Skip to content

DSV3 NVFP4 recipe on GB300 #2076

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
erhoo82 merged 4 commits into from
Jan 28, 2026
Merged

DSV3 NVFP4 recipe on GB300 #2076

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
de3d019
dsv3 nvfp4 gb300
dingqingy-nv Jan 27, 2026
daf08ab
enable fast math
dingqingy-nv Jan 27, 2026
1afc195
Merge branch 'main' into dsv3-nvfp4-recipe
ko3n1g Jan 27, 2026
0aaad4d
remove redundent override
dingqingy-nv Jan 27, 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
2 changes: 1 addition & 1 deletion scripts/performance/configs/deepseek/deepseek_llm_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,7 +60,7 @@ def deepseek_v3_pretrain_config_gb300(
pipeline_model_parallel_size=base_cfg.pipeline_model_parallel_size,
virtual_pipeline_model_parallel_size=base_cfg.virtual_pipeline_model_parallel_size,
moe_flex_dispatcher_backend=base_cfg.moe_flex_dispatcher_backend,
layout=None,
layout=base_cfg.pp_layout,
)
set_deepseek_v3_common_configs(cfg)
set_workload_base_configs(cfg, base_cfg)
Expand Down
20 changes: 18 additions & 2 deletions scripts/performance/configs/deepseek/deepseek_workload_base_configs.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,20 @@
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_BF16_V1 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V1
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_FP8_CS_V1 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V1
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_FP8_MX_V1 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V1
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_NVFP4_V1 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V1
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_NVFP4_V1 = replace(
BASE_DEEPSEEK_V3_CONFIG,
num_gpus=256,
global_batch_size=2048,
micro_batch_size=2,
pipeline_model_parallel_size=2,
virtual_pipeline_model_parallel_size=8,
pp_layout="Et*4|(t*4|)*14tmL",
expert_model_parallel_size=32,
moe_flex_dispatcher_backend="hybridep",
moe_a2a_overlap=False,
cuda_graph_scope=[],
recompute_modules=["mla_up_proj"],

@erhoo82 erhoo82 Jan 27, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We need to override only the changes?
I see that most of the lines remain the same as the baseline.

@dingqingy-nv dingqingy-nv Jan 27, 2026

Copy link
Copy Markdown
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks, updated

)


DEEPSEEK_V3_PRETRAIN_CONFIG_GB200_V1 = replace(
Expand Down Expand Up @@ -133,7 +146,10 @@
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_BF16_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V2
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_FP8_CS_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V2
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_FP8_MX_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V2
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_NVFP4_V2 = DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_V2
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_NVFP4_V2 = replace(
DEEPSEEK_V3_PRETRAIN_CONFIG_GB300_NVFP4_V1,
global_batch_size=4096,
)


DEEPSEEK_V3_PRETRAIN_CONFIG_GB200_V2 = replace(
Expand Down
2 changes: 1 addition & 1 deletion scripts/performance/utils/overrides.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -357,7 +357,7 @@ def set_post_overrides(
dp = int(num_gpus / (tp * pp * cp))
logger.info(f"DP: {dp}; TP: {tp}; PP: {pp}; CP: {cp}; VP: {vp}")
## NOTE: overlap_param_gather_with_optimizer_step causes NaN grad norm for fp8_mx. Disabling it until the issue is resolved.
if dp > 1 and pp > 1 and vp > 1 and compute_dtype != "fp8_mx":
if dp > 1 and pp > 1 and vp > 1 and compute_dtype not in ("fp8_mx", "nvfp4"):
recipe.optimizer.overlap_param_gather_with_optimizer_step = True
if hasattr(recipe, "comm_overlap") and isinstance(recipe.comm_overlap, CommOverlapConfig):
recipe.comm_overlap.overlap_param_gather_with_optimizer_step = True
Expand Down
3 changes: 3 additions & 0 deletions scripts/performance/utils/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,9 @@ class WorkloadBaseConfig:
moe_a2a_overlap: Optional[bool] = False
peft: Optional[str] = None

# Pipeline parallelism layout
pp_layout: Optional[str] = None

@erhoo82 erhoo82 Jan 27, 2026

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a nice change to make the PP layout configurable.


@property
def sequence_parallel(self) -> bool:
"""Get the sequence parallel flag."""
Expand Down
Loading