[CPU] Optimize routed FP8/MXFP4 MoE GEMM dispatch (#50949)

Signed-off-by: Li, Tianmu <tianmu.li@intel.com>
This commit is contained in:
Tianmu Li
2026-08-08 12:56:14 +08:00
committed by GitHub
parent d3621c1eb3
commit 700d39b558
3 changed files with 110 additions and 5 deletions
+5
View File
@@ -59,6 +59,11 @@ inline bool can_use_brgemm<at::Float8_e4m3fn>(int M) {
return brgemm_supported() && M > 4;
}
template <typename T>
inline bool can_use_brgemm(int M, int N) {
return brgemm_supported() && (can_use_brgemm<T>(M) || N != block_size_n());
}
// work around compiler internal error
#define BLOCK_K 128 // 4 * TILE_K
+14 -5
View File
@@ -60,9 +60,6 @@ void fused_experts_fp_kernel_impl(
const int64_t stride_e = 2 * N * packed_K;
const int64_t stride_n = packed_K;
int64_t avg_M = std::max(int64_t(1), M * topk / E);
const bool use_brgemm = can_use_brgemm<packed_t>(avg_M);
int64_t B_tmp_size_per_thread = MAX_CACHE_BLOCK_SIZE * BLOCK_N * std::max(K, N);
// here we only parallel on half of 2N to fuse silu_and_mul with gemm
@@ -70,6 +67,7 @@ void fused_experts_fp_kernel_impl(
// get local pointers
int tid = get_thread_num();
scalar_t* __restrict__ A = A_tmp + tid * BLOCK_M * K;
bool use_brgemm = false;
loop_2d<packed_t>(mb0, mb1, nb0, nb1, BLOCK_N * K, [&](int64_t mb, int64_t nb, int64_t nb_offset) {
int64_t n_size = std::min(2 * N - nb * BLOCK_N, BLOCK_N);
@@ -86,6 +84,11 @@ void fused_experts_fp_kernel_impl(
bool do_unpack = (mb == mb0) || (expert_id != pre_expert_id);
int64_t m_size = offsets[mb + 1] - offsets[mb];
if (m_size == 0) {
return;
}
const bool brg = can_use_brgemm<packed_t>(m_size, n_size);
use_brgemm |= brg;
if (nb_offset == 0) {
// 1.a load A
@@ -111,7 +114,7 @@ void fused_experts_fp_kernel_impl(
/* lda */ K,
/* ldb */ n_size,
/* ldc */ 2 * N,
/* brg */ use_brgemm,
/* brg */ brg,
/* block_size_K */ block_size_K,
/* do_unpack */ do_unpack);
});
@@ -152,10 +155,16 @@ void fused_experts_fp_kernel_impl(
parallel_2d(MB2, NB2, [&](int64_t mb0, int64_t mb1, int64_t nb0, int64_t nb1) {
int tid = get_thread_num();
alignas(64) scalar_t C[BLOCK_M * BLOCK_K];
bool use_brgemm = false;
loop_2d<packed_t>(mb0, mb1, nb0, nb1, BLOCK_N * IC, [&](int64_t mb, int64_t nb, int64_t nb_offset) {
int64_t m_size = offsets[mb + 1] - offsets[mb];
if (m_size == 0) {
return;
}
int64_t n_size = std::min(OC - nb * BLOCK_N, BLOCK_N);
const bool brg = can_use_brgemm<packed_t>(m_size, n_size);
use_brgemm |= brg;
// A ptr from ic1 of [M * topk, N] in sorted order
// so as to avoid copy A to tmp buffer again
@@ -187,7 +196,7 @@ void fused_experts_fp_kernel_impl(
/* lda */ IC,
/* ldb */ n_size,
/* ldc */ BLOCK_N,
/* brg */ use_brgemm,
/* brg */ brg,
/* block_size_K */ block_size_K,
/* do_unpack */ do_unpack);
@@ -31,6 +31,21 @@ def _prepack_experts(w: torch.Tensor) -> torch.Tensor:
return torch.ops._C.convert_weight_packed(w)
def _deterministic_expert_routes(
block_sizes: tuple[int, ...],
) -> tuple[torch.Tensor, torch.Tensor]:
"""Create top-1 routes with exact per-expert block occupancies."""
