File size: 20,098 Bytes
a402b9b
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
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
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
/* Copyright 2025 SGLang Team. All Rights Reserved.

Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at

    http://www.apache.org/licenses/LICENSE-2.0

Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
==============================================================================*/

#pragma once

#include <ATen/Tensor.h>
#include <cuda_runtime.h>
#include <torch/all.h>

#ifdef USE_ROCM
#include <hip/hip_runtime.h>
#endif

#ifdef USE_ROCM
// Adapted from flashinfer-rocm [PR#491](https://github.com/flashinfer-ai/flashinfer/pull/491)
#define _DISPATCH_CASE_F16(c_type, ...) \
  case at::ScalarType::Half: {          \
    using c_type = __half;              \
    return __VA_ARGS__();               \
  }

#define _DISPATCH_CASE_BF16(c_type, ...) \
  case at::ScalarType::BFloat16: {       \
    using c_type = __hip_bfloat16;       \
    return __VA_ARGS__();                \
  }
#endif  // USE_ROCM

#ifndef USE_ROCM
// Adapt from FlashInfer
#ifdef FLASHINFER_ENABLE_F16
#define _DISPATCH_CASE_F16(c_type, ...) \
  case at::ScalarType::Half: {          \
    using c_type = nv_half;             \
    return __VA_ARGS__();               \
  }
#else
#define _DISPATCH_CASE_F16(c_type, ...)
#endif  // FLASHINFER_ENABLE_F16

#ifdef FLASHINFER_ENABLE_BF16
#define _DISPATCH_CASE_BF16(c_type, ...) \
  case at::ScalarType::BFloat16: {       \
    using c_type = nv_bfloat16;          \
    return __VA_ARGS__();                \
  }
#else
#define _DISPATCH_CASE_BF16(c_type, ...)
#endif  // FLASHINFER_ENABLE_BF16

#ifdef FLASHINFER_ENABLE_FP8_E4M3
#define _DISPATCH_CASE_FP8_E4M3(c_type, ...) \
  case at::ScalarType::Float8_e4m3fn: {      \
    using c_type = __nv_fp8_e4m3;            \
    return __VA_ARGS__();                    \
  }
#else
#define _DISPATCH_CASE_FP8_E4M3(c_type, ...)
#endif  // FLASHINFER_ENABLE_FP8_E4M3

#ifdef FLASHINFER_ENABLE_FP8_E5M2
#define _DISPATCH_CASE_FP8_E5M2(c_type, ...) \
  case at::ScalarType::Float8_e5m2: {        \
    using c_type = __nv_fp8_e5m2;            \
    return __VA_ARGS__();                    \
  }
#else
#define _DISPATCH_CASE_FP8_E5M2(c_type, ...)
#endif  // FLASHINFER_ENABLE_FP8_E5M2

#define DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP16(pytorch_dtype, c_type, ...)                 \
  [&]() -> bool {                                                                        \
    switch (pytorch_dtype) {                                                             \
      _DISPATCH_CASE_F16(c_type, __VA_ARGS__)                                            \
      _DISPATCH_CASE_BF16(c_type, __VA_ARGS__)                                           \
      default:                                                                           \
        std::ostringstream oss;                                                          \
        oss << __PRETTY_FUNCTION__ << " failed to dispatch data type " << pytorch_dtype; \
        TORCH_CHECK(false, oss.str());                                                   \
        return false;                                                                    \
    }                                                                                    \
  }()

#define DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FP8(pytorch_dtype, c_type, ...)                      \
  [&]() -> bool {                                                                            \
    switch (pytorch_dtype) {                                                                 \
      _DISPATCH_CASE_FP8_E4M3(c_type, __VA_ARGS__)                                           \
      _DISPATCH_CASE_FP8_E5M2(c_type, __VA_ARGS__)                                           \
      default:                                                                               \
        std::ostringstream oss;                                                              \
        oss << __PRETTY_FUNCTION__ << " failed to dispatch fp8 data type " << pytorch_dtype; \
        TORCH_CHECK(false, oss.str());                                                       \
        return false;                                                                        \
    }                                                                                        \
  }()

