Skip to content
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

Add perf model configs gb200 #12140

Merged
merged 13 commits into from
Feb 14, 2025
Merged

Add perf model configs gb200 #12140

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
8cc23e8
Bug fix with generation of expert_tensor_parallel_rank
guyueh1 Feb 10, 2025
bc8441e
Fix pylint
guyueh1 Feb 10, 2025
37dbae3
Support perf script with gb200
guyueh1 Feb 11, 2025
2db23e2
fix
guyueh1 Feb 11, 2025
8d87cbf
mixtral 8x22b recipe change to mitigate a megatron.core bug
guyueh1 Feb 11, 2025
135d542
Merge branch 'main' into add_perf_model_configs_gb200
guyueh1 Feb 12, 2025
215b51e
Make column names of all csvs align; fix for 8x22b
guyueh1 Feb 12, 2025
3071d05
typing fix
guyueh1 Feb 12, 2025
bbb3947
Merge branch 'add_perf_model_configs_gb200' of github.com:guyueh1/NeM…
guyueh1 Feb 12, 2025
a953197
Adjust 405B config due to asymmetric PP
guyueh1 Feb 12, 2025
7e51eab
Add a comment
guyueh1 Feb 12, 2025
0f28241
Change 405B GBS
guyueh1 Feb 12, 2025
18fa9d4
fix gbs for gb200 finetuning
guyueh1 Feb 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
2 changes: 1 addition & 1 deletion scripts/llm/performance/finetune_llama31_405b.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -77,7 +77,7 @@ def override_recipe_configs(
recipe.trainer.plugins = bf16_with_fp8_mixed()
recipe.trainer.plugins.grad_reduce_in_fp32 = False

enable_cuda_graph = bool(args.gpu.lower() in [])
enable_cuda_graph = bool(args.gpu.lower() in ["gb200"])
malay-nagda marked this conversation as resolved.
Show resolved Hide resolved
recipe.model.config.enable_cuda_graph = enable_cuda_graph
recipe.trainer.strategy.use_te_rng_tracker = enable_cuda_graph
recipe.data.packed_sequence_specs.pad_cu_seqlens = enable_cuda_graph
Expand Down
11 changes: 9 additions & 2 deletions scripts/llm/performance/finetune_llama3_70b.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -50,7 +50,14 @@ def override_recipe_configs(
NOTE: Use fp8 precision training with caution. It might not give desirable results.
"""
finetuning_scheme = "none" if args.finetuning == "sft" else args.finetuning
recipe = finetune_recipe(peft_scheme=finetuning_scheme, performance_mode=True)
gpu_type = args.gpu.lower()
if gpu_type in ["gb200"] and finetuning_scheme == "lora":
# On GB200 for lora task, we need to enable Cuda Graph for optimal performance.
# However, Cuda Graph increases memory usage, so in order to avoid OOM, we need
# to reduce the sequence length.
recipe = finetune_recipe(peft_scheme=finetuning_scheme, performance_mode=True, seq_length=2048)
else:
recipe = finetune_recipe(peft_scheme=finetuning_scheme, performance_mode=True)
recipe = set_primary_perf_configs(
recipe,
args.tensorboard,
Expand All @@ -77,7 +84,7 @@ def override_recipe_configs(
recipe.trainer.plugins = bf16_with_fp8_mixed()
recipe.trainer.plugins.grad_reduce_in_fp32 = False

enable_cuda_graph = bool(args.gpu.lower() in [] and finetuning_scheme != "lora")
enable_cuda_graph = bool(args.gpu.lower() in ["gb200"] and finetuning_scheme == "lora")
recipe.model.config.enable_cuda_graph = enable_cuda_graph
recipe.trainer.strategy.use_te_rng_tracker = enable_cuda_graph
recipe.data.packed_sequence_specs.pad_cu_seqlens = enable_cuda_graph
Expand Down
4 changes: 2 additions & 2 deletions scripts/llm/performance/finetune_llama3_8b.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -52,7 +52,7 @@ def override_recipe_configs(
finetuning_scheme = "none" if args.finetuning == "sft" else args.finetuning

