Skip to content

70b_sft_perf_optmization #2404

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
malay-nagda merged 8 commits into from
Feb 23, 2026
Merged

70b_sft_perf_optmization #2404

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
e6ee984
70b_sft_perf_optmization
malay-nagda Feb 17, 2026
15da07a
70b sft gb200 bf16 update
malay-nagda Feb 17, 2026
69d0b3b
optimized cfgs
malay-nagda Feb 18, 2026
ec53128
sft cudnn-ln, pp chunk
malay-nagda Feb 19, 2026
ae18930
70b sft h100
malay-nagda Feb 19, 2026
5587854
405b CG
malay-nagda Feb 20, 2026
3feb149
Merge branch 'main' into malay/70b_sft_perf_optmization
malay-nagda Feb 23, 2026
f17646c
cleanup
malay-nagda Feb 23, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
22 changes: 14 additions & 8 deletions scripts/performance/configs/llama/llama3_workload_base_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -500,8 +500,8 @@
num_gpus=32,
peft="none",
tensor_model_parallel_size=1,
pipeline_model_parallel_size=4,
virtual_pipeline_model_parallel_size=5,
pipeline_model_parallel_size=2,
virtual_pipeline_model_parallel_size=20,
micro_batch_size=1,
global_batch_size=32,
cuda_graph_impl="transformer_engine",
Expand All @@ -517,17 +517,20 @@
BASE_LLAMA3_70B_CONFIG,
num_gpus=32,
peft="none",
tensor_model_parallel_size=2,
pipeline_model_parallel_size=4,
virtual_pipeline_model_parallel_size=5,
tensor_model_parallel_size=1,
pipeline_model_parallel_size=8,
virtual_pipeline_model_parallel_size=10,
micro_batch_size=1,
global_batch_size=32,
cuda_graph_impl="transformer_engine",
cuda_graph_scope="mlp",
)

LLAMA3_70B_SFT_CONFIG_GB200_BF16_V1 = _LLAMA3_70B_SFT_CONFIG_GB200
LLAMA3_70B_SFT_CONFIG_GB200_FP8_CS_V1 = _LLAMA3_70B_SFT_CONFIG_GB200
LLAMA3_70B_SFT_CONFIG_GB200_FP8_CS_V1 = replace(
_LLAMA3_70B_SFT_CONFIG_GB200,
pipeline_model_parallel_size=4,
)
LLAMA3_70B_SFT_CONFIG_GB200_FP8_MX_V1 = _LLAMA3_70B_SFT_CONFIG_GB200


Expand All @@ -543,8 +546,11 @@
)

LLAMA3_70B_SFT_CONFIG_H100_BF16_V1 = _LLAMA3_70B_SFT_CONFIG_H100
LLAMA3_70B_SFT_CONFIG_H100_FP8_CS_V1 = _LLAMA3_70B_SFT_CONFIG_H100

LLAMA3_70B_SFT_CONFIG_H100_FP8_CS_V1 = replace(
_LLAMA3_70B_SFT_CONFIG_H100,
cuda_graph_impl="transformer_engine",
cuda_graph_scope="mlp",
)

# =============================================================================
# Llama3 70B finetune (LoRA) presets - V1 (only version)
Expand Down
9 changes: 9 additions & 0 deletions scripts/performance/perf_plugins.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -292,6 +292,12 @@ def _set_model_specific_environment_variables(
if model_family_name == "deepseek":
if compute_dtype == "fp8_mx":
del_cudnn_ln = False
if model_family_name in ["llama"] and train_task in ["sft"]:
# TODO: Verify for H100 and 8b
del_cudnn_ln = False
if gpu in ["h100"] and model_recipe_name in ["llama3_70b"] and compute_dtype == "fp8_cs":
executor.env_vars["PYTORCH_CUDA_ALLOC_CONF"] = "expandable_segments:True"
executor.env_vars["NCCL_GRAPH_REGISTER"] = "0"
if del_cudnn_ln:
if "NVTE_NORM_FWD_USE_CUDNN" in executor.env_vars:
executor.env_vars.pop("NVTE_NORM_FWD_USE_CUDNN")
Expand Down Expand Up @@ -448,6 +454,9 @@ def setup(self, task: Union["run.Partial", "run.Script"], executor: "run.Executo
if self.model_recipe_name in ["llama3_70b", "llama31_405b"] and self.train_task == "pretrain"
else None
)
nccl_pp_comm_chunksize = (
2097152 if self.model_family_name in ["llama"] and self.train_task in ["sft"] else None
)
self._set_nccl_pp_comm_chunksize(task, executor, nccl_pp_comm_chunksize, pp_size)

# Configure manual garbage collection
Expand Down
Loading