[Kernel] Support DS Mamba tail copy for MTP align mode (#45473)

Signed-off-by: Sungsoo Ha <sungsooh@nvidia.com>
Co-authored-by: Thomas Parnell <tom.parnell@gmail.com>
This commit is contained in:
sungsoo ha
2026-06-16 22:50:30 +00:00
committed by GitHub
co-authored by Thomas Parnell
parent 520828789c
commit 4bf699d310
3 changed files with 182 additions and 34 deletions
@@ -303,18 +303,14 @@ def get_conv_copy_spec(
src_block_id = block_ids[cur_block_idx]
offset = num_accepted_tokens - 1
if is_conv_state_dim_first():
# DS layout: (num_blocks, dim, state_len) — state_len is last.
if offset > 0:
# Slicing along the last dim yields a non-contiguous view
# because features (dim) are strided by state_len.
raise NotImplementedError(
"DS conv state layout does not yet support speculative "
"decoding with mamba_cache_mode='align' "
"(num_accepted_tokens > 1)."
)
# DS offset > 0 is handled by the fused postprocess kernel.
assert offset == 0, (
"DS conv state with num_accepted_tokens > 1 must be handled by "
"the fused postprocess kernel, not get_conv_copy_spec"
)
src_state = state[src_block_id]
else:
# SD layout: (num_blocks, state_len, dim) — dim contiguous.
# SD layout: (num_blocks, state_len, dim), with dim contiguous.
src_state = state[src_block_id, offset:]
return MambaCopySpec(
start_addr=src_state.data_ptr(), num_elements=src_state.numel()