[KV Connector][NIXL] Reject NVFP4 KV cache with heterogeneous TP head split (#61865)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Co-authored-by: Kevin H. Luu <khluu000@gmail.com>
This commit is contained in:
Nicolò Lucchesi
2026-10-15 13:57:37 -07:00
committed by GitHub
co-authored by Claude Opus 5.5 Kevin H. Luu
parent c414bf0295
commit 33212db2c2
2 changed files with 107 additions and 0 deletions
@@ -82,6 +82,7 @@ from vllm.utils.math_utils import cdiv
from vllm.utils.network_utils import make_zmq_path
from vllm.utils.torch_utils import async_tensor_h2d
from vllm.v1.kv_cache_interface import (
AttentionSpec,
CircularBufferSpec,
FullAttentionSpec,
KVCacheLayout,
@@ -2671,6 +2672,22 @@ class NixlBaseConnectorWorker:
"Heterogeneous TP head-dimension splitting requires contiguous heads. "
"Use a block-contiguous layout (e.g. LBHNC) on the prefill side."
)
# NVFP4 pages store all heads' data before all heads' scales, so a
# head's bytes are not a contiguous slice of the page.
if (
not self.use_mla
and self.transfer_topo.local_physical_heads
!= max(1, self.transfer_topo.total_num_kv_heads // remote_tp_size)
and any(
isinstance(spec, AttentionSpec) and spec.kv_quant_mode.is_nvfp4
for spec in self._layer_specs.values()
)
):
raise RuntimeError(
"NIXL KV-head re-sharding is not supported for NVFP4 KV cache "
f"(local TP {self.transfer_topo.tp_size}, remote TP "
f"{remote_tp_size}). Use matching TP sizes on prefill and decode."
)
# Per-region block_len validation enforcing the P/D invariant.
# REPLICATE regions (MLA, or a whole-model MLA / replicated-KV transfer)