| 123456789101112131415161718192021222324252627282930313233 |
- --- baked-ours/v1/worker/gpu/attn_utils.py 2026-09-10 19:56:14.000000000 +0200
- +++ baked/v1/worker/gpu/attn_utils.py 2026-09-11 10:13:47.000000000 +0200
- @@ -20,6 +20,7 @@
- KVCacheConfig,
- KVCacheSpec,
- UniformTypeKVCacheSpecs,
- + is_dcp_replicated_spec,
- )
- from vllm.v1.worker.gpu.model_states.interface import ModelSpecificAttnMetadata
- from vllm.v1.worker.utils import (
- @@ -283,6 +284,13 @@
- for i in range(num_kv_cache_groups):
- block_table = block_tables[i]
- slot_mapping = slot_mappings[i]
- + # DCP-replicated groups hold every position on every rank: their
- + # builders must see the full seq_lens, not this rank's shard.
- + group_dcp_local_seq_lens = (
- + None
- + if is_dcp_replicated_spec(kv_cache_config.kv_cache_groups[i].kv_cache_spec)
- + else dcp_local_seq_lens
- + )
- # Per-group causal for hybrid drafters (mixed SWA/full attention).
- group_causal = (
- causal if isinstance(causal, (bool, torch.Tensor)) else causal.get(i, True)
- @@ -310,7 +318,7 @@
- block_table_tensor=block_table,
- slot_mapping=slot_mapping,
- causal=group_causal,
- - dcp_local_seq_lens=dcp_local_seq_lens,
- + dcp_local_seq_lens=group_dcp_local_seq_lens,
- positions=positions,
- is_prefilling=group_is_prefilling,
- mm_req_doc_ranges=mm_req_doc_ranges,
|
