Interface to enable fsdp-double-buffer without enabling NCCL-UB

scripts/performance/argument_parser.py

@@ -300,6 +300,14 @@ def bool_arg(arg):
         required=False,
         default=None,
     )
+    parser.add_argument(
+        "-fsdp_db",
+        "--use_fsdp_double_buffer",
+        help="Enable FSDP double buffer. Disabled by default",
+        type=bool_arg,
+        required=False,
+        default=None,
+    )
     parser.add_argument(
         "-ubr",
         "--use_user_buffer_registration",

scripts/performance/diffusion/pretrain_flux_12b.py

@@ -59,6 +59,9 @@ def override_recipe_configs(
         enable_cuda_graphs=enable_cuda_graphs,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
     )
     recipe = set_exp_logging_configs(
         recipe,

scripts/performance/helpers.py

@@ -95,6 +95,13 @@ def get_user_configs(gpu: str, task: str, model_name: str, model_size: str, args
     use_mcore_fsdp = config.get("use_mcore_fsdp") if args.use_mcore_fsdp is None else args.use_mcore_fsdp
     use_mcore_fsdp = False if use_mcore_fsdp is None else bool(int(use_mcore_fsdp))
 
+    use_fsdp_double_buffer = (
+        config.get("use_fsdp_double_buffer") if args.use_fsdp_double_buffer is None else args.use_fsdp_double_buffer
+    )
+    use_fsdp_double_buffer = False if use_fsdp_double_buffer is None else bool(int(use_fsdp_double_buffer))
+    if use_fsdp_double_buffer:
+        assert use_mcore_fsdp == True, "use_fsdp_double_buffer requires use_mcore_fsdp to be True"
+
     recompute_layers = config.get("recompute_layers") if args.recompute_layers is None else args.recompute_layers
     recompute_layers = 0 if recompute_layers is None else int(recompute_layers)
     activation_offload_layers = (
@@ -130,6 +137,7 @@ def get_user_configs(gpu: str, task: str, model_name: str, model_size: str, args
     logging.info(f"{etp_size=}")
     logging.info(f"{enable_cuda_graphs=}")
     logging.info(f"{use_mcore_fsdp=}")
+    logging.info(f"{use_fsdp_double_buffer=}")
     logging.info(f"{recompute_layers=}")
     logging.info(f"{activation_offload_layers=}")
     logging.info(f"{recompute_modules=}")
@@ -240,12 +248,22 @@ def set_perf_optimization_configs(
     recompute_layers: int,
     activation_offload_layers: int,
     recompute_modules: Optional[List[str]],
-    use_sharp: bool,
-    use_user_buffer_registration: bool,
+    use_fsdp_double_buffer: Optional[bool] = None,
+    use_user_buffer_registration: Optional[bool] = None,
+    use_sharp: Optional[bool] = None,
 ):
     # enable cross entropy fusion with TE kernel
     recipe.model.config.cross_entropy_fusion_impl = "te"
 
+    if use_fsdp_double_buffer:
+        assert use_mcore_fsdp == True, "use_fsdp_double_buffer requires use_mcore_fsdp to be True"
+
+    if use_user_buffer_registration:
+        assert use_mcore_fsdp == True, "use_user_buffer_registration requires use_mcore_fsdp to be True"
+        assert (
+            use_fsdp_double_buffer is not False
+        ), "use_fsdp_double_buffer cannot be False when use_user_buffer_registration is True"
+
     if use_mcore_fsdp and enable_cuda_graphs:
         logging.warning("Currently, cuda graphs are not supported with FSDP. Disabling cuda graphs.")
         enable_cuda_graphs = False
@@ -269,6 +287,7 @@ def set_perf_optimization_configs(
         recipe.trainer.strategy.ddp.check_for_nan_in_grad = False
         recipe.trainer.strategy.ddp.check_for_large_grads = False
         recipe.trainer.strategy.ddp.nccl_ub = bool(use_user_buffer_registration)
+        recipe.trainer.strategy.ddp.fsdp_double_buffer = bool(use_fsdp_double_buffer)
 
     return recipe
 
@@ -289,8 +308,9 @@ def set_primary_perf_configs(
     etp_size: Optional[int] = None,
     enable_cuda_graphs: bool = False,
     use_mcore_fsdp: bool = False,
-    use_user_buffer_registration: bool = False,
-    use_sharp: bool = False,
+    use_fsdp_double_buffer: Optional[bool] = None,
+    use_user_buffer_registration: Optional[bool] = None,
+    use_sharp: Optional[bool] = None,
     recompute_layers: int = 0,
     activation_offload_layers: int = 0,
     compute_dtype: str = None,
@@ -334,18 +354,19 @@ def set_primary_perf_configs(
         )
 
     recipe = set_perf_optimization_configs(
-        recipe,
-        use_mcore_fsdp,
-        enable_cuda_graphs,
-        task,
-        tp_size,
-        compute_dtype,
-        fp8_recipe,
-        recompute_layers,
-        activation_offload_layers,
-        recompute_modules,
-        use_sharp,
-        use_user_buffer_registration,
+        recipe=recipe,
+        use_mcore_fsdp=use_mcore_fsdp,
+        enable_cuda_graphs=enable_cuda_graphs,
+        task=task,
+        tp_size=tp_size,
+        compute_dtype=compute_dtype,
+        fp8_recipe=fp8_recipe,
+        recompute_layers=recompute_layers,
+        activation_offload_layers=activation_offload_layers,
+        recompute_modules=recompute_modules,
+        use_fsdp_double_buffer=use_fsdp_double_buffer,
+        use_user_buffer_registration=use_user_buffer_registration,
+        use_sharp=use_sharp,
     )
 
