[Bugfix] Normalize sparse MLA warmup compression ratios (#49392)
Signed-off-by: Wu, Xiaochang <xiaochang.wu@intel.com> Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
This commit is contained in:
@@ -1,15 +1,37 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
from types import SimpleNamespace
|
||||
|
||||
import pytest
|
||||
import torch
|
||||
|
||||
from tests.v1.attention.utils import create_vllm_config
|
||||
from vllm.v1.attention.backend import CommonAttentionMetadata
|
||||
from vllm.v1.attention.backends.mla.indexer import DeepseekV32IndexerMetadataBuilder
|
||||
from vllm.v1.attention.backends.mla.indexer import (
|
||||
BuildPrefillChunkMetadataKernel,
|
||||
DeepseekV32IndexerMetadataBuilder,
|
||||
)
|
||||
from vllm.v1.kv_cache_interface import MLAAttentionSpec
|
||||
|
||||
|
||||
def test_indexer_warmup_normalizes_zero_compress_ratios():
|
||||
config = SimpleNamespace(
|
||||
scheduler_config=SimpleNamespace(max_num_batched_tokens=8),
|
||||
model_config=SimpleNamespace(
|
||||
hf_config=SimpleNamespace(compress_ratios=[0, 0, 4, 128, 0])
|
||||
),
|
||||
parallel_config=SimpleNamespace(
|
||||
decode_context_parallel_size=1,
|
||||
cp_kv_cache_interleave_size=1,
|
||||
),
|
||||
)
|
||||
|
||||
keys = BuildPrefillChunkMetadataKernel().get_warmup_keys(config)
|
||||
|
||||
assert {key.COMPRESS_RATIO for key in keys} == {1, 4, 128}
|
||||
|
||||
|
||||
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
|
||||
def test_indexer_builder_deepseek_v4_compressed_slot_mapping_uses_storage_block_size():
|
||||
"""Regression test: DeepseekV4 compression path must compute slot_mapping from
|
||||
|
||||
@@ -328,7 +328,7 @@ class BuildPrefillChunkMetadataKernel(
|
||||
dcp_interleave = parallel_config.cp_kv_cache_interleave_size
|
||||
dcp_rank = get_dcp_group().rank_in_group if dcp_world > 1 else 0
|
||||
compress_ratios = tuple(
|
||||
int(ratio)
|
||||
max(1, int(ratio))
|
||||
for ratio in (getattr(hf_config, "compress_ratios", None) or (1,))
|
||||
)
|
||||
return self._trace_dispatch(self.dispatch)(
|
||||
|
||||
Reference in New Issue
Block a user