minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

[core] Probe set_kv_buffer / set_mla_kv_buffer slot ids for OOB (#27459)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-06-06 19:51:09 -07:00
committed by GitHub
parent fe548f36b0
commit 066b4a2180
1 changed files with 10 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/mem_cache/memory_pool.py
+10
View File
@@ -1126,6 +1126,9 @@ class MHATokenToKVPool(KVCache):
v_scale: Optional[float] = None,
layer_id_override: Optional[int] = None,
):
# Catch stale slot ids here instead of as illegal-addr / silent KV
# corruption in the store_kvcache write (gated on SGLANG_ENABLE_ASYNC_ASSERT).
maybe_detect_oob(loc, 0, self.size + self.page_size, "set_kv_buffer (MHA)")
if layer_id_override is not None:
layer_id = layer_id_override
else:
@@ -1422,6 +1425,7 @@ class MHATokenToKVPoolFP4(MHATokenToKVPool):
v_scale: Optional[float] = None,
layer_id_override: Optional[int] = None,
):
maybe_detect_oob(loc, 0, self.size + self.page_size, "set_kv_buffer (MHA-FP4)")
from sglang.srt.model_executor.cuda_graph_runner import get_is_capture_mode
if layer_id_override is not None:
@@ -1812,6 +1816,7 @@ class MLATokenToKVPool(KVCache):
cache_k: torch.Tensor,
cache_v: torch.Tensor,
):
maybe_detect_oob(loc, 0, self.size + self.page_size, "set_kv_buffer (MLA)")
layer_id = layer.layer_id
assert not self.dsa_kv_cache_store_fp8
if cache_k.dtype != self.dtype:
@@ -1831,6 +1836,7 @@ class MLATokenToKVPool(KVCache):
cache_k_nope: torch.Tensor,
cache_k_rope: torch.Tensor,
):
maybe_detect_oob(loc, 0, self.size + self.page_size, "set_mla_kv_buffer (MLA)")
layer_id = layer.layer_id
if _is_hip and self.use_dsa and self.dtype == fp8_dtype:
@@ -2004,6 +2010,7 @@ class MLATokenToKVPoolFP4(MLATokenToKVPool):
cache_k: torch.Tensor,
cache_v: torch.Tensor,
):
maybe_detect_oob(loc, 0, self.size + self.page_size, "set_kv_buffer (MLA-FP4)")
layer_id = layer.layer_id
assert not self.dsa_kv_cache_store_fp8
if cache_k.dtype != self.dtype:
@@ -2028,6 +2035,9 @@ class MLATokenToKVPoolFP4(MLATokenToKVPool):
cache_k_nope: torch.Tensor,
cache_k_rope: torch.Tensor,
):
maybe_detect_oob(
loc, 0, self.size + self.page_size, "set_mla_kv_buffer (MLA-FP4)"
)
layer_id = layer.layer_id
if self.dsa_kv_cache_store_fp8: