Whether the MLA draft layers can select an attention backend with DCP.
Runs the same selection as the draft layers, so a draft with no DCP-capable
backend can fall back to DCP=1 and a replicated KV cache. Sliding-window
drafts always do: their cache cannot shard by DCP, and NIXL pulls such a
draft group from a single DCP prefill rank.
Source code in vllm/v1/worker/gpu/spec_decode/dflash/utils.py
| def draft_attention_supports_dcp(draft_vllm_config: VllmConfig, model_tag: str) -> bool:
"""Whether the MLA draft layers can select an attention backend with DCP.
Runs the same selection as the draft layers, so a draft with no DCP-capable
backend can fall back to DCP=1 and a replicated KV cache. Sliding-window
drafts always do: their cache cannot shard by DCP, and NIXL pulls such a
draft group from a single DCP prefill rank.
"""
from vllm.compilation.backends import set_model_tag
from vllm.config import set_current_vllm_config
from vllm.v1.attention.selector import get_attn_backend
speculative_config = draft_vllm_config.speculative_config
assert speculative_config is not None
draft_model_config = speculative_config.draft_model_config
if (
not draft_model_config.use_mla
or getattr(draft_model_config.hf_text_config, "sliding_window", None)
is not None
):
return False
try:
with set_current_vllm_config(draft_vllm_config), set_model_tag(model_tag):
get_attn_backend(
draft_model_config.get_head_size(),
draft_model_config.dtype,
draft_vllm_config.cache_config.cache_dtype,
use_mla=True,
num_heads=draft_model_config.get_num_attention_heads(
draft_vllm_config.parallel_config
),
)
except ValueError:
return False
return True
|