torch_bindings.cpp 30.7 KB
Newer Older
1
2
3
#include "cache.h"
#include "cuda_utils.h"
#include "ops.h"
4
#include "core/registration.h"
5
6

#include <torch/library.h>
7
#include <torch/version.h>
8
9
10
11
12
13
14
15
16
17
18
19
20

// 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
24
25
26
27
28
29
  //

  // The default behavior in PyTorch 2.6 is "requires_contiguous", so we need
  // to override this for many GEMMs with the following tag. Otherwise,
  // torch.compile will force all input tensors to be contiguous(), which
  // will break many custom ops that require column-major weight matrices.
  // TODO: remove this for PyTorch 2.8, when the default is planned to switch
  // to match exact eager-mode strides.
  at::Tag stride_tag = at::Tag::needs_fixed_stride_order;
30

31
32
33
  ops.def("weak_ref_tensor(Tensor input) -> Tensor");
  ops.impl("weak_ref_tensor", torch::kCUDA, &weak_ref_tensor);

34
35
36
37
  ops.def("get_cuda_view_from_cpu_tensor(Tensor cpu_tensor) -> Tensor");
  ops.impl("get_cuda_view_from_cpu_tensor", torch::kCPU,
           &get_cuda_view_from_cpu_tensor);

38
39
40
41
42
43
44
45
46
  // 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,"
47
      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
48
      "    int tp_rank, int blocksparse_local_blocks,"
49
50
51
52
53
54
55
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v1", torch::kCUDA, &paged_attention_v1);

  // PagedAttention V2.
  ops.def(
      "paged_attention_v2("
56
57
      "    Tensor! out, Tensor! exp_sums, Tensor! max_logits,"
      "    Tensor! tmp_out, Tensor query, Tensor key_cache,"
58
59
60
      "    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,"
61
      "    str kv_cache_dtype, Tensor k_scale, Tensor v_scale,"
62
      "    int tp_rank, int blocksparse_local_blocks,"
63
64
65
66
      "    int blocksparse_vert_stride, int blocksparse_block_size,"
      "    int blocksparse_head_sliding_step) -> ()");
  ops.impl("paged_attention_v2", torch::kCUDA, &paged_attention_v2);

67
68
69
70
71
72
73
74
75
76
77
78
79
#ifndef USE_ROCM
  // Merge attn states
  // Implements section 2.2 of https://www.arxiv.org/pdf/2501.01005
  // can be used to combine partial attention results (in the split-KV case)
  ops.def(
      "merge_attn_states("
      "    Tensor! output,"
      "    Tensor!? output_lse,"
      "    Tensor prefix_output,"
      "    Tensor prefix_lse,"
      "    Tensor suffix_output,"
      "    Tensor suffix_lse) -> ()");
  ops.impl("merge_attn_states", torch::kCUDA, &merge_attn_states);
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102

  ops.def(
      "convert_vertical_slash_indexes("
      "   Tensor! block_count, Tensor! block_offset, "
      "   Tensor! column_count, Tensor! column_index, "
      "   Tensor q_seqlens, Tensor q_seqlens, "
      "   Tensor vertical_indexes, Tensor slash_indexes, "
      "   int context_size, int block_size_M, int block_size_N, "
      "   bool causal) -> ()");
  ops.impl("convert_vertical_slash_indexes", torch::kCUDA,
           &convert_vertical_slash_indexes);

  ops.def(
      "convert_vertical_slash_indexes_mergehead("
      "   Tensor! block_count, Tensor! block_offset, "
      "   Tensor! column_count, Tensor! column_index, "
      "   Tensor q_seqlens, Tensor q_seqlens, "
      "   Tensor vertical_indexes, Tensor slash_indexes, "
      "   Tensor vertical_indices_count, Tensor slash_indices_count, "
      "   int context_size, int block_size_M, int block_size_N, "
      "   bool causal) -> ()");
  ops.impl("convert_vertical_slash_indexes_mergehead", torch::kCUDA,
           &convert_vertical_slash_indexes_mergehead);
103
104
#endif

