common.h 35.1 KB
Newer Older
Przemek Tredak's avatar
Przemek Tredak committed
1
/*************************************************************************
2
 * Copyright (c) 2022-2025, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
Przemek Tredak's avatar
Przemek Tredak committed
3
4
5
6
7
8
9
 *
 * See LICENSE for license information.
 ************************************************************************/

#ifndef TRANSFORMER_ENGINE_COMMON_COMMON_H_
#define TRANSFORMER_ENGINE_COMMON_COMMON_H_

10
#include <cudaTypedefs.h>
11
12
#define FP4_TYPE_SUPPORTED (CUDA_VERSION >= 12080)

13
14
15
#include <cuda_bf16.h>
#include <cuda_fp16.h>
#include <cuda_fp8.h>
16
17
18
19
#if FP4_TYPE_SUPPORTED
#include <cuda_fp4.h>
#endif

20
21
22
#include <cuda_runtime_api.h>
#include <transformer_engine/transformer_engine.h>

23
#include <cstdint>
Przemek Tredak's avatar
Przemek Tredak committed
24
25
26
27
#include <functional>
#include <stdexcept>
#include <string>
#include <tuple>
Tim Moon's avatar
Tim Moon committed
28
29
#include <type_traits>
#include <unordered_map>
Przemek Tredak's avatar
Przemek Tredak committed
30
#include <vector>
Tim Moon's avatar
Tim Moon committed
31
32

#include "./nvtx.h"
33
#include "./util/cuda_driver.h"
Tim Moon's avatar
Tim Moon committed
34
#include "./util/logging.h"
Przemek Tredak's avatar
Przemek Tredak committed
35
36
37

namespace transformer_engine {

38
39
40
std::string to_string(const DType type);
std::string to_string(const NVTEScalingMode &mode);

41
42
43
44
45
46
47
48
49
50
inline bool is_tensor_scaling(const NVTEScalingMode &mode) {
  return mode == NVTE_DELAYED_TENSOR_SCALING;
}

inline bool is_block_scaling(const NVTEScalingMode &mode) { return !is_tensor_scaling(mode); }

inline bool is_delayed_tensor_scaling(const NVTEScalingMode &mode) {
  return mode == NVTE_DELAYED_TENSOR_SCALING;
}

51
52
53
54
inline bool is_nvfp4_scaling(const NVTEScalingMode &mode) { return mode == NVTE_NVFP4_1D_SCALING; }

inline bool is_mxfp8_scaling(const NVTEScalingMode &mode) { return mode == NVTE_MXFP8_1D_SCALING; }

55
56
inline bool is_mxfp_scaling(const NVTEScalingMode &mode) { return mode == NVTE_MXFP8_1D_SCALING; }

57
58
inline bool is_nvfp_scaling(const NVTEScalingMode &mode) { return mode == NVTE_NVFP4_1D_SCALING; }

59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
inline size_t product(const std::vector<size_t> &shape, const size_t begin, const size_t end) {
  NVTE_CHECK(begin <= end && end <= shape.size(), "Attempted to access entries ", begin, " to ",
             end, " in a vector with ", shape.size(), " entries");
  size_t ret = 1;
  for (size_t i = begin; i < end; ++i) {
    ret *= shape[i];
  }
  return ret;
}

inline size_t product(const std::vector<size_t> &shape) {
  size_t ret = 1;
  for (const auto &elem : shape) {
    ret *= elem;
  }
  return ret;
}

77
78
79
80
81
struct SimpleTensor {
  void *dptr;
  std::vector<size_t> shape;
  DType dtype;

82
83
  SimpleTensor(void *dptr, const std::vector<size_t> &shape, DType dtype)
      : dptr(dptr), shape(shape), dtype(dtype) {}
84
85
86
87
88
89

  SimpleTensor(const NVTEBasicTensor &tensor)  // NOLINT
      : dptr(tensor.data_ptr),
        shape(tensor.shape.data, tensor.shape.data + tensor.shape.ndim),
        dtype(static_cast<DType>(tensor.dtype)) {}

90
  SimpleTensor() : SimpleTensor(nullptr, {}, DType::kFloat32) {}
91
92

  operator NVTEBasicTensor() const {
93
94
    return {dptr, static_cast<NVTEDType>(dtype),
            nvte_make_shape(this->shape.data(), this->shape.size())};
95
96
  }

97
  size_t numel() const {
98
99
100
101
102
103
    size_t acc = 1;
    for (const auto &dim : shape) {
      acc *= dim;
    }
    return acc;
  }
104
  bool has_data() const noexcept { return dptr != nullptr && numel() > 0; }
105
106
107
108
109
110

  void clear() {
    dptr = nullptr;
    shape.resize(0);
    dtype = DType::kFloat32;
  }
111
};
Przemek Tredak's avatar
Przemek Tredak committed
112

