Skip to content

[None][perf] DFlash unified draft cache (#19733): FA2 num_splits + remove per-step host syncs #1

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Draft
viraatc wants to merge 2 commits into
base: viraatc/pr19733-base
Choose a base branch
from
Draft

[None][perf] DFlash unified draft cache (#19733): FA2 num_splits + remove per-step host syncs #1

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6a8c3a8
[None][perf] DFlash VANILLA draft attention: pass explicit FA2 num_sp…
viraatc Oct 12, 2026
adf62c6
[None][perf] DFlash unified draft cache: avoid per-step host syncs on…
viraatc Oct 12, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 5 additions & 0 deletions tensorrt_llm/_torch/models/modeling_dflash.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1838,6 +1838,11 @@ def dflash_forward(
cache_batch_idx=cache_batch_idx_i32,
causal=causal,
window_size=window_size,
# FA2's default heuristic sizes splits from k_cache.size(1),
# the max_seq_len-dense arena, not the live context, so long
# contexts run on split 0 with a few CTAs. A fixed count is
# CUDA-graph safe; splits past the live length exit early.
num_splits=128,
)
if pack_gqa:
# Undo the packing: [B, group*blk, h_kv, d] -> [B, blk, h_q, d].
Expand Down
12 changes: 6 additions & 6 deletions tensorrt_llm/_torch/speculative/dflash.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -844,8 +844,8 @@ def _prepare_managed_history(self, request_ids: list[int]) -> list[int]:
torch.tensor(
[self._req_to_slot.get(request_id, self._dummy_slot) for request_id in request_ids],
dtype=torch.long,
device=self._batch_to_slot.device,
)
pin_memory=prefer_pinned(),
).to(self._batch_to_slot.device, non_blocking=True)
)
self._managed_request_ids = tuple(
request_id for request_id in request_ids if request_id in real_request_ids
Expand Down Expand Up @@ -891,8 +891,8 @@ def snapshot_managed_draft_history(self) -> DraftHistoryUpdate | None:
slots = torch.tensor(
[self._req_to_slot[request_id] for request_id in self._managed_request_ids],
dtype=torch.long,
device=self._ctx_len.device,
)
pin_memory=prefer_pinned(),
).to(self._ctx_len.device, non_blocking=True)
lengths = self._ctx_len[slots]
positions = lengths + self._ctx_position_offset[slots]
return DraftHistoryUpdate.capture(
Expand Down Expand Up @@ -1091,8 +1091,8 @@ def _refresh_ctx_block_tables(
for rid in request_ids[:num_seqs]
],
dtype=torch.long,
device=self._ctx_block_counts.device,
)
pin_memory=prefer_pinned(),
).to(self._ctx_block_counts.device, non_blocking=True)
)
return True
src = getattr(attn_metadata, "draft_kv_cache_block_offsets", None)
Expand Down
Loading