#define DISPATCH_PYTORCH_DTYPE_TO_CTYPE(pytorch_dtype, c_type, ...)                      \
  [&]() -> bool {                                                                        \
    switch (pytorch_dtype) {                                                             \
      _DISPATCH_CASE_F16(c_type, __VA_ARGS__)                                            \
      _DISPATCH_CASE_BF16(c_type, __VA_ARGS__)                                           \
      _DISPATCH_CASE_FP8_E4M3(c_type, __VA_ARGS__)                                       \
      _DISPATCH_CASE_FP8_E5M2(c_type, __VA_ARGS__)                                       \
      default:                                                                           \
        std::ostringstream oss;                                                          \
        oss << __PRETTY_FUNCTION__ << " failed to dispatch data type " << pytorch_dtype; \
        TORCH_CHECK(false, oss.str());                                                   \
        return false;                                                                    \
    }                                                                                    \
  }()

#define _DISPATCH_SWITCH(var_name, cond, ...)                                           \
  [&]() -> bool {                                                                       \
    switch (cond) {                                                                     \
      __VA_ARGS__                                                                       \
      default:                                                                          \
        std::ostringstream oss;                                                         \
        oss << __PRETTY_FUNCTION__ << " failed to dispatch " var_name " " << int(cond); \
        TORCH_CHECK(false, oss.str());                                                  \
        return false;                                                                   \
    }                                                                                   \
  }()

#define _DISPATCH_SWITCH_U16x2(var1_name, var2_name, cond1, cond2, ...)                                             \
  [&]() -> bool {                                                                                                   \
    switch (pack_u16(cond1, cond2)) {                                                                               \
      __VA_ARGS__                                                                                                   \
      default:                                                                                                      \
        std::ostringstream oss;                                                                                     \
        oss << __PRETTY_FUNCTION__ << " failed to dispatch (" var1_name ", " var2_name "): (" << int(cond1) << ", " \
            << int(cond2) << ")";                                                                                   \
        TORCH_CHECK(false, oss.str());                                                                              \
        return false;                                                                                               \
    }                                                                                                               \
  }()

#define _DISPATCH_CASE(case_expr, case_var, ...) \
  case case_expr: {                              \
    constexpr auto case_var = case_expr;         \
    return __VA_ARGS__();                        \
  }

#define _DISPATCH_CASE_U16x2(case_expr1, case_expr2, case_var1, case_var2, ...) \
  case pack_u16(case_expr1, case_expr2): {                                      \
    constexpr auto case_var1 = case_expr1;                                      \
    constexpr auto case_var2 = case_expr2;                                      \
    return __VA_ARGS__();                                                       \
  }

#define DISPATCH_BOOL(expr, const_expr, ...) \
  [&]() -> bool {                            \
    if (expr) {                              \
      constexpr bool const_expr = true;      \
      return __VA_ARGS__();                  \
    } else {                                 \
      constexpr bool const_expr = false;     \
      return __VA_ARGS__();                  \
    }                                        \
  }()

inline void check_shape(const at::Tensor& a, const at::Tensor& b, const char* a_name, const char* b_name) {
  TORCH_CHECK(a.dim() == b.dim(), a_name, ".dim() != ", b_name, ".dim(). ", a.dim(), " vs ", b.dim());
  for (int i = 0; i < a.dim(); ++i) {
    TORCH_CHECK(a.size(i) == b.size(i), a_name, ".size(", i, ") != ", b_name, ".size(", i, ")");
  }
}

inline constexpr uint32_t pack_u16(uint16_t a, uint16_t b) {
  return (uint32_t(a) << 16) | uint32_t(b);
}

#define CHECK_GQA_HEAD_DIVISIBLE(num_qo_heads, num_kv_heads) \
  TORCH_CHECK(                                               \
      num_qo_heads % num_kv_heads == 0,                      \
      "num_qo_heads(",                                       \
      num_qo_heads,                                          \
      ") must be divisible by num_kv_heads(",                \
      num_kv_heads,                                          \
      ")")

