Skip to content

fix(truss): make the passing in the chat template more robust BT-15187 #1881

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
merged 4 commits into from
Aug 21, 2025
Merged

fix(truss): make the passing in the chat template more robust BT-15187 #1881

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
67c5914
fix(truss): explicitly pass in the chat_template flag and file
Aug 20, 2025
a4527e2
fix(truss): move deployments more robust
Aug 21, 2025
df83df0
fix(truss): unit test for vllm serve command
Aug 21, 2025
562433b
fix(truss): add comment
Aug 21, 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
17 changes: 12 additions & 5 deletions truss/cli/train/deploy_checkpoints/deploy_full_checkpoints.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,12 +14,19 @@
setup_environment_variables_and_secrets,
)

# NB(aghilan): Transformers was recently changed to save a chat_template.jinja file instead of inside the tokenizer_config.json file.
Copy link
Contributor

@rcano-baseten rcano-baseten Aug 21, 2025

Choose a reason for hiding this comment

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

what version? might be helpful to know when it's safe to remove this logic/change this logic?

# Old Models will not have this file, so we check for it and use it if it exists.
# vLLM will not automatically resolve the chat_template.jinja file, so we need to pass it to the start command.
# This logic is needed for any models trained using Transformers v4.51.3 or later
VLLM_FULL_START_COMMAND = Template(
'sh -c "{%if envvars %}{{ envvars }} {% endif %}vllm serve {{ model_path }}'
+ " --port 8000"
+ " --tensor-parallel-size {{ specify_tensor_parallelism }}"
+ " --dtype bfloat16"
+ '"'
"sh -c '{% if envvars %}{{ envvars }} {% endif %}"
'HF_TOKEN="$$(cat /secrets/hf_access_token)" && export HF_TOKEN && '
"if [ -f {{ model_path }}/chat_template.jinja ]; then "
" vllm serve {{ model_path }} --chat-template {{ model_path }}/chat_template.jinja "
" --port 8000 --tensor-parallel-size {{ specify_tensor_parallelism }} --dtype bfloat16; "
"else "
" vllm serve {{ model_path }} --port 8000 --tensor-parallel-size {{ specify_tensor_parallelism }} --dtype bfloat16; "
"fi'"
)


Expand Down
12 changes: 10 additions & 2 deletions truss/tests/cli/train/test_deploy_checkpoints.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -505,8 +505,16 @@ def test_render_vllm_full_truss_config():
)

result = render_vllm_full_truss_config(deploy_config)

expected_vllm_command = 'sh -c "HF_TOKEN=$(cat /secrets/hf_token) vllm serve /tmp/training_checkpoints/job123/rank-0/checkpoint-1 --port 8000 --tensor-parallel-size 2 --dtype bfloat16"'
expected_vllm_command = (
"sh -c 'HF_TOKEN=$(cat /secrets/hf_token) "
'HF_TOKEN="$$(cat /secrets/hf_access_token)" && export HF_TOKEN && '
"if [ -f /tmp/training_checkpoints/job123/rank-0/checkpoint-1/chat_template.jinja ]; then "
"vllm serve /tmp/training_checkpoints/job123/rank-0/checkpoint-1 "
"--chat-template /tmp/training_checkpoints/job123/rank-0/checkpoint-1/chat_template.jinja "
"--port 8000 --tensor-parallel-size 2 --dtype bfloat16; else "
"vllm serve /tmp/training_checkpoints/job123/rank-0/checkpoint-1 "
"--port 8000 --tensor-parallel-size 2 --dtype bfloat16; fi'"
)

assert isinstance(result, truss_config.TrussConfig)
assert result.model_name == "test-full-model"
Expand Down
Loading