113
struct Tensor {
114
 public:
115
  SimpleTensor data;
116
  SimpleTensor columnwise_data;
117
  SimpleTensor amax;
118
  SimpleTensor columnwise_amax;
119
120
  SimpleTensor scale;
  SimpleTensor scale_inv;
121
122
123
  SimpleTensor columnwise_scale_inv;

  NVTEScalingMode scaling_mode;
124
  NVTETensor nvte_tensor;
125

126
127
  Tensor()
      : data(),
128
        columnwise_data(),
129
        amax(nullptr, {1}, DType::kFloat32),
130
        columnwise_amax(nullptr, {1}, DType::kFloat32),
131
        scale(nullptr, {1}, DType::kFloat32),
132
133
        scale_inv(nullptr, {1}, DType::kFloat32),
        columnwise_scale_inv(nullptr, {1}, DType::kFloat32),
134
135
136
137
138
139
140
        scaling_mode(NVTE_DELAYED_TENSOR_SCALING),
        nvte_tensor(0) {}

  void clear() {
    data.clear();
    columnwise_data.clear();
    amax.clear();
141
    columnwise_amax.clear();
142
143
144
145
146
147
148
    scale.clear();
    scale_inv.clear();
    columnwise_scale_inv.clear();
    scaling_mode = NVTE_DELAYED_TENSOR_SCALING;
  }

  explicit operator NVTETensor() const noexcept { return nvte_tensor; }
149

150
  size_t numel() const {
151
    size_t acc = 1;
152
    for (const auto dim : shape()) {
153
154
155
156
157
      acc *= dim;
    }
    return acc;
  }

158
  // TODO(Tim): Change this to use data.has_data()
159
160
  bool has_data() const noexcept { return data.dptr != nullptr; }

161
  // Check for size (not just pointer) for 0-dim or no token cases.
162
  // TODO(Tim): Change this to use columnwise_data.has_data()
163
164
165
  bool has_columnwise_data() const noexcept {
    return columnwise_data.dptr != nullptr || columnwise_data.shape.size() != 0;
  }
166
167
168
169
170
171
172
173

  DType dtype() const {
    if (has_data()) return data.dtype;
    if (has_columnwise_data()) return columnwise_data.dtype;
    // Fallback, used e.g. in workspace
    return data.dtype;
  }

174
175
176
177
178
179
180
181
  size_t dim() const {
    if (!has_data() && has_columnwise_data()) {
      return columnwise_data.shape.size();
    } else {
      return data.shape.size();
    }
  }

182
183
184
185
186
187
188
189
  std::vector<size_t> shape() const {
    /* Note: We sometimes experience spurious compiler errors
     * (-Wstringop-overflow) from this function. It appears that GCC
     * has some bugs with std::vector (see
     * https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109569).
     */
    switch (scaling_mode) {
      case NVTE_DELAYED_TENSOR_SCALING:
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
      case NVTE_NVFP4_1D_SCALING: {
        // Choose data buffer based on whether it is initialized
        // Note: Uninitialized buffers currently have shape=[].
        // However, this is logically incorrect. 0-D tensors have 1
        // entry, and uninitialized tensors should have shape=[0].
        bool use_columnwise_shape = false;
        if (data.dptr != nullptr) {
          use_columnwise_shape = false;
        } else if (columnwise_data.dptr != nullptr) {
          use_columnwise_shape = true;
        } else if (data.shape.size() != 0) {
          use_columnwise_shape = false;
        } else if (columnwise_data.shape.size() != 0) {
          use_columnwise_shape = true;
        }

        // Infer shape based on data
        if (use_columnwise_shape) {
          // Column-wise data is transposed
209
210
          std::vector<size_t> ret;
          if (!columnwise_data.shape.empty()) {
211
            ret.reserve(columnwise_data.shape.size());
212
213
214
215
216
217
218
            for (size_t i = 1; i < columnwise_data.shape.size(); i++) {
              ret.push_back(columnwise_data.shape[i]);
            }
            ret.push_back(columnwise_data.shape.front());
          }
          return ret;
        }
219
220
        return data.shape;
      }
221
222
223
224
225
226
227
      case NVTE_MXFP8_1D_SCALING:
        if (!has_data() && has_columnwise_data()) {
          return columnwise_data.shape;
        } else {
          return data.shape;
        }
        break;
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
      case NVTE_BLOCK_SCALING_1D:
      case NVTE_BLOCK_SCALING_2D: {
        if (!has_data() && has_columnwise_data()) {
          std::vector<size_t> shape;
          size_t ndim = columnwise_data.shape.size();
          shape.reserve(ndim);
          for (size_t i = 0; i + 1 < ndim; ++i) {
            shape.push_back(columnwise_data.shape[i + 1]);
          }
          if (ndim > 0) {
            shape.push_back(columnwise_data.shape[0]);
          }
          return shape;
        } else {
          // NOTE: We may have removed the data pointer from
          // data by setting usage. In that case, we return
          // the non-null shape. It is our best guess at the most
          // recent shape.
          return data.shape;
        }
        break;
      }
250
251
252
253
254
255
      default:
        NVTE_ERROR("Cannot parse tensor shape with scaling mode \"", to_string(scaling_mode), "\"");
        return {};
    }
  }

