Merged
fix: 修复 QuantMatmulActivationQuant golden #13974
panzhijaAdd on 12 days ago
fix: 修复 QuantMatmulActivationQuant golden #13974
Merged
Merged fromsmdbha/ops-nnmasterintocann/ops-nnmaster
Total: 3 files changed+74-37
| @@ -49,6 +49,32 @@ def mx_dtype(dtype): | |||
| 49 | return fp8_dtype(dtype) | 49 | return fp8_dtype(dtype) |
| 50 | 50 | ||
| 51 | 51 | ||
| 52 | +def _mx_operand_dtypes( | ||
| 53 | + x1, x2, output_dtype, input_dtypes=(), x1_dtype=None, x2_dtype=None | ||
| 54 | +): | ||
| 55 | + """Resolve logical MX types before promoted inputs lose their native dtype.""" | ||
| 56 | + input_dtypes = input_dtypes or () | ||
| 57 | + declared = (x1_dtype, x2_dtype) | ||
| 58 | + operands = [] | ||
| 59 | + for index, value in enumerate((x1, x2)): | ||
| 60 | + if value is None: | ||
| 61 | + raise ValueError("MX inputs require x1 and x2") | ||
| 62 | + dtype = declared[index] | ||
| 63 | + if dtype is None: | ||
| 64 | + dtype = input_dtypes[index] if index < len(input_dtypes) else value.dtype | ||
| 65 | + if str(dtype).removeprefix("torch.") in ("float32", "float64"): | ||
| 66 | + dtype = str(value.dtype).removeprefix("torch.") | ||
| 67 | + # ACLNN may promote both the values and their dtype metadata. | ||
| 68 | + if dtype in ("float32", "float64"): | ||
| 69 | + dtype = output_dtype | ||
| 70 | + operands.append(mx_dtype(dtype)) | ||
| 71 | + if operands[0] != output_dtype: | ||
| 72 | + raise ValueError("MX output dtype must match x1") | ||
| 73 | + if (operands[0] == "float4_e2m1") != (operands[1] == "float4_e2m1"): | ||
| 74 | + raise ValueError("MXFP4 requires both inputs to be FP4") | ||
| 75 | + return tuple(operands) | ||
| 76 | + | ||
| 77 | + | ||
| 52 | _FP32_TINY = np.float32(np.finfo(np.float32).tiny) | 78 | _FP32_TINY = np.float32(np.finfo(np.float32).tiny) |
| 53 | _QUANTIZE_CHUNK_ELEMENTS = 1 << 20 | 79 | _QUANTIZE_CHUNK_ELEMENTS = 1 << 20 |
| 54 | _MX_GROUP_SIZE = 32 | 80 | _MX_GROUP_SIZE = 32 |
| @@ -941,12 +967,9 @@ def _npu_activation_golden( | |||
| 941 | import torch.nn.functional as functional | 967 | import torch.nn.functional as functional |
| 942 | import torch_npu | 968 | import torch_npu |
| 943 | 969 | ||
| 944 | - a_dtype = mx_dtype(x1.dtype if x1_dtype is None else x1_dtype) | 970 | + a_dtype, b_dtype = _mx_operand_dtypes( |
| 945 | - b_dtype = mx_dtype(x2.dtype if x2_dtype is None else x2_dtype) | 971 | + x1, x2, dtype, x1_dtype=x1_dtype, x2_dtype=x2_dtype |
| 946 | - if dtype != a_dtype: | 972 | + ) |
| 947 | - raise ValueError("MX output dtype must match x1") | ||
| 948 | - if is_fp4 and b_dtype != dtype: | ||
| 949 | - raise ValueError("MXFP4 requires both inputs to be FP4") | ||
| 950 | a = _to_npu_tensor(x1, a_dtype, device) | 973 | a = _to_npu_tensor(x1, a_dtype, device) |
| 951 | b = ( | 974 | b = ( |
| 952 | _to_npu_tensor(x2, b_dtype, device) | 975 | _to_npu_tensor(x2, b_dtype, device) |
| @@ -1164,6 +1187,9 @@ class QuantMatmulActivationQuantTestSpec: | |||
| 1164 | dtype = mx_dtype(y_dtype) | 1187 | dtype = mx_dtype(y_dtype) |
| 1165 | if quant_mode != "mx": | 1188 | if quant_mode != "mx": |
| 1166 | raise ValueError("This TestSpec covers MX quantization") | 1189 | raise ValueError("This TestSpec covers MX quantization") |
| 1190 | + a_dtype, b_dtype = _mx_operand_dtypes( | ||
| 1191 | + x1, x2, dtype, kwargs.get("input_dtypes", ()) | ||
| 1192 | + ) | ||
