Skip to content

chore: Merge fp8 args #1279

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
ko3n1g merged 33 commits into from
Nov 12, 2025
Merged

chore: Merge fp8 args #1279

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
33 commits
Select commit Hold shift + click to select a range
0720374
debuf weak scaling
malay-nagda Nov 9, 2025
f6bd524
qwen ->
malay-nagda Nov 9, 2025
750df3d
will this work?
ko3n1g Nov 8, 2025
72a08a3
test
ko3n1g Nov 8, 2025
a80da88
fix
ko3n1g Nov 8, 2025
a5b397b
fixes
ko3n1g Nov 8, 2025
e569bb9
debug
ko3n1g Nov 8, 2025
2c6d523
test
ko3n1g Nov 8, 2025
9d33528
fix
ko3n1g Nov 8, 2025
472e288
bring back default
ko3n1g Nov 10, 2025
6fe07d5
fp8
ko3n1g Nov 10, 2025
b3c9812
fix
ko3n1g Nov 10, 2025
57fe17f
fix
ko3n1g Nov 10, 2025
46486c9
fix
ko3n1g Nov 10, 2025
4fc7658
tmp
ko3n1g Nov 10, 2025
3305c73
tmp
ko3n1g Nov 10, 2025
83240e7
use attn
ko3n1g Nov 10, 2025
dd28663
fix
ko3n1g Nov 10, 2025
79c924a
switch to mlp
ko3n1g Nov 10, 2025
9feb470
fix
ko3n1g Nov 10, 2025
83df56b
fix
ko3n1g Nov 10, 2025
49a780e
NCCL_NVLS_ENABLE
ko3n1g Nov 10, 2025
4383ad4
get_workload_base_config
ko3n1g Nov 10, 2025
9c791f3
bias_dropout_fusion
ko3n1g Nov 11, 2025
fcedede
qwen3
ko3n1g Nov 11, 2025
35f7737
revert
ko3n1g Nov 11, 2025
95c800a
remove arg
ko3n1g Nov 11, 2025
746dfe0
remove cuda_graph_scope
ko3n1g Nov 11, 2025
739f84f
remove last traces of fp8
ko3n1g Nov 11, 2025
748b32d
revert
ko3n1g Nov 11, 2025
7613126
expandable_segments
ko3n1g Nov 11, 2025
b527050
Merge remote-tracking branch 'origin/main' into ko3n1g/chore/merge-fp…
ko3n1g Nov 11, 2025
b95d1fc
revert
ko3n1g Nov 12, 2025
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
15 changes: 1 addition & 14 deletions scripts/performance/argument_parser.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -134,24 +134,11 @@ def parse_cli_args():
"-c",
"--compute_dtype",
type=str,
choices=["bf16", "fp8"],
choices=["bf16", "fp8_cs", "fp8_mx", "fp8_sc"],
help="Compute precision. Options- bf16 or fp8. Defaults to bf16",
required=False,
default="bf16",
Comment thread
malay-nagda marked this conversation as resolved.
Resolved
)
fp8_recipe_msg = (
"FP8 recipe. Options- ds (per-tensor delayed scaling), cs (per-tensor current scaling), "
"mxfp8, ss (subchannel scaling). Defaults to ds"
)
parser.add_argument(
"-fr",
"--fp8_recipe",
type=str,
choices=["cs", "mx", "sc"],
help=fp8_recipe_msg,
required=False,
default="cs",
)
parser.add_argument(
"--task",
choices=["pretrain", "sft", "lora"],
Expand Down
24 changes: 12 additions & 12 deletions 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 @@ -44,16 +44,16 @@ def set_deepseek_v3_common_configs(cfg: ConfigContainer) -> None:
cfg.model.moe_router_force_load_balancing = True


def deepseek_v3_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def deepseek_v3_gb300_config(precision: str = "bf16") -> ConfigContainer:
"""GB300, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.DEEPSEEK_V3_GB300_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.DEEPSEEK_V3_GB300_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.DEEPSEEK_V3_GB300_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = pretrain_config(
mock=True,
Expand All @@ -76,16 +76,16 @@ def deepseek_v3_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") ->
return cfg


