refactor(recipes): add h100 recipes and remove perf configs

Signed-off-by: yaoyu-33 <yaoyu.094@gmail.com>

Author: yaoyu-33

README.md

@@ -13,7 +13,7 @@
 ## 📣 News
 - [06/22/2026] **Megatron Bridge 0.5.0 released!** Highlights include expanded LLM and multimodal support (Qwen3.5, DeepSeek V4, Ernie 4.5, GLM-5/4.7, StepFun Step-3.5/3.7, MiMo-V2, Gemma 4, Falcon H1, Ling MoE V2, Nemotron-3 Nano Omni, Qwen3-Omni, Qwen3-ASR, and Nemotron Diffusion), MegatronMIMO and Energon v7 training updates, evaluator backend integration, eval-time context parallelism, deterministic recipes, quantized FP8/MXFP4 export, CUDA graph/performance improvements, and Megatron Inference/tokenizer unification with Megatron-LM. Huge thanks to our community contributors: [@HowardZorn](https://github.com/HowardZorn), [@hy2826](https://github.com/hy2826), [@bo-ke](https://github.com/bo-ke), [@beccohov](https://github.com/beccohov), [@dhiaEddineRhaiem](https://github.com/dhiaEddineRhaiem), [@pavelgein](https://github.com/pavelgein), [@ccclyu](https://github.com/ccclyu), [@hbhflw2000](https://github.com/hbhflw2000), and [@HollowMan6](https://github.com/HollowMan6)! See the [full release notes](https://github.com/NVIDIA-NeMo/Megatron-Bridge/releases/tag/v0.5.0).
 
-- [06/16/2026] NVIDIA topped [MLPerf Training v6.0](https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/) across every benchmark, including the new DeepSeek-V3 and GPT-OSS MoE training workloads. Megatron Bridge serves as the packaging layer for the NeMo 26.06 training stack that integrates full-iteration CUDA graphs, HybridEP/router optimizations, all-to-all overlap, MXFP8 attention, and pipeline-layout balancing; the blog highlights [DeepSeek-V3 training](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/scripts/performance/configs/deepseek) at **1,648 TFLOPS/GPU** (**6,338 tokens/sec/GPU**) on GB300. To try related runs, start from the [performance recipes](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/scripts/performance); the corresponding container is expected with the NeMo 26.06 release soon.
+- [06/16/2026] NVIDIA topped [MLPerf Training v6.0](https://developer.nvidia.com/blog/nvidia-blackwell-tops-mlperf-training-6-0-with-industry-leading-scale-and-performance/) across every benchmark, including the new DeepSeek-V3 and GPT-OSS MoE training workloads. Megatron Bridge serves as the packaging layer for the NeMo 26.06 training stack that integrates full-iteration CUDA graphs, HybridEP/router optimizations, all-to-all overlap, MXFP8 attention, and pipeline-layout balancing; the blog highlights [DeepSeek-V3 training](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/src/megatron/bridge/perf_recipes/deepseek) at **1,648 TFLOPS/GPU** (**6,338 tokens/sec/GPU**) on GB300. To try related runs, start from the [performance recipes](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/main/scripts/performance); the corresponding container is expected with the NeMo 26.06 release soon.
 
 - [06/04/2026] [**NVIDIA Nemotron 3 Ultra**](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16) is now public! Day-0 support for the 550B-A55B hybrid Mamba-Transformer MoE model is available on the [`nemotron_3_ultra`](https://github.com/NVIDIA-NeMo/Megatron-Bridge/tree/nemotron_3_ultra) branch, including checkpoint conversion, inference, SFT, PEFT (LoRA), and pretraining examples. Read the [NVIDIA Technical Blog](https://developer.nvidia.com/blog/nvidia-nemotron-3-ultra-powers-faster-more-efficient-reasoning-for-long-running-agents/) and see the [examples README](https://github.com/NVIDIA-NeMo/Megatron-Bridge/blob/nemotron_3_ultra/examples/models/nemotron/nemotron_3/ultra/README.md) for the full walkthrough.
 
