[Bugfix][LoRA] Fall back for high-rank MoE LoRA (#55161)

Signed-off-by: Raina Zhong <yuweiz@nvidia.com>
Co-authored-by: Raina Zhong <yuweiz@nvidia.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
xiaoyu-xyz
2026-10-02 13:59:28 +03:00
committed by GitHub
co-authored by Raina Zhong mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 98b29aa99a
commit 8c60714da5
2 changed files with 94 additions and 11 deletions
+77
View File
@@ -362,6 +362,83 @@ def test_fused_moe_lora_kernel(
torch.testing.assert_close(output, output2, atol=1e-2, rtol=1e-2)
@pytest.mark.parametrize("max_lora_rank", [128, 256])
@pytest.mark.parametrize("add_inputs", [True, False])
@pytest.mark.parametrize("num_slices", [1, 2])
@pytest.mark.parametrize("device", DEVICES)
def test_fused_moe_lora_kernel_large_rank_fallback(
max_lora_rank,
add_inputs,
num_slices,
device,
):
"""Ranks above the one-shot limit fall back to shrink/expand kernels."""
torch.set_default_device(device)
set_random_seed(42)
num_tokens = 8
num_sequences = 4
top_k_num = 2
num_experts = 4
max_loras = 2
N = 256
K = 256
block_size = 16
dtype = torch.bfloat16
topk_ids, topk_weights, token_lora_mapping, lora_ids = sample_data(
num_tokens, num_sequences, max_loras, num_experts, top_k_num
)
lora_a_stacked = [
torch.rand(
(max_loras, num_experts, max_lora_rank, K),
dtype=dtype,
)
for _ in range(num_slices)
]
lora_b_stacked = [
torch.rand(
(max_loras, num_experts, N // num_slices, max_lora_rank),
dtype=dtype,
)
for _ in range(num_slices)
]
hidden_states = torch.rand((num_tokens, K), dtype=dtype)
residual = torch.randn((num_tokens, top_k_num, N), dtype=dtype)
output = residual.clone()
use_fused_moe_lora_kernel(
topk_ids,
topk_weights,
token_lora_mapping,
max_lora_rank,
top_k_num,
lora_ids,
lora_a_stacked,
lora_b_stacked,
hidden_states,
output,
max_loras,
num_experts,
block_size,
add_inputs=add_inputs,
)
expected = use_torch(
hidden_states,
token_lora_mapping,
topk_ids,
lora_a_stacked,
lora_b_stacked,
top_k_num,
num_slices,
)
if add_inputs:
expected += residual
torch.testing.assert_close(output, expected, atol=2e-2, rtol=2e-2)
def use_fused_moe_lora_kernel_naive(
topk_ids,
topk_weights,
+17 -11
View File
@@ -106,14 +106,18 @@ def _get_c_ptrs(
_LORA_PTR_DICT: dict[tuple[int, ...], torch.tensor] = {}
# The one-shot kernel keeps the rank-dimension intermediate in registers.
# Larger ranks require rank tiling, so route them through the two-kernel path.
_FUSED_MOE_LORA_ONE_SHOT_MAX_RANK = 128
# ---------------------------------------------------------------------------
# Fully-fused MoE-LoRA kernel (one-shot): shrink + expand combined into a single
# launch with the rank-dim intermediate kept in registers. Used by the fast
# path of `_fused_moe_lora` for `fully_sharded=False`. The legacy two-kernel
# path (`_fused_moe_lora_kernel` above) is retained for `fully_sharded=True`
# because that path needs to materialise the intermediate cache for an
# all_reduce / all_gather between shrink and expand.
# path of `_fused_moe_lora` for supported ranks with `fully_sharded=False`.
# The legacy two-kernel path (`_fused_moe_lora_kernel` above) is retained for
# larger ranks and for `fully_sharded=True`, where the intermediate cache must
# be materialised for an all_reduce / all_gather between shrink and expand.
# ---------------------------------------------------------------------------
@@ -349,8 +353,9 @@ def _run_fused_moe_lora_one_shot(
# regressed across all M (+8 to +40%): the (64,32) fp32 accumulator +
# widened B tile pushed register count past spill threshold, lowering
# occupancy by more than the MMA gain saved.
assert rank <= 128, (
f"fused_moe_lora_one_shot supports max_lora_rank<=128; got rank={rank}"
assert rank <= _FUSED_MOE_LORA_ONE_SHOT_MAX_RANK, (
"fused_moe_lora_one_shot supports "
f"max_lora_rank<={_FUSED_MOE_LORA_ONE_SHOT_MAX_RANK}; got rank={rank}"
)
BLOCK_R = max(triton.next_power_of_2(rank), 16)
@@ -1318,6 +1323,7 @@ def _fused_moe_lora_expand(
offset: int = 0,
use_gdc: bool = False,
use_tma: bool = False,
add_inputs: bool = True,
) -> None:
b_ptr = _get_ptr(lora_b_stacked, device)
K = max_lora_rank
@@ -1408,7 +1414,7 @@ def _fused_moe_lora_expand(
token_mapping_factor=1,
naive_block_assignment=sorted_token_ids is None,
MUL_ROUTED_WEIGHT=mul_routed_weight,
ADD_INPUTS=True,
ADD_INPUTS=add_inputs,
USE_B_L2_CACHE=True,
sort_c=False,
IS_PRIMARY=False,
@@ -1478,7 +1484,7 @@ def _fused_moe_lora(
assert top_k_num == topk_weights.shape[1]
# Fast path: single fused kernel
if not fully_sharded:
if not fully_sharded and max_lora_rank <= _FUSED_MOE_LORA_ONE_SHOT_MAX_RANK:
M_pairs = topk_weights.numel()
if (
sorted_token_ids is None
@@ -1524,9 +1530,8 @@ def _fused_moe_lora(
)
return
assert add_inputs, (
"fused_moe_lora(add_inputs=False) is only supported on the "
"fully_sharded=False fast path"
assert add_inputs or not fully_sharded, (
"fused_moe_lora(add_inputs=False) is not supported with fully_sharded=True"
)
device = qcurr_hidden_states.device
@@ -1660,6 +1665,7 @@ def _fused_moe_lora(
offset,
use_gdc=use_gdc,
use_tma=use_tma,
add_inputs=add_inputs,
)