256
257
258
259
260
261
  /*! Matrix height after tensor is flattened to 2D
   *
   * If a tensor has dimensions (D1, D2, ..., Dn), it is reinterpreted
   * as a (D1*D2*...*D(n-1), Dn) matrix.
   */
  size_t flat_first_dim() const {
262
263
264
265
266
    const auto &full_shape = shape();
    size_t ret = 1;
    if (!full_shape.empty()) {
      for (size_t i = 0; i < full_shape.size() - 1; i++) {
        ret *= full_shape[i];
267
268
      }
    }
269
    return ret;
270
271
272
273
274
275
276
277
  }

  /*! Matrix width after tensor is flattened to 2D
   *
   * If a tensor has dimensions (D1, D2, ..., Dn), it is reinterpreted
   * as a (D1*D2*...*D(n-1), Dn) matrix.
   */
  size_t flat_last_dim() const {
278
279
280
281
282
    const auto &full_shape = shape();
    if (full_shape.empty()) {
      return 1;
    } else {
      return full_shape.back();
283
284
    }
  }
Przemek Tredak's avatar
Przemek Tredak committed
285
286
};

287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
struct GroupedTensor {
 public:
  /* EXPERIMENTAL FEATURE AND SUBJECT TO CHANGE. */
  /*
  Grouped tensor is a collection of tensors with different shapes but the same dtype and scaling mode

  Shape Representation:
  - logical_shape: 2D shape representing the conceptual layouy, i.e. the shape when member tensors are flattened to 2D and stacked together (REQUIRED)
    + When all_same_shape(): [num_tensors * M, N] where each tensor is (M, N)
    + When varying_first_dim(): [~sum_of_first_dims, N] where N is common
    + When varying_last_dim(): [M, ~sum_of_last_dims] where M is common
    + When varying_both_dims(): [1, total_elements] (fully flattened)

  - first_dims and last_dims are OPTIONAL (empty if dimension is uniform)
    + Empty first_dims: all tensors have the same first dimension
    + Empty last_dims: all tensors have the same last dimension
    + Both empty: all tensors have identical shapes
    + Both set: each tensor has unique shape (first_dims[i], last_dims[i])

  Data Layout:
  - ALL data fields are stored as 1D flattened arrays (data, columnwise_data, scale_inv, etc.)
  - logical_shape provides the conceptual 2D interpretation
  - All data is stored on device in contiguous layout
  */

  SimpleTensor data;
  SimpleTensor columnwise_data;
  SimpleTensor scale_inv;
  SimpleTensor columnwise_scale_inv;
  SimpleTensor amax;
  SimpleTensor columnwise_amax;
  SimpleTensor scale;  // for FP8-DS only

  // Shape information (OPTIONAL - empty if dimension is uniform across all tensors)
  // first_dims[i] = first dimension of tensor i (empty if all tensors have same first dim)
  // last_dims[i] = last dimension of tensor i (empty if all tensors have same last dim)
  SimpleTensor first_dims;  // Device pointer to int64_t array of length num_tensors (or empty)
  SimpleTensor last_dims;   // Device pointer to int64_t array of length num_tensors (or empty)

  // Offsets for indexing into contiguous 1D layout (OPTIONAL - not needed if all_same_shape())
  // tensor_offsets[i] = element offset to start of tensor i (cumulative sum of numel for tensors 0..i-1)
  // Usage: tensor_i_ptr = (char*)data.dptr + tensor_offsets[i] * element_size
  // If empty and all_same_shape(): offset[i] = i * M * N (where M, N are common dimensions)
  SimpleTensor tensor_offsets;  // Device pointer to int64_t array of length num_tensors (or empty)

  // Logical shape: conceptual 2D shape of the grouped data (REQUIRED)
  // Represents how the 1D flattened data should be interpreted as 2D
  // Always 2D with positive dimensions
  NVTEShape logical_shape;

  NVTEScalingMode scaling_mode;
  size_t num_tensors;
  NVTEGroupedTensor nvte_tensor;