gpu_type = args.gpu.lower()
if gpu_type == "b200":
if gpu_type in ["b200", "gb200"]:
recipe = finetune_recipe(peft_scheme=finetuning_scheme, performance_mode=True, seq_length=16384)
else:
recipe = finetune_recipe(peft_scheme=finetuning_scheme, performance_mode=True)
Expand Down Expand Up @@ -82,7 +82,7 @@ def override_recipe_configs(
recipe.trainer.plugins = bf16_with_fp8_mixed()
recipe.trainer.plugins.grad_reduce_in_fp32 = False

enable_cuda_graph = bool(gpu_type in ["b200"])
enable_cuda_graph = bool(gpu_type in ["b200", "gb200"])
recipe.model.config.enable_cuda_graph = enable_cuda_graph
recipe.trainer.strategy.use_te_rng_tracker = enable_cuda_graph
recipe.data.packed_sequence_specs.pad_cu_seqlens = enable_cuda_graph
Expand Down
2 changes: 1 addition & 1 deletion scripts/llm/performance/pretrain_llama3_8b.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,7 +65,7 @@ def override_recipe_configs(
recipe.trainer.plugins = bf16_with_fp8_mixed()
recipe.trainer.plugins.grad_reduce_in_fp32 = False

enable_cuda_graph = bool(args.gpu.lower() in [])
enable_cuda_graph = bool(args.gpu.lower() in ["gb200"])
recipe.model.config.enable_cuda_graph = enable_cuda_graph
recipe.trainer.strategy.use_te_rng_tracker = enable_cuda_graph

Expand Down
8 changes: 7 additions & 1 deletion scripts/llm/performance/pretrain_mixtral_8x22b.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -13,6 +13,7 @@
# limitations under the License.

from os.path import basename, splitext
from typing import Optional

import nemo_run as run
from argument_parser import parse_cli_args
Expand All @@ -33,7 +34,7 @@ def override_recipe_configs(
cp_size: int,
vp_size: int,
ep_size: int,
etp_size: int,
etp_size: Optional[int],
):
"""
mixtral 8x22b pre-train recipe aimed at achieving best possible performance.
Expand Down Expand Up @@ -66,6 +67,11 @@ def override_recipe_configs(
recipe.trainer.plugins = bf16_with_fp8_mixed()
recipe.trainer.plugins.grad_reduce_in_fp32 = False

# to mitigate the incorrect gradient_scaling_factor calculation in megatron.core
# under scenario average_in_collective=True and tp_size != etp_size, disabling average_in_collective.
if etp_size is not None and etp_size != tp_size:
recipe.trainer.strategy.ddp.average_in_collective = False