#define CHECK_CUDA(x) TORCH_CHECK(x.is_cuda(), #x " must be a CUDA tensor")

#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_LAST_DIM_CONTIGUOUS(x) \
  TORCH_CHECK(x.strides()[x.strides().size() - 1] == 1, #x "must be contiguous at last dimension")

#define CHECK_INPUT(x) \
  CHECK_CUDA(x);       \
  CHECK_CONTIGUOUS(x)
#define CHECK_LAST_DIM_CONTIGUOUS_INPUT(x) \
  CHECK_CUDA(x);                           \
  CHECK_LAST_DIM_CONTIGUOUS(x)

#define CHECK_DIM(d, x) TORCH_CHECK(x.dim() == d, #x " must be a " #d "D tensor")

#define CHECK_SHAPE(a, b) check_shape(a, b, #a, #b)

#define CHECK_EQ(a, b) TORCH_CHECK((a) == (b), "CHECK_EQ(" #a ", " #b ") failed. ", a, " vs ", b)

#define CHECK_GE(a, b) TORCH_CHECK((a) >= (b), "CHECK_GE(" #a ", " #b ") failed. ", a, " vs ", b)

inline bool is_float8_tensor(const at::Tensor& tensor) {
  return tensor.scalar_type() == at::ScalarType::Float8_e4m3fn || tensor.scalar_type() == at::ScalarType::Float8_e5m2;
}
#endif  // USE_ROCM

struct cuda_error : public std::runtime_error {
  /**
   * @brief Constructs a `cuda_error` object with the given `message`.
   *
   * @param message The error char array used to construct `cuda_error`
   */
  cuda_error(const char* message) : std::runtime_error(message) {}
  /**
   * @brief Constructs a `cuda_error` object with the given `message` string.
   *
   * @param message The `std::string` used to construct `cuda_error`
   */
  cuda_error(std::string const& message) : cuda_error{message.c_str()} {}
};

#define CHECK_CUDA_SUCCESS(cmd)                                         \
  do {                                                                  \
    cudaError_t e = cmd;                                                \
    if (e != cudaSuccess) {                                             \
      std::stringstream _message;                                       \
      auto s = cudaGetErrorString(e);                                   \
      _message << std::string(s) + "\n" << __FILE__ << ':' << __LINE__; \
      throw cuda_error(_message.str());                                 \
    }                                                                   \
  } while (0)

#define CHECK_IS_CUDA(x) TORCH_CHECK(x.device().is_cuda(), #x " must be a CUDA tensor")
#define CHECK_IS_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_CUDA_INPUT(x) \
  CHECK_IS_CUDA(x);         \
  CHECK_IS_CONTIGUOUS(x)

inline int getSMVersion() {
  int device{-1};
  CHECK_CUDA_SUCCESS(cudaGetDevice(&device));
  int sm_major = 0;
  int sm_minor = 0;
  CHECK_CUDA_SUCCESS(cudaDeviceGetAttribute(&sm_major, cudaDevAttrComputeCapabilityMajor, device));
  CHECK_CUDA_SUCCESS(cudaDeviceGetAttribute(&sm_minor, cudaDevAttrComputeCapabilityMinor, device));
  return sm_major * 10 + sm_minor;
}

inline bool isDeviceType(const std::string& device_type) {
  int deviceCount;
  CHECK_CUDA_SUCCESS(cudaGetDeviceCount(&deviceCount));

  int device_id = -1;
  if (deviceCount >= 1) {
    CHECK_CUDA_SUCCESS(cudaGetDevice(&device_id));
  } else {
    return false;
  }

  cudaDeviceProp prop;
  CHECK_CUDA_SUCCESS(cudaGetDeviceProperties(&prop, device_id));
  if (device_type == std::string(prop.name)) {
    return true;
  }
  return false;
}

inline bool getBoolEnv(char const* name) {
  char const* env = std::getenv(name);
  return env && env[0] == '1' && env[1] == '\0';
}

inline bool getEnvEnablePDL() {
  static std::once_flag flag;
  static bool enablePDL = false;
  std::call_once(flag, [&]() {
    if (getSMVersion() >= 90) {
      // PDL will be enabled by setting the env variables `TRTLLM_ENABLE_PDL` to `1`
      enablePDL = getBoolEnv("TRTLLM_ENABLE_PDL");
    }
  });
  return enablePDL;
}