@@ -141,14 +141,14 @@ Training quickstart using pre-configured recipes:
 
 ```python
 from megatron.bridge import AutoBridge
-from megatron.bridge.recipes.llama import llama32_1b_pretrain_config
+from megatron.bridge.recipes.llama.h100 import llama32_1b_pretrain_1gpu_h100_bf16_config
 from megatron.bridge.training.gpt_step import forward_step
 from megatron.bridge.training.pretrain import pretrain
 
 if __name__ == "__main__":
     # The recipe uses the Llama 3.2 1B architecture from Hugging Face.
     # This is random-init pretraining and does not require a converted Megatron checkpoint.
-    cfg = llama32_1b_pretrain_config()
+    cfg = llama32_1b_pretrain_1gpu_h100_bf16_config()
 
     # The recipe already sets cfg.model internally using this pattern.
     # Override cfg.model to choose a different Hugging Face model ID as the architecture source.

docs/fern/versions/nightly/pages/index.mdx

@@ -132,14 +132,14 @@ Training quickstart using pre-configured recipes:
 
 ```python
 from megatron.bridge import AutoBridge
-from megatron.bridge.recipes.llama import llama32_1b_pretrain_config
+from megatron.bridge.recipes.llama.h100 import llama32_1b_pretrain_1gpu_h100_bf16_config
 from megatron.bridge.training.gpt_step import forward_step
 from megatron.bridge.training.pretrain import pretrain
 
 if __name__ == "__main__":
     # The recipe uses the Llama 3.2 1B architecture from Hugging Face.
     # This is random-init pretraining and does not require a converted Megatron checkpoint.
-    cfg = llama32_1b_pretrain_config()
+    cfg = llama32_1b_pretrain_1gpu_h100_bf16_config()
 
     # The recipe already sets cfg.model internally using this pattern.
     # Override cfg.model to choose a different Hugging Face model ID as the architecture source.

docs/fern/versions/nightly/pages/megatron-lm-to-megatron-bridge.mdx

@@ -30,11 +30,11 @@ uv run python scripts/translate_mlm_to_bridge.py \
 ```bash
 # From a Bridge recipe name (defaults exported as MLM args)
 uv run python scripts/translate_mlm_to_bridge.py --reverse \
-    --recipe llama32_1b_pretrain_config
+    --recipe llama32_1b_pretrain_1gpu_h100_bf16_config
 
 # From a recipe plus inline overrides
 uv run python scripts/translate_mlm_to_bridge.py --reverse \
-    --recipe llama32_1b_pretrain_config \
+    --recipe llama32_1b_pretrain_1gpu_h100_bf16_config \
     --args "train.train_iters=1000 model.tensor_model_parallel_size=2"
 
 # From Bridge overrides only (no recipe)
@@ -99,7 +99,7 @@ Run the generic recipe launcher and override config keys directly:
 
 ```bash
 uv run python scripts/training/run_recipe.py \
-  --recipe llama3_8b_pretrain_config \
+  --recipe llama3_8b_pretrain_2gpu_h100_bf16_config \
   --dataset llm-pretrain \
   train.micro_batch_size=2 \
   train.global_batch_size=128 \

docs/fern/versions/nightly/pages/models/gemma/gemma2.mdx

@@ -166,7 +166,7 @@ config = gemma2_27b_pretrain_config(
 
 #### Gemma 2 2B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-2b \
 --recipe gemma2_2b_sft_config \
 train.global_batch_size=64 \
@@ -188,7 +188,7 @@ config = gemma2_2b_sft_config(
 
 #### Gemma 2 9B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-9b \
 --recipe gemma2_9b_sft_config \
 train.global_batch_size=64 \
@@ -198,7 +198,7 @@ checkpoint.save=$SAVE_DIR/gemma2_9b_finetune
 
 #### Gemma 2 27B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=16 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=16 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-27b \
 --recipe gemma2_27b_sft_config \
 train.global_batch_size=64 \
@@ -210,7 +210,7 @@ checkpoint.save=$SAVE_DIR/gemma2_27b_finetune
 
 #### Gemma 2 2B
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/gemma-2-2b \
 --recipe gemma2_2b_peft_config \
 --peft_scheme lora \

docs/fern/versions/nightly/pages/models/gemma/gemma3.mdx

@@ -155,7 +155,7 @@ config = gemma3_1b_peft_config(
 
 **Full Finetuning:**
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
   --pretrained-checkpoint /models/gemma-3-1b-it \
   --recipe gemma3_1b_sft_config \
   train.global_batch_size=64 \
@@ -165,7 +165,7 @@ uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
 