| 1167 | npu_device = _npu_golden_device(device) | 1193 | npu_device = _npu_golden_device(device) |
| 1168 | weight_nz_shape = None | 1194 | weight_nz_shape = None |
| 1169 | input_formats = kwargs.get("input_formats", ()) | 1195 | input_formats = kwargs.get("input_formats", ()) |
| @@ -1175,9 +1201,8 @@ class QuantMatmulActivationQuantTestSpec: | |||
| 1175 | raise ValueError("WeightNZ kernel golden requires input_ori_shapes[1]") | 1201 | raise ValueError("WeightNZ kernel golden requires input_ori_shapes[1]") |
| 1176 | weight_nz_shape = input_ori_shapes[1] | 1202 | weight_nz_shape = input_ori_shapes[1] |
| 1177 | if npu_device is None: | 1203 | if npu_device is None: |
| 1178 | - x2 = weight_nz_to_nd(x2, weight_nz_shape, dtype) | 1204 | + x2 = weight_nz_to_nd(x2, weight_nz_shape, b_dtype) |
| 1179 | if npu_device is not None: | 1205 | if npu_device is not None: |
| 1180 | - input_dtypes = kwargs.get("input_dtypes", ()) | ||
| 1181 | return _npu_activation_golden( | 1206 | return _npu_activation_golden( |
| 1182 | x1, | 1207 | x1, |
| 1183 | x2, | 1208 | x2, |
| @@ -1193,8 +1218,8 @@ class QuantMatmulActivationQuantTestSpec: | |||
| 1193 | dst_type_max=dst_type_max, | 1218 | dst_type_max=dst_type_max, |
| 1194 | transpose_x1=transpose_x1, | 1219 | transpose_x1=transpose_x1, |
| 1195 | transpose_x2=transpose_x2, | 1220 | transpose_x2=transpose_x2, |
| 1196 | - x1_dtype=input_dtypes[0] if input_dtypes else None, | 1221 | + x1_dtype=a_dtype, |
| 1197 | - x2_dtype=input_dtypes[1] if input_dtypes else None, | 1222 | + x2_dtype=b_dtype, |
| 1198 | weight_nz_shape=weight_nz_shape, | 1223 | weight_nz_shape=weight_nz_shape, |
| 1199 | ) | 1224 | ) |
| 1200 | # TTK kernel path normalizes the first output dtype to float32, so the | 1225 | # TTK kernel path normalizes the first output dtype to float32, so the |
| @@ -1251,6 +1276,9 @@ class QuantMatmulActivationQuantAclnnTestSpec: | |||
| 1251 | if dtype is None: | 1276 | if dtype is None: |
| 1252 | dtype = str(yOut.dtype) if yOut is not None else str(x1.dtype) | 1277 | dtype = str(yOut.dtype) if yOut is not None else str(x1.dtype) |
| 1253 | dtype = mx_dtype(dtype) | 1278 | dtype = mx_dtype(dtype) |
| 1279 | + a_dtype, b_dtype = _mx_operand_dtypes( | ||
| 1280 | + x1, x2, dtype, kwargs.get("tensor_dtypes", ()) | ||
| 1281 | + ) | ||
| 1254 | npu_device = _npu_golden_device(device) | 1282 | npu_device = _npu_golden_device(device) |
| 1255 | weight_nz_shape = None | 1283 | weight_nz_shape = None |
| 1256 | if _is_fractal_nz(kwargs.get("tensor_formats", ())): | 1284 | if _is_fractal_nz(kwargs.get("tensor_formats", ())): |
| @@ -1260,9 +1288,8 @@ class QuantMatmulActivationQuantAclnnTestSpec: | |||
| 1260 | # Recover and decode that storage before _numpy() copies the view. | 1288 | # Recover and decode that storage before _numpy() copies the view. |
| 1261 | weight_nz_shape = tuple(x2.shape) | 1289 | weight_nz_shape = tuple(x2.shape) |
| 1262 | if npu_device is None: | 1290 | if npu_device is None: |
| 1263 | - x2 = weight_nz_to_nd(x2, weight_nz_shape, dtype) | 1291 | + x2 = weight_nz_to_nd(x2, weight_nz_shape, b_dtype) |
| 1264 | if npu_device is not None: | 1292 | if npu_device is not None: |
| 1265 | - tensor_dtypes = kwargs.get("tensor_dtypes", ()) | ||
| 1266 | return _npu_activation_golden( | 1293 | return _npu_activation_golden( |
| 1267 | x1, | 1294 | x1, |
| 1268 | x2, | 1295 | x2, |
| @@ -1278,8 +1305,8 @@ class QuantMatmulActivationQuantAclnnTestSpec: | |||