  GroupedTensor(NVTEScalingMode scaling_mode, size_t num_tensors)
      : data(),
        columnwise_data(),
        scale_inv(),
        columnwise_scale_inv(),
        amax(),
        columnwise_amax(),
        scale(),
        num_tensors(num_tensors),
        first_dims(nullptr, {}, DType::kInt64),
        last_dims(nullptr, {}, DType::kInt64),
        tensor_offsets(nullptr, {}, DType::kInt64),
        logical_shape(nvte_make_shape(nullptr, 0)),
        scaling_mode(scaling_mode),
        nvte_tensor(0) {}

  explicit operator NVTEGroupedTensor() const noexcept { return nvte_tensor; }

  bool has_data() const noexcept { return data.has_data(); }
  bool has_columnwise_data() const noexcept { return columnwise_data.has_data(); }

  bool all_same_first_dim() const noexcept { return !first_dims.has_data(); }
  bool all_same_last_dim() const noexcept { return !last_dims.has_data(); }
  bool all_same_shape() const noexcept { return !first_dims.has_data() && !last_dims.has_data(); }
  bool varying_both_dims() const noexcept { return first_dims.has_data() && last_dims.has_data(); }

  size_t get_common_first_dim() const {
    NVTE_CHECK(all_same_first_dim(), "First dim varies across tensors");
    NVTE_CHECK(logical_shape.ndim == 2, "Logical shape must be 2D");
    if (all_same_shape()) {
      // When both dims are uniform: logical_shape = [num_tensors * M, N]
      return logical_shape.data[0] / num_tensors;
    } else {
      // When varying last dims but not first dim: logical_shape = [M, sum_of_last_dims]
      return logical_shape.data[0];
    }
  }
  size_t get_common_last_dim() const {
    NVTE_CHECK(all_same_last_dim(), "Last dim varies across tensors");
    NVTE_CHECK(logical_shape.ndim == 2, "Logical shape must be 2D");
    // For both uniform and varying first dim cases: logical_shape[1] is the common last dim
    return logical_shape.data[1];
  }

  DType dtype() const {
    if (has_data()) return data.dtype;
    if (has_columnwise_data()) return columnwise_data.dtype;
    // Fallback, used e.g. in workspace or when allow_empty=true
    return data.dtype;
  }

  void clear() {
    data.clear();
    columnwise_data.clear();
    scale_inv.clear();
    columnwise_scale_inv.clear();
    amax.clear();
    columnwise_amax.clear();
    scale.clear();
    first_dims.clear();
    last_dims.clear();
    tensor_offsets.clear();
    logical_shape = nvte_make_shape(nullptr, 0);
    num_tensors = 0;
    scaling_mode = NVTE_DELAYED_TENSOR_SCALING;
    nvte_tensor = 0;
  }
};

410
411
412
struct QuantizationConfig {
  bool force_pow_2_scales = false;
  float amax_epsilon = 0.0f;
413
  NVTETensor noop_tensor = nullptr;
414
415
  Float8BlockScaleTensorFormat float8_block_scale_tensor_format =
      Float8BlockScaleTensorFormat::GEMM_READY;
416
417
418
  NVTETensor rng_state = nullptr;
  bool nvfp4_2d_quantization = false;
  bool stochastic_rounding = false;
419
420

  static constexpr size_t attr_sizes[] = {
421
422
423
424
425
426
427
      sizeof(bool),                          // force_pow_2_scales
      sizeof(float),                         // amax_epsilon
      sizeof(NVTETensor),                    // noop_tensor
      sizeof(Float8BlockScaleTensorFormat),  // float8_block_scale_tensor_format
      sizeof(NVTETensor),                    // rng_seed and offset
      sizeof(bool),                          // nvfp4_2d_quantization
      sizeof(bool)                           // stochastic_rounding
428
429
430
  };
};

431
432
cudaDataType_t get_cuda_dtype(const transformer_engine::DType t);

Przemek Tredak's avatar
Przemek Tredak committed
433
434
template <typename T>
constexpr T DIVUP(const T &x, const T &y) {
435
  return (((x) + ((y)-1)) / (y));
Przemek Tredak's avatar
Przemek Tredak committed
436
437
}

438
439
440
441
442
443
444
template <typename T1, typename T2>
constexpr __device__ __host__ __forceinline__ uint64_t DIVUP_TO_MULTIPLE(const T1 &N, const T2 &M) {
  static_assert(std::is_integral<T1>::value && std::is_integral<T2>::value,
                "Integral type required.");
  return DIVUP(static_cast<uint64_t>(N), static_cast<uint64_t>(M)) * M;
}

Przemek Tredak's avatar
Przemek Tredak committed
445
using byte = uint8_t;
446
using int16 = int16_t;
Przemek Tredak's avatar
Przemek Tredak committed
447
using int32 = int32_t;
448
using int64 = int64_t;
Przemek Tredak's avatar
Przemek Tredak committed
449
450
451
452
453
using fp32 = float;
using fp16 = half;
using bf16 = nv_bfloat16;
using fp8e4m3 = __nv_fp8_e4m3;
using fp8e5m2 = __nv_fp8_e5m2;
454
455
456
#if CUDA_VERSION >= 12080
using fp8e8m0 = __nv_fp8_e8m0;
#endif
457
458
#if FP4_TYPE_SUPPORTED
using fp4e2m1 = __nv_fp4_e2m1;
459
460
using fp4e2m1x2 = __nv_fp4x2_e2m1;
using fp4e2m1x4 = __nv_fp4x4_e2m1;
461
#endif
462
using e8m0_t = uint8_t;
Przemek Tredak's avatar
Przemek Tredak committed
463

Tim Moon's avatar
Tim Moon committed
464
465
466
467
namespace detail {

template <typename T>
constexpr inline const char *type_name() noexcept;
468
469
470
471
472
#define TRANSFORMER_ENGINE_TYPE_NAME(T)                  \
  template <>                                            \
  inline constexpr const char *type_name<T>() noexcept { \
    return #T;                                           \
  }
Tim Moon's avatar
Tim Moon committed
473
TRANSFORMER_ENGINE_TYPE_NAME(uint8_t)
474
TRANSFORMER_ENGINE_TYPE_NAME(int16_t)
Tim Moon's avatar
Tim Moon committed
475
TRANSFORMER_ENGINE_TYPE_NAME(int32_t)
476
TRANSFORMER_ENGINE_TYPE_NAME(int64_t)
Tim Moon's avatar
Tim Moon committed
477
478
479
480
481
TRANSFORMER_ENGINE_TYPE_NAME(float)
TRANSFORMER_ENGINE_TYPE_NAME(half)
TRANSFORMER_ENGINE_TYPE_NAME(nv_bfloat16)
TRANSFORMER_ENGINE_TYPE_NAME(__nv_fp8_e4m3)
TRANSFORMER_ENGINE_TYPE_NAME(__nv_fp8_e5m2)
482
483
484
#if CUDA_VERSION >= 12080
TRANSFORMER_ENGINE_TYPE_NAME(__nv_fp8_e8m0)
#endif
485
486
487
#if FP4_TYPE_SUPPORTED
TRANSFORMER_ENGINE_TYPE_NAME(__nv_fp4_e2m1)
#endif
Tim Moon's avatar
Tim Moon committed
488
489
#undef TRANSFORMER_ENGINE_TYPE_NAME

490
491
492
template <typename T>
struct TypeExtrema;

493
494
495
496
#if FP4_TYPE_SUPPORTED
template <>
struct TypeExtrema<fp4e2m1> {
  static constexpr float max = 6.0f;
497
  static constexpr float max_inverse = 1.0 / max;
498
499
500
};
#endif

501
502
503
template <>
struct TypeExtrema<fp8e4m3> {
  static constexpr float max = 448.0f;
504
  static constexpr float max_inverse = 1.0 / max;
505
506
507
508
509
};

template <>
struct TypeExtrema<fp8e5m2> {
  static constexpr float max = 57344.0f;
510
  static constexpr float max_inverse = 1.0 / max;
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
};

template <>
struct TypeExtrema<bf16> {
  // Hex float format of 1.(7 bits of 1) * 2 ^ 127
  static constexpr float max = 0x1.FEp127;
};

template <>
struct TypeExtrema<fp16> {
  // Hex float format of 1.(10 bits of 1) * 2 ^ 15
  static constexpr float max = 0x1.FFCp15;
};

template <typename T>
struct TypeExtrema {
  static constexpr float max = std::numeric_limits<T>::max();
};

Tim Moon's avatar
Tim Moon committed
530
}  // namespace detail
Przemek Tredak's avatar
Przemek Tredak committed
531

532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
template <typename T>
struct BitsNumber;

#if FP4_TYPE_SUPPORTED
template <>
struct BitsNumber<fp4e2m1> {
  static constexpr size_t num_bits = 4;
};
#endif

template <typename T>
struct BitsNumber {
  static constexpr size_t num_bits = 8 * sizeof(T);
};

Przemek Tredak's avatar
Przemek Tredak committed
547
template <typename T>
548
struct TypeInfo {
549
#if FP4_TYPE_SUPPORTED
550
551
552
553
554
555
  using types = std::tuple<byte, int16, int32, int64, fp32, fp16, bf16, fp8e4m3, fp8e5m2, fp4e2m1
#if CUDA_VERSION >= 12080
                           ,
                           fp8e8m0
#endif
                           >;
556
#else
557
558
559
560
561
562
  using types = std::tuple<byte, int16, int32, int64, fp32, fp16, bf16, fp8e4m3, fp8e5m2
#if CUDA_VERSION >= 12080
                           ,
                           fp8e8m0
#endif
                           >;
563
#endif
564
565
566

  template <typename U, DType current>
  struct Helper {
Przemek Tredak's avatar
Przemek Tredak committed
567
    constexpr static DType getType() {
568
569
570
571
572
573
      constexpr int i = static_cast<int>(current);
      if (std::is_same<U, typename std::tuple_element<i, types>::type>::value) {
        return current;
      } else {
        return Helper<U, static_cast<DType>(i + 1)>::getType();
      }
Przemek Tredak's avatar
Przemek Tredak committed
574
    }
575
576
577
578
579
580
  };

  template <typename U>
  struct Helper<U, DType::kNumTypes> {
    constexpr static DType getType() { return DType::kNumTypes; }
  };
Przemek Tredak's avatar
Przemek Tredak committed
581

