torch_bindings.cpp 28.7 KB
Newer Older
1
2
3
#include "cache.h"
#include "cuda_utils.h"
#include "ops.h"
4
#include "core/registration.h"
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20

#include <torch/library.h>

// Note on op signatures:
// The X_meta signatures are for the meta functions corresponding to op X.
// They must be kept in sync with the signature for X. Generally, only
// functions that return Tensors require a meta function.
//
// See the following links for detailed docs on op registration and function
// schemas.
// https://docs.google.com/document/d/1_W62p8WJOQQUzPsJYa7s701JXt0qf2OfLub2sbkHOaU/edit#heading=h.ptttacy8y1u9
// https://github.com/pytorch/pytorch/blob/main/aten/src/ATen/native/README.md#annotations

TORCH_LIBRARY_EXPAND(TORCH_EXTENSION_NAME, ops) {
  // vLLM custom ops

21
22
23
  ops.def("weak_ref_tensor(Tensor input) -> Tensor");
  ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);

24
25
26
27
28
29
30
31
32
  // Attention ops
  // Compute the attention between an input query and the cached
  // keys/values using PagedAttention.
  ops.def(
      "paged_attention_v1("
      "    Tensor! out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
33
34
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
35
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
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
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v1", torch::kCUDA, &paged_attention_v1);

  // PagedAttention V2.
  ops.def(
      "paged_attention_v2("
      "    Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
      "    Tensor! tmp_out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);

  // Compute the attention between an input query and the cached
  // keys/values using PagedAttention. (opt)
  ops.def(
      "paged_attention_v1_opt("
      "    Tensor! out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v1_opt", torch::kCUDA, &paged_attention_v1_opt);

  // PagedAttention V2 (opt). 
  ops.def(
      "paged_attention_v2_opt("
      "    Tensor! out, Tensor exp_sums, Tensor max_logits,"
      "    Tensor tmp_out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v2_opt", torch::kCUDA, &paged_attention_v2_opt);

  // Compute the attention between an input query and the cached
  // keys/values using PagedAttention. (opt)
  ops.def(
      "paged_attention_v1_opt_tc("
      "    Tensor! out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v1_opt_tc", torch::kCUDA, &paged_attention_v1_opt_tc);

  // PagedAttention V2 (opt). 
  ops.def(
      "paged_attention_v2_opt_tc("
      "    Tensor! out, Tensor exp_sums, Tensor max_logits,"
      "    Tensor tmp_out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v2_opt_tc", torch::kCUDA, &paged_attention_v2_opt_tc);


  // paged_attention with atth_masks
  ops.def(
      "paged_attention_v1_with_mask("
      "    Tensor! out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
120
121
122
      "    int blocksparse_head_sliding_step,"
      "    Tensor? attn_masks,"
      "    int attn_masks_stride) -> ()");
123
  ops.impl("paged_attention_v1_with_mask", torch::kCUDA, &paged_attention_v1_with_mask);
124
125
126

  // PagedAttention V2.
  ops.def(
127
      "paged_attention_v2_with_mask("
128
129
      "    Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
      "    Tensor! tmp_out, Tensor query, Tensor key_cache,"
130
131
132
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
133
134
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
135
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
136
137
138
      "    int blocksparse_head_sliding_step,"
      "    Tensor? attn_masks,"
      "    int attn_masks_stride) -> ()");
139
  ops.impl("paged_attention_v2_with_mask", torch::kCUDA, &paged_attention_v2_with_mask);
140

zhuwenwen's avatar
zhuwenwen committed
141
142
143
  // Compute the attention between an input query and the cached
  // keys/values using PagedAttention. (opt)
  ops.def(
144
      "paged_attention_v1_opt_with_mask("
zhuwenwen's avatar
zhuwenwen committed
145
146
147
148
149
150
151
      "    Tensor! out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
152
153
154
      "    int blocksparse_head_sliding_step,"
      "    Tensor? attn_masks,"
      "    int attn_masks_stride) -> ()");
155
  ops.impl("paged_attention_v1_opt_with_mask", torch::kCUDA, &paged_attention_v1_opt_with_mask);
zhuwenwen's avatar
zhuwenwen committed
156
157
158