// SGLANG_SHFL_XOR_* adapted from https://github.com/vllm-project/vllm/blob/v0.7.3/csrc/cuda_compat.h#L19-L28
#ifndef USE_ROCM
#define SGLANG_SHFL_XOR_SYNC(mask, var, lane_mask) __shfl_xor_sync((mask), (var), (lane_mask))
#define SGLANG_SHFL_XOR_SYNC_WIDTH(mask, var, lane_mask, width) __shfl_xor_sync((mask), (var), (lane_mask), (width))
#else
#define SGLANG_SHFL_XOR_SYNC(mask, var, lane_mask) __shfl_xor((var), (lane_mask))
#define SGLANG_SHFL_XOR_SYNC_WIDTH(mask, var, lane_mask, width) __shfl_xor((var), (lane_mask), (width))
#endif

#define DISPATCH_PYTORCH_DTYPE_TO_CTYPE_FLOAT_FP16(pytorch_dtype, c_type, ...)           \
  [&]() -> bool {                                                                        \
    switch (pytorch_dtype) {                                                             \
      case at::ScalarType::Float: {                                                      \
        using c_type = float;                                                            \
        return __VA_ARGS__();                                                            \
      }                                                                                  \
        _DISPATCH_CASE_F16(c_type, __VA_ARGS__)                                          \
        _DISPATCH_CASE_BF16(c_type, __VA_ARGS__)                                         \
      default:                                                                           \
        std::ostringstream oss;                                                          \
        oss << __PRETTY_FUNCTION__ << " failed to dispatch data type " << pytorch_dtype; \
        TORCH_CHECK(false, oss.str());                                                   \
        return false;                                                                    \
    }                                                                                    \
  }()

#define DISPATCH_CASE_INTEGRAL_TYPES(...)              \
  AT_DISPATCH_CASE(at::ScalarType::Byte, __VA_ARGS__)  \
  AT_DISPATCH_CASE(at::ScalarType::Char, __VA_ARGS__)  \
  AT_DISPATCH_CASE(at::ScalarType::Short, __VA_ARGS__) \
  AT_DISPATCH_CASE(at::ScalarType::Int, __VA_ARGS__)   \
  AT_DISPATCH_CASE(at::ScalarType::Long, __VA_ARGS__)

#define DISPATCH_INTEGRAL_TYPES(TYPE, NAME, ...) \
  AT_DISPATCH_SWITCH(TYPE, NAME, DISPATCH_CASE_INTEGRAL_TYPES(__VA_ARGS__))

#define DISPATCH_CASE_FLOAT_TYPES(...)                 \
  AT_DISPATCH_CASE(at::ScalarType::Float, __VA_ARGS__) \
  AT_DISPATCH_CASE(at::ScalarType::Half, __VA_ARGS__)  \
  AT_DISPATCH_CASE(at::ScalarType::BFloat16, __VA_ARGS__)

#define DISPATCH_FLOAT_TYPES(TYPE, NAME, ...) AT_DISPATCH_SWITCH(TYPE, NAME, DISPATCH_CASE_FLOAT_TYPES(__VA_ARGS__))

#define CEILDIV(x, y) (((x) + (y) - 1) / (y))

#ifndef USE_ROCM
#define WARP_SIZE 32
#else
#if defined(__GFX9__) || !defined(__HIP_DEVICE_COMPILE__)
#define WARP_SIZE 64
#else
#define WARP_SIZE 32
#endif
#endif

#ifdef USE_ROCM

#include "hip/hip_math_def.h"
#include "hip/hip_vec_dtypes.h"

#else

template <typename srcDtype>
__device__ __forceinline__ float castToFloat(srcDtype val) {
  return static_cast<srcDtype>(val);
}

template <typename dstDtype>
__device__ __forceinline__ dstDtype castFromFloat(float val) {
  return static_cast<dstDtype>(val);
}