582
583
584
585
586
587
  template <typename U>
  constexpr static DType getType() {
    return Helper<U, DType::kByte>::getType();
  }

  constexpr static DType dtype = getType<T>();
588
  constexpr static size_t size = BitsNumber<T>::num_bits;
589
  constexpr static float max_finite_value = detail::TypeExtrema<T>::max;
590
  constexpr static const char *name = detail::type_name<T>();
Przemek Tredak's avatar
Przemek Tredak committed
591
592
};

593
594
595
596
597
598
599
600
601
602
#if FP4_TYPE_SUPPORTED
#define SWITCH_FP4_TYPE_HANDLE(type, ...) \
  case DType::kFloat4E2M1: {              \
    using type = fp4e2m1;                 \
    { __VA_ARGS__ }                       \
  } break;
#else
#define SWITCH_FP4_TYPE_HANDLE(type, ...)  // do nothing
#endif

Przemek Tredak's avatar
Przemek Tredak committed
603
#define TRANSFORMER_ENGINE_TYPE_SWITCH_ALL(dtype, type, ...) \
604
605
606
607
608
609
  switch (dtype) {                                           \
    using namespace transformer_engine;                      \
    case DType::kByte: {                                     \
      using type = unsigned char;                            \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
610
611
612
613
    case DType::kInt16: {                                    \
      using type = int16_t;                                  \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
614
    case DType::kInt32: {                                    \
615
616
617
618
619
      using type = int32_t;                                  \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
    case DType::kInt64: {                                    \
      using type = int64_t;                                  \
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
      { __VA_ARGS__ }                                        \
    } break;                                                 \
    case DType::kFloat32: {                                  \
      using type = float;                                    \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
    case DType::kFloat16: {                                  \
      using type = fp16;                                     \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
    case DType::kBFloat16: {                                 \
      using type = bf16;                                     \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
    case DType::kFloat8E4M3: {                               \
      using type = fp8e4m3;                                  \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
    case DType::kFloat8E5M2: {                               \
      using type = fp8e5m2;                                  \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
642
643
644
645
    case DType::kFloat8E8M0: {                               \
      using type = byte;                                     \
      { __VA_ARGS__ }                                        \
    } break;                                                 \
646
      SWITCH_FP4_TYPE_HANDLE(type, __VA_ARGS__)              \
647
648
649
    default:                                                 \
      NVTE_ERROR("Invalid type.");                           \
  }
Przemek Tredak's avatar
Przemek Tredak committed
650

651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
#define TRANSFORMER_ENGINE_TYPE_SWITCH_FLOAT(dtype, type, ...) \
  switch (dtype) {                                             \
    using namespace transformer_engine;                        \
    case DType::kFloat32: {                                    \
      using type = float;                                      \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kFloat16: {                                    \
      using type = fp16;                                       \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kBFloat16: {                                   \
      using type = bf16;                                       \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kFloat8E4M3: {                                 \
      using type = fp8e4m3;                                    \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kFloat8E5M2: {                                 \
      using type = fp8e5m2;                                    \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    default:                                                   \
      NVTE_ERROR("Invalid type.");                             \
  }

Przemek Tredak's avatar
Przemek Tredak committed
678
#define TRANSFORMER_ENGINE_TYPE_SWITCH_OUTPUT(dtype, type, ...) \
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
  switch (dtype) {                                              \
    using namespace transformer_engine;                         \
    case DType::kFloat32: {                                     \
      using type = float;                                       \
      { __VA_ARGS__ }                                           \
    } break;                                                    \
    case DType::kFloat16: {                                     \
      using type = fp16;                                        \
      { __VA_ARGS__ }                                           \
    } break;                                                    \
    case DType::kBFloat16: {                                    \
      using type = bf16;                                        \
      { __VA_ARGS__ }                                           \
    } break;                                                    \
    case DType::kFloat8E5M2: {                                  \
      using type = fp8e5m2;                                     \
      { __VA_ARGS__ }                                           \
    } break;                                                    \
    case DType::kFloat8E4M3: {                                  \
      using type = fp8e4m3;                                     \
      { __VA_ARGS__ }                                           \
    } break;                                                    \
    default:                                                    \
      NVTE_ERROR("Invalid type.");                              \
  }
Przemek Tredak's avatar
Przemek Tredak committed
704

