perf_precision(compute_dtype)
@@ -183,4 +377,4 @@ def _llama2_benchmark_common(cfg: ConfigContainer) -> None:
     cfg.checkpoint.load_optim = False
     cfg.checkpoint.load_rng = False
     cfg.tokenizer.tokenizer_type = "HuggingFaceTokenizer"
-    cfg.tokenizer.tensor_model_parallel_size = 1
\ No newline at end of file
+    cfg.tokenizer.tensor_model_parallel_size = 1
diff --git a/src/megatron/bridge/perf_recipes/llama/gb200/llama31.py b/src/megatron/bridge/perf_recipes/llama/gb200/llama31.py
index 664729ae19..a31adf999b 100644
--- a/src/megatron/bridge/perf_recipes/llama/gb200/llama31.py
+++ b/src/megatron/bridge/perf_recipes/llama/gb200/llama31.py
@@ -17,13 +17,194 @@
 from megatron.bridge.perf_recipes.llama.common import (
     ConfigContainer,
     _llama_benchmark_common,
+    _llama31_8b_common,
     _perf_precision,
+    llama31_8b_pretrain_config,
     llama31_405b_pretrain_config,
     userbuffers_bf16_b200_h16384_tp4_cp2_mbs1_seqlen8192,
     userbuffers_fp8_b200_h16384_tp4_cp2_mbs1_seqlen8192,
 )
 
 
+def llama31_8b_pretrain_8gpu_gb200_nvfp4_config() -> ConfigContainer:
+    """Llama3.1 8B pretrain: 8× GB200, NVFP4."""
+    cfg = llama31_8b_pretrain_config()
+    cfg.mixed_precision = _perf_precision("nvfp4")
+    cfg.mixed_precision.num_layers_at_start_in_bf16 = 0
+    cfg.mixed_precision.num_layers_at_end_in_bf16 = 0
+    cfg.mixed_precision.first_last_layers_bf16 = False
+    cfg.mixed_precision.fp4_param_gather = False
+    cfg.mixed_precision.fp4_param = False
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.context_parallel_size = 1
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.sequence_parallel = False
+    cfg.train.global_batch_size = 16
+    cfg.train.micro_batch_size = 2
+
+    cfg.model.cuda_graph_impl = "local"
+    cfg.model.cuda_graph_scope = ["full_iteration"]
+
+    _llama31_8b_common(cfg)
+    cfg.optimizer.lr = 0.0004
+    cfg.optimizer.min_lr = 4e-05
+    cfg.model.tp_comm_bootstrap_backend = 'mpi'
+    cfg.model.fp8_dot_product_attention = True
+    cfg.mixed_precision.fp8_dot_product_attention = True
+    cfg.model.use_transformer_engine_op_fuser = True
+    cfg.validation.eval_global_batch_size = 16
+    cfg.validation.eval_interval = 768
+    cfg.validation.eval_iters = 64
+    cfg.train.eval_iters = 64
+    cfg.validation.eval_micro_batch_size = 2
+    cfg.scheduler.lr_decay_iters = 1199984
+    cfg.scheduler.lr_decay_steps = 19199744
+    cfg.scheduler.lr_warmup_iters = 16
+    cfg.scheduler.lr_warmup_steps = 256
+    cfg.scheduler.wd_incr_steps = 19200000
+    cfg.load_main_params_from_ckpt = False
+    # Keep process settings next to the recipe so users can see the exact benchmark environment.
+    cfg.env_vars = {
+        **COMMON_PERF_ENV_VARS,
+        # CUDA stream scheduling for this model and parallel layout.
+        "CUDA_DEVICE_MAX_CONNECTIONS": 1,
+        # CUDA graph and allocator behavior for this recipe.
+        "NCCL_GRAPH_REGISTER": 0,
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+        "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # NCCL user-buffer and launch settings.
+        "NCCL_NVLS_ENABLE": 0,
+        # Transformer Engine overlap settings for this model.
+        "NVTE_BWD_LAYERNORM_SM_MARGIN": 16,
+        "NVTE_FWD_LAYERNORM_SM_MARGIN": 16,
+        # NVFP4 fast-math path.