 **LoRA Finetuning:**
 ```bash
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
   --pretrained-checkpoint /models/gemma-3-1b-it \
   --recipe gemma3_1b_peft_config \
   --peft_scheme lora \

docs/fern/versions/nightly/pages/models/glm/glm45.mdx

@@ -214,7 +214,7 @@ config = glm45_355b_peft_config(
 
 ```bash
 # GLM 4.5 Air - LoRA finetuning on single node (8 GPUs)
-uv run python -m torch.distributed.run --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/glm45-air-106b \
 --recipe glm45_air_106b_peft_config \
 --peft_scheme lora \
@@ -223,7 +223,7 @@ train.train_iters=1000 \
 checkpoint.save=$SAVE_DIR/glm45_air_lora
 
 # GLM 4.5 355B - Full finetuning (256 GPUs)
-uv run python -m torch.distributed.run --nnodes=32 --nproc-per-node=8 run/run_recipe.py \
+uv run python -m torch.distributed.run --nnodes=32 --nproc-per-node=8 scripts/training/run_recipe.py \
 --pretrained-checkpoint /models/glm45-355b \
 --recipe glm45_355b_sft_config \
 train.global_batch_size=256 \

docs/fern/versions/nightly/pages/models/llama/llama3.mdx

@@ -86,7 +86,7 @@ See: [bridge.recipes.llama.llama3](../../apidocs/bridge/bridge.recipes.llama.lla
   - `llama32_1b_pretrain_config`, `llama32_3b_pretrain_config`: Llama 3.2 (1B, 3B)
   - `llama3_8b_pretrain_config`: Llama 3 8B with 8K context
   - `llama3_8b_16k_pretrain_config`, `llama3_8b_64k_pretrain_config`, `llama3_8b_128k_pretrain_config`: Llama 3 8B with extended context (16K/64K/128K)
-  - `llama3_8b_low_precision_pretrain_config`: Llama 3 8B with low precision (FP8/MXFP8/NVFP4)
+  - `llama3_8b_pretrain_2gpu_h100_fp8cs_config`, `llama3_8b_pretrain_2gpu_h100_fp8mx_config`, `llama3_8b_pretrain_2gpu_h100_nvfp4_config`: Llama 3 8B H100 low precision (FP8/MXFP8/NVFP4)
   - `llama3_70b_pretrain_config`, `llama3_70b_16k_pretrain_config`, `llama3_70b_64k_pretrain_config`: Llama 3 70B (8K/16K/64K context)
   - `llama31_8b_pretrain_config`, `llama31_70b_pretrain_config`, `llama31_405b_pretrain_config`: Llama 3.1 (8B/70B/405B, 128K context)
 

docs/fern/versions/nightly/pages/recipe-usage.mdx

@@ -12,7 +12,7 @@ This guide will cover the next steps to make use of a training recipe, including
 
 ## Choosing a recipe or a new config
 
-Start from an exported recipe when the model family and workflow already exist in `megatron.bridge.recipes`. Recipe functions such as `llama3_8b_pretrain_config`, `llama32_1b_sft_config`, and `qwen3_8b_peft_config` provide model, optimizer, scheduler, precision, dataset, logger, and checkpoint defaults in one `ConfigContainer`. Override those defaults for your dataset, checkpoint paths, run length, parallelism, or precision before creating a new recipe.
+Start from an exported recipe when the model family and workflow already exist in `megatron.bridge.recipes`. Hardware-specific H100 recipes such as `llama3_8b_pretrain_2gpu_h100_bf16_config`, `llama32_1b_sft_1gpu_h100_bf16_config`, and `qwen3_8b_lora_1gpu_h100_bf16_config` are the canonical recipe names and share the same `<model>_<task>_<num>gpu_<gpu>_<dtype>_config` naming strategy as performance recipes. The older recipe names remain available as compatibility wrappers. Override those defaults for your dataset, checkpoint paths, run length, parallelism, or precision before creating a new recipe.
 