  // PagedAttention V2 (opt). 
  ops.def(
159
      "paged_attention_v2_opt_with_mask("
zhuwenwen's avatar
zhuwenwen committed
160
161
162
163
164
165
166
167
      "    Tensor! out, Tensor exp_sums, Tensor max_logits,"
      "    Tensor tmp_out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
168
169
170
      "    int blocksparse_head_sliding_step,"
      "    Tensor? attn_masks,"
      "    int attn_masks_stride) -> ()");
171
  ops.impl("paged_attention_v2_opt_with_mask", torch::kCUDA, &paged_attention_v2_opt_with_mask);
zhuwenwen's avatar
zhuwenwen committed
172

173
  // Compute the attention between an input query and the cached
zhuwenwen's avatar
zhuwenwen committed
174
  // keys/values using PagedAttention. (opt)
175
  ops.def(
176
      "paged_attention_v1_opt_tc_with_mask("
177
178
179
180
      "    Tensor! out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
zhuwenwen's avatar
zhuwenwen committed
181
182
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
183
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
184
185
186
      "    int blocksparse_head_sliding_step,"
      "    Tensor? attn_masks,"
      "    int attn_masks_stride) -> ()");
187
  ops.impl("paged_attention_v1_opt_tc_with_mask", torch::kCUDA, &paged_attention_v1_opt_tc_with_mask);
188

zhuwenwen's avatar
zhuwenwen committed
189
  // PagedAttention V2 (opt). 
190
  ops.def(
191
      "paged_attention_v2_opt_tc_with_mask("
192
193
194
195
196
      "    Tensor! out, Tensor exp_sums, Tensor max_logits,"
      "    Tensor tmp_out, Tensor query, Tensor key_cache,"
      "    Tensor value_cache, int num_kv_heads, float scale,"
      "    Tensor block_tables, Tensor seq_lens, int block_size,"
      "    int max_seq_len, Tensor? alibi_slopes,"
zhuwenwen's avatar
zhuwenwen committed
197
198
      "    str kv_cache_dtype, float k_scale, float v_scale,"
      "    int tp_rank, int blocksparse_local_blocks,"
199
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
200
201
202
      "    int blocksparse_head_sliding_step,"
      "    Tensor? attn_masks,"
      "    int attn_masks_stride) -> ()");
203
204
  ops.impl("paged_attention_v2_opt_tc_with_mask", torch::kCUDA, &paged_attention_v2_opt_tc_with_mask);

205

206
207
208
209
210
211
212
213
214
215
216
217
218
  // Activation ops
  // Activation function used in SwiGLU.
  ops.def("silu_and_mul(Tensor! out, Tensor input) -> ()");
  ops.impl("silu_and_mul", torch::kCUDA, &silu_and_mul);

  // Activation function used in GeGLU with `none` approximation.
  ops.def("gelu_and_mul(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_and_mul", torch::kCUDA, &gelu_and_mul);

  // Activation function used in GeGLU with `tanh` approximation.
  ops.def("gelu_tanh_and_mul(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_tanh_and_mul", torch::kCUDA, &gelu_tanh_and_mul);

zhuwenwen's avatar
zhuwenwen committed
219
220
221
222
223
224
225
226
227
228
229
230
  // Activation function used in SwiGLU. (opt)
  ops.def("silu_and_mul_opt(Tensor! out, Tensor input) -> ()");
  ops.impl("silu_and_mul_opt", torch::kCUDA, &silu_and_mul);

  // Activation function used in GeGLU with `none` approximation. (opt)
  ops.def("gelu_and_mul_opt(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_and_mul_opt", torch::kCUDA, &gelu_and_mul);

  // Activation function used in GeGLU with `tanh` approximation. (opt)
  ops.def("gelu_tanh_and_mul_opt(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_tanh_and_mul_opt", torch::kCUDA, &gelu_tanh_and_mul);