return recipe


Expand Down
54 changes: 27 additions & 27 deletions scripts/llm/performance/recommended_model_configs/model_configs_b200.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,27 +1,27 @@
task,model,size,system,dtype,num_gpus,seq_len,tp_size,pp_size,cp_size,dp_size,ep_size,etp_size,vp_size,mbs,gbs,etp_size
lora,llama3,70b,b200,fp8,8.0,4096.0,1.0,4.0,1.0,2.0,1.0,,20.0,1.0,32.0,
lora,llama3,70b,b200,bf16,8.0,4096.0,1.0,4.0,1.0,2.0,1.0,,20.0,1.0,32.0,
lora,llama3,8b,b200,bf16,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
lora,llama3,8b,b200,fp8,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
lora,llama31,405b,b200,fp8,24.0,2048.0,4.0,6.0,1.0,1.0,1.0,,7.0,1.0,24.0,
lora,llama31,405b,b200,bf16,24.0,4096.0,4.0,6.0,1.0,1.0,1.0,,7.0,1.0,24.0,
pre_train,gpt3,175b,b200,bf16,128.0,2048.0,4.0,4.0,1.0,8.0,1.0,,12.0,2.0,256.0,
pre_train,gpt3,175b,b200,fp8,128.0,2048.0,4.0,4.0,1.0,8.0,1.0,,12.0,2.0,256.0,
pre_train,llama3,8b,b200,bf16,8.0,8192.0,1.0,1.0,1.0,8.0,1.0,,1.0,2.0,128.0,
pre_train,llama3,8b,b200,fp8,8.0,8192.0,1.0,1.0,1.0,8.0,1.0,,1.0,2.0,128.0,
pre_train,llama3,70b,b200,fp8,64.0,8192.0,2.0,4.0,2.0,4.0,1.0,,5.0,1.0,128.0,
pre_train,llama3,70b,b200,bf16,64.0,8192.0,2.0,4.0,2.0,4.0,1.0,,5.0,1.0,128.0,
pre_train,llama31,405b,b200,bf16,144.0,8192.0,4.0,9.0,2.0,2.0,1.0,,7.0,1.0,36.0,
pre_train,llama31,405b,b200,fp8,144.0,8192.0,4.0,9.0,2.0,2.0,1.0,,7.0,1.0,36.0,
pre_train,mixtral,8x7b,b200,bf16,64.0,4096.0,1.0,1.0,1.0,64.0,8.0,,1.0,2.0,256.0,
pre_train,mixtral,8x7b,b200,fp8,64.0,4096.0,1.0,1.0,1.0,64.0,8.0,,1.0,2.0,256.0,
pre_train,mixtral,8x22b,b200,bf16,256.0,65536.0,2.0,4.0,8.0,4.0,8.0,1.0,14.0,1.0,64.0,1.0
pre_train,mixtral,8x22b,b200,fp8,256.0,65536.0,2.0,4.0,8.0,4.0,8.0,1.0,14.0,1.0,64.0,1.0
pre_train,nemotron4,15b,b200,bf16,64.0,4096.0,1.0,1.0,1.0,32.0,1.0,,1.0,2.0,256.0,
pre_train,nemotron4,340b,b200,bf16,128.0,4096.0,8.0,4.0,1.0,4.0,1.0,,12.0,1.0,32.0,
pre_train,nemotron4,340b,b200,fp8,128.0,4096.0,8.0,4.0,1.0,4.0,1.0,,12.0,1.0,32.0,
pre_train,nemotron4,15b,b200,fp8,64.0,4096.0,1.0,1.0,1.0,32.0,1.0,,1.0,2.0,256.0,
sft,llama3,8b,b200,fp8,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
sft,llama3,70b,b200,bf16,32.0,4096.0,2.0,4.0,1.0,4.0,1.0,,5.0,1.0,32.0,
sft,llama3,70b,b200,fp8,32.0,4096.0,2.0,4.0,1.0,4.0,1.0,,5.0,1.0,32.0,
sft,llama3,8b,b200,bf16,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
task,model,size,system,dtype,num_gpus,seq_len,tp_size,pp_size,cp_size,dp_size,ep_size,vp_size,mbs,gbs,etp_size
lora,llama3,70b,b200,fp8,8.0,4096.0,1.0,4.0,1.0,2.0,1.0,20.0,1.0,32.0,
lora,llama3,70b,b200,bf16,8.0,4096.0,1.0,4.0,1.0,2.0,1.0,20.0,1.0,32.0,
lora,llama3,8b,b200,bf16,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
lora,llama3,8b,b200,fp8,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
lora,llama31,405b,b200,fp8,24.0,2048.0,4.0,6.0,1.0,1.0,1.0,7.0,1.0,24.0,
