| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Tests for CPU FP8 scaled-mm GEMM kernels. |
| 4 | |
| 5 | Covers: |
| 6 | * fp8_scaled_mm_cpu (W8A16) |
| 7 | * fp8_scaled_mm_with_quant (W8A8) |
| 8 | |
| 9 | Run `pytest tests/kernels/quantization/test_cpu_fp8_scaled_mm.py -v`. |
| 10 | """ |
| 11 | |
| 12 | import pytest |
| 13 | import torch |
| 14 | |
| 15 | from vllm import _custom_ops as ops |
| 16 | from vllm.model_executor.kernels.linear.scaled_mm.cpu import ( |
| 17 | CPUFp8PerTensorScaledMMLinearKernel, |
| 18 | ) |
| 19 | from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import ( |
| 20 | FP8ScaledMMLinearLayerConfig, |
| 21 | ) |
| 22 | from vllm.model_executor.layers.quantization.utils.quant_utils import ( |
| 23 | kFp8StaticTensorSym, |
| 24 | ) |
| 25 | from vllm.platforms import current_platform |
| 26 | |
| 27 | if not current_platform.is_cpu(): |
| 28 | pytest.skip("skipping CPU-only tests", allow_module_level=True) |
| 29 | |
| 30 | requires_cpu_fp8_w8a16 = pytest.mark.skipif( |
| 31 | not ops._supports_cpu_fp8_w8a16, reason="fp8_scaled_mm_cpu op not available" |
| 32 | ) |
| 33 | requires_cpu_fp8_w8a8 = pytest.mark.skipif( |
| 34 | not ops.cpu_has_amx_fp8(), |
| 35 | reason="requires native AMX-FP8 hardware", |
| 36 | ) |
| 37 | |
| 38 | if not torch.cpu._is_amx_tile_supported(): |
| 39 | pytest.skip("requires AMX tile support", allow_module_level=True) |
| 40 | |
| 41 | FP8_MAX = torch.finfo(torch.float8_e4m3fn).max |
| 42 | BLOCK_SIZE = [128, 128] |
| 43 | GROUP_SIZE = 128 |
| 44 | |
| 45 | |
| 46 | def cdiv(a: int, b: int) -> int: |
| 47 | return -(a // -b) |
| 48 | |
| 49 | |
| 50 | # --------------------------------------------------------------------------- |
| 51 | # W8A16 helpers |
| 52 | # --------------------------------------------------------------------------- |
| 53 | |
| 54 | |
| 55 | def quantize_weight_block_fp8( |
| 56 | weight: torch.Tensor, |
| 57 | block_size: list[int], |
| 58 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 59 | """Quantize weight [N, K] to FP8 with block scales. |
| 60 | |
| 61 | Returns: |
| 62 | fp8_weight: [N, K] float8_e4m3fn |
| 63 | scales: [n_tiles, k_tiles] float32 |
| 64 | |
| 65 | """ |
| 66 | N, K = weight.shape |
| 67 | block_n, block_k = block_size |
| 68 | fp8_max = torch.finfo(torch.float8_e4m3fn).max |
| 69 | |
| 70 | n_tiles = cdiv(N, block_n) |
| 71 | k_tiles = cdiv(K, block_k) |
| 72 | |
| 73 | # Pad for even blocking |
| 74 | pad_N = (block_n - (N % block_n)) % block_n |
| 75 | pad_K = (block_k - (K % block_k)) % block_k |
| 76 | if pad_N > 0 or pad_K > 0: |
| 77 | weight = torch.nn.functional.pad(weight, (0, pad_K, 0, pad_N)) |
| 78 | |
| 79 | # Reshape into blocks |
| 80 | w_blocks = weight.view(n_tiles, block_n, k_tiles, block_k) |
| 81 | w_blocks = w_blocks.permute(0, 2, 1, 3).contiguous() |
| 82 | |
| 83 | # Per-block scale |
| 84 | abs_max = w_blocks.abs().amax(dim=(-2, -1), keepdim=True) |
