[PCP][Spec Decode] Adds PCP support for single-module MTP and replicated DSpark. (#56107)

Signed-off-by: Summer Yang <girasoleyang@gmail.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Summer Yang <girasoleyang@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
Lucas Wilkinson
2026-09-10 22:07:53 -07:00
committed by GitHub
co-authored by Summer Yang OpenAI Codex OpenAI Codex
parent e7edf17cea
commit 980c16c8e4
11 changed files with 159 additions and 79 deletions
@@ -38,6 +38,13 @@ class DFlashSpeculator(DraftModelSpeculator):
_speculator_name = "DFlash" # For logging, so we can share methods with subclasses
def __init__(self, vllm_config: VllmConfig, device: torch.device):
parallel_config = vllm_config.parallel_config
if parallel_config.prefill_context_parallel_size > 1:
vllm_config = copy.copy(vllm_config)
vllm_config.parallel_config = replace(
parallel_config,
prefill_context_parallel_size=1,
)
super().__init__(vllm_config, device)
self.hidden_states = torch.zeros(
@@ -388,6 +395,11 @@ class DFlashSpeculator(DraftModelSpeculator):
)
return self.draft_tokens[:num_reqs]
if self.pcp_manager is not None and not dummy_run:
self.block_tables.gather_block_tables(
input_batch.idx_mapping, num_reqs_padded=num_reqs
)
# The query slot mapping is written into the shared BlockTables slot_mappings.
# That buffer's address is what the captured CUDA graph reads from at replay.
assert self.draft_kv_cache_group_id >= 0