231
232
233
234
  // FATReLU implementation.
  ops.def("fatrelu_and_mul(Tensor! out, Tensor input, float threshold) -> ()");
  ops.impl("fatrelu_and_mul", torch::kCUDA, &fatrelu_and_mul);

235
236
237
238
239
240
241
242
  // GELU implementation used in GPT-2.
  ops.def("gelu_new(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_new", torch::kCUDA, &gelu_new);

  // Approximate GELU implementation.
  ops.def("gelu_fast(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_fast", torch::kCUDA, &gelu_fast);

243
244
245
246
  // Quick GELU implementation.
  ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_quick", torch::kCUDA, &gelu_quick);

247
  // prepare_inputs advance_step
248
  ops.def(
249
      "advance_step_flashattn(int num_seqs, int num_queries, int block_size, "
250
251
252
      "Tensor! input_tokens, Tensor sampled_token_ids, "
      "Tensor! input_positions, Tensor! seq_lens, Tensor! slot_mapping, "
      "Tensor block_tables) -> ()");
253
254
255
256
257
258
259
260
261
262
263
264
  ops.impl("advance_step_flashattn", torch::kCUDA, &advance_step_flashattn);

  ops.def(
      "advance_step_flashinfer("
      "    int num_seqs, int num_queries, int block_size,"
      "    Tensor! input_tokens, Tensor sampled_token_ids,"
      "    Tensor! input_positions, Tensor! seq_lens, Tensor! slot_mapping,"
      "    Tensor block_tables, Tensor! paged_kv_indices,"
      "    Tensor! paged_kv_indptr, Tensor! paged_kv_last_page_len,"
      "    Tensor! block_table_bounds"
      ") -> ()");
  ops.impl("advance_step_flashinfer", torch::kCUDA, &advance_step_flashinfer);
265

266
267
268
  // Layernorm
  // Apply Root Mean Square (RMS) Normalization to the input tensor.
  ops.def(
269
      "rms_norm(Tensor! result, Tensor input, Tensor weight, float epsilon) -> "
270
271
272
273
274
275
276
277
278
      "()");
  ops.impl("rms_norm", torch::kCUDA, &rms_norm);

  // In-place fused Add and RMS Normalization.
  ops.def(
      "fused_add_rms_norm(Tensor! input, Tensor! residual, Tensor weight, "
      "float epsilon) -> ()");
  ops.impl("fused_add_rms_norm", torch::kCUDA, &fused_add_rms_norm);

zhuwenwen's avatar
zhuwenwen committed
279
280
281
282
283
284
285
286
287
288
289
290
  // Apply Root Mean Square (RMS) Normalization to the input tensor. (opt)
  ops.def(
      "rms_norm_opt(Tensor! out, Tensor input, Tensor weight, float epsilon) -> "
      "()");
  ops.impl("rms_norm_opt", torch::kCUDA, &rms_norm_opt);

  // In-place fused Add and RMS Normalization. (opt)
  ops.def(
      "fused_add_rms_norm_opt(Tensor! input, Tensor! residual, Tensor weight, "
      "float epsilon) -> ()");
  ops.impl("fused_add_rms_norm_opt", torch::kCUDA, &fused_add_rms_norm_opt);

zhuwenwen's avatar
zhuwenwen committed
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
  // Layernorm-quant
  // Apply Root Mean Square (RMS) Normalization to the input tensor.
//   ops.def(
//       "rms_norm_static_fp8_quant(Tensor! result, Tensor input, Tensor weight, "
//       "Tensor scale, float epsilon) -> "
//       "()");
//   ops.impl("rms_norm_static_fp8_quant", torch::kCUDA,
//            &rms_norm_static_fp8_quant);

  // In-place fused Add and RMS Normalization.
//   ops.def(
//       "fused_add_rms_norm_static_fp8_quant(Tensor! result, Tensor input, "
//       "Tensor! residual, Tensor weight, "
//       "Tensor scale, float epsilon) -> ()");
//   ops.impl("fused_add_rms_norm_static_fp8_quant", torch::kCUDA,
//            &fused_add_rms_norm_static_fp8_quant);
307