105
106
  // Activation ops
  // Activation function used in SwiGLU.
107
  ops.def("silu_and_mul(Tensor! result, Tensor input) -> ()");
108
109
  ops.impl("silu_and_mul", torch::kCUDA, &silu_and_mul);

zhuwenwen's avatar
zhuwenwen committed
110
111
112
//   ops.def(
//       "silu_and_mul_quant(Tensor! result, Tensor input, Tensor scale) -> ()");
//   ops.impl("silu_and_mul_quant", torch::kCUDA, &silu_and_mul_quant);
113

114
115
116
  ops.def("mul_and_silu(Tensor! out, Tensor input) -> ()");
  ops.impl("mul_and_silu", torch::kCUDA, &mul_and_silu);

117
118
119
120
121
122
123
124
  // 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);

125
126
127
128
  // FATReLU implementation.
  ops.def("fatrelu_and_mul(Tensor! out, Tensor input, float threshold) -> ()");
  ops.impl("fatrelu_and_mul", torch::kCUDA, &fatrelu_and_mul);

129
130
131
132
133
134
135
136
  // 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);

137
138
139
140
  // Quick GELU implementation.
  ops.def("gelu_quick(Tensor! out, Tensor input) -> ()");
  ops.impl("gelu_quick", torch::kCUDA, &gelu_quick);

141
  // prepare_inputs advance_step
142
  ops.def(
143
      "advance_step_flashattn(int num_seqs, int num_queries, int block_size, "
144
145
146
      "Tensor! input_tokens, Tensor sampled_token_ids, "
      "Tensor! input_positions, Tensor! seq_lens, Tensor! slot_mapping, "
      "Tensor block_tables) -> ()");
147
148
149
150
151
152
153
154
155
156
157
158
  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);
159

160
161
162
  // Layernorm
  // Apply Root Mean Square (RMS) Normalization to the input tensor.
  ops.def(
163
      "rms_norm(Tensor! result, Tensor input, Tensor weight, float epsilon) -> "
164
165
166
167
168
169
170
171
172
      "()");
  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);

173
174
175
176
177
178
179
  // Apply repetition penalties to logits in-place
  ops.def(
      "apply_repetition_penalties_(Tensor! logits, Tensor prompt_mask, "
      "Tensor output_mask, Tensor repetition_penalties) -> ()");
  ops.impl("apply_repetition_penalties_", torch::kCUDA,
           &apply_repetition_penalties_);

180
181
  // Layernorm-quant
  // Apply Root Mean Square (RMS) Normalization to the input tensor.
182
183
184
185
186
187
//   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);
188
189

  // In-place fused Add and RMS Normalization.
190
191
192
193
194
195
//   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);
196

197
  // Fused Layernorm + Quant kernels
198
199
200
201
202
203
//   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);
204

205
206
207
208
  // Rotary embedding
  // Apply GPT-NeoX or GPT-J style rotary embedding to query and key.
  ops.def(
      "rotary_embedding(Tensor positions, Tensor! query,"
209
      "                 Tensor!? key, int head_size,"
210
211
212
213
214
215
216
      "                 Tensor cos_sin_cache, bool is_neox) -> ()");
  ops.impl("rotary_embedding", torch::kCUDA, &rotary_embedding);

  // 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,"
217
      "                         Tensor!? key, int head_size,"
218
219
220
221
222
      "                         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);

223
224
225
226
  // 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);
  
227
228
229
  // Quantization ops
#ifndef USE_ROCM
  // Quantized GEMM for AQLM.
230
231
232
  ops.def(
      "aqlm_gemm(Tensor input, Tensor codes, Tensor codebooks, "
      "Tensor scales, int[] codebook_partition_sizes, Tensor? bias) "
233
234
      "-> Tensor",
      {stride_tag});
235
236
237
  ops.impl("aqlm_gemm", torch::kCUDA, &aqlm_gemm);

