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

Signed-off-by: Zihua Wu <13583761+lucifer1004@users.noreply.github.com>
Co-authored-by: Kimi Code <noreply@moonshot.cn>
Co-authored-by: Yongye Zhu <zyy1102000@gmail.com>
This commit is contained in:
Gabriel Wu
2026-09-23 04:52:28 -07:00
committed by GitHub
co-authored by Kimi Code Yongye Zhu
parent 15ed1262e7
commit 157bcb7c48
11 changed files with 629 additions and 28 deletions
@@ -256,6 +256,7 @@ class DFlashSpeculator(DraftModelSpeculator):
num_tokens_across_dp=num_tokens_across_dp,
slot_mapping=slot_mappings,
batch_descriptor=batch_descriptor,
is_padding=self.input_buffers.is_padding[:num_tokens],
):
last_hidden_states = self.model(
input_ids=self.input_buffers.input_ids[:num_tokens],
@@ -487,6 +488,7 @@ def _prepare_dflash_inputs_kernel(
# Outputs
out_input_ids_ptr,
out_query_positions_ptr,
out_is_padding_ptr,
out_query_start_loc_ptr,
out_seq_lens_ptr,
out_query_slot_mapping_ptr,
@@ -617,6 +619,7 @@ def _prepare_dflash_inputs_kernel(
)
tl.store(out_input_ids_ptr + query_idx, input_id, mask=is_query)
tl.store(out_is_padding_ptr + query_idx, False, mask=is_query)
clamped_query_pos = tl.minimum(query_pos, max_model_len - 1)
tl.store(out_query_positions_ptr + query_idx, clamped_query_pos, mask=is_query)
tl.store(out_query_slot_mapping_ptr + query_idx, q_slot, mask=is_query)
@@ -679,6 +682,9 @@ def _prepare_dflash_inputs_kernel(
for i in range(q_pad_start, max_num_tokens, BLOCK_SIZE):
block = i + tl.arange(0, BLOCK_SIZE)
mask = block < max_num_tokens
tl.store(out_input_ids_ptr + block, 0, mask=mask)
tl.store(out_query_positions_ptr + block, 0, mask=mask)
tl.store(out_is_padding_ptr + block, True, mask=mask)
tl.store(out_query_slot_mapping_ptr + block, PAD_SLOT_ID, mask=mask)
@@ -730,6 +736,7 @@ def prepare_dflash_inputs(
_prepare_dflash_inputs_kernel[(num_reqs, num_blocks)](
input_buffers.input_ids,
input_buffers.positions,
input_buffers.is_padding,
input_buffers.query_start_loc,
input_buffers.seq_lens,
query_slot_mapping,