lora,llama31,405b,b200,bf16,24.0,4096.0,4.0,6.0,1.0,1.0,1.0,7.0,1.0,24.0,
pre_train,gpt3,175b,b200,bf16,128.0,2048.0,4.0,4.0,1.0,8.0,1.0,12.0,2.0,256.0,
pre_train,gpt3,175b,b200,fp8,128.0,2048.0,4.0,4.0,1.0,8.0,1.0,12.0,2.0,256.0,
pre_train,llama3,8b,b200,bf16,8.0,8192.0,1.0,1.0,1.0,8.0,1.0,1.0,2.0,128.0,
pre_train,llama3,8b,b200,fp8,8.0,8192.0,1.0,1.0,1.0,8.0,1.0,1.0,2.0,128.0,
pre_train,llama3,70b,b200,fp8,64.0,8192.0,2.0,4.0,2.0,4.0,1.0,5.0,1.0,128.0,
pre_train,llama3,70b,b200,bf16,64.0,8192.0,2.0,4.0,2.0,4.0,1.0,5.0,1.0,128.0,
pre_train,llama31,405b,b200,bf16,512.0,8192.0,4.0,8.0,2.0,8.0,1.0,8.0,1.0,256.0,
pre_train,llama31,405b,b200,fp8,512.0,8192.0,4.0,8.0,2.0,8.0,1.0,8.0,1.0,256.0,
pre_train,mixtral,8x7b,b200,bf16,64.0,4096.0,1.0,1.0,1.0,64.0,8.0,1.0,2.0,256.0,
pre_train,mixtral,8x7b,b200,fp8,64.0,4096.0,1.0,1.0,1.0,64.0,8.0,1.0,2.0,256.0,
pre_train,mixtral,8x22b,b200,bf16,256.0,65536.0,2.0,4.0,8.0,4.0,8.0,14.0,1.0,64.0,1.0
pre_train,mixtral,8x22b,b200,fp8,256.0,65536.0,2.0,4.0,8.0,4.0,8.0,14.0,1.0,64.0,1.0
pre_train,nemotron4,15b,b200,bf16,64.0,4096.0,1.0,1.0,1.0,32.0,1.0,1.0,2.0,256.0,
pre_train,nemotron4,340b,b200,bf16,128.0,4096.0,8.0,4.0,1.0,4.0,1.0,12.0,1.0,32.0,
pre_train,nemotron4,340b,b200,fp8,128.0,4096.0,8.0,4.0,1.0,4.0,1.0,12.0,1.0,32.0,
pre_train,nemotron4,15b,b200,fp8,64.0,4096.0,1.0,1.0,1.0,32.0,1.0,1.0,2.0,256.0,
sft,llama3,8b,b200,fp8,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
sft,llama3,70b,b200,bf16,32.0,4096.0,2.0,4.0,1.0,4.0,1.0,5.0,1.0,32.0,
sft,llama3,70b,b200,fp8,32.0,4096.0,2.0,4.0,1.0,4.0,1.0,5.0,1.0,32.0,
sft,llama3,8b,b200,bf16,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
11 changes: 11 additions & 0 deletions scripts/llm/performance/recommended_model_configs/model_configs_gb200.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,11 @@
task,model,size,system,dtype,num_gpus,seq_len,tp_size,pp_size,cp_size,dp_size,ep_size,vp_size,mbs,gbs,etp_size
lora,llama3,70b,gb200,fp8,8.0,2048.0,1.0,4.0,1.0,2.0,1.0,20.0,1.0,64.0,
lora,llama3,70b,gb200,bf16,8.0,2048.0,1.0,4.0,1.0,2.0,1.0,20.0,1.0,64.0,
lora,llama3,8b,gb200,bf16,8.0,16384.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
lora,llama3,8b,gb200,fp8,8.0,16384.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
lora,llama31,405b,gb200,fp8,24.0,2048.0,4.0,6.0,1.0,1.0,1.0,7.0,1.0,24.0,
lora,llama31,405b,gb200,bf16,24.0,2048.0,4.0,6.0,1.0,1.0,1.0,7.0,1.0,24.0,
pre_train,llama3,8b,gb200,bf16,8.0,8192.0,1.0,1.0,1.0,8.0,1.0,1.0,2.0,128.0,
pre_train,llama3,8b,gb200,fp8,8.0,8192.0,1.0,1.0,1.0,8.0,1.0,1.0,2.0,128.0,
sft,llama3,8b,gb200,fp8,8.0,16384.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
sft,llama3,8b,gb200,bf16,8.0,16384.0,1.0,1.0,1.0,8.0,1.0,1.0,1.0,8.0,
4 changes: 2 additions & 2 deletions scripts/llm/performance/recommended_model_configs/model_configs_h100.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,8 +8,8 @@ lora,llama31,405b,h100,fp8,24.0,2048.0,4.0,6.0,1.0,1.0,1.0,7.0,1.0,24.0,