+        "NVTE_USE_FAST_MATH": 0,
+    }
+    return cfg
+
+
+def llama31_8b_pretrain_72gpu_gb200_nvfp4_config() -> ConfigContainer:
+    """Llama3.1 8B pretrain: 72× GB200, NVFP4."""
+    cfg = llama31_8b_pretrain_config()
+    cfg.mixed_precision = _perf_precision("nvfp4")
+    cfg.mixed_precision.num_layers_at_start_in_bf16 = 0
+    cfg.mixed_precision.num_layers_at_end_in_bf16 = 0
+    cfg.mixed_precision.first_last_layers_bf16 = False
+    cfg.mixed_precision.fp4_param_gather = False
+    cfg.mixed_precision.fp4_param = False
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.context_parallel_size = 1
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.sequence_parallel = False
+    cfg.train.global_batch_size = 72
+    cfg.train.micro_batch_size = 1
+
+    cfg.model.cuda_graph_impl = "local"
+    cfg.model.cuda_graph_scope = ["full_iteration"]
+
+    _llama31_8b_common(cfg)
+    cfg.optimizer.lr = 0.0008
+    cfg.optimizer.min_lr = 8e-05
+    cfg.model.tp_comm_bootstrap_backend = 'mpi'
+    cfg.model.fp8_dot_product_attention = True
+    cfg.mixed_precision.fp8_dot_product_attention = True
+    cfg.model.use_transformer_engine_op_fuser = True
+    cfg.validation.eval_global_batch_size = 72
+    cfg.validation.eval_interval = 171
+    cfg.validation.eval_iters = 15
+    cfg.validation.eval_micro_batch_size = 1
+    cfg.train.eval_iters = 15
+    cfg.scheduler.lr_decay_iters = 1199936
+    cfg.scheduler.lr_decay_steps = 86395392
+    cfg.scheduler.lr_warmup_iters = 64
+    cfg.scheduler.lr_warmup_steps = 4608
+    cfg.scheduler.wd_incr_steps = 86400000
+    cfg.load_main_params_from_ckpt = False
+    # Keep process settings next to the recipe so users can see the exact benchmark environment.
+    cfg.env_vars = {
+        **COMMON_PERF_ENV_VARS,
+        # CUDA stream scheduling for this model and parallel layout.
+        "CUDA_DEVICE_MAX_CONNECTIONS": 1,
+        # CUDA graph and allocator behavior for this recipe.
+        "NCCL_GRAPH_REGISTER": 0,
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+        "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # NCCL user-buffer and launch settings.
+        "NCCL_NVLS_ENABLE": 0,
+        # Transformer Engine overlap settings for this model.
+        "NVTE_BWD_LAYERNORM_SM_MARGIN": 16,
+        "NVTE_FWD_LAYERNORM_SM_MARGIN": 16,
+        # NVFP4 fast-math path.
+        "NVTE_USE_FAST_MATH": 0,
+    }
+    return cfg
+
+
+def llama31_8b_pretrain_512gpu_gb200_fp8cs_config() -> ConfigContainer:
+    """Llama3.1 8B pretrain: 512× GB200, FP8 current-scaling."""
+    cfg = llama31_8b_pretrain_config()
+    cfg.mixed_precision = _perf_precision("fp8_cs")
+    cfg.mixed_precision.num_layers_at_start_in_bf16 = 0
+    cfg.mixed_precision.num_layers_at_end_in_bf16 = 0
+    cfg.mixed_precision.first_last_layers_bf16 = False
+    cfg.mixed_precision.fp4_param_gather = False
+    cfg.mixed_precision.fp4_param = False
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.context_parallel_size = 1
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.sequence_parallel = False
+    cfg.train.global_batch_size = 64
+    cfg.train.micro_batch_size = 1
+
+    cfg.model.cuda_graph_impl = "local"
+    cfg.model.cuda_graph_scope = ["full_iteration"]
+
+    _llama31_8b_common(cfg)
+    cfg.optimizer.lr = 0.0008
+    cfg.optimizer.min_lr = 8e-05
+    cfg.ddp.fp8_param_gather = True
+    cfg.model.tp_comm_bootstrap_backend = 'gloo'
+    cfg.model.tp_comm_overlap = True
+    cfg.validation.eval_global_batch_size = 64
+    cfg.validation.eval_interval = 192
+    cfg.validation.eval_iters = 16
+    cfg.validation.eval_micro_batch_size = 1
+    cfg.train.eval_iters = 16
+    cfg.scheduler.lr_decay_iters = 1199936
+    cfg.scheduler.lr_decay_steps = 76795904
+    cfg.scheduler.lr_warmup_iters = 64
+    cfg.scheduler.lr_warmup_steps = 4096
+    cfg.scheduler.wd_incr_steps = 76800000
+    cfg.load_main_params_from_ckpt = True
+    # Keep process settings next to the recipe so users can see the exact benchmark environment.
