Skip to content

[Fix] Avoid pickling entire LLMEngine for Ray workers #3207

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
Yard1 merged 4 commits into from
Mar 6, 2024
Merged

[Fix] Avoid pickling entire LLMEngine for Ray workers #3207

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
96a0707
[Fix] Avoid pickling entire LLMEngine for Ray workers
njhill Mar 5, 2024
0893465
Add __reduce__ method to LLMEngine per @simon-mo's suggestion
njhill Mar 5, 2024
ee9644f
Update vllm/engine/llm_engine.py
Yard1 Mar 5, 2024
3085f15
Update vllm/engine/llm_engine.py
Yard1 Mar 5, 2024
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
21 changes: 14 additions & 7 deletions vllm/engine/llm_engine.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,6 +158,11 @@ def __init__(
if USE_RAY_COMPILED_DAG:
self.forward_dag = self._compiled_ray_dag()

def __reduce__(self):
# This is to ensure that the LLMEngine is not referenced in
# the closure used to initialize Ray worker actors
raise RuntimeError("LLMEngine should not be pickled!")

def get_tokenizer_for_seq(self, sequence: Sequence):
return self.tokenizer.get_lora_tokenizer(sequence.lora_request)

Expand Down Expand Up @@ -280,6 +285,8 @@ def _init_workers_ray(self, placement_group: "PlacementGroup",
parallel_config = copy.deepcopy(self.parallel_config)
scheduler_config = copy.deepcopy(self.scheduler_config)
device_config = copy.deepcopy(self.device_config)
lora_config = copy.deepcopy(self.lora_config)
kv_cache_dtype = self.cache_config.cache_dtype

for rank, (worker, (node_id,
_)) in enumerate(zip(self.workers,
Expand All @@ -295,22 +302,22 @@ def _init_workers_ray(self, placement_group: "PlacementGroup",
local_rank,
rank,
distributed_init_method,
lora_config=self.lora_config,
kv_cache_dtype=self.cache_config.cache_dtype,
lora_config=lora_config,
kv_cache_dtype=kv_cache_dtype,
))

driver_rank = 0
driver_local_rank = node_workers[driver_node_id].index(driver_rank)
self.driver_worker = Worker(
model_config,
parallel_config,
scheduler_config,
device_config,
self.model_config,
self.parallel_config,
self.scheduler_config,
self.device_config,
driver_local_rank,
driver_rank,
distributed_init_method,
lora_config=self.lora_config,
kv_cache_dtype=self.cache_config.cache_dtype,
kv_cache_dtype=kv_cache_dtype,
is_driver_worker=True,
)

Expand Down