| 85 | scales = abs_max / fp8_max |
| 86 | scales = torch.where(scales == 0, torch.ones_like(scales), scales) |
| 87 | |
| 88 | # Quantize |
| 89 | q_fp8 = (w_blocks / scales).clamp(-fp8_max, fp8_max).to(torch.float8_e4m3fn) |
| 90 | |
| 91 | # Reshape back |
| 92 | fp8_weight = ( |
| 93 | q_fp8.permute(0, 2, 1, 3) |
| 94 | .contiguous() |
| 95 | .view(N + pad_N, K + pad_K)[:N, :K] |
| 96 | .contiguous() |
| 97 | ) |
| 98 | |
| 99 | scales = scales.view(n_tiles, k_tiles) |
| 100 | return fp8_weight, scales |
| 101 | |
| 102 | |
| 103 | def dequant_weight_block_fp8( |
| 104 | fp8_weight: torch.Tensor, |
| 105 | scales: torch.Tensor, |
| 106 | block_size: list[int], |
| 107 | out_dtype: torch.dtype, |
| 108 | ) -> torch.Tensor: |
| 109 | """Dequantize FP8 weight back to float for reference computation.""" |
| 110 | N, K = fp8_weight.shape |
| 111 | block_n, block_k = block_size |
| 112 | n_tiles, k_tiles = scales.shape |
| 113 | |
| 114 | pad_N = (block_n - (N % block_n)) % block_n |
| 115 | pad_K = (block_k - (K % block_k)) % block_k |
| 116 | if pad_N > 0 or pad_K > 0: |
| 117 | fp8_padded = torch.nn.functional.pad(fp8_weight.float(), (0, pad_K, 0, pad_N)) |
| 118 | else: |
| 119 | fp8_padded = fp8_weight.float() |
| 120 | |
| 121 | w_blocks = fp8_padded.view(n_tiles, block_n, k_tiles, block_k) |
| 122 | w_blocks = w_blocks.permute(0, 2, 1, 3).contiguous() |
| 123 | dq = w_blocks * scales.view(n_tiles, k_tiles, 1, 1) |
| 124 | dq = dq.permute(0, 2, 1, 3).contiguous().view(N + pad_N, K + pad_K) |
| 125 | return dq[:N, :K].to(out_dtype) |
| 126 | |
| 127 | |
| 128 | def ref_fp8_block_scaled_mm( |
| 129 | x: torch.Tensor, |
| 130 | fp8_weight: torch.Tensor, |
| 131 | scales: torch.Tensor, |
| 132 | block_size: list[int], |
| 133 | bias: torch.Tensor | None, |
| 134 | out_dtype: torch.dtype, |
| 135 | ) -> torch.Tensor: |
| 136 | """Reference: dequant FP8→float32, matmul in float32, cast to out_dtype.""" |
| 137 | w_dq = dequant_weight_block_fp8(fp8_weight, scales, block_size, torch.float32) |
| 138 | out = torch.mm(x.float(), w_dq.t()) |
| 139 | if bias is not None: |
| 140 | out = out + bias.float() |
| 141 | return out.to(out_dtype) |
| 142 | |
| 143 | |
| 144 | # --------------------------------------------------------------------------- |
| 145 | # W8A16 test: fp8_scaled_mm_cpu |
| 146 | # --------------------------------------------------------------------------- |
| 147 | M_SIZES = [1, 4, 16, 64, 128] |
| 148 | # (N, K) — weight shape is [N, K], output has N columns. |
| 149 | NK_SIZES = [ |
| 150 | (128, 256), |
| 151 | (256, 512), |
| 152 | (512, 1024), |
| 153 | (1024, 2048), |
| 154 | (5120, 5120), |
| 155 | (17408, 5120), |
| 156 | (5120, 17408), |
| 157 | ] |
| 158 | |
| 159 | |
| 160 | @requires_cpu_fp8_w8a16 |
| 161 | @pytest.mark.parametrize("M", M_SIZES) |
| 162 | @pytest.mark.parametrize("N,K", NK_SIZES) |
