feat(cpu): add CPU support for draft model speculative decoding (#32662)

Signed-off-by: R <Ganesh.R@amd.com>
This commit is contained in:
Ganesh R
2026-04-10 11:49:52 +08:00
committed by GitHub
parent 55d037e2e5
commit 445a2a4d1a
7 changed files with 945 additions and 7 deletions
+15 -1
View File
@@ -11,6 +11,20 @@ from vllm.v1.attention.backends.utils import (
PADDING_SLOT_ID = -1
def next_power_of_2(n: int) -> int:
"""Return the smallest power of 2 >= n."""
if n <= 0:
return 1
n -= 1
n |= n >> 1
n |= n >> 2
n |= n >> 4
n |= n >> 8
n |= n >> 16
n |= n >> 32
return n + 1
@triton.jit
def eagle_step_slot_mapping_metadata_kernel(
positions_ptr, # [batch_size] - current positions (1D view for M-RoPE)
@@ -102,8 +116,8 @@ def eagle_step_update_slot_mapping_and_metadata(
batch_size = positions_1d.shape[0]
if input_batch_size is None:
input_batch_size = batch_size
n_blocks_per_req = block_table_tensor.shape[1]
n_blocks_per_req = block_table_tensor.shape[1]
eagle_step_slot_mapping_metadata_kernel[(input_batch_size,)](
positions_1d,
block_table_tensor,