[DSA] k-pool: 256-token logical page so index page id = logical page id (#42178)
Co-authored-by: Cursor <cursoragent@cursor.com> Co-authored-by: Ke Bao <ispobaoke@gmail.com>
This commit is contained in:
co-authored by
Cursor
Ke Bao
parent
c765f8818a
commit
61ba98443d
@@ -24,7 +24,7 @@ register_cuda_ci(est_time=15, stage="base-b-kernel-unit", runner_config="1-gpu-l
|
||||
@unittest.skipUnless(torch.cuda.is_available(), "Test requires CUDA")
|
||||
class TestDsaKpoolMultiPool(CustomTestCase):
|
||||
POOL_SIZE = 4
|
||||
PAGE_SIZE = 64
|
||||
PAGE_SIZE = 256
|
||||
SLOTS_PER_PAGE = 64
|
||||
NUM_DRAFT_TOKENS = 6
|
||||
|
||||
@@ -64,10 +64,10 @@ class TestDsaKpoolMultiPool(CustomTestCase):
|
||||
dtype=torch.int32,
|
||||
device="cuda",
|
||||
)
|
||||
real_page_table[:num_draft_tokens, 0] = 2
|
||||
real_page_table[:num_draft_tokens, 4] = 3
|
||||
real_page_table[num_draft_tokens:, 0] = 5
|
||||
real_page_table[num_draft_tokens:, 4] = 6
|
||||
real_page_table[:num_draft_tokens, 0] = 2 * self.POOL_SIZE
|
||||
real_page_table[:num_draft_tokens, 4] = 3 * self.POOL_SIZE
|
||||
real_page_table[num_draft_tokens:, 0] = 5 * self.POOL_SIZE
|
||||
real_page_table[num_draft_tokens:, 4] = 6 * self.POOL_SIZE
|
||||
|
||||
update_kpool_write_plan_cuda_graph(
|
||||
write_start=write_start,
|
||||
|
||||
@@ -24,7 +24,7 @@ class TestKpoolPreshuffledCacheLayout(CustomTestCase):
|
||||
"""A layout mismatch between the two writers never raises; it silently
|
||||
returns the wrong top-k."""
|
||||
torch.manual_seed(0)
|
||||
pool = SimpleNamespace(page_size=PAGE_SIZE)
|
||||
pool = SimpleNamespace(page_size=4 * PAGE_SIZE, slots_per_page=PAGE_SIZE)
|
||||
num_pages, seq_len = 2, PAGE_SIZE + 5
|
||||
buf = torch.zeros(
|
||||
num_pages,
|
||||
|
||||
@@ -671,6 +671,7 @@ class TestKPoolPerRequestChunking(CustomTestCase):
|
||||
get_seqlens_expanded=lambda: seqlens_expanded,
|
||||
topk_transform_method=TopkTransformMethod.RAGGED,
|
||||
attn_metadata=SimpleNamespace(
|
||||
page_size=64,
|
||||
topk_indices_offset=torch.arange(token_nums, dtype=torch.int32) + 100,
|
||||
),
|
||||
)
|
||||
@@ -689,7 +690,8 @@ class TestKPoolPerRequestChunking(CustomTestCase):
|
||||
index_kpool=self.POOL,
|
||||
)
|
||||
with _chunking_patches(
|
||||
rows_per_chunk=rows_per_chunk, kv_pool=SimpleNamespace(page_size=64)
|
||||
rows_per_chunk=rows_per_chunk,
|
||||
kv_pool=SimpleNamespace(page_size=256, slots_per_page=64),
|
||||
):
|
||||
result = IndexerKPool._get_topk_ragged_kpool(
|
||||
backend,
|
||||
|
||||
@@ -268,6 +268,8 @@ class TestDraftSidecarPoolDispatch(CustomTestCase):
|
||||
draft_kv_pool.layer_num = 1
|
||||
draft_kv_pool.size = 800
|
||||
draft_kv_pool.index_head_dim = 128
|
||||
draft_kv_pool.page_size = 64
|
||||
draft_kv_pool.slots_per_page = 64
|
||||
draft_kv_pool.skip_topk_layers = [False]
|
||||
draft_kv_pool.index_key_cache = SimpleNamespace(buffer=[object()])
|
||||
draft_host_pool = SimpleNamespace(layer_num=1)
|
||||
@@ -333,6 +335,8 @@ def _dsa_pool_stub(*, layer_num: int, size: int = 4096, shard: tuple | None = No
|
||||
pool.qk_rope_head_dim = 64
|
||||
pool.kv_cache_dim = 576
|
||||
pool.index_head_dim = 128
|
||||
pool.page_size = 64
|
||||
pool.slots_per_page = 64
|
||||
pool.skip_topk_layers = [False] * layer_num
|
||||
pool.index_key_cache = SimpleNamespace(buffer=[object()] * layer_num)
|
||||
pool.layer_shard_enabled = shard is not None
|
||||
|
||||
@@ -483,6 +483,8 @@ class TestDSAIndexerPoolDecl(CustomTestCase):
|
||||
qk_rope_head_dim=64,
|
||||
index_head_dim=128,
|
||||
quant_block_size=128,
|
||||
page_size=64,
|
||||
slots_per_page=64,
|
||||
skip_topk_layers=[False] * 5,
|
||||
)
|
||||
|
||||
|
||||
@@ -83,6 +83,8 @@ class TestMLAHostDedupPrimitives(unittest.TestCase):
|
||||
end_layer=2,
|
||||
index_head_dim=8,
|
||||
quant_block_size=4,
|
||||
page_size=2,
|
||||
slots_per_page=2,
|
||||
skip_topk_layers=[False] * 2,
|
||||
)
|
||||
indexer_host = DSAIndexerPoolHost(
|
||||
|
||||
Reference in New Issue
Block a user