308
309
310
311
312
313
314
315
  // Fused Layernorm + Quant kernels
  ops.def(
      "rms_norm_dynamic_per_token_quant(Tensor! result, Tensor input, "
      "Tensor weight, Tensor! scale, float epsilon, "
      "Tensor? scale_ub, Tensor!? residual) -> ()");
  ops.impl("rms_norm_dynamic_per_token_quant", torch::kCUDA,
           &rms_norm_dynamic_per_token_quant);

316
317
318
319
320
321
322
323
  // Rotary embedding
  // Apply GPT-NeoX or GPT-J style rotary embedding to query and key.
  ops.def(
      "rotary_embedding(Tensor positions, Tensor! query,"
      "                 Tensor! key, int head_size,"
      "                 Tensor cos_sin_cache, bool is_neox) -> ()");
  ops.impl("rotary_embedding", torch::kCUDA, &rotary_embedding);

huangwb's avatar
huangwb committed
324
325
326
327
328
329
330
331
332
  // Rotary embedding TGI for TGI
  // Apply GPT-NeoX or GPT-J style rotary embedding to query and key.
  ops.def(
      "rotary_embedding_tgi(Tensor! query, Tensor! key,"
      "                 int head_size, Tensor cos_cache,"
      "                 Tensor sin_cache, bool is_neox) -> ()");
//   ops.def("rotary_embedding_tgi",&rotary_embedding_tgi);
  ops.impl("rotary_embedding_tgi", torch::kCUDA, &rotary_embedding_tgi);

333
334
335
336
337
338
339
340
341
342
  // Apply GPT-NeoX or GPT-J style rotary embedding to query and key
  // (supports multiple loras).
  ops.def(
      "batched_rotary_embedding(Tensor positions, Tensor! query,"
      "                         Tensor! key, int head_size,"
      "                         Tensor cos_sin_cache, bool is_neox,"
      "                         int rot_dim,"
      "                         Tensor cos_sin_cache_offsets) -> ()");
  ops.impl("batched_rotary_embedding", torch::kCUDA, &batched_rotary_embedding);

zhuwenwen's avatar
zhuwenwen committed
343
344
345
346
  // trans w16
  ops.def("trans_w16_gemm(Tensor! dst, Tensor src, int row, int col) -> ()");
  ops.impl("trans_w16_gemm", torch::kCUDA, &trans_w16_gemm);

347
348
349
  // Quantization ops
#ifndef USE_ROCM
  // Quantized GEMM for AQLM.
350
351
352
353
  ops.def(
      "aqlm_gemm(Tensor input, Tensor codes, Tensor codebooks, "
      "Tensor scales, int[] codebook_partition_sizes, Tensor? bias) "
      "-> Tensor");
354
355
356
  ops.impl("aqlm_gemm", torch::kCUDA, &aqlm_gemm);

  // Decompression method for AQLM.
357
358
359
  ops.def(
      "aqlm_dequant(Tensor codes, Tensor codebooks, "
      "int[] codebook_partition_sizes) -> Tensor");
360
361
362
  ops.impl("aqlm_dequant", torch::kCUDA, &aqlm_dequant);

  // Quantized GEMM for AWQ.
363
364
  ops.def(
      "awq_gemm(Tensor _in_feats, Tensor _kernel, Tensor _scaling_factors, "
365
      "Tensor _zeros, SymInt split_k_iters) -> Tensor");
366
367
368
  ops.impl("awq_gemm", torch::kCUDA, &awq_gemm);

  // Dequantization for AWQ.
369
370
  ops.def(
      "awq_dequantize(Tensor _kernel, Tensor _scaling_factors, "
371
      "Tensor _zeros, SymInt split_k_iters, int thx, int thy) -> Tensor");
372
373
  ops.impl("awq_dequantize", torch::kCUDA, &awq_dequantize);