| 163 | @pytest.mark.parametrize("use_bias", [False, True]) |
| 164 | def test_cpu_fp8_scaled_mm(M: int, N: int, K: int, use_bias: bool): |
| 165 | """fp8_scaled_mm_cpu correctness against float reference.""" |
| 166 | torch.manual_seed(42) |
| 167 | out_dtype = torch.bfloat16 |
| 168 | block_size = BLOCK_SIZE |
| 169 | |
| 170 | x = torch.randn(M, K, dtype=out_dtype) / (K**0.5) |
| 171 | w_f32 = torch.randn(N, K, dtype=torch.float32) / (K**0.5) |
| 172 | fp8_weight, scales = quantize_weight_block_fp8(w_f32, block_size) |
| 173 | |
| 174 | bias = torch.randn(N, dtype=torch.float32) * 0.1 if use_bias else None |
| 175 | |
| 176 | ref_out = ref_fp8_block_scaled_mm( |
| 177 | x, fp8_weight, scales, block_size, bias, out_dtype |
| 178 | ) |
| 179 | |
| 180 | packed_weight = torch.ops._C.convert_weight_packed(fp8_weight) |
| 181 | kernel_out = ops.fp8_scaled_mm_cpu( |
| 182 | x, |
| 183 | packed_weight, |
| 184 | scales, |
| 185 | block_size, |
| 186 | bias, |
| 187 | out_dtype, |
| 188 | True, |
| 189 | ) |
| 190 | |
| 191 | assert kernel_out.dtype == out_dtype |
| 192 | torch.testing.assert_close(kernel_out, ref_out, rtol=0.02, atol=0.01) |
| 193 | |
| 194 | |
| 195 | def quantize_weight_per_tensor_fp8( |
| 196 | weight: torch.Tensor, |
| 197 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 198 | """Quantize weight [N, K] to FP8 with a single per-tensor scale.""" |
| 199 | fp8_max = torch.finfo(torch.float8_e4m3fn).max |
| 200 | scale = weight.abs().amax() / fp8_max |
| 201 | scale = torch.where(scale == 0, torch.ones_like(scale), scale) |
| 202 | q = (weight / scale).clamp(-fp8_max, fp8_max).to(torch.float8_e4m3fn) |
| 203 | return q, scale |
| 204 | |
| 205 | |
| 206 | def ref_fp8_per_tensor_scaled_mm( |
| 207 | x: torch.Tensor, |
| 208 | fp8_weight: torch.Tensor, |
| 209 | scale: torch.Tensor, |
| 210 | bias: torch.Tensor | None, |
| 211 | out_dtype: torch.dtype, |
| 212 | ) -> torch.Tensor: |
| 213 | w_dq = fp8_weight.float() * scale |
| 214 | out = torch.mm(x.float(), w_dq.t()) |
| 215 | if bias is not None: |
| 216 | out = out + bias.float() |
| 217 | return out.to(out_dtype) |
| 218 | |
| 219 | |
| 220 | # --------------------------------------------------------------------------- |
| 221 | # W8A8 helpers |
| 222 | # --------------------------------------------------------------------------- |
| 223 | |
| 224 | |
| 225 | def quantize_weight_per_tensor( |
| 226 | weight: torch.Tensor, |
| 227 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 228 | """Quantize weight [N, K] to FP8 with a single per-tensor scale, broadcast |
| 229 | to [N, 1] (the shape float8_linear_prepack_cpu expects).""" |
| 230 | N = weight.shape[0] |
| 231 | abs_max = weight.abs().max() |
| 232 | scale = (abs_max / FP8_MAX).clamp(min=1e-7) |
| 233 | q = (weight.float() / scale).clamp(-FP8_MAX, FP8_MAX).to(torch.float8_e4m3fn) |
| 234 | return q, scale.expand(N).contiguous().view(N, 1) |
| 235 | |
| 236 | |