  // Decompression method for AQLM.
238
239
  ops.def(
      "aqlm_dequant(Tensor codes, Tensor codebooks, "
240
241
      "int[] codebook_partition_sizes) -> Tensor",
      {stride_tag});
242
243
244
  ops.impl("aqlm_dequant", torch::kCUDA, &aqlm_dequant);

  // Quantized GEMM for AWQ.
245
246
  ops.def(
      "awq_gemm(Tensor _in_feats, Tensor _kernel, Tensor _scaling_factors, "
247
248
      "Tensor _zeros, SymInt split_k_iters) -> Tensor",
      {stride_tag});
249
250
251
  ops.impl("awq_gemm", torch::kCUDA, &awq_gemm);

  // Dequantization for AWQ.
252
253
  ops.def(
      "awq_dequantize(Tensor _kernel, Tensor _scaling_factors, "
254
255
      "Tensor _zeros, SymInt split_k_iters, int thx, int thy) -> Tensor",
      {stride_tag});
256
257
  ops.impl("awq_dequantize", torch::kCUDA, &awq_dequantize);

258
259
260
261
262
263
264
265
266
267
268
269
270
271
  // 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

272
  // Marlin (Dense) Optimized Quantized GEMM for GPTQ.
273
274
  ops.def(
      "marlin_gemm(Tensor a, Tensor b_q_weight, Tensor b_scales, "
275
      "Tensor! workspace, SymInt size_m, SymInt size_n, SymInt size_k) -> "
276
277
      "Tensor",
      {stride_tag});
278
  // conditionally compiled so impl in source file
279
280

  // Marlin_24 (Sparse) Optimized Quantized GEMM for GPTQ.
281
282
283
  ops.def(
      "gptq_marlin_24_gemm(Tensor a, Tensor b_q_weight, Tensor b_meta, "
      "Tensor b_scales, Tensor workspace, "
284
      "int b_q_type, "
285
286
      "SymInt size_m, SymInt size_n, SymInt size_k) -> Tensor",
      {stride_tag});
287
  //  conditionally compiled so impl in source file
288

289
  // Machete (Dense) Optimized Mixed Precision GEMM for Hopper.
290
  ops.def(
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
      "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"
312
313
      ") -> Tensor",
      {stride_tag});
314
315
316
317
318
319
320
  ops.def(
      "machete_prepack_B("
      "   Tensor B,"
      "   ScalarType a_type,"
      "   int b_type,"
      "   ScalarType? group_scales_type"
      ") -> Tensor");
321
  // conditionally compiled so impl registration is in source file
322

323
324
325
  ops.def("permute_cols(Tensor A, Tensor perm) -> Tensor");
  ops.impl("permute_cols", torch::kCUDA, &permute_cols);

326
  // gptq_marlin Optimized Quantized GEMM for GPTQ.
327
  ops.def(
328
      "gptq_marlin_gemm(Tensor a, Tensor? c_or_none, Tensor b_q_weight, "
329
330
      "Tensor b_scales, Tensor? global_scale, Tensor? b_zeros_or_none, Tensor? "
      "g_idx_or_none, Tensor? perm_or_none, Tensor workspace, int b_q_type, "
331
      "SymInt size_m, SymInt size_n, SymInt size_k, bool is_k_full, "
332
      "bool use_atomic_add, bool use_fp32_reduce, bool is_zp_float) -> Tensor",
333
      {stride_tag});
334
  // conditionally compiled so impl registration is in source file
335
336

  // gptq_marlin repack from GPTQ.
337
338
339
  ops.def(
      "gptq_marlin_repack(Tensor b_q_weight, Tensor perm, "
      "SymInt size_k, SymInt size_n, int num_bits) -> Tensor");
340
  // conditionally compiled so impl registrations are in source file
341

