| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Test comparing Marlin INT4 MoE vs FlashInfer TRT-LLM MXINT4 MoE.""" |
| 4 | |
| 5 | import pytest |
| 6 | import torch |
| 7 | |
| 8 | from vllm.model_executor.layers.fused_moe.activation import MoEActivation |
| 9 | from vllm.model_executor.layers.fused_moe.experts.marlin_moe import ( |
| 10 | fused_marlin_moe, |
| 11 | ) |
| 12 | from vllm.model_executor.layers.fused_moe.experts.trtllm_mxint4_moe import ( |
| 13 | TrtLlmMxint4ExpertsMonolithic, |
| 14 | ) |
| 15 | from vllm.model_executor.layers.fused_moe.router.grouped_topk_router import ( |
| 16 | grouped_topk, |
| 17 | ) |
| 18 | from vllm.model_executor.layers.quantization.utils.flashinfer_mxint4_moe import ( |
| 19 | prepare_static_weights_for_trtllm_mxint4_moe, |
| 20 | ) |
| 21 | from vllm.platforms import current_platform |
| 22 | from vllm.scalar_type import scalar_types |
| 23 | from vllm.utils.torch_utils import set_random_seed |
| 24 | |
| 25 | |
| 26 | def mxint4_quantize( |
| 27 | x: torch.Tensor, sf_vec_size: int = 32 |
| 28 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 29 | """Quantize BF16 tensor to MXINT4 with block scaling (group_size=sf_vec_size). |
| 30 | |
| 31 | Returns: |
| 32 | - uint8 packed (2 INT4/byte): [..., k//2] - stores SIGNED INT4 [-8, 7] |
| 33 | - scales in BF16: [..., k//sf_vec_size] |
| 34 | |
| 35 | """ |
| 36 | x_reshaped = x.reshape(-1, sf_vec_size) |
| 37 | x_max = x_reshaped.max(dim=-1, keepdim=True)[0].to(torch.float32) |
| 38 | x_min = x_reshaped.min(dim=-1, keepdim=True)[0].to(torch.float32) |
| 39 | x_max = x_max * 8.0 / 7.0 |
| 40 | amax = torch.where(x_max > -x_min, x_max, -x_min) |
| 41 | scales = amax / 8.0 |
| 42 | x_scaled = x_reshaped * scales.reciprocal() |
| 43 | x_int8 = ( |
| 44 | x_scaled.round().clamp(-8, 7).to(torch.int8).reshape(-1, sf_vec_size // 2, 2) |
| 45 | ) |
| 46 | x_int4 = (x_int8[..., 0] & 0x0F) | ((x_int8[..., 1] & 0x0F) << 4) |
| 47 | return ( |
| 48 | x_int4.to(torch.uint8).reshape(*x.shape[:-1], x.shape[-1] // 2), |
| 49 | scales.to(x.dtype).reshape(*x.shape[:-1], x.shape[-1] // sf_vec_size), |
| 50 | ) |
| 51 | |
| 52 | |
| 53 | def mxint4_quantize_moe_weights( |
| 54 | weights_bf16: torch.Tensor, group_size: int = 32 |
| 55 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 56 | """Quantize MoE weights [e, n, k] to MxInt4 format. |
| 57 | |
| 58 | Args: |
| 59 | weights_bf16: BF16 weights of shape [num_experts, out_features, in_features] |
| 60 | group_size: Quantization group size (default: 32) |
| 61 | |
| 62 | Returns: |
| 63 | - weights_mxint4: Quantized weights [e, n, k//2] uint8 |
| 64 | - scales_mxint4: Quantization scales [e, n, k//group_size] bf16 |
| 65 | |
| 66 | """ |
| 67 | e = weights_bf16.shape[0] |
| 68 | weight_list = [] |
| 69 | scale_list = [] |
| 70 | |
| 71 | for i in range(e): |
| 72 | w_q, w_s = mxint4_quantize(weights_bf16[i], sf_vec_size=group_size) |
| 73 | weight_list.append(w_q) |
| 74 | scale_list.append(w_s) |
| 75 | |
| 76 | return torch.stack(weight_list), torch.stack(scale_list) |
| 77 | |
| 78 | |
| 79 | __all__ = [ |
| 80 | "mxint4_quantize", |
