| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for the CUTLASS-based W4A8 grouped GEMM kernel and the full MoE layer.""" |
| 4 | |
| 5 | import random |
| 6 | from dataclasses import dataclass |
| 7 | |
| 8 | import pytest |
| 9 | import torch |
| 10 | |
| 11 | from vllm import _custom_ops as ops |
| 12 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 13 | pack_rows, |
| 14 | quantize_weights, |
| 15 | ) |
| 16 | from vllm.platforms import current_platform |
| 17 | from vllm.scalar_type import ScalarType, scalar_types |
| 18 | from vllm.utils.torch_utils import set_random_seed |
| 19 | |
| 20 | capability = current_platform.get_device_capability() |
| 21 | IS_SUPPORTED_BY_GPU = ( |
| 22 | current_platform.is_cuda() and capability is not None and capability[0] >= 9 |
| 23 | ) |
| 24 | |
| 25 | |
| 26 | def to_fp8(tensor: torch.Tensor) -> torch.Tensor: |
| 27 | finfo = torch.finfo(torch.float8_e4m3fn) |
| 28 | return tensor.clamp(min=finfo.min, max=finfo.max).to(dtype=torch.float8_e4m3fn) |
| 29 | |
| 30 | |
| 31 | def cutlass_quantize( |
| 32 | atype: torch.dtype, |
| 33 | w: torch.Tensor, |
| 34 | wtype: ScalarType, |
| 35 | stype: torch.dtype | None, |
| 36 | group_size: int | None, |
| 37 | zero_points: bool = False, |
| 38 | ): |
| 39 | """Quantize weights into W4 and compute reference dequantized weights. |
| 40 | |
| 41 | Encoding/reordering of weights and packing of scales is deferred |
| 42 | until after all experts are combined. |
| 43 | """ |
| 44 | assert wtype.is_integer(), "TODO: support floating point weights" |
| 45 | |
| 46 | w_ref, w_q, w_s, w_zp = quantize_weights( |
| 47 | w, wtype, group_size=group_size, zero_points=zero_points |
| 48 | ) |
| 49 | |
| 50 | # Since scales are later cast to fp8, recompute w_ref in atype here. |
| 51 | w_ref = ( |
| 52 | w_q.to(torch.float32) |
| 53 | * w_s.to(atype).to(torch.float32).repeat_interleave(group_size, dim=0) |
| 54 | ).to(atype) |
| 55 | |
| 56 | # Bit mask prevents sign extension of int4 when packing. |
| 57 | w_q = pack_rows(w_q & 0x0F, wtype.size_bits, *w_q.shape) |
| 58 | # Make weights row-major (N, K). |
| 59 | w_q = w_q.t().contiguous() |
| 60 | |
| 61 | return w_ref, w_q, w_s.to(atype), w_zp |
| 62 | |
| 63 | |
| 64 | def cutlass_preprocess( |
| 65 | w_q_experts: list[torch.Tensor], w_s_experts: list[torch.Tensor] |
| 66 | ): |
| 67 | """Reorder/encode expert weights and pack scales. |
| 68 | |
| 69 | Returns: |
| 70 | w_q_packed: Packed/encoded int4 weights for all experts. |
| 71 | w_s_packed: Packed fp8 scales for all experts. |
| 72 | packed_layout: Layout/stride metadata for grouped GEMM. |
| 73 | |
| 74 | """ |
| 75 | w_s_packed = ops.cutlass_pack_scale_fp8(torch.stack(w_s_experts)) |
| 76 | w_q_packed, packed_layout = ops.cutlass_encode_and_reorder_int4b_grouped( |
| 77 | torch.stack(w_q_experts) |
| 78 | ) # expects dim 3 |