#endif

// add FP8 support
#ifndef USE_ROCM
#include <c10/util/Float8_e4m3fn.h>
using FP8_TYPE = c10::Float8_e4m3fn;
C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
#else  // USE_ROCM
#if HIP_FP8_TYPE_FNUZ
#include <c10/util/Float8_e4m3fnuz.h>
using FP8_TYPE = c10::Float8_e4m3fnuz;
constexpr auto FP8_E4M3_MAX = 224.0f;
#else
#if HIP_FP8_TYPE_E4M3
#include <c10/util/Float8_e4m3fn.h>
using FP8_TYPE = c10::Float8_e4m3fn;
C10_HOST_DEVICE constexpr auto FP8_E4M3_MAX = std::numeric_limits<FP8_TYPE>::max();
#else
#error "fp8 is not supported in this processor (arch < gfx942)."
#endif  // HIP_FP8_TYPE_E4M3
#endif  // HIP_FP8_TYPE_FNUZ
#endif  // USE_ROCM

#define FULL_MASK 0xffffffff

__device__ __forceinline__ float atomicMaxFloat(float* addr, float value) {
#ifndef USE_ROCM
  float old;
  old = (value >= 0) ? __int_as_float(atomicMax((int*)addr, __float_as_int(value)))
                     : __uint_as_float(atomicMin((unsigned int*)addr, __float_as_uint(value)));
  return old;
#else
  int* addr_as_i = (int*)addr;
  int old = *addr_as_i, assumed;
  do {
    assumed = old;
    old = atomicCAS(addr_as_i, assumed, __float_as_int(fmaxf(value, __int_as_float(assumed))));
  } while (assumed != old);
  return __int_as_float(old);
#endif
}

__device__ __forceinline__ float warpReduceMax(float value) {
  value = fmaxf(value, __shfl_xor_sync(FULL_MASK, value, 16));
  value = fmaxf(value, __shfl_xor_sync(FULL_MASK, value, 8));
  value = fmaxf(value, __shfl_xor_sync(FULL_MASK, value, 4));
  value = fmaxf(value, __shfl_xor_sync(FULL_MASK, value, 2));
  value = fmaxf(value, __shfl_xor_sync(FULL_MASK, value, 1));
  return value;
}

__device__ __forceinline__ float blockReduceMax(float value) {
  static __shared__ float warpLevelMaxs[WARP_SIZE];
  const int laneId = threadIdx.x % WARP_SIZE;
  const int warpId = threadIdx.x / WARP_SIZE;

  value = warpReduceMax(value);

  if (laneId == 0) warpLevelMaxs[warpId] = value;
  __syncthreads();

  value = (threadIdx.x < blockDim.x / WARP_SIZE) ? warpLevelMaxs[laneId] : 0;
  if (warpId == 0) value = warpReduceMax(value);

  return value;
}

// Pads to a multiple of `alignment` rows.
inline torch::Tensor pad_tensor(const torch::Tensor& tensor, int64_t alignment = 4, bool is_column_major = false) {
  int64_t rows = tensor.size(0);
  int64_t cols = tensor.size(1);
  int64_t pad_rows = (alignment - (rows % alignment)) % alignment;  // Compute padding size

  if (pad_rows == 0) {
    return tensor;  // Already aligned
  }

  torch::Tensor padding = torch::zeros({pad_rows, cols}, tensor.options());
  torch::Tensor tensor_padded = torch::cat({tensor, padding}, 0);  // Pad along rows

  // Ensure column-major layout
  if (is_column_major) {
    return tensor_padded.t().contiguous().t();
  }
  return tensor_padded;
}

// Get the next power of 2 of a number
inline uint32_t next_pow2(uint32_t x) noexcept {
  if (x <= 1) return 1;
  return 1u << (32 - __builtin_clz(x - 1));
}

/*
 * LDG Support
 */
#ifndef USE_ROCM
#define SGLANG_LDG(arg) __ldg(arg)
#else
#define SGLANG_LDG(arg) *(arg)
#endif