| 81 | "mxint4_quantize_moe_weights", |
| 82 | "marlin_quantize_moe_weights", |
| 83 | ] |
| 84 | |
| 85 | |
| 86 | def test_trtllm_mxint4_activation_supports_vllm_gated_silu(): |
| 87 | assert TrtLlmMxint4ExpertsMonolithic._supports_activation(MoEActivation.SILU) |
| 88 | assert TrtLlmMxint4ExpertsMonolithic._supports_activation(MoEActivation.SWIGLUOAI) |
| 89 | assert not TrtLlmMxint4ExpertsMonolithic._supports_activation( |
| 90 | MoEActivation.RELU2_NO_MUL |
| 91 | ) |
| 92 | |
| 93 | |
| 94 | def marlin_quantize_moe_weights( |
| 95 | weights_bf16: torch.Tensor, group_size: int = 32 |
| 96 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 97 | """Quantize MoE weights [e, n, k] to Marlin INT4 format. |
| 98 | |
| 99 | Args: |
| 100 | weights_bf16: BF16 weights of shape [num_experts, out_features, in_features] |
| 101 | group_size: Quantization group size (default: 32) |
| 102 | |
| 103 | Returns: |
| 104 | - weights_marlin: Marlin quantized weights [e, k//8, n] int32 |
| 105 | - scales_marlin: Marlin quantization scales [e, k//group_size, n] bf16 |
| 106 | |
| 107 | """ |
| 108 | from vllm.model_executor.layers.quantization.utils.marlin_utils_test import ( |
| 109 | marlin_quantize, |
| 110 | ) |
| 111 | |
| 112 | e, n, k = weights_bf16.shape |
| 113 | weight_list = [] |
| 114 | scale_list = [] |
| 115 | |
| 116 | for i in range(e): |
| 117 | # Transpose for Marlin: [n, k] → [k, n] |
| 118 | w_t = weights_bf16[i].T.contiguous() |
| 119 | _, w_q, w_s = marlin_quantize(w_t, scalar_types.uint4b8, group_size) |
| 120 | weight_list.append(w_q) |
| 121 | scale_list.append(w_s) |
| 122 | |
| 123 | # Stack to get [e, ...] shape |
| 124 | weights_marlin = torch.stack(weight_list) # [e, k // 8, n] |
| 125 | scales_marlin = torch.stack(scale_list) # [e, k // group_size, n] |
| 126 | |
| 127 | return weights_marlin, scales_marlin |
| 128 | |
| 129 | |
| 130 | TRTLLM_GEN_AVAILABLE = ( |
| 131 | current_platform.is_cuda() and current_platform.is_device_capability_family(100) |
| 132 | ) |
| 133 | |
| 134 | |
| 135 | @pytest.mark.skipif(not TRTLLM_GEN_AVAILABLE, reason="Skip for non SM100") |
| 136 | @pytest.mark.parametrize("m", [1, 33]) |
| 137 | @pytest.mark.parametrize("n", [7168]) |
| 138 | @pytest.mark.parametrize("k", [512]) |
| 139 | @pytest.mark.parametrize("e", [384]) |
| 140 | @pytest.mark.parametrize("topk", [8]) |
| 141 | @pytest.mark.parametrize("group_size", [32]) |
| 142 | def test_marlin_vs_trtllm_mxint4_moe_kimik2(monkeypatch, m, n, k, e, topk, group_size): |
| 143 | """Compare Marlin INT4 MoE vs FlashInfer TRT-LLM MXINT4 MoE. |
| 144 | |
| 145 | Uses mxint4_quantize() to generate common INT4 weights + BF16 scales, |
| 146 | then runs both Marlin and TRT-LLM kernels and compares outputs. |
| 147 | """ |
| 148 | pytest.importorskip("flashinfer") |
| 149 | monkeypatch.setenv("VLLM_USE_FLASHINFER_MOE_INT4", "1") |
| 150 | |
| 151 | set_random_seed(0) |
| 152 | |
| 153 | dtype = torch.bfloat16 |
| 154 | |
| 155 | # DeepSeekV3 routing config (from Kimi-K2-Thinking config.json) |
| 156 | n_group = 1 # n_group from model config |