705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
#define TRANSFORMER_ENGINE_TYPE_SWITCH_NON_FP8ONLY(dtype, type, ...) \
  switch (dtype) {                                                   \
    using namespace transformer_engine;                              \
    case DType::kFloat32: {                                          \
      using type = float;                                            \
      { __VA_ARGS__ }                                                \
    } break;                                                         \
    case DType::kFloat16: {                                          \
      using type = fp16;                                             \
      { __VA_ARGS__ }                                                \
    } break;                                                         \
    case DType::kBFloat16: {                                         \
      using type = bf16;                                             \
      { __VA_ARGS__ }                                                \
    } break;                                                         \
    default:                                                         \
      NVTE_ERROR("Invalid type.");                                   \
  }

724
725
726
727
728
729
730
731
732
733
734
735
// Add a pack_size argument to select the packed type for FP4
#define TRANSFORMER_ENGINE_TYPE_SWITCH_FP4x2_ONLY(dtype, pack_size, type, ...) \
  switch (dtype) {                                                             \
    using namespace transformer_engine;                                        \
    case DType::kFloat4E2M1: {                                                 \
      using type = __nv_fp4x2_storage_t;                                       \
      { __VA_ARGS__ }                                                          \
    } break;                                                                   \
    default:                                                                   \
      NVTE_ERROR("Invalid type.");                                             \
  }

Przemek Tredak's avatar
Przemek Tredak committed
736
#define TRANSFORMER_ENGINE_TYPE_SWITCH_FP8ONLY(dtype, type, ...) \
737
738
739
740
741
742
743
744
745
746
747
748
749
  switch (dtype) {                                               \
    using namespace transformer_engine;                          \
    case DType::kFloat8E5M2: {                                   \
      using type = fp8e5m2;                                      \
      { __VA_ARGS__ }                                            \
    } break;                                                     \
    case DType::kFloat8E4M3: {                                   \
      using type = fp8e4m3;                                      \
      { __VA_ARGS__ }                                            \
    } break;                                                     \
    default:                                                     \
      NVTE_ERROR("Invalid type.");                               \
  }
Przemek Tredak's avatar
Przemek Tredak committed
750
751

#define TRANSFORMER_ENGINE_TYPE_SWITCH_INPUT(dtype, type, ...) \
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
  switch (dtype) {                                             \
    using namespace transformer_engine;                        \
    case DType::kFloat32: {                                    \
      using type = float;                                      \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kFloat16: {                                    \
      using type = fp16;                                       \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kBFloat16: {                                   \
      using type = bf16;                                       \
      { __VA_ARGS__ }                                          \
    } break;                                                   \
    case DType::kFloat8E5M2:                                   \
    case DType::kFloat8E4M3: {                                 \
      NVTE_ERROR("FP8 type not instantiated for input.");      \
    } break;                                                   \
770
771
772
    case DType::kFloat4E2M1: {                                 \
      NVTE_ERROR("FP4 type not instantiated for input.");      \
    } break;                                                   \
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
    default:                                                   \
      NVTE_ERROR("Invalid type.");                             \
  }

#define TRANSFORMER_ENGINE_TYPE_SWITCH_16BIT(dtype, type, ...) \
  switch (dtype) {                                             \
    using namespace transformer_engine;                        \
    case DType::kFloat16: {                                    \
      using type = fp16;                                       \
      __VA_ARGS__;                                             \
      break;                                                   \
    }                                                          \
    case DType::kBFloat16: {                                   \
      using type = bf16;                                       \
      __VA_ARGS__;                                             \
      break;                                                   \
    }                                                          \
    default:                                                   \
      NVTE_ERROR("Invalid type for 16 bit.");                  \
  }
793

794
795
796
797
798
799
800
801
802
803
804
805
806
#define TRANSFORMER_ENGINE_MX_SCALE_DIM_SWITCH(SCALE_DIM, DIM, ...) \
  switch (SCALE_DIM) {                                              \
    case 1: {                                                       \
      constexpr size_t DIM = 1;                                     \
      { __VA_ARGS__ }                                               \
    } break;                                                        \
    case 32: {                                                      \
      constexpr size_t DIM = 32;                                    \
      { __VA_ARGS__ }                                               \
    } break;                                                        \
    default: {                                                      \
      NVTE_ERROR("Invalid size of the MX scaling factor.");         \
    }                                                               \
807
  }
808

809
810
811
812
813
814
815
816
817
#define TRANSFORMER_ENGINE_SWITCH_CONDITION(CONDITION, FLAG, ...) \
  if (CONDITION) {                                                \
    constexpr bool FLAG = true;                                   \
    { __VA_ARGS__ }                                               \
  } else {                                                        \
    constexpr bool FLAG = false;                                  \
    { __VA_ARGS__ }                                               \
  }

818
////////////////////////////////////////////////////////////////////////////////////////////////////
Przemek Tredak's avatar
Przemek Tredak committed
819

820
inline int log2_ceil(int value) {
821
822
823
  int log2_value = 0;
  while ((1 << log2_value) < value) ++log2_value;
  return log2_value;
824
825
}

