Skip to content

Add max-model-len param for vLLM #502

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

Open
oyilmaz-nvidia wants to merge 3 commits into
base: main
Choose a base branch
from
Open

Add max-model-len param for vLLM #502

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
6ffc9b3
Add max-model-len param for vLLM
oyilmaz-nvidia Nov 4, 2025
6326dd2
Merge branch 'main' into onur/add_max_model_len
oyilmaz-nvidia Nov 5, 2025
4d16afa
Fix minor test
oyilmaz-nvidia Nov 5, 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
4 changes: 4 additions & 0 deletions nemo_export/vllm_exporter.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -116,6 +116,7 @@ def export(
cpu_offload_gb: float = 0,
enforce_eager: bool = False,
max_seq_len_to_capture: int = 8192,
max_model_len: int = 8192,
Copy link
Contributor

@athitten athitten Nov 4, 2025

Choose a reason for hiding this comment

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

Thank you @oyilmaz-nvidia . For my understanding, why is this param needed now ? Is it new introduced by vLLM

Copy link
Contributor Author

@oyilmaz-nvidia oyilmaz-nvidia Nov 4, 2025

Choose a reason for hiding this comment

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

We really didn't need to set the parameter up until now but some of the large models like llama 70B might need some tuning to fit model into the GPUs. And CI is giving error now to fit this model (it was working before but with the newer versions of vllm, we might need to tune it).

task: Literal["auto", "generate", "embedding"] = "auto",
):
"""
Expand All @@ -135,6 +136,7 @@ def export(
cpu_offload_gb (float, optional): Amount of CPU offload memory (in GB). Defaults to 0.
enforce_eager (bool, optional): Whether to enforce eager execution. Defaults to False.
max_seq_len_to_capture (int, optional): Maximum sequence length to capture. Defaults to 8192.
max_model_len (int, optional): Maximum model length. Defaults to 8192.
task (Literal["auto", "generate", "embedding"], optional): Task type for vLLM. Defaults to "auto".

Raises:
Expand Down Expand Up @@ -171,6 +173,7 @@ def export(
cpu_offload_gb=cpu_offload_gb,
enforce_eager=enforce_eager,
max_seq_len_to_capture=max_seq_len_to_capture,
max_model_len=max_model_len,
task=task,
)
else:
Expand All @@ -188,6 +191,7 @@ def export(
cpu_offload_gb=cpu_offload_gb,
enforce_eager=enforce_eager,
max_seq_len_to_capture=max_seq_len_to_capture,
max_model_len=max_model_len,
task=task,
)

Expand Down
8 changes: 8 additions & 0 deletions scripts/deploy/nlp/deploy_vllm_triton.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -95,6 +95,13 @@ def get_args(argv):
type=float,
help="GPU memory utilization percentage for vLLM.",
)
parser.add_argument(
"-mml",
"--max_model_len",
default=8192,
type=int,
help="Max model length for vLLM.",
)
parser.add_argument(
"-sp",
"--swap_space",
Expand Down Expand Up @@ -198,6 +205,7 @@ def nemo_deploy(argv):
cpu_offload_gb=args.cpu_offload_gb,
enforce_eager=args.enforce_eager,
max_seq_len_to_capture=args.max_seq_len_to_capture,
max_model_len=args.max_model_len,
task="generate",
)

Expand Down