| 79 | return w_q_packed, w_s_packed, packed_layout |
| 80 | |
| 81 | |
| 82 | GROUP_SIZE = 128 |
| 83 | # (num_experts, N, K) |
| 84 | TEST_SHAPES = [ |
| 85 | (8, 512, 2048), |
| 86 | (8, 2048, 2048), |
| 87 | (64, 512, 1024), |
| 88 | (64, 2048, 2048), |
| 89 | (4, 2048, 768), |
| 90 | (8, 768, 2048), |
| 91 | (64, 1536, 2048), |
| 92 | (128, 8192, 4096), # test overflow int32 |
| 93 | ] |
| 94 | ALIGNMENT = 16 # torch._scaled_mm alignment for M, needed for reference check |
| 95 | |
| 96 | |
| 97 | @dataclass |
| 98 | class MoETestSetup: |
| 99 | num_experts: int |
| 100 | K: int |
| 101 | N: int |
| 102 | Ms: list[int] |
| 103 | M_full: int |
| 104 | a: torch.Tensor |
| 105 | a_ref: torch.Tensor |
| 106 | a_strides: torch.Tensor |
| 107 | out: torch.Tensor |
| 108 | c_strides: torch.Tensor |
| 109 | per_tok_scales: torch.Tensor |
| 110 | per_chan_scales: torch.Tensor |
| 111 | w_refs: list[torch.Tensor] |
| 112 | w_q_packed: torch.Tensor |
| 113 | w_s_packed: torch.Tensor |
| 114 | problem_sizes: torch.Tensor |
| 115 | expert_offsets: torch.Tensor |
| 116 | b_strides: torch.Tensor |
| 117 | group_scale_strides: torch.Tensor |
| 118 | |
| 119 | |
| 120 | def make_moe_test_setup( |
| 121 | num_experts: int, |
| 122 | K: int, |
| 123 | N: int, |
| 124 | *, |
| 125 | alignment: int = ALIGNMENT, |
| 126 | max_blocks: int = 64, |
| 127 | device: str = "cuda", |
| 128 | random_zero: bool = False, |
| 129 | ) -> MoETestSetup: |
| 130 | """Create a full set of tensors for testing cutlass_w4a8_moe_mm.""" |
| 131 | assert K % GROUP_SIZE == 0 |
| 132 | # Token counts per expert (multiples of `alignment`). |
| 133 | Ms = [alignment * random.randint(1, max_blocks) for _ in range(num_experts)] |
| 134 | |
| 135 | # set random experts to 0 tokens |
| 136 | if random_zero and num_experts > 1: |
| 137 | num_zero = max(1, num_experts // 8) |
| 138 | zero_indices = random.sample(range(num_experts), k=num_zero) |
| 139 | for idx in zero_indices: |
| 140 | Ms[idx] = 0 |
| 141 | |
| 142 | M_full = sum(Ms) |
| 143 | assert M_full > 0 |
| 144 | |
| 145 | # Activations. |
| 146 | a = to_fp8(torch.randn((M_full, K), device=device)) |
| 147 | a_ref = a.to(torch.float32) |
| 148 | a_strides = torch.full((num_experts,), K, dtype=torch.int64, device=device) |
| 149 | |
| 150 | # Output buffer. |
| 151 | out = torch.empty((M_full, N), dtype=torch.bfloat16, device=device) |
| 152 | c_strides = torch.full((num_experts,), N, dtype=torch.int64, device=device) |
| 153 | |
| 154 | # Channel/token scales. |
| 155 | per_tok_scales = torch.randn((M_full, 1), dtype=torch.float32, device=device) |
| 156 | per_chan_scales = torch.randn( |
| 157 | (num_experts, N, 1), dtype=torch.float32, device=device |
| 158 | ) |
| 159 | |
| 160 | # Expert weights and scales. |
| 161 | wtype = scalar_types.int4 |
| 162 | atype = stype = torch.float8_e4m3fn |
| 163 | w_refs, w_qs, w_ss = [], [], [] |