374
375
376
377
378
379
380
381
382
383
384
385
386
387
  // Note about marlin kernel 'workspace' arguments:
  // Technically these should be mutable since they are modified by the kernel.
  // But since they are set back to zero once the kernel is finished we can
  // hand wave and say that they have no net effect.
  //
  // The reason to mark 'workspace' as immutable is so that they don't interfere
  // with using ScalarType arguments in the ops. If they are marked as mutable,
  // pytorch throws an assert in
  // 'torch._higher_order_ops._register_effectful_op' that prevents these
  // kernels from being torch.compile'd.
  // See the following document for more info on custom types and ops that use
  // custom types:
  // https://docs.google.com/document/d/18fBMPuOJ0fY5ZQ6YyrHUppw9FA332CpNtgB6SOIgyuA

388
  // Marlin (Dense) Optimized Quantized GEMM for GPTQ.
389
390
  ops.def(
      "marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
391
392
      "Tensor! workspace, SymInt size_m, SymInt size_n, SymInt size_k) -> "
      "Tensor");
393
  // conditionally compiled so impl in source file
394
395

  // Marlin_24 (Sparse) Optimized Quantized GEMM for GPTQ.
396
397
398
  ops.def(
      "gptq_marlin_24_gemm(Tensor a, Tensor b_q_weight, Tensor b_meta, "
      "Tensor b_scales, Tensor workspace, "
399
400
      "int b_q_type, "
      "SymInt size_m, SymInt size_n, SymInt size_k) -> Tensor");
401
  //  conditionally compiled so impl in source file
402

403
404
  // Machete (Dense) Optimized Mixed Precision GEMM for Hopper.
  ops.def(
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
      "machete_supported_schedules("
      "   ScalarType a_type,"
      "   int b_type,"
      "   ScalarType? maybe_group_scales_type,"
      "   ScalarType? maybe_group_zeros_type,"
      "   ScalarType? maybe_channel_scales_type,"
      "   ScalarType? maybe_token_scales_type,"
      "   ScalarType? maybe_out_type"
      ") -> str[]");
  ops.def(
      "machete_mm("
      "   Tensor A,"
      "   Tensor B,"
      "   int b_type,"
      "   ScalarType? out_type,"
      "   Tensor? group_scales,"
      "   Tensor? group_zeros,"
      "   int?    group_size,"
      "   Tensor? channel_scales,"
      "   Tensor? token_scales,"
      "   str?    schedule"
      ") -> Tensor");
  ops.def(
      "machete_prepack_B("
      "   Tensor B,"
      "   ScalarType a_type,"
      "   int b_type,"
      "   ScalarType? group_scales_type"
      ") -> Tensor");
434
  // conditionally compiled so impl registration is in source file
435

436
437
438
  ops.def("permute_cols(Tensor A, Tensor perm) -> Tensor");
  ops.impl("permute_cols", torch::kCUDA, &permute_cols);

439
  // gptq_marlin Optimized Quantized GEMM for GPTQ.
440
441
442
  ops.def(
      "gptq_marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
      "Tensor b_zeros, Tensor g_idx, Tensor perm, Tensor workspace, "
443
444
      "int b_q_type, "
      "SymInt size_m, SymInt size_n, SymInt size_k, bool is_k_full, "
445
      "bool has_zp, bool use_fp32_reduce, bool is_zp_float) -> Tensor");
446
  // conditionally compiled so impl registration is in source file
447
448

  // gptq_marlin repack from GPTQ.
449
450
451
  ops.def(
      "gptq_marlin_repack(Tensor b_q_weight, Tensor perm, "
      "SymInt size_k, SymInt size_n, int num_bits) -> Tensor");
452
  // conditionally compiled so impl registrations are in source file
453

454
  // awq_marlin repack from AWQ.
455
456
457
  ops.def(
      "awq_marlin_repack(Tensor b_q_weight, SymInt size_k, "
      "SymInt size_n, int num_bits) -> Tensor");
458
  // conditionally compiled so impl registrations are in source file
459
#endif
460

461
  // Dequantization for GGML.
462
  ops.def("ggml_dequantize(Tensor W, int type, SymInt m, SymInt n) -> Tensor");
463
464
465
  ops.impl("ggml_dequantize", torch::kCUDA, &ggml_dequantize);