342
  // awq_marlin repack from AWQ.
343
344
345
  ops.def(
      "awq_marlin_repack(Tensor b_q_weight, SymInt size_k, "
      "SymInt size_n, int num_bits) -> Tensor");
346
  // conditionally compiled so impl registrations are in source file
347
#endif
348

349
  // Dequantization for GGML.
350
351
352
  ops.def(
      "ggml_dequantize(Tensor W, int type, SymInt m, SymInt n, ScalarType? "
      "dtype) -> Tensor");
353
354
355
  ops.impl("ggml_dequantize", torch::kCUDA, &ggml_dequantize);

  // mmvq kernel for GGML.
356
  ops.def(
357
      "ggml_mul_mat_vec_a8(Tensor W, Tensor X, int type, SymInt row) "
358
      "-> Tensor");
359
360
361
  ops.impl("ggml_mul_mat_vec_a8", torch::kCUDA, &ggml_mul_mat_vec_a8);

  // mmq kernel for GGML.
362
363
  ops.def(
      "ggml_mul_mat_a8(Tensor W, Tensor X, int type, SymInt row) -> Tensor");
364
365
  ops.impl("ggml_mul_mat_a8", torch::kCUDA, &ggml_mul_mat_a8);

366
367
368
369
370
371
372
373
  // moe kernel for GGML.
  ops.def(
      "ggml_moe_a8(Tensor X, Tensor W, "
      "Tensor sorted_token_ids, Tensor expert_ids, Tensor "
      "num_tokens_post_padded, "
      "int type, SymInt row, SymInt top_k, SymInt tokens) -> Tensor");
  ops.impl("ggml_moe_a8", torch::kCUDA, &ggml_moe_a8);

374
375
376
377
378
379
  ops.def(
      "ggml_moe_a8_vec(Tensor X, Tensor W, "
      "Tensor topk_ids, int top_k, "
      "int type, SymInt row, SymInt tokens) -> Tensor");
  ops.impl("ggml_moe_a8_vec", torch::kCUDA, &ggml_moe_a8_vec);

380
381
  ops.def("ggml_moe_get_block_size", &ggml_moe_get_block_size);

382
#ifndef USE_ROCM
383
  // marlin_qqq_gemm for QQQ.
384
385
386
  ops.def(
      "marlin_qqq_gemm(Tensor a, Tensor b_q_weight, "
      "Tensor s_tok, Tensor s_ch, Tensor s_group, "
387
      "Tensor! workspace, SymInt size_m, SymInt size_n, "
388
389
      "SymInt size_k) -> Tensor",
      {stride_tag});
390
  // conditionally compiled so impl registration is in source file
391

392
393
394
395
  // CUTLASS nvfp4 block scaled GEMM
  ops.def(
      "cutlass_scaled_fp4_mm(Tensor! out, Tensor a, Tensor b,"
      "                      Tensor block_scale_a, Tensor block_scale_b,"
396
397
      "                      Tensor alpha) -> ()",
      {stride_tag});
398
399
  ops.impl("cutlass_scaled_fp4_mm", torch::kCUDA, &cutlass_scaled_fp4_mm);

400
401
402
403
404
405
  // cutlass blockwise scaledgroup GEMM
  ops.def(
      "cutlass_blockwise_scaled_grouped_mm(Tensor! output, Tensor a, Tensor b, "
      "Tensor scales_a, Tensor scales_b, "
      "Tensor problem_sizes, Tensor expert_offsets) -> ()",
      {stride_tag});
406
  // conditionally compiled so impl registration is in source file
407

408
409
410
411
412
413
414
415
  // cutlass nvfp4 block scaled group GEMM
  ops.def(
      "cutlass_fp4_group_mm(Tensor! out, Tensor a, Tensor b,"
      " Tensor a_blockscale, Tensor b_blockscales, Tensor alphas,"
      " Tensor problem_sizes, Tensor expert_offsets, Tensor sf_offsets) -> ()",
      {stride_tag});
  ops.impl("cutlass_fp4_group_mm", torch::kCUDA, &cutlass_fp4_group_mm);

