biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[V0 Deprecation] Remove pooling model support in V0 (#23434)

Browse Source 
Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
This commit is contained in:
Maximilien de Bayser
2025-08-29 04:04:02 -03:00
committed by GitHub
parent 934bebf192
commit 2554b27baa
38 changed files with 99 additions and 808 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
tests/worker/test_model_input.py
View File

@@ -9,10 +9,7 @@ from vllm.attention import AttentionMetadata, AttentionMetadataBuilder
from vllm.attention.backends.abstract import AttentionBackend
from vllm.attention.backends.utils import CommonAttentionState
from vllm.model_executor import SamplingMetadata
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.worker.model_runner import ModelInputForGPUWithSamplingMetadata
from vllm.worker.pooling_model_runner import (
ModelInputForGPUWithPoolingMetadata)
class MockAttentionBackend(AttentionBackend):
@@ -114,54 +111,3 @@ def test_model_runner_input():
assert (received_model_input.sampling_metadata.selected_token_indices ==
sampling_metadata.selected_token_indices)
assert received_model_input.sampling_metadata.seq_groups is None
def test_embedding_model_runner_input():
pooling_metadata = PoolingMetadata(
seq_groups=[[0]],
seq_data={},
prompt_lens=[1],
)
attn_metadata = AttentionMetadata(
num_prefills=1,
num_prefill_tokens=2,
num_decode_tokens=3,
slot_mapping=torch.zeros(1),
multi_modal_placeholder_index_maps=None,
enable_kv_scales_calculation=True,
)
model_input = ModelInputForGPUWithPoolingMetadata(
input_tokens=torch.ones(10),
input_positions=torch.ones(10),
pooling_metadata=pooling_metadata,
attn_metadata=attn_metadata)
assert isinstance(model_input, ModelInputForGPUWithPoolingMetadata)
# Test round trip serialization.
tensor_dict = model_input.as_broadcastable_tensor_dict()
attn_backend = MockAttentionBackend()
received_model_input = (
ModelInputForGPUWithPoolingMetadata.from_broadcasted_tensor_dict(
tensor_dict, attn_backend=attn_backend))
# Check that received copy has correct values.
assert isinstance(received_model_input,
ModelInputForGPUWithPoolingMetadata)
assert received_model_input.input_tokens is not None
assert (
received_model_input.input_tokens == model_input.input_tokens).all()
assert received_model_input.input_positions is not None
assert (received_model_input.input_positions == model_input.input_positions
).all()
assert received_model_input.multi_modal_kwargs is None
assert (received_model_input.multi_modal_kwargs ==
model_input.multi_modal_kwargs)
assert received_model_input.lora_requests is None
assert received_model_input.lora_requests == model_input.lora_requests
assert received_model_input.lora_mapping is None
assert received_model_input.lora_mapping == model_input.lora_mapping
for field in dataclasses.fields(AttentionMetadata):
assert getattr(received_model_input.attn_metadata, field.name,
None) == getattr(attn_metadata, field.name, None)
# Pooling metadata is not broadcast.
assert received_model_input.pooling_metadata is None