[Bugfix][MRV2] Keep GDN prefill checkpoint metadata local to each cache group (#59536)

Signed-off-by: Jingqiao Zhang <ai-jz@users.noreply.github.com>
Signed-off-by: zjy0516 <riverclouds.zhu@qq.com>
Co-authored-by: Jingqiao Zhang <ai-jz@users.noreply.github.com>
Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
This commit is contained in:
ai-jz
2026-10-01 22:59:15 +08:00
committed by GitHub
co-authored by Jingqiao Zhang zjy0516
parent 5d9214d6b2
commit b538d807ff
3 changed files with 83 additions and 18 deletions
@@ -17,6 +17,7 @@ from tests.v1.attention.utils import (
)
from vllm.config import SpeculativeConfig
from vllm.config.compilation import CUDAGraphMode
from vllm.v1.attention.backend import CommonAttentionMetadata
from vllm.v1.attention.backends.gdn_attn import (
GDNAttentionMetadata,
GDNAttentionMetadataBuilder,
@@ -383,14 +384,9 @@ def test_cudagraph_capture_batch_stays_decode_only():
torch.testing.assert_close(staged, common_attn_metadata.block_table_tensor[:, 0])
@pytest.mark.parametrize("num_spec", [0, 3])
@pytest.mark.parametrize("prefix_match_unit, expected_offset", [(16, 96), (8, 96)])
def test_checkpoint_metadata_preserves_non_spec_order(
num_spec, prefix_match_unit, expected_offset
):
"""Only eligible non-spec rows get a checkpoint in their reserved page."""
from vllm.v1.attention.backends.utils import NULL_BLOCK_ID
def _create_checkpoint_builder_and_batch(
num_spec: int, prefix_match_unit: int = 16
) -> tuple[GDNAttentionMetadataBuilder, CommonAttentionMetadata, dict]:
vllm_config = create_vllm_config(
model_name="Qwen/Qwen3.5-0.8B",
block_size=BLOCK_SIZE,
@@ -431,7 +427,20 @@ def test_checkpoint_metadata_preserves_non_spec_order(
num_decode_draft_tokens_cpu=torch.tensor([2, -1, -1], dtype=torch.int32),
num_accepted_tokens=torch.ones(3, dtype=torch.int32),
)
return builder, common, kwargs
@pytest.mark.parametrize("num_spec", [0, 3])
@pytest.mark.parametrize("prefix_match_unit, expected_offset", [(16, 96), (8, 96)])
def test_checkpoint_metadata_preserves_non_spec_order(
num_spec, prefix_match_unit, expected_offset
):
"""Only eligible non-spec rows get a checkpoint in their reserved page."""
from vllm.v1.attention.backends.utils import NULL_BLOCK_ID
builder, common, kwargs = _create_checkpoint_builder_and_batch(
num_spec, prefix_match_unit
)
actual = builder.build(0, common, **kwargs)
assert actual.checkpoint is not None
offsets = [expected_offset, 0] if num_spec else [0, expected_offset, 0]
@@ -444,3 +453,34 @@ def test_checkpoint_metadata_preserves_non_spec_order(
torch.testing.assert_close(
actual.checkpoint.state_indices, torch.tensor(slots, dtype=torch.int32)
)
@pytest.mark.parametrize("num_spec", [0, 3])
def test_update_block_table_regathers_checkpoint(num_spec):
"""Checkpoint pages come from the target group's block table, not the
source group's (otherwise the target group's checkpoint page is never
written while the prefix cache treats it as valid)."""
src, common, kwargs = _create_checkpoint_builder_and_batch(num_spec)
dst, _, _ = _create_checkpoint_builder_and_batch(num_spec)
ref, _, _ = _create_checkpoint_builder_and_batch(num_spec)
other_table = common.block_table_tensor + 100
other = common.replace(block_table_tensor=other_table)
dst.mamba_aligned_state_indices = mamba_get_block_table_tensor(
other_table, other.seq_lens, dst.kv_cache_spec, "align"
)
source = src.build(0, common, **kwargs)
expected = ref.build(0, other, **kwargs)
actual = dst.update_block_table(source, other_table, other.slot_mapping)
assert source.checkpoint is not None and expected.checkpoint is not None
assert actual.checkpoint is not None
torch.testing.assert_close(
actual.checkpoint.state_indices, expected.checkpoint.state_indices
)
torch.testing.assert_close(
actual.checkpoint.checkpoint_offsets, expected.checkpoint.checkpoint_offsets
)
assert not torch.equal(
actual.checkpoint.state_indices, source.checkpoint.state_indices
)