Skip to content

[rollout] feat: Add vllm logprob mode and default processed_logprob #4755

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
vermouth1992 merged 4 commits into from
Dec 31, 2025
+7 −0
Merged

[rollout] feat: Add vllm logprob mode and default processed_logprob #4755

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
7701dde
[rollout] feat: add logprobs mode and default processed_logprobs for …
Dec 31, 2025
2f9a399
Merge branch 'volcengine:main' into add_vllm_logprob_mode
RobotGF Dec 31, 2025
6f1308f
Update verl/workers/config/rollout.py
RobotGF Dec 31, 2025
7264045
add default in generate config yaml
Dec 31, 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
1 change: 1 addition & 0 deletions verl/trainer/config/_generated_ppo_megatron_trainer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -220,6 +220,7 @@ actor_rollout_ref:
max_num_seqs: 1024
enable_chunked_prefill: true
enable_prefix_caching: true
logprobs_mode: processed_logprobs
load_format: dummy
log_prob_micro_batch_size: null
log_prob_micro_batch_size_per_gpu: null
Expand Down
1 change: 1 addition & 0 deletions verl/trainer/config/_generated_ppo_trainer.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -209,6 +209,7 @@ actor_rollout_ref:
max_num_seqs: 1024
enable_chunked_prefill: true
enable_prefix_caching: true
logprobs_mode: processed_logprobs
load_format: dummy
log_prob_micro_batch_size: null
log_prob_micro_batch_size_per_gpu: null
Expand Down
3 changes: 3 additions & 0 deletions verl/trainer/config/rollout/rollout.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -73,6 +73,9 @@ enable_chunked_prefill: True
# Prefix caching kv-cache blocks is a popular optimization in LLM inference to avoid redundant prompt computations.
enable_prefix_caching: True

# logprobs mode for rollout logprobs
logprobs_mode: processed_logprobs

# Which loader to use for rollout model weights: dummy, hf, megatron, etc.
# safetensors (for huge model, and set use_shm=True); dummy: randomly init model weight
load_format: dummy
Expand Down
1 change: 1 addition & 0 deletions verl/workers/config/rollout.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -149,6 +149,7 @@ class RolloutConfig(BaseConfig):
tensor_model_parallel_size: int = 2
pipeline_model_parallel_size: int = 1
max_num_batched_tokens: int = 8192
logprobs_mode: Optional[str] = "processed_logprobs"

# TODO: enable train_kwargs
# train_sampling_config: SamplingConfig = field(default_factory=SamplingConfig)
Expand Down
1 change: 1 addition & 0 deletions verl/workers/rollout/vllm_rollout/vllm_async_server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -299,6 +299,7 @@ async def launch_server(self, master_address: str = None, master_port: int = Non
"max_num_batched_tokens": self.config.max_num_batched_tokens,
"enable_prefix_caching": self.config.enable_prefix_caching,
"enable_sleep_mode": self.config.enable_sleep_mode,
"logprobs_mode": self.config.logprobs_mode,
"disable_custom_all_reduce": True,
"enforce_eager": self.config.enforce_eager,
"gpu_memory_utilization": self.config.gpu_memory_utilization,
Expand Down
Loading