[Core] Add MRV2 virtual-batch PCP for MLA (#46570)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-07-19 02:53:15 +00:00
committed by GitHub
co-authored by Codex OpenAI Codex Matthew Bonanni
parent 9243e0124e
commit b6ff8a2f50
39 changed files with 1204 additions and 196 deletions
+7
View File
@@ -578,6 +578,7 @@ def build_attn_metadata(
seq_lens_cpu_upper_bound: torch.Tensor | None = None,
dcp_local_seq_lens: torch.Tensor | None = None,
positions: torch.Tensor | None = None,
is_prefilling: torch.Tensor | None = None,
mm_req_doc_ranges: dict[int, list[tuple[int, int]]] | None = None,
model_specific_attn_metadata: ModelSpecificAttnMetadata | None = None,
for_cudagraph_capture: bool = False,
@@ -605,6 +606,11 @@ def build_attn_metadata(
if model_specific_attn_metadata is not None
else {}
)
# Model-specific metadata (e.g. Mamba hybrid) may supply its own
# padding-aware is_prefilling, which takes precedence over the default.
group_is_prefilling = common_attn_metadata_extra_kwargs.pop(
"is_prefilling", is_prefilling
)
common_attn_metadata = CommonAttentionMetadata(
query_start_loc=query_start_loc_gpu,
query_start_loc_cpu=query_start_loc_cpu,
@@ -619,6 +625,7 @@ def build_attn_metadata(
causal=group_causal,
dcp_local_seq_lens=dcp_local_seq_lens,
positions=positions,
is_prefilling=group_is_prefilling,
mm_req_doc_ranges=mm_req_doc_ranges,
rswa_prefix_lens=rswa_prefix_lens,
**common_attn_metadata_extra_kwargs,