| 237 | def quantize_weight_per_channel( |
| 238 | weight: torch.Tensor, |
| 239 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 240 | """Quantize weight [N, K] to FP8 with per-channel (per-row) scales [N, 1].""" |
| 241 | abs_max = weight.abs().amax(dim=1, keepdim=True) |
| 242 | scale = (abs_max / FP8_MAX).clamp(min=1e-7) |
| 243 | q = (weight.float() / scale).clamp(-FP8_MAX, FP8_MAX).to(torch.float8_e4m3fn) |
| 244 | return q, scale.float() |
| 245 | |
| 246 | |
| 247 | def quantize_weight_per_group( |
| 248 | weight: torch.Tensor, |
| 249 | group_size: int = GROUP_SIZE, |
| 250 | ) -> tuple[torch.Tensor, torch.Tensor]: |
| 251 | """Quantize weight [N, K] to FP8 with per-group (128-block) scales [N, G].""" |
| 252 | N, K = weight.shape |
| 253 | G = K // group_size |
| 254 | w_groups = weight.view(N, G, group_size) |
| 255 | abs_max = w_groups.abs().amax(dim=2, keepdim=True) |
| 256 | scale = (abs_max / FP8_MAX).clamp(min=1e-7) |
| 257 | q = (w_groups.float() / scale).clamp(-FP8_MAX, FP8_MAX).to(torch.float8_e4m3fn) |
| 258 | return q.view(N, K), scale.squeeze(2).float() |
| 259 | |
| 260 | |
| 261 | def quantize_act_per_token(x: torch.Tensor) -> tuple[torch.Tensor, torch.Tensor]: |
| 262 | """Reference per-token (row) dynamic FP8 quantization of activations.""" |
| 263 | x_f = x.float() |
| 264 | scale = (x_f.abs().amax(dim=1, keepdim=True) / FP8_MAX).clamp(min=1e-7) |
| 265 | scaled = (x_f / scale).clamp(-FP8_MAX, FP8_MAX) |
| 266 | # Round via fp16 to match the AVX10.2 kernel path. |
| 267 | q = scaled.half().to(torch.float8_e4m3fn) |
| 268 | return q, scale |
| 269 | |
| 270 | |
| 271 | def reference_static_fp8_linear( |
| 272 | x: torch.Tensor, # BF16 [M, K] |
| 273 | weight_fp8: torch.Tensor, # FP8 [N, K] |
| 274 | weight_scale: torch.Tensor, # float32 [N, 1] |
| 275 | act_scale: torch.Tensor, # float32 scalar |
| 276 | bias: torch.Tensor | None, |
| 277 | out_dtype: torch.dtype, |
| 278 | ) -> torch.Tensor: |
| 279 | """Static per-tensor: quantize/dequant x with act_scale, dequant weight, matmul.""" |
| 280 | x_scaled = (x.float() / act_scale.item()).clamp(-FP8_MAX, FP8_MAX) |
| 281 | # Round via fp16 to match the AVX10.2 kernel path. |
| 282 | x_q = x_scaled.half().to(torch.float8_e4m3fn) |
| 283 | x_dq = x_q.float() * act_scale.item() |
| 284 | w_dq = weight_fp8.float() * weight_scale.view(-1, 1) |
| 285 | out = torch.mm(x_dq, w_dq.t()) |
| 286 | if bias is not None: |
| 287 | out = out + bias.float() |
| 288 | return out.to(out_dtype) |
| 289 | |
| 290 | |
| 291 | NK_SIZES_PER_TENSOR = [ |
| 292 | (32, 64), |
| 293 | (5120, 5120), |
| 294 | ] |
| 295 | |
| 296 | |
| 297 | @pytest.mark.parametrize("M", [1, 64]) |
| 298 | @pytest.mark.parametrize("N,K", NK_SIZES_PER_TENSOR) |
| 299 | @pytest.mark.parametrize("use_bias", [False, True]) |
| 300 | def test_cpu_fp8_per_tensor_scaled_mm_kernel( |
| 301 | M: int, N: int, K: int, use_bias: bool, default_vllm_config |
