[1/N][warmup][DSv4] Migrate generic MLA metadata and indexing kernels (#50175)

Signed-off-by: LopezCastroRoberto <rocastro@redhat.com>
Co-authored-by: Codex <codex@openai.com>
This commit is contained in:
Roberto L. Castro
2026-08-31 23:26:05 -04:00
committed by GitHub
co-authored by Codex
parent 188716ace7
commit e16b5e518d
17 changed files with 1261 additions and 617 deletions
@@ -8,10 +8,16 @@ import torch
from tests.v1.attention.utils import create_vllm_config
from vllm.v1.attention.backend import CommonAttentionMetadata
from vllm.v1.attention.backends.mla.compressor_utils import (
CompressedSlotMappingKernel,
)
from vllm.v1.attention.backends.mla.indexer import (
BuildPrefillChunkMetadataKernel,
DeepseekV32IndexerMetadataBuilder,
)
from vllm.v1.attention.backends.mla.sparse_utils import (
ConvertReqIndexToGlobalIndexKernel,
)
from vllm.v1.kv_cache_interface import MLAAttentionSpec
from vllm.v1.worker.block_table import get_block_table_width
@@ -20,7 +26,7 @@ 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])
hf_config=SimpleNamespace(compress_ratios=[0, 0, 4, 128, 0], index_kpool=32)
),
parallel_config=SimpleNamespace(
decode_context_parallel_size=1,
@@ -30,7 +36,42 @@ def test_indexer_warmup_normalizes_zero_compress_ratios():
keys = BuildPrefillChunkMetadataKernel().get_warmup_keys(config)
assert {key.COMPRESS_RATIO for key in keys} == {1, 4, 128}
assert {key.compress_ratio for key in keys} == {1, 4, 32, 128}
assert {(key.query_slice_start, key.query_slice_stop) for key in keys} == {
(query_slice_start, query_slice_stop)
for query_slice_start in (1, 2, 16)
for query_slice_stop in (1, 2, 16)
}
def test_compressed_slot_mapping_warmup_includes_index_kpool():
config = SimpleNamespace(
cache_config=SimpleNamespace(block_size=256),
model_config=SimpleNamespace(hf_config=SimpleNamespace(index_kpool=32)),
)
keys = CompressedSlotMappingKernel().get_warmup_keys(config)
assert {(key.compress_ratio, key.block_size) for key in keys} == {(32, 2)}
def test_index_conversion_warmup_uses_physical_block_stride():
config = SimpleNamespace(
cache_config=SimpleNamespace(block_size=64),
model_config=SimpleNamespace(
max_model_len=1024,
hf_config=SimpleNamespace(index_topk=2048),
),
parallel_config=SimpleNamespace(
decode_context_parallel_size=1,
cp_kv_cache_interleave_size=1,
),
)
keys = ConvertReqIndexToGlobalIndexKernel().get_warmup_keys(
config,
block_stride_rows=4096,
)
assert {key.block_stride_rows for key in keys} == {4096}
@pytest.mark.skipif(not torch.cuda.is_available(), reason="requires CUDA")
@@ -7,6 +7,7 @@ from typing import Any
import torch
from vllm.model_executor.warmup.jit_warmup import JitWarmupRegistry
from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT
from vllm.v1.worker.gpu import eplb_utils as eplb
from vllm.v1.worker.gpu import model_runner as mrv2
@@ -84,6 +85,7 @@ def _make_runner(**overrides: Any) -> Any:
post_forward=lambda *_, **__: None,
)
runner.eplb = eplb.EPLBController(runner.parallel_config, runner.device)
runner.jit_warmup_registry = JitWarmupRegistry(runner.vllm_config)
runner.pooling_runner = None
runner.execute_model_state = None
for key, value in overrides.items():