[GLM-5.3-Flash] Route kpool indexer top-k through the shared SparseIndexerTopk dispatcher (#57546)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
Chauncey
2026-09-20 11:17:59 +00:00
committed by GitHub
parent f648eed23d
commit bf01fc4a31
4 changed files with 94 additions and 83 deletions
+23 -8
View File
@@ -1871,17 +1871,32 @@ def test_sparse_indexer_topk_backend_resolution() -> None:
has_coop = _has_cooperative_topk()
has_fi = _has_flashinfer_topk()
# "auto" is exactly the pre-existing chain: cooperative -> persistent ->
# per_row. It must never select the opt-in backends by itself.
# "auto" is the chain cooperative -> persistent -> per_row, with the
# cooperative/persistent switch at cooperative_topk's cluster-wave row
# limit. It must never select the opt-in backends by itself.
if has_coop:
assert resolve("auto") == "cooperative"
from vllm.model_executor.layers.indexer_topk import (
AUTO_COOPERATIVE_MAX_ROWS,
)
# cooperative for every batch within its cluster-wave row limit...
assert (
resolve("auto", k=512, num_rows=AUTO_COOPERATIVE_MAX_ROWS) == "cooperative"
)
assert resolve("auto", k=512, num_rows=8) == "cooperative"
# Past cooperative's 64-row limit -> persistent, even where
# DeepSelect would be applicable.
# ...persistent past the row limit, even where DeepSelect would be
# applicable.
assert (
resolve("auto", k=512, num_rows=AUTO_COOPERATIVE_MAX_ROWS + 1)
== "persistent"
)
assert resolve("auto", k=512, num_rows=128) == "persistent"
# 16B-aligned (cooperative TMA ok) but not DeepSelect-aligned:
# still cooperative at <= 64 rows...
assert resolve("auto", unaligned_logits, num_rows=64) == "cooperative"
# 16B-aligned (cooperative TMA ok) but not DeepSelect-aligned: the
# same row-limit logic applies.
assert (
resolve("auto", unaligned_logits, num_rows=AUTO_COOPERATIVE_MAX_ROWS)
== "cooperative"
)
# ...and persistent past the row limit.
assert resolve("auto", unaligned_logits, num_rows=128) == "persistent"
# Unsupported topk (outside {512, 1024, 2048} for the workspace kernels)
@@ -1,11 +1,11 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Focused unit tests for the GLM decode top-k dispatch."""
"""Tests for the GLM kpool indexer's top-k backend wiring."""
import pytest
import torch
from vllm.models.glm5next.nvidia.sparse_indexer import _use_cooperative_topk
from vllm.config import VllmConfig, set_current_vllm_config
from vllm.platforms import current_platform
pytestmark = pytest.mark.skipif(
@@ -13,29 +13,31 @@ pytestmark = pytest.mark.skipif(
)
@pytest.mark.parametrize("num_rows", [1, 64], ids=["1row", "64rows"])
@torch.inference_mode()
def test_cooperative_topk_is_selected_for_supported_batch(num_rows: int):
logits = torch.empty(num_rows, 8192, dtype=torch.float32, device="cuda")
assert _use_cooperative_topk(logits, 512, num_rows)
def _require_deep_gemm() -> None:
from vllm.utils.deep_gemm import has_deep_gemm
if not has_deep_gemm():
pytest.skip("kpool indexer requires DeepGEMM")
@pytest.mark.parametrize(
("num_rows", "select_k", "stride_padding"),
[
(65, 512, 0),
(1, 256, 0),
(1, 512, 1),
],
ids=["too_many_rows", "unsupported_k", "unaligned_stride"],
)
@torch.inference_mode()
def test_cooperative_topk_falls_back_on_unsupported_batch(
num_rows: int, select_k: int, stride_padding: int
):
logits = torch.empty(
num_rows, 8192 + stride_padding, dtype=torch.float32, device="cuda"
)
if stride_padding:
logits = logits[:, :8192]
assert not _use_cooperative_topk(logits, select_k, num_rows)
@pytest.mark.parametrize("backend", ["auto", "persistent", "cooperative", "torch"])
def test_kpool_indexer_dispatches_through_shared_topk_backend(backend: str) -> None:
"""The kpool indexer must read kernel_config.sparse_indexer_topk_backend
and hand it to the shared SparseIndexerTopk dispatcher, rather than
hard-coding its own cooperative/persistent/per_row choice."""
_require_deep_gemm()
from vllm.models.glm5next.nvidia.sparse_indexer import SparseAttnIndexerKpool
cfg = VllmConfig(kernel_config={"sparse_indexer_topk_backend": backend})
with set_current_vllm_config(cfg):
op = SparseAttnIndexerKpool(
k_cache=None,
quant_block_size=128,
scale_fmt="ue8m0",
topk_tokens=2048,
head_dim=128,
max_pool_len=4096,
max_total_seq_len=8192,
topk_indices_buffer=torch.empty(8, 2176, dtype=torch.int32, device="cuda"),
)
assert op.topk_backend == backend