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

Enable Delayed Sampling by default #937

Open
wants to merge 2 commits into
base: habana_main
Choose a base branch
from
Open

Enable Delayed Sampling by default #937

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e831365
Enable Delayed Sampling by default in case multi step scheduler is no…
mswiniarsk Mar 20, 2025
b6d7a59
Disable delayed sampling for fake_hpu mode
mswiniarsk Mar 20, 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
14 changes: 7 additions & 7 deletions vllm/worker/hpu_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -82,7 +82,7 @@
LORA_WARMUP_RANK = 8

VLLM_DELAYED_SAMPLING = os.environ.get('VLLM_DELAYED_SAMPLING',
'false').lower() == 'true'
'true').lower() == 'true'
DUMMY_TOKEN_ID = -1


Expand Down Expand Up @@ -729,6 +729,8 @@ def __init__(
"Speculative decoding is not supported with "
"contiguous PA, please set VLLM_CONTIGUOUS_PA=false")
# For both multi-step scheduling and delayed sampling
self.is_single_step = \
self.vllm_config.scheduler_config.num_scheduler_steps == 1
self.cached_step_outputs: List[torch.Tensor] = []
self.is_pooler = False
# For delayed sampling
Expand Down Expand Up @@ -1873,9 +1875,7 @@ def warmup_scenario(self,
profiler.start()
for _ in range(times):
inputs = self.prepare_model_input(seqs)
is_single_step = \
self.vllm_config.scheduler_config.num_scheduler_steps == 1
if is_prompt or is_single_step:
if is_prompt or self.is_single_step:
self.execute_model(inputs, kv_caches, warmup_mode=True)
else: # decode with multi-step
inputs = dataclasses.replace(inputs,
Expand Down Expand Up @@ -2441,9 +2441,9 @@ def execute_model(
previous_hidden_states: Optional[torch.Tensor] = None,
seqs=None,
) -> Optional[Union[List[SamplerOutput], IntermediateTensors]]:
use_delayed_sampling = VLLM_DELAYED_SAMPLING and not warmup_mode
assert not (use_delayed_sampling and num_steps != 1), \
'Delayed sampling is not compatible with MSS!'
# Delayed sampling is only supported for single step scheduling
use_delayed_sampling = VLLM_DELAYED_SAMPLING and not warmup_mode \
and self.is_single_step and not is_fake_hpu()
assert model_input.input_tokens is not None
if use_delayed_sampling and not model_input.is_prompt and \
self.is_driver_worker:
Expand Down
Loading