def deepseek_v3_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def deepseek_v3_gb200_config(precision: str = "bf16") -> ConfigContainer:
"""GB200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.DEEPSEEK_V3_GB200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.DEEPSEEK_V3_GB200_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.DEEPSEEK_V3_GB200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = pretrain_config(
mock=True,
Expand All @@ -108,16 +108,16 @@ def deepseek_v3_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") ->
return cfg


def deepseek_v3_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def deepseek_v3_b200_config(precision: str = "bf16") -> ConfigContainer:
"""B200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.DEEPSEEK_V3_B200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.DEEPSEEK_V3_B200_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.DEEPSEEK_V3_B200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = pretrain_config(
mock=True,
Expand All @@ -135,16 +135,16 @@ def deepseek_v3_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") ->
return cfg


def deepseek_v3_h100_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def deepseek_v3_h100_config(precision: str = "bf16") -> ConfigContainer:
"""H100, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.DEEPSEEK_V3_H100_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.DEEPSEEK_V3_H100_FP8_CS_BASE_CONFIG
if fp8_recipe == "sc":
if precision == "fp8_sc":
base_cfg = base_cfgs.DEEPSEEK_V3_H100_FP8_SC_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = pretrain_config(
mock=True,
Expand Down
16 changes: 8 additions & 8 deletions scripts/performance/configs/gpt_oss/gpt_oss_llm_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -36,14 +36,14 @@ def set_gpt_oss_common_configs(cfg: ConfigContainer) -> None:
cfg.model.moe_router_force_load_balancing = True


def gpt_oss_120b_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def gpt_oss_120b_gb300_config(precision: str = "bf16") -> ConfigContainer:
"""GB300, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.GPT_OSS_120B_GB300_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.GPT_OSS_120B_GB300_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = gpt_oss_120b_pretrain_config(
mock=True,
Expand All @@ -55,14 +55,14 @@ def gpt_oss_120b_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") -
return cfg


def gpt_oss_120b_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def gpt_oss_120b_gb200_config(precision: str = "bf16") -> ConfigContainer:
"""GB200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.GPT_OSS_120B_GB200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.GPT_OSS_120B_GB200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = gpt_oss_120b_pretrain_config(
mock=True,
Expand All @@ -74,14 +74,14 @@ def gpt_oss_120b_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") -
return cfg


def gpt_oss_120b_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def gpt_oss_120b_b200_config(precision: str = "bf16") -> ConfigContainer:
"""B200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.GPT_OSS_120B_B200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.GPT_OSS_120B_B200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = gpt_oss_120b_pretrain_config(
mock=True,
Expand All @@ -93,14 +93,14 @@ def gpt_oss_120b_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") ->
return cfg


def gpt_oss_120b_h100_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def gpt_oss_120b_h100_config(precision: str = "bf16") -> ConfigContainer:
"""H100, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.GPT_OSS_120B_H100_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.GPT_OSS_120B_H100_FP8_CS_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = gpt_oss_120b_pretrain_config(
mock=True,
Expand Down
44 changes: 22 additions & 22 deletions scripts/performance/configs/llama3/llama3_llm_pretrain.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,17 +50,17 @@ def set_llama3_common_configs(cfg: ConfigContainer) -> None:
# Llama3 70B configs ---------------------------------------------------------


def llama3_70b_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_70b_gb300_config(precision: str = "bf16") -> ConfigContainer:
"""GB300, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_70B_GB300_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_bf16_b200_h8192_tp2_mbs1_seqlen8192
else:
base_cfg = base_cfgs.LLAMA3_70B_GB300_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.LLAMA3_70B_GB300_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_fp8_b200_h8192_tp2_mbs1_seqlen8192

cfg = llama3_70b_pretrain_config(mock=True, precision_config=precision_config)
Expand All @@ -77,17 +77,17 @@ def llama3_70b_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") ->
return cfg


