[PCP][DCP] Enable PCP+DCP on sparse-MLA models (#56157)

Signed-off-by: patryk.saffer <patryk.saffer@blc-login-1.blc1.mistralcompute.ai>
Signed-off-by: patryk.saffer <patryk.saffer@blc-login-0.blc1.mistralcompute.ai>
Signed-off-by: PatrykSaffer <patryk.saffer@mistral.ai>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: patryk.saffer <patryk.saffer@blc-login-1.blc1.mistralcompute.ai>
Co-authored-by: patryk.saffer <patryk.saffer@blc-login-0.blc1.mistralcompute.ai>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Patryk Saffer <PatrykSaffer@users.noreply.github.com>
Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
This commit is contained in:
PatrykSaffer
2026-09-12 23:24:44 +00:00
committed by GitHub
co-authored by patryk.saffer patryk.saffer Lucas Wilkinson Patryk Saffer Claude Fable 5.1
parent 72d4d83157
commit ebe1dec2da
21 changed files with 917 additions and 138 deletions
@@ -420,6 +420,9 @@ class MLAAttention(nn.Module, AttentionLayerBase):
"""
supports_dense_mha_prefill: ClassVar[bool] = True
# Under PCP+DCP only the decode rows carry an LSE; the base forward
# merges a full-batch LSE, so subclasses opt in with their own forward.
supports_pcp_dcp: ClassVar[bool] = False
def __init__(
self,
@@ -660,6 +663,10 @@ class MLAAttention(nn.Module, AttentionLayerBase):
_COMPUTE_PREFILL_METADATA_KERNEL.register_warmup()
if self.use_pcp and self.impl.dcp_world_size > 1 and not self.supports_pcp_dcp:
raise NotImplementedError(
f"{type(self).__name__} does not support PCP+DCP."
)
self.dcp_manager: MLADCPManager | None = None
if self.impl.dcp_world_size > 1:
query_dtype = (
@@ -190,7 +190,8 @@ class SparseMLACommonMetadataBuilder(AttentionMetadataBuilder[T]):
pin_memory=PIN_MEMORY,
)
parallel_config = vllm_config.parallel_config
self.use_pcp = parallel_config.prefill_context_parallel_size > 1
self.pcp_world_size = parallel_config.prefill_context_parallel_size
self.use_pcp = self.pcp_world_size > 1
try:
self.dcp_world_size = get_dcp_group().world_size
except AssertionError:
@@ -409,8 +410,12 @@ class SparseMLACommonMetadataBuilder(AttentionMetadataBuilder[T]):
q_data_type=self.model_config.dtype,
output_dtype=self.model_config.dtype,
prefill_backend=self._prefill_backend,
use_dense_mha=_use_dense_mha_prefill(
self.vllm_config, prefill_max_seq_len, self.topk_tokens
use_dense_mha=(
_use_dense_mha_prefill(
self.vllm_config, prefill_max_seq_len, self.topk_tokens
)
# Dense MHA cannot be used under PCP+DCP.
and not (self.use_pcp and self.dcp_world_size > 1)
),
topk_mask_workspace=self.topk_mask_workspace,
host_staging_plan=staging_plan,