| 164 | for _ in range(num_experts): |
| 165 | b = to_fp8(torch.randn((K, N), device=device)) |
| 166 | w_ref, w_q, w_s, _ = cutlass_quantize( |
| 167 | atype, b.to(torch.float16), wtype, stype, GROUP_SIZE, zero_points=False |
| 168 | ) |
| 169 | w_refs.append(w_ref) |
| 170 | w_qs.append(w_q) |
| 171 | w_ss.append(w_s) |
| 172 | |
| 173 | w_q_packed, w_s_packed, packed_layout = cutlass_preprocess(w_qs, w_ss) |
| 174 | |
| 175 | problem_sizes = torch.tensor( |
| 176 | [[N, M, K] for M in Ms], dtype=torch.int32, device=device |
| 177 | ) |
| 178 | |
| 179 | expert_offsets = torch.cat( |
| 180 | [ |
| 181 | torch.tensor([0], dtype=torch.int64), |
| 182 | torch.cumsum(torch.tensor(Ms, dtype=torch.int64), dim=0)[:-1], |
| 183 | ] |
| 184 | ).to(device=device) |
| 185 | |
| 186 | # B strides and group scale strides. |
| 187 | b_strides = packed_layout |
| 188 | group_scale_strides = torch.zeros( |
| 189 | (num_experts, 2), dtype=torch.int64, device=device |
| 190 | ) |
| 191 | group_scale_strides[:, 0] = N |
| 192 | |
| 193 | return MoETestSetup( |
| 194 | num_experts=num_experts, |
| 195 | K=K, |
| 196 | N=N, |
| 197 | Ms=Ms, |
| 198 | M_full=M_full, |
| 199 | a=a, |
| 200 | a_ref=a_ref, |
| 201 | a_strides=a_strides, |
| 202 | out=out, |
| 203 | c_strides=c_strides, |
| 204 | per_tok_scales=per_tok_scales, |
| 205 | per_chan_scales=per_chan_scales, |
| 206 | w_refs=w_refs, |
| 207 | w_q_packed=w_q_packed, |
| 208 | w_s_packed=w_s_packed, |
| 209 | problem_sizes=problem_sizes, |
| 210 | expert_offsets=expert_offsets, |
| 211 | b_strides=b_strides, |
| 212 | group_scale_strides=group_scale_strides, |
| 213 | ) |
| 214 | |
| 215 | |
| 216 | def compute_moe_reference_output(setup: MoETestSetup) -> torch.Tensor: |
| 217 | """Compute reference output using torch._scaled_mm per expert.""" |
| 218 | out_ref = torch.empty_like(setup.out) |
| 219 | |
| 220 | ends = torch.cumsum(torch.tensor(setup.Ms), 0).tolist() |
| 221 | starts = setup.expert_offsets.cpu().tolist() |
| 222 | |
| 223 | for i in range(setup.num_experts): |
| 224 | start, end = starts[i], ends[i] |
| 225 | if start == end: |
| 226 | continue |
| 227 | |
| 228 | out_ref_i = torch._scaled_mm( |
| 229 | setup.a_ref[start:end].to(torch.float8_e4m3fn), |
| 230 | setup.w_refs[i].to(torch.float8_e4m3fn).t().contiguous().t(), |
| 231 | setup.per_tok_scales[start:end], # (M, 1) |
| 232 | setup.per_chan_scales[i].reshape(1, -1), # (1, N) |
| 233 | out_dtype=torch.bfloat16, |
| 234 | use_fast_accum=True, |
| 235 | ) |
| 236 | out_ref[start:end] = out_ref_i |
| 237 | |
| 238 | return out_ref |
| 239 | |
| 240 | |
| 241 | @pytest.mark.skipif( |
| 242 | not IS_SUPPORTED_BY_GPU, |
| 243 | reason="W4A8 Grouped GEMM is not supported on this GPU type.", |
| 244 | ) |
| 245 | @pytest.mark.parametrize("shape", TEST_SHAPES) |
| 246 | @pytest.mark.parametrize("random_zero", [True, False]) |