def llama3_70b_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_70b_gb200_config(precision: str = "bf16") -> ConfigContainer:
"""GB200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_70B_GB200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_bf16_b200_h8192_tp2_mbs1_seqlen8192
else:
base_cfg = base_cfgs.LLAMA3_70B_GB200_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.LLAMA3_70B_GB200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_fp8_b200_h8192_tp2_mbs1_seqlen8192

cfg = llama3_70b_pretrain_config(mock=True, precision_config=precision_config)
Expand All @@ -104,17 +104,17 @@ def llama3_70b_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") ->
return cfg


def llama3_70b_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_70b_b200_config(precision: str = "bf16") -> ConfigContainer:
"""B200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_70B_B200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_bf16_b200_h8192_tp2_mbs1_seqlen8192
else:
base_cfg = base_cfgs.LLAMA3_70B_B200_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.LLAMA3_70B_B200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_fp8_b200_h8192_tp2_mbs1_seqlen8192

cfg = llama3_70b_pretrain_config(mock=True, precision_config=precision_config)
Expand All @@ -131,15 +131,15 @@ def llama3_70b_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> C
return cfg


def llama3_70b_h100_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_70b_h100_config(precision: str = "bf16") -> ConfigContainer:
"""H100, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_70B_H100_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_bf16_h100_h8192_tp4_mbs1_seqlen8192
else:
base_cfg = base_cfgs.LLAMA3_70B_H100_FP8_CS_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)
comm_overlap_cfg = userbuffers_fp8_h100_h8192_tp4_mbs1_seqlen8192

cfg = llama3_70b_pretrain_config(mock=True, precision_config=precision_config)
Expand All @@ -154,16 +154,16 @@ def llama3_70b_h100_config(precision: str = "bf16", fp8_recipe: str = "cs") -> C
# Llama3 8B configs ---------------------------------------------------------


def llama3_8b_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_8b_gb300_config(precision: str = "bf16") -> ConfigContainer:
"""GB300, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_8B_GB300_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.LLAMA3_8B_GB300_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.LLAMA3_8B_GB300_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = llama3_8b_pretrain_config(mock=True, precision_config=precision_config)
set_llama3_common_configs(cfg)
Expand All @@ -174,16 +174,16 @@ def llama3_8b_gb300_config(precision: str = "bf16", fp8_recipe: str = "cs") -> C
return cfg


def llama3_8b_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_8b_gb200_config(precision: str = "bf16") -> ConfigContainer:
"""GB200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_8B_GB200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.LLAMA3_8B_GB200_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.LLAMA3_8B_GB200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = llama3_8b_pretrain_config(mock=True, precision_config=precision_config)
set_llama3_common_configs(cfg)
Expand All @@ -194,16 +194,16 @@ def llama3_8b_gb200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> C
return cfg


def llama3_8b_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_8b_b200_config(precision: str = "bf16") -> ConfigContainer:
"""B200, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_8B_B200_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.LLAMA3_8B_B200_FP8_CS_BASE_CONFIG
if fp8_recipe == "mx":
if precision == "fp8_mx":
base_cfg = base_cfgs.LLAMA3_8B_B200_FP8_MX_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = llama3_8b_pretrain_config(mock=True, precision_config=precision_config)
set_llama3_common_configs(cfg)
Expand All @@ -214,14 +214,14 @@ def llama3_8b_b200_config(precision: str = "bf16", fp8_recipe: str = "cs") -> Co
return cfg


def llama3_8b_h100_config(precision: str = "bf16", fp8_recipe: str = "cs") -> ConfigContainer:
def llama3_8b_h100_config(precision: str = "bf16") -> ConfigContainer:
"""H100, baseline config."""
if precision == "bf16":
base_cfg = base_cfgs.LLAMA3_8B_H100_BF16_BASE_CONFIG
precision_config = get_precision_config(precision)
else:
base_cfg = base_cfgs.LLAMA3_8B_H100_FP8_CS_BASE_CONFIG
precision_config = get_precision_config(precision, fp8_recipe)
precision_config = get_precision_config(precision)

cfg = llama3_8b_pretrain_config(mock=True, precision_config=precision_config)
set_llama3_common_configs(cfg)
Expand Down
Loading
Loading