     return recipe

scripts/performance/llm/finetune_deepseek_v3.py

@@ -83,6 +83,8 @@ def override_recipe_configs(
         enable_cuda_graphs=enable_cuda_graphs,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         nccl_communicator_config_path=args.nccl_communicator_config_path,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,

scripts/performance/llm/finetune_llama31_405b.py

@@ -85,12 +85,13 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
         nccl_communicator_config_path=args.nccl_communicator_config_path,
-        use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
     )
     recipe = set_exp_logging_configs(

scripts/performance/llm/finetune_llama3_70b.py

@@ -92,12 +92,13 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
         nccl_communicator_config_path=args.nccl_communicator_config_path,
-        use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
     )
     recipe = set_exp_logging_configs(

scripts/performance/llm/finetune_llama3_8b.py

@@ -78,8 +78,10 @@ def override_recipe_configs(
         enable_cuda_graphs=enable_cuda_graphs,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
-        nccl_communicator_config_path=args.nccl_communicator_config_path,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
+        nccl_communicator_config_path=args.nccl_communicator_config_path,
         use_sharp=args.use_sharp,
     )
     recipe = set_exp_logging_configs(

scripts/performance/llm/finetune_llama4_e128.py

@@ -86,6 +86,8 @@ def override_recipe_configs(
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_deepseek_v3.py

@@ -86,6 +86,7 @@ def override_recipe_configs(
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         recompute_layers=recompute_layers,

scripts/performance/llm/pretrain_gpt3_175b.py

@@ -70,6 +70,7 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         recompute_layers=recompute_layers,

scripts/performance/llm/pretrain_llama31_405b.py

@@ -70,6 +70,7 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         recompute_layers=recompute_layers,

scripts/performance/llm/pretrain_llama3_70b.py

@@ -71,6 +71,7 @@ def override_recipe_configs(
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
         use_user_buffer_registration=args.use_user_buffer_registration,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_sharp=args.use_sharp,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,

scripts/performance/llm/pretrain_llama3_8b.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
         nccl_communicator_config_path=args.nccl_communicator_config_path,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
     )

scripts/performance/llm/pretrain_llama4_e128.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_llama4_e16.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_mixtral_8x22b.py

@@ -65,6 +65,7 @@ def override_recipe_configs(
         etp_size,
         enable_cuda_graphs,
         use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         recompute_layers=recompute_layers,

scripts/performance/llm/pretrain_mixtral_8x7b.py

@@ -61,6 +61,8 @@ def override_recipe_configs(
         ep_size,
         etp_size,
         enable_cuda_graphs,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_nemotron3_22b.py

@@ -61,6 +61,7 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         recompute_layers=recompute_layers,

scripts/performance/llm/pretrain_nemotron3_8b.py

@@ -60,6 +60,8 @@ def override_recipe_configs(
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
         nccl_communicator_config_path=args.nccl_communicator_config_path,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
     )

scripts/performance/llm/pretrain_nemotron4_15b.py

@@ -61,6 +61,8 @@ def override_recipe_configs(
         vp_size,
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_nemotron4_340b.py

@@ -68,11 +68,12 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
-        use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         nccl_communicator_config_path=args.nccl_communicator_config_path,
     )

scripts/performance/llm/pretrain_nemotronh_47b.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_nemotronh_56b.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_nemotronh_8b.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         enable_cuda_graphs=enable_cuda_graphs,
         use_mcore_fsdp=use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
         recompute_layers=recompute_layers,
         activation_offload_layers=activation_offload_layers,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_qwen3_235b_a22b.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         compute_dtype=args.compute_dtype,

scripts/performance/llm/pretrain_qwen3_30b_a3b.py

@@ -62,6 +62,8 @@ def override_recipe_configs(
         ep_size,
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
         use_user_buffer_registration=args.use_user_buffer_registration,
         use_sharp=args.use_sharp,
         compute_dtype=args.compute_dtype,

scripts/performance/vlm/finetune_neva_8b.py

@@ -60,6 +60,9 @@ def override_recipe_configs(
         enable_cuda_graphs=enable_cuda_graphs,
         compute_dtype=args.compute_dtype,
         fp8_recipe=args.fp8_recipe,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
     )
     recipe = set_exp_logging_configs(
         recipe,

scripts/performance/vlm/pretrain_vlm_llama4_e128.py

@@ -65,6 +65,9 @@ def override_recipe_configs(
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
         compute_dtype=args.compute_dtype,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
     )
     recipe = set_exp_logging_configs(
         recipe,

scripts/performance/vlm/pretrain_vlm_llama4_e16.py

@@ -65,6 +65,9 @@ def override_recipe_configs(
         etp_size,
         enable_cuda_graphs=enable_cuda_graphs,
         compute_dtype=args.compute_dtype,
+        use_mcore_fsdp=args.use_mcore_fsdp,
+        use_fsdp_double_buffer=args.use_fsdp_double_buffer,
+        use_user_buffer_registration=args.use_user_buffer_registration,
     )
     recipe = set_exp_logging_configs(
         recipe,