+    cfg.env_vars = {
+        **COMMON_PERF_ENV_VARS,
+        # CUDA stream scheduling for this model and parallel layout.
+        "CUDA_DEVICE_MAX_CONNECTIONS": 1,
+        # CUDA graph and allocator behavior for this recipe.
+        "NCCL_GRAPH_REGISTER": 0,
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+        "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # NCCL user-buffer and launch settings.
+        "NCCL_NVLS_ENABLE": 0,
+        # Transformer Engine overlap settings for this model.
+        "NVTE_BWD_LAYERNORM_SM_MARGIN": 16,
+        "NVTE_FWD_LAYERNORM_SM_MARGIN": 16,
+        # NVFP4 fast-math path.
+        "NVTE_USE_FAST_MATH": 0,
+    }
+    return cfg
+
+
 def llama31_405b_pretrain_128gpu_gb200_bf16_config() -> ConfigContainer:
     """Llama3.1 405B pretrain: 128× GB200, BF16, TP=4 PP=8 CP=2."""
     cfg = llama31_405b_pretrain_config()
diff --git a/src/megatron/bridge/perf_recipes/llama/gb300/llama31.py b/src/megatron/bridge/perf_recipes/llama/gb300/llama31.py
index e2a7e05d6a..e7b0320476 100644
--- a/src/megatron/bridge/perf_recipes/llama/gb300/llama31.py
+++ b/src/megatron/bridge/perf_recipes/llama/gb300/llama31.py
@@ -17,13 +17,194 @@
 from megatron.bridge.perf_recipes.llama.common import (
     ConfigContainer,
     _llama_benchmark_common,
+    _llama31_8b_common,
     _perf_precision,
+    llama31_8b_pretrain_config,
     llama31_405b_pretrain_config,
     userbuffers_bf16_b200_h16384_tp4_cp2_mbs1_seqlen8192,
     userbuffers_fp8_b200_h16384_tp4_cp2_mbs1_seqlen8192,
 )
 
 
+def llama31_8b_pretrain_8gpu_gb300_nvfp4_config() -> ConfigContainer:
+    """Llama3.1 8B pretrain: 8× GB200, NVFP4."""
+    cfg = llama31_8b_pretrain_config()
+    cfg.mixed_precision = _perf_precision("nvfp4")
+    cfg.mixed_precision.num_layers_at_start_in_bf16 = 0
+    cfg.mixed_precision.num_layers_at_end_in_bf16 = 0
+    cfg.mixed_precision.first_last_layers_bf16 = False
+    cfg.mixed_precision.fp4_param_gather = False
+    cfg.mixed_precision.fp4_param = False
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.context_parallel_size = 1
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.sequence_parallel = False
+    cfg.train.global_batch_size = 16
+    cfg.train.micro_batch_size = 2
+
+    cfg.model.cuda_graph_impl = "local"
+    cfg.model.cuda_graph_scope = ["full_iteration"]
+
+    _llama31_8b_common(cfg)
+    cfg.optimizer.lr = 0.0004
+    cfg.optimizer.min_lr = 4e-05
+    cfg.model.tp_comm_bootstrap_backend = 'mpi'
+    cfg.model.fp8_dot_product_attention = True
+    cfg.model.use_transformer_engine_op_fuser = True
+    cfg.mixed_precision.fp8_dot_product_attention = True
+    cfg.validation.eval_global_batch_size = 16
+    cfg.validation.eval_interval = 768
+    cfg.validation.eval_iters = 64
+    cfg.validation.eval_micro_batch_size = 2
+    cfg.train.eval_iters = 64
+    cfg.scheduler.lr_decay_iters = 1199984
+    cfg.scheduler.lr_decay_steps = 19199744
+    cfg.scheduler.lr_warmup_iters = 16
+    cfg.scheduler.lr_warmup_steps = 256
+    cfg.scheduler.wd_incr_steps = 19200000
+    cfg.load_main_params_from_ckpt = False
+    # Keep process settings next to the recipe so users can see the exact benchmark environment.
