Revert "[DSpark] Support pipeline-parallel targets in aggregated serving (#56956)" (#58484)

This commit is contained in:
Nick Hill
2026-09-24 08:48:36 -07:00
committed by GitHub
parent d636456b1f
commit 09fe178dba
11 changed files with 28 additions and 629 deletions
-26
View File
@@ -555,11 +555,6 @@ def _post_update_kernel(
all_token_ids_ptr,
all_token_ids_stride,
total_len_ptr,
broadcast_drafts_ptr,
broadcast_drafts_stride,
draft_tokens_ptr,
draft_tokens_stride,
num_spec,
):
req_id = tl.program_id(0)
req_state_idx = tl.load(idx_mapping_ptr + req_id)
@@ -567,17 +562,6 @@ def _post_update_kernel(
# Filter rows with negative index entries.
return
if broadcast_drafts_ptr is not None:
# PP path: adopt the draft tokens proposed by the last rank's
# speculator so the next verification step embeds the real drafts.
for i in range(num_spec):
token_id = tl.load(
broadcast_drafts_ptr + req_id * broadcast_drafts_stride + i
)
tl.store(
draft_tokens_ptr + req_state_idx * draft_tokens_stride + i, token_id
)
total_len = tl.load(total_len_ptr + req_state_idx)
num_sampled = tl.load(num_sampled_ptr + req_id)
if num_sampled > 0:
@@ -638,11 +622,6 @@ def post_update(
all_token_ids: torch.Tensor,
# [max_num_reqs]
total_len: torch.Tensor,
# [num_reqs, num_spec]; drafts broadcast from the last PP rank. Only
# passed on non-last PP ranks, which never run the speculator.
broadcast_drafts: torch.Tensor | None = None,
# [max_num_reqs, num_spec]
draft_tokens_out: torch.Tensor | None = None,
) -> None:
num_reqs = idx_mapping.shape[0]
_post_update_kernel[(num_reqs,)](
@@ -659,11 +638,6 @@ def post_update(
all_token_ids,
all_token_ids.stride(0),
total_len,
broadcast_drafts,
broadcast_drafts.stride(0) if broadcast_drafts is not None else 0,
draft_tokens_out,
draft_tokens_out.stride(0) if draft_tokens_out is not None else 0,
broadcast_drafts.shape[1] if broadcast_drafts is not None else 0,
num_warps=1,
)