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 6 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
7 changes: 6 additions & 1 deletion 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 +50,12 @@
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)

Check warning

Code scanning / CodeQL

Variable defined multiple times Warning

This assignment to 'recipe' is unnecessary as it is
redefined
Loading
before this value is used.
This assignment to 'recipe' is unnecessary as it is
redefined
Loading
before this value is used.
gpu_type = args.gpu.lower()
if gpu_type in ["gb200"] and finetuning_scheme == "lora":
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 +82,7 @@
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
4 changes: 4 additions & 0 deletions scripts/llm/performance/pretrain_mixtral_8x22b.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,6 +66,10 @@ 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 etp_size>1, disabling average_in_collective.
malay-nagda marked this conversation as resolved.
Show resolved Hide resolved
recipe.trainer.strategy.ddp.average_in_collective = False

return recipe


Expand Down
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,etp_size,vp_size,mbs,gbs,etp_size
malay-nagda marked this conversation as resolved.
Show resolved Hide resolved
lora,llama3,70b,gb200,fp8,8.0,2048.0,1.0,4.0,1.0,2.0,1.0,,20.0,1.0,32.0,
lora,llama3,70b,gb200,bf16,8.0,2048.0,1.0,4.0,1.0,2.0,1.0,,20.0,1.0,32.0,
lora,llama3,8b,gb200,bf16,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
lora,llama3,8b,gb200,fp8,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,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,bf16,24.0,4096.0,4.0,6.0,1.0,1.0,1.0,,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,fp8,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
sft,llama3,8b,gb200,bf16,8.0,4096.0,1.0,1.0,1.0,8.0,1.0,,1.0,1.0,8.0,
Loading