| 302 | ): |
| 303 | """CPUFp8PerTensorScaledMMLinearKernel correctness against float reference. |
| 304 | |
| 305 | Exercises the full kernel class (not just the raw op), including the |
| 306 | weight-orientation fixup: Fp8LinearMethod stores `layer.weight` as |
| 307 | [K, N] (torch._scaled_mm convention) before calling |
| 308 | process_weights_after_loading, so the kernel must transpose back to |
| 309 | [N, K] before VNNI-packing. |
| 310 | """ |
| 311 | torch.manual_seed(0) |
| 312 | out_dtype = torch.bfloat16 |
| 313 | |
| 314 | x = torch.randn(M, K, dtype=out_dtype) / (K**0.5) |
| 315 | w_f32 = torch.randn(N, K, dtype=torch.float32) / (K**0.5) |
| 316 | fp8_weight, scale = quantize_weight_per_tensor_fp8(w_f32) |
| 317 | bias = torch.randn(N, dtype=torch.float32) * 0.1 if use_bias else None |
| 318 | |
| 319 | ref_out = ref_fp8_per_tensor_scaled_mm(x, fp8_weight, scale, bias, out_dtype) |
| 320 | |
| 321 | config = FP8ScaledMMLinearLayerConfig( |
| 322 | weight_quant_key=kFp8StaticTensorSym, |
| 323 | activation_quant_key=kFp8StaticTensorSym, |
| 324 | weight_shape=(N, K), |
| 325 | input_dtype=out_dtype, |
| 326 | out_dtype=out_dtype, |
| 327 | ) |
| 328 | kernel = CPUFp8PerTensorScaledMMLinearKernel( |
| 329 | config, |
| 330 | layer_param_names=["weight", "weight_scale", "input_scale", "input_scale_ub"], |
| 331 | ) |
| 332 | |
| 333 | layer = torch.nn.Module() |
| 334 | # Fp8LinearMethod stores the weight transposed to [K, N] before calling |
| 335 | # process_weights_after_loading. |
| 336 | layer.register_parameter( |
| 337 | "weight", torch.nn.Parameter(fp8_weight.t().contiguous(), requires_grad=False) |
| 338 | ) |
| 339 | layer.register_parameter( |
| 340 | "weight_scale", torch.nn.Parameter(scale.clone(), requires_grad=False) |
| 341 | ) |
| 342 | kernel.process_weights_after_loading(layer) |
| 343 | |
| 344 | kernel_out = kernel.apply_weights(layer, x, bias) |
| 345 | |
| 346 | assert kernel_out.dtype == out_dtype |
| 347 | torch.testing.assert_close(kernel_out, ref_out, rtol=0.02, atol=0.01) |
| 348 | |
| 349 | |
| 350 | @pytest.mark.parametrize("n", [16, 48]) |
| 351 | def test_cpu_fp8_per_tensor_kernel_rejects_non_multiple_of_32_n(n: int): |
| 352 | """The AMX tinygemm kernel tiles N in chunks of 32 and cannot handle a |
| 353 | remainder tile smaller than that, so can_implement must reject any N |
| 354 | that isn't a multiple of 32 rather than let it crash the process. |
| 355 | """ |
| 356 | config = FP8ScaledMMLinearLayerConfig( |
| 357 | weight_quant_key=kFp8StaticTensorSym, |
| 358 | activation_quant_key=kFp8StaticTensorSym, |
| 359 | weight_shape=(n, 64), |
| 360 | input_dtype=torch.bfloat16, |
| 361 | out_dtype=torch.bfloat16, |
| 362 | ) |
| 363 | supported, _ = CPUFp8PerTensorScaledMMLinearKernel.can_implement(config) |
| 364 | assert not supported |
| 365 | |
| 366 | |
| 367 | def reference_dynamic_fp8_linear( |
| 368 | x: torch.Tensor, # BF16 [M, K] |