 Create a new recipe or config when the base model architecture is not represented by an existing model provider, the checkpoint conversion needs a new bridge, the forward step or dataset provider is model-specific, or you need a reusable configuration that will be shared across jobs. If the Hugging Face model is already supported by `AutoBridge`, you usually only need to start from the closest recipe and override the model provider or `hf_path`.
 

docs/fern/versions/nightly/pages/training/data-preparation.mdx

@@ -41,8 +41,8 @@ dataset = GPTDatasetConfig(
 The CLI-friendly `data_path` field is converted to Megatron Core's `blend` field during config finalization. For weighted multi-dataset training, use either a flattened `data_path` list with weights and prefixes or set `blend`/`blend_per_split` directly.
 
 ```bash
-uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-    --recipe llama32_1b_pretrain_config \
+uv run python -m torch.distributed.run --nproc_per_node=1 scripts/training/run_recipe.py \
+    --recipe llama32_1b_pretrain_1gpu_h100_bf16_config \
     --dataset llm-pretrain \
     dataset.data_path=/data/dclm/preprocessed_text_document \
     dataset.seq_length=8192
@@ -81,8 +81,8 @@ dataset = FinetuningDatasetConfig(
 Launch the generic recipe runner with the preloaded local JSONL dataset type:
 
 ```bash
-uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-    --recipe llama32_1b_sft_config \
+uv run python -m torch.distributed.run --nproc_per_node=1 scripts/training/run_recipe.py \
+    --recipe llama32_1b_sft_1gpu_h100_bf16_config \
     --dataset llm-finetune-preloaded \
     dataset.dataset_root=/data/sft_jsonl \
     dataset.seq_length=4096 \
@@ -114,8 +114,8 @@ If `dataset_root` is omitted, the processed JSONL is cached under the NeMo datas
 The generic launcher provides preset Hugging Face text datasets through `--dataset llm-finetune`:
 
 ```bash
-uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-    --recipe llama32_1b_peft_config \
+uv run python -m torch.distributed.run --nproc_per_node=1 scripts/training/run_recipe.py \
+    --recipe llama32_1b_lora_1gpu_h100_bf16_config \
     --dataset llm-finetune \
     dataset.dataset_name=gsm8k \
     checkpoint.pretrained_checkpoint=/checkpoints/base_model
@@ -128,8 +128,8 @@ VLM recipes usually use a dataset provider instead of `FinetuningDatasetConfig`.
 For Energon/WebDataset data, create tar shards plus `.nv-meta` metadata and pass the dataset root to the recipe provider:
 
 ```bash
-uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-    --recipe qwen3_vl_8b_peft_energon_config \
+uv run python -m torch.distributed.run --nproc_per_node=1 scripts/training/run_recipe.py \
+    --recipe qwen3_vl_8b_lora_1gpu_h100_bf16_energon_config \
     --dataset vlm-energon \
     --step_func qwen3_vl_step \
     dataset.path=/data/vlm_energon \
@@ -143,8 +143,8 @@ For preloaded VLM JSON or JSONL, use records with `messages` or `conversations`
 ```
 
 ```bash
-uv run python -m torch.distributed.run --nproc_per_node=8 scripts/training/run_recipe.py \
-    --recipe qwen3_vl_8b_peft_config \
+uv run python -m torch.distributed.run --nproc_per_node=1 scripts/training/run_recipe.py \
+    --recipe qwen3_vl_8b_lora_1gpu_h100_bf16_config \
     --dataset vlm-preloaded \
     --step_func qwen3_vl_step \
     dataset.train_data_path=/data/vlm/train.jsonl \

docs/fern/versions/nightly/pages/training/mixed-precision.mdx

@@ -129,12 +129,12 @@ cfg.mixed_precision = "bf16_with_nvfp4_mixed"
 cfg.checkpoint.pretrained_checkpoint = "/checkpoints/base_model"
 ```
 
-Some model families expose tuned NVFP4 recipes. For example, Llama 3 8B has a low-precision pretraining helper:
+Some model families expose tuned NVFP4 recipes. For example, Llama 3 8B has a hardware-specific low-precision pretraining recipe:
 
 ```python
-from megatron.bridge.recipes.llama import llama3_8b_low_precision_pretrain_config
+from megatron.bridge.recipes.llama.h100 import llama3_8b_pretrain_2gpu_h100_nvfp4_config
 
-cfg = llama3_8b_low_precision_pretrain_config("bf16_with_nvfp4_mixed")
+cfg = llama3_8b_pretrain_2gpu_h100_nvfp4_config()
 ```
 
 NVFP4 is related to FP8 in that both use Transformer Engine low-precision kernels and BF16 fallback for non-quantized state, but the configuration knobs are separate. Use FP8 recipes on Hopper or Blackwell when FP8 is the target precision; use NVFP4 recipes on Blackwell when the model and hardware have been validated for 4-bit training.