fix(perf): remove Nemotron 3.5 Lightning B-series FSDP recipes

Signed-off-by: Raghav Hrishikeshan Mukundan <rmukundan@nvidia.com>

Author: rhmukundan

scripts/performance/dump_perf_configs.py

@@ -286,11 +286,9 @@
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "gb200", "nvfp4"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b300", "bf16"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b300", "fp8_mx"),
-    ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b300", "fp8_mx", "fsdp"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b300", "nvfp4"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b200", "bf16"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b200", "fp8_mx"),
-    ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b200", "fp8_mx", "fsdp"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 8, "b200", "nvfp4"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 16, "h100", "bf16"),
     ("nemotronh", "nemotron_3_5_lightning", "pretrain", 16, "h100", "fp8_cs"),

src/megatron/bridge/perf_recipes/nemotronh/__init__.py

@@ -1,7 +1,6 @@
 from megatron.bridge.perf_recipes.nemotronh.b200.nemotronh import (
     nemotron_3_5_lightning_pretrain_8gpu_b200_bf16_config,
     nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_config,
-    nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_fsdp_config,
     nemotron_3_5_lightning_pretrain_8gpu_b200_nvfp4_config,
     nemotron_3_nano_pretrain_8gpu_b200_bf16_config,
     nemotron_3_nano_pretrain_8gpu_b200_fp8mx_config,
@@ -16,7 +15,6 @@
 from megatron.bridge.perf_recipes.nemotronh.b300.nemotronh import (
     nemotron_3_5_lightning_pretrain_8gpu_b300_bf16_config,
     nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_config,
-    nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_fsdp_config,
     nemotron_3_5_lightning_pretrain_8gpu_b300_nvfp4_config,
     nemotron_3_nano_pretrain_8gpu_b300_bf16_config,
     nemotron_3_nano_pretrain_8gpu_b300_fp8mx_config,

src/megatron/bridge/perf_recipes/nemotronh/b200/nemotronh.py

@@ -391,41 +391,9 @@ def nemotron_3_5_lightning_pretrain_8gpu_b200_bf16_config() -> ConfigContainer:
     return _apply_nemotron_3_5_lightning_defaults(nemotron_3_nano_pretrain_8gpu_b200_bf16_config())
 
 
-def _build_nemotron_3_5_lightning_b200_mxfp8() -> ConfigContainer:
-    return _apply_nemotron_3_5_lightning_defaults(nemotron_3_nano_pretrain_8gpu_b200_fp8mx_config())
-
-
 def nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_config() -> ConfigContainer:
     """Nemotron 3.5 Lightning pretrain: 8× B200, MXFP8."""
-    return _build_nemotron_3_5_lightning_b200_mxfp8()
-
-
-def nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_fsdp_config() -> ConfigContainer:
-    """Nemotron 3.5 Lightning pretrain: 8× B200, MXFP8, Megatron FSDP."""
-    cfg = _build_nemotron_3_5_lightning_b200_mxfp8()
-
-    # Megatron FSDP registers module hooks that Transformer Engine CUDA graph
-    # capture rejects.
-    cfg.train.global_batch_size = 384
-    cfg.train.micro_batch_size = 3
-    cfg.model.cuda_graph_impl = "none"
-    set_cuda_graph_modules(cfg.model, [])
-
-    cfg.model.init_model_with_meta_device = True
-    cfg.mixed_precision.reuse_grad_buf_for_mxfp8_param_ag = False
-    cfg.dist.use_megatron_fsdp = True
-    cfg.ddp.use_megatron_fsdp = True
-    cfg.ddp.num_distributed_optimizer_instances = 1
-    cfg.ddp.data_parallel_sharding_strategy = "optim_grads_params"
-    cfg.ddp.outer_dp_sharding_strategy = "no_shard"
-    cfg.ddp.average_in_collective = False
-    cfg.ddp.keep_fp8_transpose_cache = False
-    cfg.ddp.reuse_grad_buf_for_mxfp8_param_ag = False
-    cfg.optimizer.reuse_grad_buf_for_mxfp8_param_ag = False
-
-    cfg.checkpoint.load = None
-    cfg.checkpoint.ckpt_format = "fsdp_dtensor"
-    return cfg
+    return _apply_nemotron_3_5_lightning_defaults(nemotron_3_nano_pretrain_8gpu_b200_fp8mx_config())
 
 
 def nemotron_3_5_lightning_pretrain_8gpu_b200_nvfp4_config() -> ConfigContainer:

src/megatron/bridge/perf_recipes/nemotronh/b300/nemotronh.py

@@ -405,41 +405,9 @@ def nemotron_3_5_lightning_pretrain_8gpu_b300_bf16_config() -> ConfigContainer:
     return _apply_nemotron_3_5_lightning_defaults(nemotron_3_nano_pretrain_8gpu_b300_bf16_config())
 