+    cfg.env_vars = {
+        **COMMON_PERF_ENV_VARS,
+        # CUDA stream scheduling for this model and parallel layout.
+        "CUDA_DEVICE_MAX_CONNECTIONS": 1,
+        # CUDA graph and allocator behavior for this recipe.
+        "NCCL_GRAPH_REGISTER": 0,
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+        "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # NCCL user-buffer and launch settings.
+        "NCCL_NVLS_ENABLE": 0,
+        # Transformer Engine overlap settings for this model.
+        "NVTE_BWD_LAYERNORM_SM_MARGIN": 16,
+        "NVTE_FWD_LAYERNORM_SM_MARGIN": 16,
+        # NVFP4 fast-math path.
+        "NVTE_USE_FAST_MATH": 0,
+    }
+    return cfg
+
+
+def llama31_8b_pretrain_72gpu_gb300_nvfp4_config() -> ConfigContainer:
+    """Llama3.1 8B pretrain: 72× GB300, NVFP4."""
+    cfg = llama31_8b_pretrain_config()
+    cfg.mixed_precision = _perf_precision("nvfp4")
+    cfg.mixed_precision.num_layers_at_start_in_bf16 = 0
+    cfg.mixed_precision.num_layers_at_end_in_bf16 = 0
+    cfg.mixed_precision.first_last_layers_bf16 = False
+    cfg.mixed_precision.fp4_param_gather = False
+    cfg.mixed_precision.fp4_param = False
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.context_parallel_size = 1
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.sequence_parallel = False
+    cfg.train.global_batch_size = 72
+    cfg.train.micro_batch_size = 1
+
+    cfg.model.cuda_graph_impl = "local"
+    cfg.model.cuda_graph_scope = ["full_iteration"]
+
+    _llama31_8b_common(cfg)
+    cfg.optimizer.lr = 0.0008
+    cfg.optimizer.min_lr = 8e-05
+    cfg.model.tp_comm_bootstrap_backend = 'mpi'
+    cfg.model.fp8_dot_product_attention = True
+    cfg.mixed_precision.fp8_dot_product_attention = True
+    cfg.model.use_transformer_engine_op_fuser = True
+    cfg.validation.eval_global_batch_size = 72
+    cfg.validation.eval_interval = 171
+    cfg.validation.eval_iters = 15
+    cfg.validation.eval_micro_batch_size = 1
+    cfg.train.eval_iters = 15
+    cfg.scheduler.lr_decay_iters = 1199936
+    cfg.scheduler.lr_decay_steps = 86395392
+    cfg.scheduler.lr_warmup_iters = 64
+    cfg.scheduler.lr_warmup_steps = 4608
+    cfg.scheduler.wd_incr_steps = 86400000
+    cfg.load_main_params_from_ckpt = False
+    # Keep process settings next to the recipe so users can see the exact benchmark environment.
+    cfg.env_vars = {
+        **COMMON_PERF_ENV_VARS,
+        # CUDA stream scheduling for this model and parallel layout.
+        "CUDA_DEVICE_MAX_CONNECTIONS": 1,
+        # CUDA graph and allocator behavior for this recipe.
+        "NCCL_GRAPH_REGISTER": 0,
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+        "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # NCCL user-buffer and launch settings.
+        "NCCL_NVLS_ENABLE": 0,
+        # Transformer Engine overlap settings for this model.
+        "NVTE_BWD_LAYERNORM_SM_MARGIN": 16,
+        "NVTE_FWD_LAYERNORM_SM_MARGIN": 16,
+        # NVFP4 fast-math path.
+        "NVTE_USE_FAST_MATH": 0,
+    }
+    return cfg
+
+
+def llama31_8b_pretrain_512gpu_gb300_fp8cs_config() -> ConfigContainer:
+    """Llama3.1 8B pretrain: 512× GB300, FP8 current-scaling."""