| 157 | topk_group = 1 # topk_group from model config |
| 158 | routed_scaling = 2.827 # routed_scaling_factor from model config |
| 159 | |
| 160 | # Input - realistic activation range for LLM (after LayerNorm: mean~0, std~1) |
| 161 | a = torch.randn((m, k), device="cuda", dtype=dtype) * 0.5 |
| 162 | |
| 163 | # Generate routing logits and bias (DeepSeekV3 expects float logits) |
| 164 | # Realistic ranges: logits typically [-3, 3], bias [-2, 2] |
| 165 | routing_logits = torch.randn((m, e), device="cuda", dtype=torch.float32) * 1.5 |
| 166 | routing_bias = torch.randn(e, device="cuda", dtype=torch.float32) * 0.8 |
| 167 | |
| 168 | # 1. Generate BF16 weights (SHARED between both paths) |
| 169 | # Realistic weight initialization: Xavier/Glorot uniform scaling |
| 170 | # std = sqrt(2 / (fan_in + fan_out)) |
| 171 | std_w1 = (2.0 / (k + 2 * n)) ** 0.5 |
| 172 | std_w2 = (2.0 / (n + k)) ** 0.5 |
| 173 | w1_bf16 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) * std_w1 |
| 174 | w2_bf16 = torch.randn((e, k, n), device="cuda", dtype=dtype) * std_w2 |
| 175 | |
| 176 | # === Path 1: TRT-LLM FlashInfer MXINT4 MoE === |
| 177 | # Similar to: if self.use_flashinfer_mxint4_moe |
| 178 | # Quantize using MXINT4 method (signed INT4) |
| 179 | w1_int4, w1_scales = mxint4_quantize_moe_weights(w1_bf16, group_size) |
| 180 | w2_int4, w2_scales = mxint4_quantize_moe_weights(w2_bf16, group_size) |
| 181 | |
| 182 | trtllm_weights = prepare_static_weights_for_trtllm_mxint4_moe( |
| 183 | gemm1_weights=w1_int4, |
| 184 | gemm1_scales=w1_scales, |
| 185 | gemm2_weights=w2_int4, |
| 186 | gemm2_scales=w2_scales, |
| 187 | ) |
| 188 | |
| 189 | from flashinfer import RoutingMethodType |
| 190 | from flashinfer.fused_moe import trtllm_mxint4_block_scale_moe |
| 191 | |
| 192 | # Routing handled internally by trtllm_mxint4_block_scale_moe |
| 193 | trtllm_output = trtllm_mxint4_block_scale_moe( |
| 194 | routing_logits=routing_logits, |
| 195 | routing_bias=routing_bias.to(torch.bfloat16), |
| 196 | hidden_states=a, |
| 197 | gemm1_weights=trtllm_weights["gemm1_weights"], |
| 198 | gemm1_weights_scale=trtllm_weights["gemm1_scales"], |
| 199 | gemm1_alpha=None, |
| 200 | gemm1_beta=None, |
| 201 | gemm1_clamp_limit=None, |
| 202 | gemm2_weights=trtllm_weights["gemm2_weights"], |
| 203 | gemm2_weights_scale=trtllm_weights["gemm2_scales"], |
| 204 | num_experts=e, |
| 205 | top_k=topk, |
| 206 | n_group=n_group, |
| 207 | topk_group=topk_group, |
| 208 | intermediate_size=n, |
| 209 | local_expert_offset=0, |
| 210 | local_num_experts=e, |
| 211 | routed_scaling_factor=routed_scaling, |
| 212 | routing_method_type=RoutingMethodType.DeepSeekV3, |
| 213 | enable_pdl=None, |
| 214 | output=None, |
| 215 | tune_max_num_tokens=8192, |
| 216 | ).to(dtype) |
| 217 | |
| 218 | # === Path 2: Marlin INT4 MoE === |
| 219 | # Similar to: else (non-flashinfer path) |
| 220 | # Quantize using Marlin's method (UINT4b8) |
| 221 | w1_marlin, w1_scales_marlin = marlin_quantize_moe_weights(w1_bf16, group_size) |
| 222 | w2_marlin, w2_scales_marlin = marlin_quantize_moe_weights(w2_bf16, group_size) |
| 223 | |
