[Bugfix] Shard UniformTypeKVCacheSpecs block table width under DCP (#50823)

Signed-off-by: Mikhail Kostryukov <mike@triptrack.net>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
drakosha
2026-08-03 17:34:29 +00:00
committed by GitHub
co-authored by Claude Matthew Bonanni
parent 5df9999fcf
commit f0de1a604c
2 changed files with 42 additions and 0 deletions
+28
View File
@@ -1257,6 +1257,34 @@ def test_project_kv_cache_groups_to_worker():
assert set(proj_spec.kv_cache_specs.keys()) == {"layer1", "layer3"}
@pytest.mark.parametrize(
"layer_type,dcp_size,expected_width",
[
("mla", 1, 64),
("mla", 2, 32),
# Mamba state is replicated, not DCP-sharded, and its width is the
# resident state block count rather than cdiv(max_len, block_size).
("mamba", 2, 3),
],
)
def test_uniform_type_spec_block_table_width_matches_layer_spec(
layer_type, dcp_size, expected_width
):
# The runner sizes the block table from the group spec while the metadata
# builders are constructed from the per-layer spec, so the aggregate must
# report the same width as the layers it wraps.
vllm_config = VllmConfig(model_config=ModelConfig(max_model_len=1024))
vllm_config.parallel_config.decode_context_parallel_size = dcp_size
layer_spec = new_mla_spec() if layer_type == "mla" else new_mamba_spec()
uniform_spec = UniformTypeKVCacheSpecs(
block_size=layer_spec.block_size,
kv_cache_specs={"layer1": layer_spec, "layer2": layer_spec},
)
assert layer_spec.max_num_blocks_per_req(vllm_config, 1024) == expected_width
assert uniform_spec.max_num_blocks_per_req(vllm_config, 1024) == expected_width
def test_merge_kv_cache_spec():
same_layer_specs = [
new_kv_cache_spec(num_kv_heads=32),