  // mmvq kernel for GGML.
466
  ops.def(
467
      "ggml_mul_mat_vec_a8(Tensor W, Tensor X, int type, SymInt row) "
468
      "-> Tensor");
469
470
471
  ops.impl("ggml_mul_mat_vec_a8", torch::kCUDA, &ggml_mul_mat_vec_a8);

  // mmq kernel for GGML.
472
473
  ops.def(
      "ggml_mul_mat_a8(Tensor W, Tensor X, int type, SymInt row) -> Tensor");
474
475
  ops.impl("ggml_mul_mat_a8", torch::kCUDA, &ggml_mul_mat_a8);

476
#ifndef USE_ROCM
477
  // fp8_marlin Optimized Quantized GEMM for FP8 weight-only.
478
479
  ops.def(
      "fp8_marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
480
481
      "Tensor! workspace, int num_bits, SymInt size_m, SymInt size_n, "
      "SymInt size_k) -> Tensor");
482
  // conditionally compiled so impl registration is in source file
483

484
  // marlin_qqq_gemm for QQQ.
485
486
487
  ops.def(
      "marlin_qqq_gemm(Tensor a, Tensor b_q_weight, "
      "Tensor s_tok, Tensor s_ch, Tensor s_group, "
488
489
      "Tensor! workspace, SymInt size_m, SymInt size_n, "
      "SymInt size_k) -> Tensor");
490
  // conditionally compiled so impl registration is in source file
491

492
  // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
493
  // quantization, as well as bias
494
  ops.def(
495
496
      "cutlass_scaled_mm(Tensor! out, Tensor a,"
      "                  Tensor b, Tensor a_scales,"
497
      "                  Tensor b_scales, Tensor? bias) -> ()");
498
  ops.impl("cutlass_scaled_mm", torch::kCUDA, &cutlass_scaled_mm);
499

500
501
502
503
504
505
506
507
508
  // CUTLASS w8a8 GEMM, supporting asymmetric per-tensor or per-row/column
  // quantization.
  ops.def(
      "cutlass_scaled_mm_azp(Tensor! out, Tensor a,"
      "                  Tensor b, Tensor a_scales,"
      "                  Tensor b_scales, Tensor azp_adj,"
      "                  Tensor? azp, Tensor? bias) -> ()");
  ops.impl("cutlass_scaled_mm_azp", torch::kCUDA, &cutlass_scaled_mm_azp);

509
510
  // Check if cutlass scaled_mm is supported for CUDA devices of the given
  // capability
511
512
513
  ops.def("cutlass_scaled_mm_supports_fp8(int cuda_device_capability) -> bool");
  ops.impl("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);

514
515
516
517
  // Mamba selective scan kernel
  ops.def(
      "selective_scan_fwd(Tensor! u, Tensor! delta,"
      "Tensor! A, Tensor! B, Tensor! C,"
518
      "Tensor? D_, Tensor!? z_, Tensor? delta_bias_,"
519
      "bool delta_softplus,"
520
521
522
      "Tensor? query_start_loc,"
      "Tensor? cache_indices,"
      "Tensor? has_initial_state,"
523
524
      "Tensor! ssm_states,"
      "int pad_slot_id) -> ()");
525
526
527
528
529
530
  ops.impl("selective_scan_fwd", torch::kCUDA, &selective_scan_fwd);

  ops.def(
      "causal_conv1d_update(Tensor! x,"
      "Tensor! conv_state,"
      "Tensor! weight,"
531
      "Tensor? bias_,"
532
      "bool silu_activation,"
533
      "Tensor? cache_seqlens_,"
534
535
      "Tensor? conv_state_indices,"
      "int pad_slot_id) -> ()");
536
537
538
539
540
  ops.impl("causal_conv1d_update", torch::kCUDA, &causal_conv1d_update);

  ops.def(
      "causal_conv1d_fwd(Tensor! x, Tensor! weight,"
      "Tensor? bias_,"
541
542
543
544
      "Tensor!? conv_states,"
      "Tensor? query_start_loc,"
      "Tensor? cache_indices,"
      "Tensor? has_initial_state,"
545
546
      "bool silu_activation,"
      "int pad_slot_id) -> ()");
547
  ops.impl("causal_conv1d_fwd", torch::kCUDA, &causal_conv1d_fwd);
548
549
550
#endif

