[PD] Fix block scale registration for mixed full/SWA KV dtypes (#41924)

This commit is contained in:
Eric Zhang
2026-10-01 14:25:04 +08:00
committed by GitHub
parent 77b64e353f
commit 947fee9384
2 changed files with 65 additions and 13 deletions
+10 -13
View File
@@ -1389,19 +1389,16 @@ def setup_state_kv_args(
# so they inherit the index payload of the KV they describe.
# Only the concrete SWAKVPool owns a full sub-pool; other
# BaseSWAKVPool implementations describe their state per entry.
if isinstance(token_to_kv_pool, SWAKVPool) and isinstance(
token_to_kv_pool.full_kv_pool, MHATokenToKVPoolMXFP8
):
append_state_component(
kv_args,
StateType.BLOCK_SCALE,
*token_to_kv_pool.get_kv_scale_buf_infos(),
)
append_state_component(
kv_args,
StateType.BLOCK_SCALE_SWA,
*token_to_kv_pool.get_swa_kv_scale_buf_infos(),
)
if isinstance(token_to_kv_pool, SWAKVPool):
# Global and SWA caches may use different dtypes.
for pool, state_type in (
(token_to_kv_pool.full_kv_pool, StateType.BLOCK_SCALE),
(token_to_kv_pool.swa_kv_pool, StateType.BLOCK_SCALE_SWA),
):
if isinstance(pool, MHATokenToKVPoolMXFP8):
append_state_component(
kv_args, state_type, *pool.get_kv_scale_buf_infos()
)
# unified_kv: the SWA ring lives in the unified buffers (no separate
# swa_kv_pool) and is addressed per-row, so ship it as SWA_RING.
if getattr(token_to_kv_pool, "_unified_kv", False) and hasattr(
@@ -70,6 +70,61 @@ from sglang.test.test_utils import CustomTestCase
register_cpu_ci(est_time=11, suite="base-a-test-cpu")
class TestMixedSWAKVBlockScales(CustomTestCase):
def test_swa_block_scales_follow_each_subpools_dtype(self):
from sglang.srt.mem_cache.memory_pool import (
MHATokenToKVPool,
MHATokenToKVPoolMXFP8,
)
from sglang.srt.mem_cache.swa_memory_pool import SWAKVPool
swa_infos = ([0x1000, 0x2000], [1024, 1024], [64, 64])
full_scale_infos = ([0x3000, 0x4000], [128, 128], [8, 8])
swa_scale_infos = ([0x5000, 0x6000], [128, 128], [8, 8])
for full_quantized, swa_quantized in (
(False, False),
(True, False),
(False, True),
(True, True),
):
with self.subTest(full=full_quantized, swa=swa_quantized):
target = object.__new__(SWAKVPool)
for name, quantized, scale_infos in (
("full_kv_pool", full_quantized, full_scale_infos),
("swa_kv_pool", swa_quantized, swa_scale_infos),
):
pool = object.__new__(
MHATokenToKVPoolMXFP8 if quantized else MHATokenToKVPool
)
pool.get_contiguous_buf_infos = Mock(return_value=swa_infos)
if quantized:
pool.get_kv_scale_buf_infos = Mock(return_value=scale_infos)
setattr(target, name, pool)
kv_args = KVArgs()
setup_state_kv_args(kv_args, target)
expected_types = [StateType.SWA]
expected_infos = [swa_infos]
if full_quantized:
expected_types.append(StateType.BLOCK_SCALE)
expected_infos.append(full_scale_infos)
if swa_quantized:
expected_types.append(StateType.BLOCK_SCALE_SWA)
expected_infos.append(swa_scale_infos)
self.assertEqual(kv_args.state_types, expected_types)
self.assertEqual(
kv_args.state_data_ptrs, [infos[0] for infos in expected_infos]
)
self.assertEqual(
kv_args.state_data_lens, [infos[1] for infos in expected_infos]
)
self.assertEqual(
kv_args.state_item_lens, [infos[2] for infos in expected_infos]
)
class TestDisaggregationWire(unittest.TestCase):
def test_sender_clear_keeps_abort_ack_until_writes_drain(self):
manager = object.__new__(MooncakeKVManager)