[Bugfix][KVConnector] Disable cross-layer KV blocks for per-token-head quant (#49226)
Signed-off-by: Achyuthan Sivasankar <achyuthan.sivasankar@gmail.com> Co-authored-by: Or Ozeri <oro@il.ibm.com>
achyuthan.s committed
Jul 26, 2026 at 06:24 UTC
48ebd6f2f164057043edb8301771a17dcaebd61d
1 file changed
+4
vllm/v1/worker/kv_connector_model_runner_mixin.py
+4
@@ -155,6 +155,10 @@ class KVConnectorModelRunnerMixin:
155
kv_cache_spec = attn_group.kv_cache_spec
156
if not isinstance(kv_cache_spec, AttentionSpec):
157
return False
158
+ # Per-token-head quant carves inline-scale views that assume per-layer
159
+ # contiguous KV buffers; the cross-layer layout breaks this and corrupts KV.
160
+ if kv_cache_spec.kv_quant_mode.is_per_token_head:
161
+ return False
162
return kv_cache_spec.indexes_kv_by_block_stride
163
164
@staticmethod