 
-def _build_nemotron_3_5_lightning_b300_mxfp8() -> ConfigContainer:
-    return _apply_nemotron_3_5_lightning_defaults(nemotron_3_nano_pretrain_8gpu_b300_fp8mx_config())
-
-
 def nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_config() -> ConfigContainer:
     """Nemotron 3.5 Lightning pretrain: 8× B300, MXFP8."""
-    return _build_nemotron_3_5_lightning_b300_mxfp8()
-
-
-def nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_fsdp_config() -> ConfigContainer:
-    """Nemotron 3.5 Lightning pretrain: 8× B300, MXFP8, Megatron FSDP."""
-    cfg = _build_nemotron_3_5_lightning_b300_mxfp8()
-
-    # Megatron FSDP registers module hooks that Transformer Engine CUDA graph
-    # capture rejects.
-    cfg.train.global_batch_size = 384
-    cfg.train.micro_batch_size = 3
-    cfg.model.cuda_graph_impl = "none"
-    set_cuda_graph_modules(cfg.model, [])
-
-    cfg.model.init_model_with_meta_device = True
-    cfg.mixed_precision.reuse_grad_buf_for_mxfp8_param_ag = False
-    cfg.dist.use_megatron_fsdp = True
-    cfg.ddp.use_megatron_fsdp = True
-    cfg.ddp.num_distributed_optimizer_instances = 1
-    cfg.ddp.data_parallel_sharding_strategy = "optim_grads_params"
-    cfg.ddp.outer_dp_sharding_strategy = "no_shard"
-    cfg.ddp.average_in_collective = False
-    cfg.ddp.keep_fp8_transpose_cache = False
-    cfg.ddp.reuse_grad_buf_for_mxfp8_param_ag = False
-    cfg.optimizer.reuse_grad_buf_for_mxfp8_param_ag = False
-
-    cfg.checkpoint.load = None
-    cfg.checkpoint.ckpt_format = "fsdp_dtensor"
-    return cfg
+    return _apply_nemotron_3_5_lightning_defaults(nemotron_3_nano_pretrain_8gpu_b300_fp8mx_config())
 
 
 def nemotron_3_5_lightning_pretrain_8gpu_b300_nvfp4_config() -> ConfigContainer:

tests/unit_tests/recipes/test_nemotron_3_5_lightning_perf_recipes.py

@@ -24,11 +24,9 @@
 from megatron.bridge.perf_recipes.nemotronh import (
     nemotron_3_5_lightning_pretrain_8gpu_b200_bf16_config,
     nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_config,
-    nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_fsdp_config,
     nemotron_3_5_lightning_pretrain_8gpu_b200_nvfp4_config,
     nemotron_3_5_lightning_pretrain_8gpu_b300_bf16_config,
     nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_config,
-    nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_fsdp_config,
     nemotron_3_5_lightning_pretrain_8gpu_b300_nvfp4_config,
     nemotron_3_5_lightning_pretrain_8gpu_gb200_bf16_config,
     nemotron_3_5_lightning_pretrain_8gpu_gb200_fp8mx_config,
@@ -77,10 +75,6 @@
     nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_config,
     nemotron_3_5_lightning_pretrain_8gpu_b300_nvfp4_config,
 )
-_B_FSDP_RECIPES = (
-    nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_fsdp_config,
-    nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_fsdp_config,
-)
 _GB200_RECIPES = (
     nemotron_3_5_lightning_pretrain_8gpu_gb200_bf16_config,
     nemotron_3_5_lightning_pretrain_8gpu_gb200_fp8mx_config,
@@ -95,7 +89,6 @@
     nemotron_3_5_lightning_pretrain_8gpu_gb200_fp8mx_fsdp_config,
     nemotron_3_5_lightning_pretrain_8gpu_gb300_fp8mx_fsdp_config,
 )
-_ALL_FSDP_RECIPES = (*_B_FSDP_RECIPES, *_GB_FSDP_RECIPES)
 _NEMOTRON_3_RECIPES = (
     nemotron_3_nano_pretrain_16gpu_h100_bf16_config,
     nemotron_3_nano_pretrain_16gpu_h100_fp8cs_config,
@@ -204,10 +197,6 @@
         "megatron.bridge.perf_recipes.nemotronh.b200.nemotronh",
         "nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_config",
     ),
-    (
-        "megatron.bridge.perf_recipes.nemotronh.b200.nemotronh",
-        "nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_fsdp_config",
-    ),
     (
         "megatron.bridge.perf_recipes.nemotronh.b200.nemotronh",
         "nemotron_3_5_lightning_pretrain_8gpu_b200_nvfp4_config",
@@ -220,10 +209,6 @@
         "megatron.bridge.perf_recipes.nemotronh.b300.nemotronh",
         "nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_config",
     ),
-    (
-        "megatron.bridge.perf_recipes.nemotronh.b300.nemotronh",
-        "nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_fsdp_config",
-    ),
     (
         "megatron.bridge.perf_recipes.nemotronh.b300.nemotronh",
         "nemotron_3_5_lightning_pretrain_8gpu_b300_nvfp4_config",
@@ -285,7 +270,7 @@ def test_standard_perf_recipes_do_not_expose_mtp_flag(recipe_factory: Callable[[
 