| 1278 | dst_type_max=dstTypeMax, | 1305 | dst_type_max=dstTypeMax, |
| 1279 | transpose_x1=transposeX1, | 1306 | transpose_x1=transposeX1, |
| 1280 | transpose_x2=transposeX2, | 1307 | transpose_x2=transposeX2, |
| 1281 | - x1_dtype=tensor_dtypes[0] if tensor_dtypes else None, | 1308 | + x1_dtype=a_dtype, |
| 1282 | - x2_dtype=tensor_dtypes[1] if tensor_dtypes else None, | 1309 | + x2_dtype=b_dtype, |
| 1283 | weight_nz_shape=weight_nz_shape, | 1310 | weight_nz_shape=weight_nz_shape, |
| 1284 | ) | 1311 | ) |
| 1285 | result = reference( | 1312 | result = reference( |
| @@ -1317,18 +1344,36 @@ class QuantMatmulActivationQuantE2ETestSpec: | |||
| 1317 | **kwargs, | 1344 | **kwargs, |
| 1318 | ): | 1345 | ): |
| 1319 | # PyTorch views already express the logical matrix: no second transpose. | 1346 | # PyTorch views already express the logical matrix: no second transpose. |
| 1347 | + tensor_dtypes = kwargs.get("tensor_dtypes", ()) | ||
| 1320 | dtype = output_dtype | 1348 | dtype = output_dtype |
| 1321 | if dtype is None: | 1349 | if dtype is None: |
| 1322 | - dtype = x1_dtype if x1_dtype is not None else str(x1.dtype) | 1350 | + dtype = x1_dtype |
| 1351 | + if dtype is None: | ||
| 1352 | + dtype = tensor_dtypes[0] if tensor_dtypes else x1.dtype | ||
| 1323 | dtype = mx_dtype(dtype) | 1353 | dtype = mx_dtype(dtype) |
| 1354 | + a_dtype, b_dtype = _mx_operand_dtypes( | ||
| 1355 | + x1, x2, dtype, tensor_dtypes, x1_dtype, x2_dtype | ||
| 1356 | + ) | ||
| 1357 | + quant_mode = kwargs.get("quant_mode", "mx") | ||
| 1358 | + if quant_mode != "mx": | ||
| 1359 | + raise ValueError("This TestSpec covers MX quantization") | ||
| 1360 | + group_sizes = kwargs.get("group_sizes") | ||
| 1361 | + if group_sizes is not None and tuple(group_sizes) not in ( | ||
| 1362 | + (), | ||
| 1363 | + (0, 0, 0), | ||
| 1364 | + (1, 1, 32), | ||
| 1365 | + ): | ||
| 1366 | + raise ValueError("Invalid MX group_sizes") | ||
| 1367 | + scale_alg = 0 if scale_alg is None else scale_alg | ||
| 1324 | npu_device = _npu_golden_device(device) | 1368 | npu_device = _npu_golden_device(device) |
| 1325 | if npu_device is not None: | 1369 | if npu_device is not None: |
| 1326 | - group_sizes = kwargs.get("group_sizes") | 1370 | + if _is_fractal_nz(kwargs.get("tensor_formats", ())): |
| 1327 | - if group_sizes is not None and tuple(group_sizes) not in ( | 1371 | + import torch_npu |
| 1328 | - (0, 0, 0), | 1372 | + |
| 1329 | - (1, 1, 32), | 1373 | + # Match the NZ binding applied before the actual E2E operator call. |
| 1330 | - ): | 1374 | + x2 = torch_npu.npu_format_cast( |
| 1331 | - raise ValueError("Invalid MX group_sizes") | 1375 | + _to_npu_tensor(x2, b_dtype, npu_device), 29 |
| 1376 | + ) | ||
| 1332 | return _npu_activation_golden( | 1377 | return _npu_activation_golden( |
| 1333 | x1, | 1378 | x1, |
| 1334 | x2, | 1379 | x2, |
| @@ -1338,23 +1383,21 @@ class QuantMatmulActivationQuantE2ETestSpec: | |||
| 1338 | device=npu_device, | 1383 | device=npu_device, |
| 1339 | output_dtype=dtype, | 1384 | output_dtype=dtype, |
| 1340 | activation_type=activation_type, | 1385 | activation_type=activation_type, |
| 1341 | - quant_mode=kwargs.get("quant_mode", "mx"), | 1386 | + quant_mode=quant_mode, |
| 1342 | round_mode=round_mode, | 1387 | round_mode=round_mode, |
| 1343 | scale_alg=scale_alg, | 1388 | scale_alg=scale_alg, |
| 1344 | dst_type_max=dst_type_max, | 1389 | dst_type_max=dst_type_max, |