| 369 | weight_fp8: torch.Tensor, # FP8 [N, K] |
| 370 | weight_scale: torch.Tensor, # float32 [N, G] |
| 371 | bias: torch.Tensor | None, |
| 372 | out_dtype: torch.dtype, |
| 373 | ) -> torch.Tensor: |
| 374 | """Dynamic per-token act x per-group weight: quantize x per-token, |
| 375 | dequant weight per-group, matmul.""" |
| 376 | N, K = weight_fp8.shape |
| 377 | G = weight_scale.shape[1] |
| 378 | group_size = K // G |
| 379 | x_q, x_scale = quantize_act_per_token(x) |
| 380 | x_dq = x_q.float() * x_scale |
| 381 | w_dq = ( |
| 382 | weight_fp8.float().view(N, G, group_size) * weight_scale.view(N, G, 1) |
| 383 | ).view(N, K) |
| 384 | out = torch.mm(x_dq, w_dq.t()) |
| 385 | if bias is not None: |
| 386 | out = out + bias.float() |
| 387 | return out.to(out_dtype) |
| 388 | |
| 389 | |
| 390 | # --------------------------------------------------------------------------- |
| 391 | # W8A8 tests: fp8_scaled_mm_with_quant |
| 392 | # --------------------------------------------------------------------------- |
| 393 | W8A8_M_SIZES = [1, 4, 16, 64] |
| 394 | W8A8_NK_SIZES = [(128, 256), (256, 512), (64, 128)] |
| 395 | |
| 396 | |
| 397 | @requires_cpu_fp8_w8a8 |
| 398 | @pytest.mark.parametrize("static_activation", [True, False]) |
| 399 | @pytest.mark.parametrize("M", W8A8_M_SIZES) |
| 400 | @pytest.mark.parametrize("N,K", W8A8_NK_SIZES) |
| 401 | @pytest.mark.parametrize("use_bias", [False, True]) |
| 402 | def test_cpu_fp8_w8a8_per_tensor( |
| 403 | M: int, N: int, K: int, use_bias: bool, static_activation: bool |
| 404 | ): |
| 405 | """CPUFP8W8A8ScaledMMLinearKernel with per-tensor weight scale.""" |
| 406 | torch.manual_seed(42) |
| 407 | out_dtype = torch.bfloat16 |
| 408 | |
| 409 | x = torch.randn(M, K, dtype=torch.bfloat16) |
| 410 | weight = torch.randn(N, K, dtype=torch.bfloat16) |
| 411 | bias = torch.randn(N, dtype=torch.float32) if use_bias else None |
| 412 | |
| 413 | w_fp8, w_scale = quantize_weight_per_tensor(weight) |
| 414 | packed_weight, packed_scale = torch.ops._C.float8_linear_prepack_cpu(w_fp8, w_scale) |
| 415 | |
| 416 | if static_activation: |
| 417 | act_scale = (x.float().abs().max() / FP8_MAX).clamp(min=1e-7).view(1) |
| 418 | kernel_out = ops.fp8_scaled_mm_with_quant( |
| 419 | x, act_scale, False, packed_weight, packed_scale, bias, out_dtype |
| 420 | ) |
| 421 | ref_out = reference_static_fp8_linear( |
| 422 | x, w_fp8, w_scale, act_scale, bias, out_dtype |
| 423 | ) |
| 424 | atol = 0.05 |
| 425 | else: |
| 426 | kernel_out = ops.fp8_scaled_mm_with_quant( |
| 427 | x, None, True, packed_weight, packed_scale, bias, out_dtype |
| 428 | ) |
| 429 | ref_out = reference_dynamic_fp8_linear(x, w_fp8, w_scale, bias, out_dtype) |
| 430 | atol = 0.05 |
| 431 | |
| 432 | assert kernel_out.dtype == out_dtype |
| 433 | torch.testing.assert_close(kernel_out, ref_out, rtol=0.05, atol=atol) |
| 434 | |
| 435 | |
| 436 | @requires_cpu_fp8_w8a8 |