  // Quantized GEMM for GPTQ.
551
552
  // Note: even though the C++ inferred schema is correct for this op, it seems
  // to prevent the meta function registry.
zhuwenwen's avatar
zhuwenwen committed
553
554
555
556
//   ops.def(
//       "gptq_gemm(Tensor a, Tensor b_q_weight, Tensor b_gptq_qzeros, "
//       "Tensor b_gptq_scales, Tensor b_g_idx, bool use_exllama, int bit) "
//       "-> Tensor");
557
//   ops.impl("gptq_gemm", torch::kCUDA, &gptq_gemm);
558
559

  // Post processing for GPTQ.
560
561
//   ops.def("gptq_shuffle(Tensor! q_weight, Tensor q_perm, int bit) -> ()");
//   ops.impl("gptq_shuffle", torch::kCUDA, &gptq_shuffle);
562
563

  // Compute FP8 quantized tensor for given scaling factor.
zhuwenwen's avatar
zhuwenwen committed
564
//   ops.def(
zhuwenwen's avatar
zhuwenwen committed
565
566
//       "static_scaled_fp8_quant(Tensor! result, Tensor input, Tensor scale) -> "
//       "()");
zhuwenwen's avatar
zhuwenwen committed
567
//   ops.impl("static_scaled_fp8_quant", torch::kCUDA, &static_scaled_fp8_quant);
568

zhuwenwen's avatar
zhuwenwen committed
569
//   // Compute dynamic-per-tensor FP8 quantized tensor and scaling factor.
zhuwenwen's avatar
zhuwenwen committed
570
//   ops.def(
zhuwenwen's avatar
zhuwenwen committed
571
572
//       "dynamic_scaled_fp8_quant(Tensor! result, Tensor input, Tensor! scale) "
//       "-> "
zhuwenwen's avatar
zhuwenwen committed
573
574
//       "()");
//   ops.impl("dynamic_scaled_fp8_quant", torch::kCUDA, &dynamic_scaled_fp8_quant);
575

zhuwenwen's avatar
zhuwenwen committed
576
//   // Compute dynamic-per-token FP8 quantized tensor and scaling factor.
577
//   ops.def(
zhuwenwen's avatar
zhuwenwen committed
578
//       "dynamic_per_token_scaled_fp8_quant(Tensor! result, Tensor input, "
zhuwenwen's avatar
zhuwenwen committed
579
//       "Tensor! scale, Tensor? scale_ub) -> "
580
581
582
//       "()");
//   ops.impl("dynamic_per_token_scaled_fp8_quant", torch::kCUDA,
//            &dynamic_per_token_scaled_fp8_quant);
583

584
585
  // Compute int8 quantized tensor for given scaling factor.
  ops.def(
586
      "static_scaled_int8_quant(Tensor! result, Tensor input, Tensor scale,"
587
      "Tensor? azp) -> ()");
588
589
590
591
  ops.impl("static_scaled_int8_quant", torch::kCUDA, &static_scaled_int8_quant);

  // Compute int8 quantized tensor and scaling factor
  ops.def(
592
      "dynamic_scaled_int8_quant(Tensor! result, Tensor input, Tensor! scale, "
593
      "Tensor!? azp) -> ()");
594
595
596
597
598
599
600
601
602
603
604
605
606
  ops.impl("dynamic_scaled_int8_quant", torch::kCUDA,
           &dynamic_scaled_int8_quant);
}

TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cache_ops), cache_ops) {
  // Cache ops
  // Swap in (out) the cache blocks from src to dst.
  cache_ops.def(
      "swap_blocks(Tensor src, Tensor! dst, Tensor block_mapping) -> ()");
  cache_ops.impl("swap_blocks", torch::kCUDA, &swap_blocks);