| 1345 | - x1_dtype=x1_dtype, | 1390 | + x1_dtype=a_dtype, |
| 1346 | - x2_dtype=x2_dtype, | 1391 | + x2_dtype=b_dtype, |
| 1347 | pack_fp4_output=dtype == "float4_e2m1", | 1392 | pack_fp4_output=dtype == "float4_e2m1", |
| 1348 | ) | 1393 | ) |
| 1349 | is_fp4 = dtype == "float4_e2m1" | 1394 | is_fp4 = dtype == "float4_e2m1" |
| 1350 | if is_fp4: | 1395 | if is_fp4: |
| 1351 | - if mx_dtype(x1_dtype) != dtype or mx_dtype(x2_dtype) != dtype: | ||
| 1352 | - raise ValueError("MXFP4 E2E inputs require FP4 dtype attributes") | ||
| 1353 | x1_value = _unpack_fp4(x1) | 1396 | x1_value = _unpack_fp4(x1) |
| 1354 | x2_value = _unpack_fp4(x2) | 1397 | x2_value = _unpack_fp4(x2) |
| 1355 | else: | 1398 | else: |
| 1356 | - x1_value = _numpy_mx(x1, x1.dtype if x1_dtype is None else x1_dtype) | 1399 | + x1_value = _numpy_mx(x1, a_dtype) |
| 1357 | - x2_value = _numpy_mx(x2, x2.dtype if x2_dtype is None else x2_dtype) | 1400 | + x2_value = _numpy_mx(x2, b_dtype) |
| 1358 | result = reference( | 1401 | result = reference( |
| 1359 | x1_value, | 1402 | x1_value, |
| 1360 | x2_value, | 1403 | x2_value, |
Mmatmul/quant_matmul_activation_quant/tests/st/arch35/ttk_e2e_quant_matmul_activation_quant_st.csv+3-0
| @@ -49,3 +49,6 @@ gelu_tanh_fp4_stride_tt_scale2_max7,cann_ops_nn.ops.quant_matmul_activation_quan | |||
| 49 | gelu_tanh_fp4_byte_scales,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 33), (1, 66, 2), (4, 1, 2), (66,))","((32, 1), (33, 1), (132, 2, 1), (2, 2, 1), (1,))","('uint8', 'uint8', 'uint8', 'uint8', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_tanh', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 0, 'dst_type_max': 0.0}",(),"((0, 255), (0, 255), (126, 128), (126, 128), (-0.25, 0.25))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True | 49 | gelu_tanh_fp4_byte_scales,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 33), (1, 66, 2), (4, 1, 2), (66,))","((32, 1), (33, 1), (132, 2, 1), (2, 2, 1), (1,))","('uint8', 'uint8', 'uint8', 'uint8', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_tanh', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 0, 'dst_type_max': 0.0}",(),"((0, 255), (0, 255), (126, 128), (126, 128), (-0.25, 0.25))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True |
| 50 | gelu_erf_fp4_scale2_max7_float32_attr,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 33), (1, 66, 2), (4, 1, 2), (66,))","((32, 1), (33, 1), (132, 2, 1), (2, 2, 1), (1,))","('uint8', 'uint8', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_erf', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 2, 'dst_type_max': 7.00000001}",(),"((0, 0), (0, 0), (1.0, 1.0), (1.0, 1.0), (1.1754943508222875e-38, 1.1754943508222875e-38))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True | 50 | gelu_erf_fp4_scale2_max7_float32_attr,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 33), (1, 66, 2), (4, 1, 2), (66,))","((32, 1), (33, 1), (132, 2, 1), (2, 2, 1), (1,))","('uint8', 'uint8', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_erf', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 2, 'dst_type_max': 7.00000001}",(),"((0, 0), (0, 0), (1.0, 1.0), (1.0, 1.0), (1.1754943508222875e-38, 1.1754943508222875e-38))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True |