| 247 | def test_cutlass_w4a8_moe_mm_end_to_end(shape, random_zero): |
| 248 | num_experts, N, K = shape |
| 249 | set_random_seed(42) |
| 250 | setup = make_moe_test_setup( |
| 251 | num_experts=num_experts, K=K, N=N, max_blocks=64, random_zero=random_zero |
| 252 | ) |
| 253 | |
| 254 | ops.cutlass_w4a8_moe_mm( |
| 255 | setup.out, |
| 256 | setup.a, |
| 257 | setup.w_q_packed, |
| 258 | setup.per_tok_scales, |
| 259 | setup.per_chan_scales, |
| 260 | setup.w_s_packed, |
| 261 | GROUP_SIZE, |
| 262 | setup.expert_offsets, |
| 263 | setup.problem_sizes, |
| 264 | setup.a_strides, |
| 265 | setup.b_strides, |
| 266 | setup.c_strides, |
| 267 | setup.group_scale_strides, |
| 268 | ) |
| 269 | torch.accelerator.synchronize() |
| 270 | |
| 271 | out_ref = compute_moe_reference_output(setup) |
| 272 | torch.testing.assert_close(setup.out, out_ref, rtol=1e-2, atol=1e-2) |
| 273 | |
| 274 | |
| 275 | class W4A8MoELayer(torch.nn.Module): |
| 276 | """Minimal wrapper module to test cuda graphs.""" |
| 277 | |
| 278 | def __init__(self, setup: MoETestSetup): |
| 279 | super().__init__() |
| 280 | self.setup = setup |
| 281 | |
| 282 | def forward(self, a: torch.Tensor) -> torch.Tensor: |
| 283 | s = self.setup |
| 284 | ops.cutlass_w4a8_moe_mm( |
| 285 | s.out, |
| 286 | a, |
| 287 | s.w_q_packed, |
| 288 | s.per_tok_scales, |
| 289 | s.per_chan_scales, |
| 290 | s.w_s_packed, |
| 291 | GROUP_SIZE, |
| 292 | s.expert_offsets, |
| 293 | s.problem_sizes, |
| 294 | s.a_strides, |
| 295 | s.b_strides, |
| 296 | s.c_strides, |
| 297 | s.group_scale_strides, |
| 298 | ) |
| 299 | return s.out |
| 300 | |
| 301 | |
| 302 | @pytest.mark.skipif( |
| 303 | not IS_SUPPORTED_BY_GPU, |
| 304 | reason="W4A8 Grouped GEMM is not supported on this GPU type.", |
| 305 | ) |
| 306 | def test_cutlass_w4a8_moe_mm_cuda_graph(): |
| 307 | set_random_seed(42) |
| 308 | # Fixed config for CUDA graph test (single parameter point). |
| 309 | num_experts = 8 |
| 310 | K = 512 |
| 311 | N = 2048 |
| 312 | |
| 313 | setup = make_moe_test_setup( |
| 314 | num_experts=num_experts, |
| 315 | K=K, |
| 316 | N=N, |
| 317 | max_blocks=32, |
| 318 | ) |
| 319 | |
| 320 | # Construct model that calls the grouped GEMM kernel. |
| 321 | model = W4A8MoELayer(setup) |
| 322 | |
| 323 | # Build reference output once. |
| 324 | out_ref = compute_moe_reference_output(setup) |
| 325 | |
| 326 | # Capture and run the model in a CUDA graph. |
| 327 | a_static = setup.a.clone() # static input tensor for graph replay |
| 328 | |
| 329 | stream = torch.cuda.Stream() |
| 330 | with torch.cuda.stream(stream): |
| 331 | g = torch.cuda.CUDAGraph() |
| 332 | with torch.cuda.graph(g): |
| 333 | out_static = model(a_static) |
| 334 | |
| 335 | out_static.zero_() |
| 336 | g.replay() |
| 337 | |
| 338 | torch.testing.assert_close(out_static, out_ref, rtol=1e-2, atol=1e-2) |