topk_ids = torch.cat(
[
torch.full((size,), expert, dtype=torch.int32)
for expert, size in enumerate(block_sizes)
]
)
topk_ids = topk_ids.view(-1, 1)
topk_weights = torch.ones(topk_ids.shape, dtype=torch.float32)
return topk_weights, topk_ids
# ===========================================================================
# FP8 W8A16 MoE
# ===========================================================================
@@ -219,6 +234,40 @@ def test_w8a16_block_fp8_cpu_fused_moe(M, N, K, E, topk, seed):
torch.testing.assert_close(out_inplace, out, atol=0, rtol=0)
def test_w8a16_block_fp8_cpu_fused_moe_small_expert_blocks():
"""Test FP8 BRGEMM at exact and multi-block expert boundaries."""
set_random_seed(0)
block_sizes = (4, 5, 33)
N, K, E = 128, 128, len(block_sizes)
M = sum(block_sizes)
a = torch.randn(M, K, dtype=torch.bfloat16) / math.sqrt(K)
w1, w2, w1_s, w2_s = _make_fp8_moe_weights(E, N, K, BLOCK_SIZE)
topk_weight, topk_ids = _deterministic_expert_routes(block_sizes)
ref_out = ref_w8a16_block_fp8_moe(
a, w1, w2, w1_s, w2_s, topk_weight, topk_ids, BLOCK_SIZE
)
pw1, pw2 = _prepack_experts(w1), _prepack_experts(w2)
out = ops.fused_experts_cpu(
a,
pw1,
pw2,
topk_weight,
topk_ids,
False,
ops.CPUQuantMethod.FP8_W8A16,
w1_s,
w2_s,
None,
None,
BLOCK_SIZE,
is_vnni=True,
)
torch.testing.assert_close(ref_out.bfloat16(), out, atol=1e-2, rtol=1e-2)
# ===========================================================================
# MXFP4 W4A16 MoE
# ===========================================================================
@@ -438,6 +487,48 @@ def test_mxfp4_cpu_fused_moe(M, N, K, E, topk, seed):
torch.testing.assert_close(ref_out.bfloat16(), out, atol=1e-2, rtol=1e-2)
def test_mxfp4_cpu_fused_moe_small_expert_blocks():
"""Test MXFP4 BRGEMM at exact and multi-block expert boundaries."""
set_random_seed(0)
block_sizes = (4, 5, 33)
N, K, E = 128, 128, len(block_sizes)
M = sum(block_sizes)
dtype = torch.bfloat16
a = torch.randn(M, K, dtype=dtype) / 10
w1_bf16 = torch.randn(E, 2 * N, K, dtype=dtype) / 10
w1q, w1s = MXFP4QuantizeUtil.quantize(w1_bf16)
w1s = w1s.reshape(E, 2 * N, K // 32)
w1dq = MXFP4QuantizeUtil.dequantize(w1q, dtype, w1s)
w2_bf16 = torch.randn(E, K, N, dtype=dtype) / 10
w2q, w2s = MXFP4QuantizeUtil.quantize(w2_bf16)
w2s = w2s.reshape(E, K, N // 32)
w2dq = MXFP4QuantizeUtil.dequantize(w2q, dtype, w2s)
topk_weight, topk_ids = _deterministic_expert_routes(block_sizes)
ref_out = ref_mxfp4_fused_moe(a, w1dq, w2dq, topk_weight, topk_ids, 1)
pw1, pw1s = _prepack_mxfp4_experts(w1q, w1s)
pw2, pw2s = _prepack_mxfp4_experts(w2q, w2s)
out = ops.fused_experts_cpu(
a,
pw1,
pw2,
topk_weight,
topk_ids,
False,
ops.CPUQuantMethod.MXFP4,
pw1s,
pw2s,
None,
None,
None,
)
torch.testing.assert_close(ref_out.bfloat16(), out, atol=1e-2, rtol=1e-2)
@pytest.mark.parametrize("M", [1, 32])
@pytest.mark.parametrize("N,K,E,topk", [(128, 128, 4, 2), (64, 64, 4, 2)])
@pytest.mark.parametrize("seed", [0])