| 51 | gelu_tanh_fp4_ocp_ignores_dst_type_max,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 33), (1, 66, 2), (4, 1, 2), (66,))","((32, 1), (33, 1), (132, 2, 1), (2, 2, 1), (1,))","('uint8', 'uint8', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_tanh', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 0, 'dst_type_max': 2.0}",(),"((0, 255), (0, 255), (0.5, 2.0), (0.5, 2.0), (-0.25, 0.25))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True | 51 | gelu_tanh_fp4_ocp_ignores_dst_type_max,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 33), (1, 66, 2), (4, 1, 2), (66,))","((32, 1), (33, 1), (132, 2, 1), (2, 2, 1), (1,))","('uint8', 'uint8', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'ND', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_tanh', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 0, 'dst_type_max': 2.0}",(),"((0, 255), (0, 255), (0.5, 2.0), (0.5, 2.0), (-0.25, 0.25))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True |
| 52 | +silu_nz_transposed_view_e4m3,cann_ops_nn.ops.quant_matmul_activation_quant,"((3, 128), (128, 256), (2, 256, 2), (3, 2, 2), None)","((128, 1), (1, 128), (512, 2, 1), (4, 2, 1), None)","('float8_e4m3fn', 'float8_e4m3fn', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'FRACTAL_NZ', 'ND', 'ND', 'ND')","{'output_dtype': 24, 'group_sizes': [1, 1, 32], 'activation_type': 'silu', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 0}",(),"((-0.5, 0.5), (-0.25, 0.75), (0.5, 2.0), (0.5, 2.0), (None, None))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True | ||
| 53 | +silu_nz_optional_defaults_e5m2_e4m3,cann_ops_nn.ops.quant_matmul_activation_quant,"((3, 64), (64, 128), (1, 128, 2), (3, 1, 2), None)","((64, 1), (128, 1), (256, 2, 1), (2, 2, 1), None)","('float8_e5m2', 'float8_e4m3fn', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'FRACTAL_NZ', 'ND', 'ND', 'ND')","{'output_dtype': None, 'group_sizes': [], 'activation_type': 'silu', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': None}",(),"((-0.5, 0.5), (-0.25, 0.75), (0.5, 2.0), (0.5, 2.0), (None, None))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True | ||
| 54 | +gelu_tanh_fp4_nz_e2e,cann_ops_nn.ops.quant_matmul_activation_quant,"((4, 32), (64, 64), (1, 128, 2), (4, 1, 2), None)","((32, 1), (64, 1), (256, 2, 1), (2, 2, 1), None)","('uint8', 'uint8', 'float8_e8m0', 'float8_e8m0', 'float32')","('ND', 'FRACTAL_NZ', 'ND', 'ND', 'ND')","{'output_dtype': 296, 'x1_dtype': 296, 'x2_dtype': 296, 'x1scale_dtype': 293, 'x2scale_dtype': 293, 'group_sizes': [1, 1, 32], 'activation_type': 'gelu_tanh', 'quant_mode': 'mx', 'round_mode': 'rint', 'scale_alg': 0, 'dst_type_max': 0.0}",(),"((0, 255), (0, 255), (0.5, 2.0), (0.5, 2.0), (None, None))","((0.0, 0.0), (0.0, 0.0))","(0.0, 0.0)",True | ||
🟡 Medium Priority
新增的
_mx_operand_dtypes在检测到 TTK 把“输入值 + dtype 元数据”同时提升为 FP32 时(第 65-69 行),对 x1 和 x2 都 回退到output_dtype。证据链:
for index, value in enumerate((x1, x2))共用同一段兜底逻辑,dtype = output_dtype与index无关;output_dtype只代表输出(= x1)的逻辑类型,代码后面第 71-72 行也明说 “MX output dtype must match x1”。因此对 index=1 用 output_dtype 是一个未经声明的猜测;silu_nz_optional_defaults_e5m2_e4m3(x1=e5m2、x2=e4m3)、18/19/21/35/44/51 行等。一旦这些用例在 TTK 提升元数据的路径上运行,b_dtype会被判成 e5m2 而不是 e4m3。失败模式:
触发条件:TTK 同时提升值与 dtype 元数据到 FP32,且该用例 x1 与 x2 逻辑类型不同(CSV 中已有多条此类用例)。
建议:只在 index==0(x1,与输出同类型)时回退到 output_dtype;对 index==1 仅在输出为 FP4(此时两个输入必须同为 FP4)时才允许回退,其余情况抛出明确错误(如 "x2 logical dtype is ambiguous after FP32 promotion, please declare x2_dtype"),避免静默按错误类型量化。同时建议在 FP4 CPU 分支保持"输入必须是 packed uint8 载体"的断言,防止 FP32 值被当成 nibble 解包。