| 224 | # Use production routing kernel (same as router.select_experts internally uses) |
| 225 | topk_weights, topk_ids = grouped_topk( |
| 226 | hidden_states=a, |
| 227 | gating_output=routing_logits, |
| 228 | topk=topk, |
| 229 | renormalize=False, # DeepSeekV3 doesn't renormalize |
| 230 | num_expert_group=n_group, |
| 231 | topk_group=topk_group, |
| 232 | scoring_func="sigmoid", # DeepSeekV3 uses sigmoid |
| 233 | routed_scaling_factor=routed_scaling, |
| 234 | e_score_correction_bias=routing_bias, |
| 235 | ) |
| 236 | |
| 237 | marlin_output = fused_marlin_moe( |
| 238 | hidden_states=a, |
| 239 | w1=w1_marlin, |
| 240 | w2=w2_marlin, |
| 241 | bias1=None, |
| 242 | bias2=None, |
| 243 | w1_scale=w1_scales_marlin, |
| 244 | w2_scale=w2_scales_marlin, |
| 245 | topk_weights=topk_weights, |
| 246 | topk_ids=topk_ids, |
| 247 | quant_type_id=scalar_types.uint4b8.id, |
| 248 | global_num_experts=e, |
| 249 | expert_map=None, |
| 250 | global_scale1=None, |
| 251 | global_scale2=None, |
| 252 | input_global_scale1=None, |
| 253 | input_global_scale2=None, |
| 254 | w1_zeros=None, |
| 255 | w2_zeros=None, |
| 256 | input_dtype=dtype, |
| 257 | ) |
| 258 | |
| 259 | # Sanity check: manually compute BF16 reference for comparison |
| 260 | # Use same routing as Marlin path for consistency |
| 261 | bf16_output = torch.zeros((m, k), device="cuda", dtype=dtype) |
| 262 | for token_idx in range(m): |
| 263 | for expert_rank in range(topk): |
| 264 | expert_id = topk_ids[token_idx, expert_rank].item() |
| 265 | weight = topk_weights[token_idx, expert_rank].item() |
| 266 | # w1: [2*n, k] @ [k] -> [2*n] |
| 267 | up_gate = a[token_idx] @ w1_bf16[expert_id].T # [2*n] |
| 268 | gate, up = up_gate.chunk(2, dim=0) |
| 269 | intermediate = torch.nn.functional.silu(gate) * up # [n] |
| 270 | # w2: [k, n] @ [n] -> [k] |
| 271 | expert_out = intermediate @ w2_bf16[expert_id].T # [k] |
| 272 | bf16_output[token_idx] += weight * expert_out |
| 273 | # Compare against BF16 reference. |
| 274 | torch.testing.assert_close(marlin_output, bf16_output, atol=0.3, rtol=1.0) |
| 275 | torch.testing.assert_close(trtllm_output, bf16_output, atol=0.3, rtol=1.0) |
| 276 | |
| 277 | # Compare against each other for sanity. |
| 278 | # Note: Different quantization schemes (UINT4b8 vs signed MXINT4) cause |
| 279 | # some differences |
| 280 | torch.testing.assert_close(marlin_output, trtllm_output, atol=0.3, rtol=6.0) |
| 281 | |
| 282 | |
| 283 | @pytest.mark.skipif(not TRTLLM_GEN_AVAILABLE, reason="Skip for non SM100") |
| 284 | @pytest.mark.parametrize("m", [1, 33]) |
| 285 | @pytest.mark.parametrize("n", [7168]) |
| 286 | @pytest.mark.parametrize("k", [512]) |
| 287 | @pytest.mark.parametrize("e", [384]) |
| 288 | @pytest.mark.parametrize("topk", [8]) |
| 289 | @torch.inference_mode() |
| 290 | def test_flashinfer_trtllm_mxint4_moe_wrapper(m, n, k, e, topk): |
| 291 | """Test that the flashinfer_trtllm_mxint4_moe wrapper matches the raw |
| 292 | trtllm_mxint4_block_scale_moe kernel call.""" |
| 293 | pytest.importorskip("flashinfer") |
| 294 | from flashinfer import RoutingMethodType |
