Skip to content

[CB] Support pseudo batch size 1 for decode, adjust warmup #287

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 14 commits into from
Jul 29, 2025
Merged

[CB] Support pseudo batch size 1 for decode, adjust warmup #287

Show file tree
Hide file tree
Changes from 5 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
a52c7ec
:test_tube: add test for cb @ bs 1
joerunde Jul 7, 2025
1797ea3
:bug: update max batch size to 2 for cb abort test
joerunde Jul 7, 2025
9561a9d
remove assertion batch size >= 2 (for warmup)
yannicks1 Jul 8, 2025
10aed51
fix batch size 1 warmup
yannicks1 Jul 8, 2025
78d864a
sorting imports
yannicks1 Jul 8, 2025
ff7b32a
Merge branch 'main' into cb-batch-size-1-test
yannicks1 Jul 9, 2025
55f4efa
Merge branch 'main' into cb-batch-size-1-test
yannicks1 Jul 10, 2025
b392e23
Merge branch 'main' into cb-batch-size-1-test
yannicks1 Jul 16, 2025
3959678
Merge branch 'main' into cb-batch-size-1-test
yannicks1 Jul 17, 2025
d55a23f
add test for max_num_seqs = 1
yannicks1 Jul 18, 2025
b2c96dd
Merge branch 'main' into cb-batch-size-1-test
yannicks1 Jul 24, 2025
c8ebad6
fix! setting VLLM_DT_MAX_BATCH_SIZE env for compiler to 2 !
yannicks1 Jul 24, 2025
c423480
Merge branch 'main' into cb-batch-size-1-test
yannicks1 Jul 25, 2025
9a2c516
revert adding test to reduce number of runs
yannicks1 Jul 25, 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
2 changes: 1 addition & 1 deletion tests/e2e/test_spyre_async_llm.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,7 +86,7 @@ async def test_abort(
model=model,
tokenizer=model,
max_model_len=128,
max_num_seqs=8,
max_num_seqs=2,
block_size=2048,
))
has_unfinished_requests = \
Expand Down
2 changes: 1 addition & 1 deletion tests/e2e/test_spyre_basic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -228,7 +228,7 @@ def test_full_batch_scheduling(model: str, backend: str, monkeypatch):
engine_args = EngineArgs(model=model,
tokenizer=model,
max_num_batched_tokens=max_batched_tokens,
max_num_seqs=4)
max_num_seqs=batch_size)
vllm_config = engine_args.create_engine_config()
executor_class = Executor.get_class(vllm_config)
engine_core = EngineCore(vllm_config=vllm_config,
Expand Down
2 changes: 1 addition & 1 deletion tests/e2e/test_spyre_cb.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -23,7 +23,7 @@
@pytest.mark.cb
@pytest.mark.parametrize("model", get_spyre_model_list())
@pytest.mark.parametrize("backend", get_spyre_backend_list())
@pytest.mark.parametrize("max_num_seqs", [2, 4],
@pytest.mark.parametrize("max_num_seqs", [1, 2, 4],
ids=lambda val: f"max_num_seqs({val})")
def test_cb_output(
model: str,
Expand Down
5 changes: 0 additions & 5 deletions vllm_spyre/v1/worker/spyre_model_runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -669,11 +669,6 @@ def __init__(
super().__init__(vllm_config=vllm_config,
is_driver_worker=is_driver_worker)

# TODO: remove this limitation once we update the warm-up logic to
# support batch_size=1
assert vllm_config.scheduler_config.max_num_seqs >= 2, "Currently, " \
"continuous batching needs config to set batch_size >= 2"

self.block_size = 64

# TODO: move to a KV cache manager
Expand Down
26 changes: 26 additions & 0 deletions vllm_spyre/v1/worker/spyre_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,7 @@
import vllm_spyre.perf_metrics as perf_metrics
from vllm_spyre.model_executor.model_loader import spyre_setup
from vllm_spyre.platform import SpyrePlatform
from vllm_spyre.v1.worker.spyre_input_batch import InputBatch
from vllm_spyre.v1.worker.spyre_model_runner import (
ContinuousBatchingSpyreModelRunner, StaticBatchingSpyreModelRunner)

Expand Down Expand Up @@ -317,6 +318,18 @@ def _warmup_spyre_dynamic_size(self, special_token_ids):
prompt_len = 42
num_decode_tokens = 2

# Fix for batch size 1: set input batch to fit 2 requests for warmup
if model_runner.vllm_config.scheduler_config.max_num_seqs == 1:
model_runner.input_batch = InputBatch(
Copy link
Collaborator

@joerunde joerunde Jul 8, 2025

Choose a reason for hiding this comment

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

Alternatively, could the InputBatch construct itself with:

self.max_num_reqs = min(max_num_reqs, 2)

since we know that it'll always need at least 2, and then we avoid reconstructing it in the worker here? That way we have a much smaller diff to back out once we can lift this bs>=2 restriction

Copy link
Collaborator Author

@yannicks1 yannicks1 Jul 8, 2025

Choose a reason for hiding this comment

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

not sure if I follow here. it has to be >=2 for the warmup. with the min(1,2) we would still fail?

Copy link
Collaborator

@sducouedic sducouedic Jul 8, 2025

Choose a reason for hiding this comment

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

would that work if you directly set model_runner.input_batch.max_num_reqs = 2, instead of instantiating a new InputBatch?

Copy link
Collaborator Author

@yannicks1 yannicks1 Jul 9, 2025

Choose a reason for hiding this comment

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

no, because InputBatch initialization gets model_runner.input_batch.max_num_reqs..

Copy link
Collaborator Author

@yannicks1 yannicks1 Jul 10, 2025

Choose a reason for hiding this comment

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

max_num_seqs occurs 17 times in the init of the InputBatch. It is not a single attribute, but used to construct several attributes. So re-initializing is simpler...

max_num_reqs=2,
max_model_len=model_runner.vllm_config.model_config.
max_model_len,
device=model_runner.device,
pin_memory=model_runner.pin_memory,
vocab_size=model_runner.vllm_config.model_config.
get_vocab_size(),
)

# Sample from the valid token ids
warmup_tokens_tensor = valid_token_ids_tensor[torch.randint(
0, len(valid_token_ids_tensor), (batch_size + 1, prompt_len))]
Expand Down Expand Up @@ -412,6 +425,19 @@ def _warmup_spyre_dynamic_size(self, special_token_ids):
self.execute_model(scheduler_output)
self._cleanup_model_runner(request=[add_dummy_request])

# Fix for batch size 1: reset input batch to fit max_num_seqs requests
if model_runner.vllm_config.scheduler_config.max_num_seqs == 1:
model_runner.input_batch = InputBatch(
max_num_reqs=model_runner.vllm_config.scheduler_config.
max_num_seqs,
max_model_len=model_runner.vllm_config.model_config.
max_model_len,
device=model_runner.device,
pin_memory=model_runner.pin_memory,
vocab_size=model_runner.vllm_config.model_config.
get_vocab_size(),
)

# get the number or pages from the actual Spyre card after the warmup
# and set it accordingly in the model runner and the kv cache size
n_blocks_avail = self._get_num_blocks_available()
Expand Down
Loading