[PD] Send DSA indexer state only past decode_prefix_len (#44140)

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Shangming Cai <csmthu@gmail.com>
This commit is contained in:
Hrithvik Alex
2026-10-15 11:25:18 +08:00
committed by GitHub
co-authored by Claude Opus 5.5 Shangming Cai
parent 0cf683ea8d
commit 2528cd2b91
3 changed files with 45 additions and 2 deletions
+14 -1
View File
@@ -96,6 +96,7 @@ from sglang.srt.mem_cache.common import (
page_align_floor,
release_kv_cache,
restore_kv_cache,
suffix_page_indices,
)
from sglang.srt.mem_cache.deepseek_v4_memory_pool import DeepSeekV4TokenToKVPool
from sglang.srt.mem_cache.hicache_storage import PoolName, PoolTransfer
@@ -1659,6 +1660,18 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
device_page_size = self.token_to_kv_pool.page_size
return kv_to_page_indices(kv_indices_full, device_page_size)
def _dsa_pages_payload():
# Mirror of prefill _dsa_pages_payload: the cached prefix keeps
# its own indexer rows, and rows still being restored by
# HiCache are not written until the restore commits.
return suffix_page_indices(
self.req_to_token_pool.req_to_token,
decode_req.req.kv.req_pool_idx,
total_prefix_len,
seq_len,
self.token_to_kv_pool.page_size,
)
def _dsa_tail_payload():
return get_dsa_tail_state_indices(
self.token_to_kv_pool,
@@ -1699,7 +1712,7 @@ class DecodePreallocQueue(DecodeHiCachePreallocMixin):
StateType.QSA_PENDING: _qsa_pending_payload,
StateType.QSA_COMPRESSED: _full_kv_pages_payload,
StateType.SWA: _swa_payload,
StateType.DSA: _full_kv_pages_payload,
StateType.DSA: _dsa_pages_payload,
StateType.DSA_TAIL: _dsa_tail_payload,
StateType.MINIMAX_INDEX_K: _full_kv_pages_payload,
StateType.MINIMAX_DENSE_KV: _full_kv_pages_payload,
+15 -1
View File
@@ -78,6 +78,7 @@ from sglang.srt.mem_cache.common import (
kv_to_page_indices,
kv_to_page_num,
release_kv_cache,
suffix_page_indices,
)
from sglang.srt.mem_cache.deepseek_v4_memory_pool import DeepSeekV4TokenToKVPool
from sglang.srt.mem_cache.memory_pool import HybridLinearKVPool
@@ -1549,6 +1550,19 @@ class SchedulerDisaggregationPrefillMixin:
]
return kv_to_page_indices(kv_indices_full, page_size)
def _dsa_pages_payload():
# Decode already holds the indexer rows of the prefix it cached
# (decode_prefix_len, page-aligned), and its rows for a prefix
# still being restored point at no page yet: send only the
# suffix, positionally matching decode's _dsa_pages_payload.
return suffix_page_indices(
self.req_to_token_pool.req_to_token,
req.kv.req_pool_idx,
req.disagg_decode_prefix_len,
seq_len,
page_size,
)
def _dsa_tail_payload():
return get_dsa_tail_state_indices(
self.token_to_kv_pool_allocator.get_kvcache(),
@@ -1588,7 +1602,7 @@ class SchedulerDisaggregationPrefillMixin:
StateType.QSA_PENDING: _qsa_pending_payload,
StateType.QSA_COMPRESSED: _full_kv_pages_payload,
StateType.SWA: _swa_payload,
StateType.DSA: _full_kv_pages_payload,
StateType.DSA: _dsa_pages_payload,
StateType.DSA_TAIL: _dsa_tail_payload,
StateType.MINIMAX_INDEX_K: _full_kv_pages_payload,
StateType.MINIMAX_DENSE_KV: _full_kv_pages_payload,
+16
View File
@@ -47,6 +47,22 @@ def kv_to_page_indices(kv_indices: torch.Tensor, page_size: int) -> np.ndarray:
return (kv_indices[::page_size] // page_size).cpu().numpy()
def suffix_page_indices(
req_to_token: torch.Tensor,
req_pool_idx: int,
prefix_len: int,
seq_len: int,
page_size: int,
) -> np.ndarray:
"""Pages of a request's tokens [prefix_len, seq_len).
PD prefill and decode both call this with the same page-aligned prefix
(decode_prefix_len), so their lists pair up page by page even though their
rows point at different physical pages.
"""
return kv_to_page_indices(req_to_token[req_pool_idx, prefix_len:seq_len], page_size)
def kv_to_page_num(num_kv_indices: int, page_size: int):
return (num_kv_indices + page_size - 1) // page_size