826
827
828
829
830
831
832
833
template <size_t B>
inline size_t alignTo(size_t x) {
  size_t r = x % B;
  if (r == 0) return x;

  return x + B - r;
}

Przemek Tredak's avatar
Przemek Tredak committed
834
835
836
837
838
839
840
841
842
template <typename T>
struct is_fp8 : std::false_type {};

template <>
struct is_fp8<fp8e4m3> : std::true_type {};

template <>
struct is_fp8<fp8e5m2> : std::true_type {};

843
844
845
846
847
848
849
850
template <typename T>
struct is_fp4 : std::false_type {};

#if FP4_TYPE_SUPPORTED
template <>
struct is_fp4<fp4e2m1> : std::true_type {};
#endif

851
852
853
854
855
856
// [128,4] rowwise and [4,128] colwise alignment requirements for the tensor with scaling factors
constexpr size_t scale_tensor_alignment_X_rowwise = 4;
constexpr size_t scale_tensor_alignment_Y_rowwise = 128;
constexpr size_t scale_tensor_alignment_X_colwise = 128;
constexpr size_t scale_tensor_alignment_Y_colwise = 4;

857
// Alignment requirements for the Tensor Memory Accelerator (TMA)
858
859
constexpr size_t TMA_GMEM_ALIGNMENT = 16;    // global memory address alignment
constexpr size_t TMA_SHMEM_ALIGNMENT = 128;  // shared memory address alignment
860
861
862
863
864
865
866
867
868

inline bool is_aligned_ptr(const void *ptr, size_t alignment) {
  return reinterpret_cast<uintptr_t>(ptr) % alignment == 0;
}

inline bool is_aligned_tensor_data(const Tensor &t, size_t alignment) {
  return is_aligned_ptr(static_cast<const void *>(t.data.dptr), alignment);
}

Przemek Tredak's avatar
Przemek Tredak committed
869
size_t typeToSize(const DType type);
870
871
872
873
874
size_t typeToNumBits(const DType type);

size_t get_buffer_size_bytes(const size_t N, const DType buffer_dtype);
size_t get_buffer_size_bytes(const size_t dim_first, const size_t dim_last,
                             const DType buffer_dtype);
Przemek Tredak's avatar
Przemek Tredak committed
875

876
void CheckNoopTensor(const Tensor &t, const std::string &name);
877
878
879
void CheckInputTensor(const Tensor &t, const std::string &name);
void CheckOutputTensor(const Tensor &t, const std::string &name, bool allow_empty = false);

880
881
882
883
884
885
886
/*! \brief Update a tensor's FP8 scale-inverse
 *
 * The FP8 scale-inverse (dequantization scaling factor) is updated
 * with the reciprocal of the FP8 scale (quantization scaling factor).
 */
void update_tensor_scale_inv(Tensor *t, cudaStream_t stream);

887
#define NVTE_API_CALL(api_name) \
888
  transformer_engine::nvtx::NVTXWrapper _##api_name##_nvtx_wrapper(#api_name);
889

890
891
892
893
894
void checkCuDriverContext(CUstream stream);

CUtensorMapDataType get_CUtensorMapDataType(DType dtype);

// Set up parameters to create TMA descriptor.
895
896
897
898
899
void create_2D_tensor_map(
    CUtensorMap &tensorMap, const SimpleTensor &tensor, const uint64_t globalY,
    const uint64_t globalX, const uint32_t shmemY, const uint32_t shmemX,
    const uint32_t stride_elems, const uint32_t offset_elems, const size_t type_num_bits,
    const CUtensorMapSwizzle swizzle = CUtensorMapSwizzle::CU_TENSOR_MAP_SWIZZLE_NONE);
900
901
902

bool is_supported_by_CC_100();

903
904
905
std::vector<std::vector<Tensor *>> convert_tensor_array(NVTETensor **nvte_tensors,
                                                        size_t outer_size, size_t inner_size);

906
907
Tensor *convertNVTETensor(const NVTETensor tensor);
Tensor *convertNVTETensorCheck(const NVTETensor tensor);
908
909
910
911
912
913
914
915
916
917

GroupedTensor *convertNVTEGroupedTensor(const NVTEGroupedTensor tensor);
GroupedTensor *convertNVTEGroupedTensorCheck(const NVTEGroupedTensor tensor);

// Helper functions for GroupedTensor validation
void CheckGroupedTensorShapeArrays(const GroupedTensor &t, const std::string &name);
void CheckInputGroupedTensor(const GroupedTensor &t, const std::string &name);
void CheckOutputGroupedTensor(const GroupedTensor &t, const std::string &name,
                              bool allow_empty = false);

Przemek Tredak's avatar
Przemek Tredak committed
918
919
920
}  // namespace transformer_engine

#endif  // TRANSFORMER_ENGINE_COMMON_COMMON_H_