 @pytest.mark.parametrize(
     "recipe_factory",
-    (*_H100_RECIPES, *_B200_RECIPES, *_B300_RECIPES, *_GB200_RECIPES, *_GB300_RECIPES, *_ALL_FSDP_RECIPES),
+    (*_H100_RECIPES, *_B200_RECIPES, *_B300_RECIPES, *_GB200_RECIPES, *_GB300_RECIPES, *_GB_FSDP_RECIPES),
     ids=lambda recipe: recipe.__name__,
 )
 def test_perf_recipes_enable_mtp(recipe_factory: Callable[[], ConfigContainer]) -> None:
@@ -513,9 +498,9 @@ def test_gb300_perf_recipe_topology(recipe_factory: Callable[[], ConfigContainer
     assert cfg.env_vars["USE_MNNVL"] == 1
 
 
-@pytest.mark.parametrize("recipe_factory", _ALL_FSDP_RECIPES, ids=lambda recipe: recipe.__name__)
-def test_mxfp8_fsdp_perf_recipe_defaults(recipe_factory: Callable[[], ConfigContainer]) -> None:
-    """The Lightning FSDP variants retain their measured 8-GPU performance settings."""
+@pytest.mark.parametrize("recipe_factory", _GB_FSDP_RECIPES, ids=lambda recipe: recipe.__name__)
+def test_gb_fsdp_perf_recipe_defaults(recipe_factory: Callable[[], ConfigContainer]) -> None:
+    """The GB FSDP variants retain their measured 8-GPU performance settings."""
     cfg = recipe_factory()
 
     assert cfg.train.global_batch_size == 384
@@ -537,23 +522,3 @@ def test_mxfp8_fsdp_perf_recipe_defaults(recipe_factory: Callable[[], ConfigCont
     assert cfg.optimizer.reuse_grad_buf_for_mxfp8_param_ag is False
     assert cfg.checkpoint.load is None
     assert cfg.checkpoint.ckpt_format == "fsdp_dtensor"
-
-
-@pytest.mark.parametrize(
-    ("recipe_factory", "expect_b300_affinity"),
-    (
-        (nemotron_3_5_lightning_pretrain_8gpu_b200_fp8mx_fsdp_config, False),
-        (nemotron_3_5_lightning_pretrain_8gpu_b300_fp8mx_fsdp_config, True),
-    ),
-    ids=lambda value: value.__name__ if callable(value) else str(value),
-)
-def test_b200_b300_mxfp8_fsdp_topology(
-    recipe_factory: Callable[[], ConfigContainer], expect_b300_affinity: bool
-) -> None:
-    """B200 and B300 FSDP variants preserve their NVL8 topology settings."""
-    cfg = recipe_factory()
-
-    assert cfg.model.expert_model_parallel_size == 8
-    assert cfg.env_vars["NVLINK_DOMAIN_SIZE"] == 8
-    assert cfg.env_vars["USE_MNNVL"] == 0
-    assert ("NCCL_IGNORE_CPU_AFFINITY" in cfg.env_vars) is expect_b300_affinity