[DSv4 Perf] Skip empty c128 kernel launch, around 2x kernel performance improvement. (#48957)

Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
Wentao Ye
2026-07-22 10:55:20 -04:00
committed by GitHub
parent 2dc5a72e7e
commit 37e370fe93
2 changed files with 53 additions and 2 deletions
+21
View File
@@ -13,6 +13,7 @@ These tests cover:
"""
import math
from types import SimpleNamespace
import pytest
import torch
@@ -27,6 +28,7 @@ from vllm.models.deepseek_v4.common.ops.fused_compress_quant_cache import (
_fused_kv_compress_norm_rope_insert_indexer_mxfp4_attn,
_launch_two_stage_sparse_attn_compressor,
)
from vllm.models.deepseek_v4.compressor import _get_c128_boundary
from vllm.platforms import current_platform
from .test_fused_indexer_q_rope_quant import quantize_to_mxfp4
@@ -58,6 +60,25 @@ def _ue8m0_reference(x: torch.Tensor, block_size: int, fp8_max: float):
return x_fp8, scales
@pytest.mark.parametrize(
("starts", "query_start_loc", "expected"),
[
([0], [0, 127], False),
([0], [0, 128], True),
([127], [0, 1], True),
([128], [0, 127], False),
([1, 255], [0, 1, 2], True),
(None, [0, 1], None),
],
)
def test_get_c128_boundary(starts, query_start_loc, expected):
metadata = SimpleNamespace(
_num_computed_tokens_cpu=None if starts is None else torch.tensor(starts),
query_start_loc_cpu=torch.tensor(query_start_loc),
)
assert _get_c128_boundary(metadata) is expected
# ── Test A: DeepseekV4 Attention path ──────────────────────────────────────────────