forked from Karylab-cklius/vllm
146 lines
4.1 KiB
Python
146 lines
4.1 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
import pytest
|
|
import torch
|
|
|
|
|
|
def test_sparse_flashmla_metadata_smoke():
|
|
import vllm.v1.attention.ops.flashmla as fm
|
|
|
|
ok, reason = fm.is_flashmla_sparse_supported()
|
|
if not ok:
|
|
pytest.skip(reason)
|
|
|
|
device = torch.device("cuda")
|
|
batch_size = 1
|
|
seqlen_q = 1
|
|
num_heads_q = 128
|
|
num_heads_k = 1
|
|
q_seq_per_hk = seqlen_q * num_heads_q // num_heads_k
|
|
topk = 128
|
|
|
|
cache_seqlens = torch.zeros(batch_size, dtype=torch.int32, device=device)
|
|
|
|
tile_md, num_splits = fm.get_mla_metadata(
|
|
cache_seqlens,
|
|
q_seq_per_hk,
|
|
num_heads_k,
|
|
num_heads_q=num_heads_q,
|
|
topk=topk,
|
|
is_fp8_kvcache=True,
|
|
)
|
|
assert isinstance(tile_md, fm.FlashMLASchedMeta)
|
|
assert tile_md.tile_scheduler_metadata is None
|
|
assert tile_md.num_splits is None
|
|
assert num_splits is None
|
|
|
|
|
|
def test_sparse_flashmla_decode_smoke():
|
|
import vllm.v1.attention.ops.flashmla as fm
|
|
|
|
ok, reason = fm.is_flashmla_sparse_supported()
|
|
if not ok:
|
|
pytest.skip(reason)
|
|
|
|
device = torch.device("cuda")
|
|
batch_size = 1
|
|
seqlen_q = 1
|
|
num_heads_q = 64
|
|
head_dim_k = 576
|
|
head_dim_v = 512
|
|
num_heads_k = 1
|
|
page_block_size = 64
|
|
bytes_per_token = 656
|
|
topk = 128
|
|
|
|
# Metadata
|
|
q_seq_per_hk = seqlen_q * num_heads_q // num_heads_k
|
|
# q_heads_per_hk = num_heads_q // num_heads_k
|
|
cache_seqlens = torch.zeros(batch_size, dtype=torch.int32, device=device)
|
|
tile_md, num_splits = fm.get_mla_metadata(
|
|
cache_seqlens,
|
|
q_seq_per_hk,
|
|
num_heads_k,
|
|
num_heads_q=num_heads_q,
|
|
topk=topk,
|
|
is_fp8_kvcache=True,
|
|
)
|
|
|
|
# Inputs
|
|
q = torch.zeros(
|
|
(batch_size, seqlen_q, num_heads_q, head_dim_k),
|
|
dtype=torch.bfloat16,
|
|
device=device,
|
|
)
|
|
k_cache = torch.zeros(
|
|
(1, page_block_size, num_heads_k, bytes_per_token),
|
|
dtype=torch.uint8,
|
|
device=device,
|
|
)
|
|
indices = torch.zeros(
|
|
(batch_size, seqlen_q, topk), dtype=torch.int32, device=device
|
|
)
|
|
|
|
block_table = torch.zeros((batch_size, 128), dtype=torch.int32, device=device)
|
|
out, lse = fm.flash_mla_with_kvcache(
|
|
q,
|
|
k_cache,
|
|
block_table,
|
|
cache_seqlens,
|
|
head_dim_v,
|
|
tile_md,
|
|
num_splits,
|
|
indices=indices,
|
|
is_fp8_kvcache=True,
|
|
)
|
|
assert out.shape[0] == batch_size
|
|
assert out.shape[-1] == head_dim_v
|
|
assert lse.shape[0] == batch_size
|
|
|
|
|
|
def test_sparse_flashmla_prefill_smoke():
|
|
import vllm.v1.attention.ops.flashmla as fm
|
|
|
|
ok, reason = fm.is_flashmla_sparse_supported()
|
|
if not ok:
|
|
pytest.skip(reason)
|
|
|
|
device = torch.device("cuda")
|
|
s_q = 1
|
|
s_kv = 1
|
|
h_q = 64 # kernel expects multiple of 64
|
|
h_kv = 1
|
|
d_qk = 576
|
|
d_v = 512
|
|
topk = 128
|
|
|
|
q = torch.zeros((s_q, h_q, d_qk), dtype=torch.bfloat16, device=device)
|
|
kv = torch.zeros((s_kv, h_kv, d_qk), dtype=torch.bfloat16, device=device)
|
|
indices = torch.zeros((s_q, h_kv, topk), dtype=torch.int32, device=device)
|
|
|
|
out, max_logits, lse = fm.flash_mla_sparse_fwd(q, kv, indices, 1.0, d_v)
|
|
assert out.shape == (s_q, h_q, d_v)
|
|
assert max_logits.shape == (s_q, h_q)
|
|
assert lse.shape == (s_q, h_q)
|
|
|
|
|
|
def test_deepseek_v4_prefill_chunk_planning_expands_for_short_sequences():
|
|
from vllm.v1.attention.backends.mla.sparse_swa import DeepseekSparseSWAMetadata
|
|
|
|
metadata = DeepseekSparseSWAMetadata(
|
|
block_table=torch.empty(0, dtype=torch.int32),
|
|
slot_mapping=torch.empty(0, dtype=torch.int32),
|
|
block_size=64,
|
|
num_prefills=5,
|
|
prefill_seq_lens_cpu=torch.tensor([80, 96, 112, 128, 144], dtype=torch.int32),
|
|
prefill_query_lens_cpu=torch.tensor([4, 4, 4, 4, 4], dtype=torch.int32),
|
|
prefill_window_size=64,
|
|
prefill_max_model_len=1024,
|
|
prefill_max_num_batched_tokens=128,
|
|
)
|
|
|
|
chunk_plan = metadata.get_prefill_chunk_plan(compress_ratio=4, prefill_chunk_size=4)
|
|
|
|
# the adaptive plan keeps all 5 in one chunk
|
|
assert chunk_plan == [(0, 5, 36, 103)]
|