+    cfg = llama31_8b_pretrain_config()
+    cfg.mixed_precision = _perf_precision("fp8_cs")
+    cfg.mixed_precision.num_layers_at_start_in_bf16 = 0
+    cfg.mixed_precision.num_layers_at_end_in_bf16 = 0
+    cfg.mixed_precision.first_last_layers_bf16 = False
+    cfg.mixed_precision.fp4_param_gather = False
+    cfg.mixed_precision.fp4_param = False
+    cfg.mixed_precision.grad_reduce_in_fp32 = False
+
+    cfg.model.tensor_model_parallel_size = 1
+    cfg.model.pipeline_model_parallel_size = 1
+    cfg.model.context_parallel_size = 1
+    cfg.model.virtual_pipeline_model_parallel_size = None
+    cfg.model.sequence_parallel = False
+    cfg.train.global_batch_size = 64
+    cfg.train.micro_batch_size = 1
+
+    cfg.model.cuda_graph_impl = "local"
+    cfg.model.cuda_graph_scope = ["full_iteration"]
+
+    _llama31_8b_common(cfg)
+    cfg.optimizer.lr = 0.0008
+    cfg.optimizer.min_lr = 8e-05
+    cfg.ddp.fp8_param_gather = True
+    cfg.model.tp_comm_bootstrap_backend = 'gloo'
+    cfg.model.tp_comm_overlap = True
+    cfg.validation.eval_global_batch_size = 64
+    cfg.validation.eval_interval = 192
+    cfg.validation.eval_iters = 16
+    cfg.validation.eval_micro_batch_size = 1
+    cfg.train.eval_iters = 16
+    cfg.scheduler.lr_decay_iters = 1199936
+    cfg.scheduler.lr_decay_steps = 76795904
+    cfg.scheduler.lr_warmup_iters = 64
+    cfg.scheduler.lr_warmup_steps = 4096
+    cfg.scheduler.wd_incr_steps = 76800000
+    cfg.load_main_params_from_ckpt = True
+    # Keep process settings next to the recipe so users can see the exact benchmark environment.
+    cfg.env_vars = {
+        **COMMON_PERF_ENV_VARS,
+        # CUDA stream scheduling for this model and parallel layout.
+        "CUDA_DEVICE_MAX_CONNECTIONS": 1,
+        # CUDA graph and allocator behavior for this recipe.
+        "NCCL_GRAPH_REGISTER": 0,
+        "PYTORCH_CUDA_ALLOC_CONF": "expandable_segments:True",
+        "TORCH_NCCL_AVOID_RECORD_STREAMS": 1,
+        # NCCL user-buffer and launch settings.
+        "NCCL_NVLS_ENABLE": 0,
+        # Transformer Engine overlap settings for this model.
+        "NVTE_BWD_LAYERNORM_SM_MARGIN": 16,
+        "NVTE_FWD_LAYERNORM_SM_MARGIN": 16,
+        # NVFP4 fast-math path.
+        "NVTE_USE_FAST_MATH": 0,
+    }
+    return cfg
+
+
 def llama31_405b_pretrain_128gpu_gb300_bf16_config() -> ConfigContainer:
     """Llama3.1 405B pretrain: 128× GB300, BF16, FSDP."""
     cfg = llama31_405b_pretrain_config()
diff --git a/src/megatron/bridge/recipes/llama/h100/llama3.py b/src/megatron/bridge/recipes/llama/h100/llama3.py
index e9c31b16a5..0636e7ea01 100644
--- a/src/megatron/bridge/recipes/llama/h100/llama3.py
+++ b/src/megatron/bridge/recipes/llama/h100/llama3.py
@@ -1060,7 +1060,6 @@ def llama31_8b_pretrain_2gpu_h100_bf16_config() -> ConfigContainer:
     cfg.model.cuda_graph_scope = "full"
     cfg.model.cuda_graph_warmup_steps = 3
 
-    cfg.model.attention_backend = None
     cfg.model.cross_entropy_loss_fusion = True
     cfg.model.cross_entropy_fusion_impl = "te"