pre_train,gpt3,175b,h100,fp8,512.0,2048.0,4.0,8.0,1.0,16.0,1.0,6.0,2.0,2048.0,
pre_train,gpt3,175b,h100,fp8,128.0,2048.0,4.0,8.0,1.0,4.0,1.0,6.0,1.0,256.0,
pre_train,gpt3,175b,h100,bf16,128.0,2048.0,4.0,8.0,1.0,4.0,1.0,6.0,1.0,256.0,
pre_train,llama31,405b,h100,fp8,576.0,8192.0,8.0,9.0,2.0,4.0,1.0,7.0,1.0,252.0,
pre_train,llama31,405b,h100,bf16,576.0,8192.0,8.0,9.0,2.0,4.0,1.0,7.0,1.0,252.0,
pre_train,llama31,405b,h100,fp8,512.0,8192.0,8.0,8.0,2.0,4.0,1.0,8.0,1.0,256.0,
pre_train,llama31,405b,h100,bf16,512.0,8192.0,8.0,8.0,2.0,4.0,1.0,8.0,1.0,256.0,
pre_train,llama3,70b,h100,fp8,64.0,8192.0,4.0,8.0,1.0,2.0,1.0,5.0,1.0,128.0,
pre_train,llama3,70b,h100,bf16,64.0,8192.0,4.0,4.0,2.0,2.0,1.0,5.0,1.0,128.0,
pre_train,llama3,8b,h100,fp8,8.0,8192.0,1.0,1.0,2.0,4.0,1.0,1.0,1.0,128.0,
Expand Down
22 changes: 11 additions & 11 deletions scripts/llm/performance/recommended_model_configs/strong_scaling_model_configs_h100.csv
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,11 +1,11 @@
task,model,size,system,dtype,num_gpus,seq_len,tp_size,pp_size,cp_size,dp_size,ep_size,vp_size,mbs,gbs
pre_train,nemotron4,15b,h100,fp8,128.0,4096.0,4.0,1.0,1.0,32.0,1.0,1.0,4.0,1024.0
pre_train,nemotron4,15b,h100,fp8,256.0,4096.0,4.0,1.0,1.0,64.0,1.0,1.0,4.0,1024.0
pre_train,nemotron4,15b,h100,fp8,64.0,4096.0,4.0,1.0,1.0,16.0,1.0,1.0,4.0,1024.0
pre_train,nemotron4,15b,h100,fp8,16.0,4096.0,4.0,1.0,1.0,4.0,1.0,1.0,4.0,1024.0
pre_train,nemotron4,15b,h100,fp8,512.0,4096.0,4.0,1.0,1.0,128.0,1.0,1.0,4.0,1024.0
pre_train,nemotron4,15b,h100,fp8,1024.0,4096.0,4.0,1.0,1.0,256.0,1.0,1.0,4.0,1024.0
pre_train,nemotron4,340b,h100,fp8,128.0,4096.0,8.0,8.0,1.0,2.0,1.0,12.0,1.0,512.0
pre_train,nemotron4,340b,h100,fp8,256.0,4096.0,8.0,8.0,1.0,4.0,1.0,12.0,1.0,512.0
pre_train,nemotron4,340b,h100,fp8,512.0,4096.0,8.0,8.0,1.0,8.0,1.0,12.0,1.0,512.0
pre_train,nemotron4,340b,h100,fp8,1024.0,4096.0,8.0,8.0,1.0,16.0,1.0,12.0,1.0,512.0
task,model,size,system,dtype,num_gpus,seq_len,tp_size,pp_size,cp_size,dp_size,ep_size,vp_size,mbs,gbs,etp_size
pre_train,nemotron4,15b,h100,fp8,128.0,4096.0,4.0,1.0,1.0,32.0,1.0,1.0,4.0,1024.0,
pre_train,nemotron4,15b,h100,fp8,256.0,4096.0,4.0,1.0,1.0,64.0,1.0,1.0,4.0,1024.0,
pre_train,nemotron4,15b,h100,fp8,64.0,4096.0,4.0,1.0,1.0,16.0,1.0,1.0,4.0,1024.0,
pre_train,nemotron4,15b,h100,fp8,16.0,4096.0,4.0,1.0,1.0,4.0,1.0,1.0,4.0,1024.0,
pre_train,nemotron4,15b,h100,fp8,512.0,4096.0,4.0,1.0,1.0,128.0,1.0,1.0,4.0,1024.0,
pre_train,nemotron4,15b,h100,fp8,1024.0,4096.0,4.0,1.0,1.0,256.0,1.0,1.0,4.0,1024.0,
pre_train,nemotron4,340b,h100,fp8,128.0,4096.0,8.0,8.0,1.0,2.0,1.0,12.0,1.0,512.0,
pre_train,nemotron4,340b,h100,fp8,256.0,4096.0,8.0,8.0,1.0,4.0,1.0,12.0,1.0,512.0,
pre_train,nemotron4,340b,h100,fp8,512.0,4096.0,8.0,8.0,1.0,8.0,1.0,12.0,1.0,512.0,
pre_train,nemotron4,340b,h100,fp8,1024.0,4096.0,8.0,8.0,1.0,16.0,1.0,12.0,1.0,512.0,
Loading