| 295 | from flashinfer.fused_moe import trtllm_mxint4_block_scale_moe |
| 296 | |
| 297 | from vllm.model_executor.layers.quantization.utils.flashinfer_mxint4_moe import ( |
| 298 | flashinfer_trtllm_mxint4_moe, |
| 299 | ) |
| 300 | |
| 301 | set_random_seed(0) |
| 302 | dtype = torch.bfloat16 |
| 303 | |
| 304 | a = torch.randn((m, k), device="cuda", dtype=dtype) * 0.5 |
| 305 | router_logits = torch.randn((m, e), device="cuda", dtype=torch.float32) * 1.5 |
| 306 | routing_bias = torch.randn(e, device="cuda", dtype=torch.float32) * 0.8 |
| 307 | |
| 308 | std_w1 = (2.0 / (k + 2 * n)) ** 0.5 |
| 309 | std_w2 = (2.0 / (n + k)) ** 0.5 |
| 310 | w1_bf16 = torch.randn((e, 2 * n, k), device="cuda", dtype=dtype) * std_w1 |
| 311 | w2_bf16 = torch.randn((e, k, n), device="cuda", dtype=dtype) * std_w2 |
| 312 | |
| 313 | w1_int4, w1_scales = mxint4_quantize_moe_weights(w1_bf16) |
| 314 | w2_int4, w2_scales = mxint4_quantize_moe_weights(w2_bf16) |
| 315 | |
| 316 | prepared = prepare_static_weights_for_trtllm_mxint4_moe( |
| 317 | gemm1_weights=w1_int4, |
| 318 | gemm1_scales=w1_scales, |
| 319 | gemm2_weights=w2_int4, |
| 320 | gemm2_scales=w2_scales, |
| 321 | ) |
| 322 | |
| 323 | # Raw kernel call (reference) |
| 324 | raw_out = trtllm_mxint4_block_scale_moe( |
| 325 | routing_logits=router_logits.to(torch.float32), |
| 326 | routing_bias=routing_bias.to(torch.bfloat16), |
| 327 | hidden_states=a, |
| 328 | gemm1_weights=prepared["gemm1_weights"].data, |
| 329 | gemm1_weights_scale=prepared["gemm1_scales"].data, |
| 330 | gemm1_alpha=None, |
| 331 | gemm1_beta=None, |
| 332 | gemm1_clamp_limit=None, |
| 333 | gemm2_weights=prepared["gemm2_weights"].data, |
| 334 | gemm2_weights_scale=prepared["gemm2_scales"].data, |
| 335 | num_experts=e, |
| 336 | top_k=topk, |
| 337 | n_group=1, |
| 338 | topk_group=1, |
| 339 | intermediate_size=n, |
| 340 | local_expert_offset=0, |
| 341 | local_num_experts=e, |
| 342 | routed_scaling_factor=None, |
| 343 | routing_method_type=RoutingMethodType.DeepSeekV3, |
| 344 | enable_pdl=None, |
| 345 | output=None, |
| 346 | tune_max_num_tokens=8192, |
| 347 | ) |
| 348 | if not isinstance(raw_out, torch.Tensor): |
| 349 | raw_out = raw_out[0] |
| 350 | raw_out = raw_out.to(dtype) |
| 351 | |
| 352 | # Wrapper call |
| 353 | wrapper_out = flashinfer_trtllm_mxint4_moe( |
| 354 | x=a, |
| 355 | router_logits=router_logits, |
| 356 | w13_weight_packed=prepared["gemm1_weights"], |
| 357 | w13_weight_scale=prepared["gemm1_scales"], |
| 358 | w2_weight_packed=prepared["gemm2_weights"], |
| 359 | w2_weight_scale=prepared["gemm2_scales"], |
| 360 | global_num_experts=e, |
| 361 | top_k=topk, |
| 362 | intermediate_size_per_partition=n, |
| 363 | local_num_experts=e, |
| 364 | ep_rank=0, |
| 365 | num_expert_group=1, |
| 366 | topk_group=1, |
| 367 | e_score_correction_bias=routing_bias, |
| 368 | routing_method_type=RoutingMethodType.DeepSeekV3, |
| 369 | ) |
| 370 | |
| 371 | assert wrapper_out.shape == (m, k) |
| 372 | assert wrapper_out.dtype == dtype |
| 373 | torch.testing.assert_close(wrapper_out, raw_out, atol=0.0, rtol=0.0) |