| 437 | @pytest.mark.parametrize("static_activation", [True, False]) |
| 438 | @pytest.mark.parametrize("M", W8A8_M_SIZES) |
| 439 | @pytest.mark.parametrize("N,K", W8A8_NK_SIZES) |
| 440 | @pytest.mark.parametrize("use_bias", [False, True]) |
| 441 | def test_cpu_fp8_w8a8_per_channel( |
| 442 | M: int, N: int, K: int, use_bias: bool, static_activation: bool |
| 443 | ): |
| 444 | """CPUFP8W8A8ScaledMMLinearKernel with per-channel weight scale.""" |
| 445 | torch.manual_seed(42) |
| 446 | out_dtype = torch.bfloat16 |
| 447 | |
| 448 | x = torch.randn(M, K, dtype=torch.bfloat16) |
| 449 | weight = torch.randn(N, K, dtype=torch.bfloat16) |
| 450 | bias = torch.randn(N, dtype=torch.float32) if use_bias else None |
| 451 | |
| 452 | w_fp8, w_scale = quantize_weight_per_channel(weight) |
| 453 | packed_weight, packed_scale = torch.ops._C.float8_linear_prepack_cpu(w_fp8, w_scale) |
| 454 | |
| 455 | if static_activation: |
| 456 | act_scale = (x.float().abs().max() / FP8_MAX).clamp(min=1e-7).view(1) |
| 457 | kernel_out = ops.fp8_scaled_mm_with_quant( |
| 458 | x, act_scale, False, packed_weight, packed_scale, bias, out_dtype |
| 459 | ) |
| 460 | ref_out = reference_static_fp8_linear( |
| 461 | x, w_fp8, w_scale, act_scale, bias, out_dtype |
| 462 | ) |
| 463 | atol = 0.05 |
| 464 | else: |
| 465 | kernel_out = ops.fp8_scaled_mm_with_quant( |
| 466 | x, None, True, packed_weight, packed_scale, bias, out_dtype |
| 467 | ) |
| 468 | ref_out = reference_dynamic_fp8_linear(x, w_fp8, w_scale, bias, out_dtype) |
| 469 | atol = 0.05 |
| 470 | |
| 471 | assert kernel_out.dtype == out_dtype |
| 472 | torch.testing.assert_close(kernel_out, ref_out, rtol=0.05, atol=atol) |
| 473 | |
| 474 | |
| 475 | @requires_cpu_fp8_w8a8 |
| 476 | @pytest.mark.parametrize("M", W8A8_M_SIZES) |
| 477 | @pytest.mark.parametrize("N,K", W8A8_NK_SIZES) |
| 478 | @pytest.mark.parametrize("use_bias", [False, True]) |
| 479 | def test_cpu_fp8_w8a8_per_group(M: int, N: int, K: int, use_bias: bool): |
| 480 | """CPUFp8W8A8BlockScaledMMKernel: per-group (128-block) weight scale + |
| 481 | dynamic per-token activation scale.""" |
| 482 | torch.manual_seed(42) |
| 483 | out_dtype = torch.bfloat16 |
| 484 | |
| 485 | x = torch.randn(M, K, dtype=torch.bfloat16) |
| 486 | weight = torch.randn(N, K, dtype=torch.bfloat16) |
| 487 | bias = torch.randn(N, dtype=torch.float32) if use_bias else None |
| 488 | |
| 489 | w_fp8, w_scale = quantize_weight_per_group(weight) |
| 490 | packed_weight, packed_scale = torch.ops._C.float8_linear_prepack_cpu(w_fp8, w_scale) |
| 491 | kernel_out = ops.fp8_scaled_mm_with_quant( |
| 492 | x, None, True, packed_weight, packed_scale, bias, out_dtype |
| 493 | ) |
| 494 | ref_out = reference_dynamic_fp8_linear(x, w_fp8, w_scale, bias, out_dtype) |
| 495 | |
| 496 | assert kernel_out.dtype == out_dtype |
| 497 | torch.testing.assert_close(kernel_out, ref_out, rtol=0.05, atol=0.05) |