[Nixl][PD] DCP support for MLA models (#50611)

Signed-off-by: NickLucche <nicolo.lucchesi@mistral.ai>
This commit is contained in:
Nicolò Lucchesi
2026-08-29 16:03:53 +00:00
committed by GitHub
parent 4fc943b867
commit 7f4793eaa3
23 changed files with 818 additions and 104 deletions
@@ -3157,9 +3157,6 @@ class MLACommonImpl(MLACommonBaseImpl[M], Generic[M]):
parallel_config = get_current_vllm_config().parallel_config
# Avoid requiring an initialized DCP group in tests.
self.dcp_world_size: int = parallel_config.decode_context_parallel_size
self.cp_kv_cache_interleave_size: int = (
parallel_config.cp_kv_cache_interleave_size
)
@abstractmethod
def forward_mqa(
@@ -771,14 +771,30 @@ class SparseAttnIndexer(CustomOp):
parallel_config = get_current_vllm_config().parallel_config
self.dcp_world_size = parallel_config.decode_context_parallel_size
self.dcp_rank = get_dcp_group().rank_in_group if self.dcp_world_size > 1 else 0
self.cp_kv_cache_interleave_size = parallel_config.cp_kv_cache_interleave_size
self.use_pcp = parallel_config.prefill_context_parallel_size > 1
self._cp_kv_cache_interleave_size: int | None = None
if current_platform.is_cuda() and not has_deep_gemm():
raise RuntimeError(
"Sparse Attention Indexer CUDA op requires DeepGEMM support in "
"the current vLLM environment."
)
@property
def cp_kv_cache_interleave_size(self) -> int:
"""With PD+DCP, the real value isn't known until block_size is finalized,
which happens after this layer is built. Safe to cache after the first access,
as long as the adjustment always runs before any forward pass
(it's set up in Worker.initialize_from_config, ahead of warmup/serving).
"""
if self._cp_kv_cache_interleave_size is None:
value = (
get_current_vllm_config().parallel_config.cp_kv_cache_interleave_size
)
if isinstance(get_forward_context().attn_metadata, dict):
self._cp_kv_cache_interleave_size = value
return value
return self._cp_kv_cache_interleave_size
def forward_native(
self,
hidden_states: torch.Tensor,