[Bugfix][Kimi-K3] Give the AMD packed KDA decode kernel the state-index stride (#51682)

Signed-off-by: Lyu, Xudong <xudong.lyu@amd.com>
This commit is contained in:
Lyu, Xudong
2026-08-11 00:05:20 +08:00
committed by GitHub
parent 640a09086d
commit 0e2d78028c
2 changed files with 18 additions and 4 deletions
+13 -1
View File
@@ -15,6 +15,9 @@ from vllm.model_executor.layers.mamba.ops.causal_conv1d import causal_conv1d_upd
from vllm.model_executor.layers.mamba.ops.gather_initial_states import (
gather_initial_states,
)
from vllm.models.kimi_k3.amd.ops.third_party.kda import (
fused_recurrent_kda_packed_decode as fused_recurrent_kda_packed_decode_amd,
)
from vllm.models.kimi_k3.nvidia.kda import (
is_flashkda_supported,
is_fused_kda_decode_supported,
@@ -31,6 +34,13 @@ from vllm.third_party.flash_linear_attention.ops.l2norm import l2norm_fwd
DEVICE = "cuda"
# The AMD and NVIDIA copies of the KDA kernels are vendored separately and are
# free to diverge, so the shared-semantics tests below run against both.
PACKED_DECODE_IMPLS = {
"nvidia": fused_recurrent_kda_packed_decode,
"amd": fused_recurrent_kda_packed_decode_amd,
}
@torch.inference_mode()
def test_gather_initial_states_correctness():
@@ -273,11 +283,13 @@ def test_chunk_kda_fused_gate_cumsum_matches_unfused(
@pytest.mark.parametrize("num_seqs", [1, 8, 32])
@pytest.mark.parametrize("lower_bound", [-5.0, None])
@pytest.mark.parametrize("state_indices_stride", [1, 8])
@pytest.mark.parametrize("impl", PACKED_DECODE_IMPLS.keys())
@torch.inference_mode()
def test_packed_kda_decode_correctness(
num_seqs: int,
lower_bound: float | None,
state_indices_stride: int,
impl: str,
):
H, D = 8, 128
torch.manual_seed(321)
@@ -361,7 +373,7 @@ def test_packed_kda_decode_correctness(
use_qk_l2norm_in_kernel=True,
)
packed_state = state
packed_out, _ = fused_recurrent_kda_packed_decode(
packed_out, _ = PACKED_DECODE_IMPLS[impl](
mixed_qkv=mixed_qkv,
raw_g=raw_g,
raw_beta=raw_beta,