  // Copy the cache blocks from src to dst.
  cache_ops.def(
607
608
      "copy_blocks(Tensor(a!)[] key_caches, Tensor[](b!) value_caches, "
      "Tensor block_mapping) -> ()");
609
610
611
612
613
614
615
616
  cache_ops.impl("copy_blocks", torch::kCUDA, &copy_blocks);

  // Reshape the key and value tensors and cache them.
  cache_ops.def(
      "reshape_and_cache(Tensor key, Tensor value,"
      "                  Tensor! key_cache, Tensor! value_cache,"
      "                  Tensor slot_mapping,"
      "                  str kv_cache_dtype,"
617
      "                  float k_scale, float v_scale) -> ()");
618
619
620
621
622
623
624
625
  cache_ops.impl("reshape_and_cache", torch::kCUDA, &reshape_and_cache);

  // Reshape the key and value tensors and cache them.
  cache_ops.def(
      "reshape_and_cache_flash(Tensor key, Tensor value,"
      "                        Tensor! key_cache,"
      "                        Tensor! value_cache,"
      "                        Tensor slot_mapping,"
626
627
      "                        str kv_cache_dtype,"
      "                        float k_scale, float v_scale) -> ()");
628
629
630
  cache_ops.impl("reshape_and_cache_flash", torch::kCUDA,
                 &reshape_and_cache_flash);

631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
  // read key and value form kv cache
  cache_ops.def(
      "read_cache(Tensor keys, Tensor values,"
      "                  Tensor[]! key_caches, Tensor[]! value_caches,"
      "                  Tensor slot_mapping,"
      "                  str kv_cache_dtype) -> ()");
  cache_ops.impl("read_cache", torch::kCUDA, &read_cache);

  // write multi-layers key and value to kv cache
  cache_ops.def(
      "write_cache_multi_layers(Tensor keys, Tensor values,"
      "                  Tensor[]! key_caches, Tensor[]! value_caches,"
      "                  Tensor slot_mapping,"
      "                  str kv_cache_dtype) -> ()");
  cache_ops.impl("write_cache_multi_layers", torch::kCUDA, &write_cache_multi_layers);

647
648
  // Convert the key and value cache to fp8 data type.
  cache_ops.def(
649
650
      "convert_fp8(Tensor! dst_cache, Tensor src_cache, float scale, "
      "str kv_cache_dtype) -> ()");
651
652
653
654
655
656
657
  cache_ops.impl("convert_fp8", torch::kCUDA, &convert_fp8);
}

TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _cuda_utils), cuda_utils) {
  // Cuda utils

  // Gets the specified device attribute.
658
659
  cuda_utils.def("get_device_attribute(int attribute, int device_id) -> int");
  cuda_utils.impl("get_device_attribute", &get_device_attribute);
660
661

  // Gets the maximum shared memory per block device attribute.
662
663
  cuda_utils.def(
      "get_max_shared_memory_per_block_device_attribute(int device_id) -> int");
664
665
666
667
668
669
670
  cuda_utils.impl("get_max_shared_memory_per_block_device_attribute",
                  &get_max_shared_memory_per_block_device_attribute);
}

#ifndef USE_ROCM
TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
  // Custom all-reduce kernels
671
  custom_ar.def(
672
673
      "init_custom_ar(int[] ipc_tensors, Tensor rank_data, "
      "int rank, bool full_nvlink) -> int");
674
675
  custom_ar.impl("init_custom_ar", torch::kCUDA, &init_custom_ar);
  custom_ar.def(
676
677
678
      "all_reduce(int fa, Tensor inp, Tensor! out, int reg_buffer, "
      "int reg_buffer_sz_bytes) -> ()");
  custom_ar.impl("all_reduce", torch::kCUDA, &all_reduce);
679
680
681
682

  custom_ar.def("dispose", &dispose);
  custom_ar.def("meta_size", &meta_size);

683
  custom_ar.def("register_buffer", &register_buffer);
684
685
686
687
688
689
  custom_ar.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta);
  custom_ar.def("register_graph_buffers", &register_graph_buffers);
}
#endif

REGISTER_EXTENSION(TORCH_EXTENSION_NAME)