416
  // CUTLASS w8a8 GEMM, supporting symmetric per-tensor or per-row/column
417
  // quantization, as well as bias
418
  ops.def(
419
420
      "cutlass_scaled_mm(Tensor! out, Tensor a,"
      "                  Tensor b, Tensor a_scales,"
421
422
      "                  Tensor b_scales, Tensor? bias) -> ()",
      {stride_tag});
423
  ops.impl("cutlass_scaled_mm", torch::kCUDA, &cutlass_scaled_mm);
424

425
426
427
428
429
430
  // 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,"
431
432
      "                  Tensor? azp, Tensor? bias) -> ()",
      {stride_tag});
433
434
  ops.impl("cutlass_scaled_mm_azp", torch::kCUDA, &cutlass_scaled_mm_azp);

435
436
  // Check if cutlass scaled_mm is supported for CUDA devices of the given
  // capability
437
438
439
  ops.def("cutlass_scaled_mm_supports_fp8(int cuda_device_capability) -> bool");
  ops.impl("cutlass_scaled_mm_supports_fp8", &cutlass_scaled_mm_supports_fp8);

440
441
442
443
444
445
446
447
448
449
  // Check if cutlass grouped gemm is supported for CUDA devices of the given
  // capability
  ops.def("cutlass_group_gemm_supported(int cuda_device_capability) -> bool");
  ops.impl("cutlass_group_gemm_supported", &cutlass_group_gemm_supported);

  // CUTLASS w8a8 grouped GEMM
  ops.def(
      "cutlass_moe_mm(Tensor! out_tensors, Tensor a_tensors, Tensor b_tensors, "
      "               Tensor a_scales, Tensor b_scales, Tensor expert_offsets, "
      "               Tensor problem_sizes, Tensor a_strides, "
450
451
      "               Tensor b_strides, Tensor c_strides, bool per_act_token, "
      "               bool per_out_ch) -> ()",
452
453
454
455
456
457
458
459
460
461
462
463
464
465
      {stride_tag});
  ops.impl("cutlass_moe_mm", torch::kCUDA, &cutlass_moe_mm);

  // A function that computes data required to run fused MoE with w8a8 grouped
  // GEMM. It takes topk_ids as an input, and computes expert_offsets
  // (token start indices of each expert). In addition to this, it computes
  // problem sizes for each expert's multiplication used by the two mms called
  // from fused MoE operation, and arrays with permutations required to shuffle
  // and de-shuffle the input/output of the fused operation.
  ops.def(
      "get_cutlass_moe_mm_data(Tensor topk_ids, Tensor! expert_offsets, "
      "                        Tensor! problem_sizes1, Tensor! problem_sizes2, "
      "                        Tensor! input_permutation, "
      "                        Tensor! output_permutation, int num_experts, "
466
      "                        int n, int k, Tensor? blockscale_offsets) -> ()",
467
468
469
      {stride_tag});
  ops.impl("get_cutlass_moe_mm_data", torch::kCUDA, &get_cutlass_moe_mm_data);

470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
  // A function that computes data required to run fused MoE with w8a8 grouped
  // GEMM and PPLX. It takes expert_num_tokens and non_zero_expert_idxs
  // as an input, and computes expert_offsets (token start indices of each
  // expert). In addition to this, it computes problem sizes for each expert's
  // multiplication used by the two mms called from fused MoE operation.
  ops.def(
      "get_cutlass_pplx_moe_mm_data(Tensor! expert_offsets, "
      "                             Tensor! problem_sizes1, "
      "                             Tensor! problem_sizes2, "
      "                             Tensor expert_num_tokens, "
      "                             int num_local_experts, int padded_m, "
      "                             int n, int k) -> ()",
      {stride_tag});
  ops.impl("get_cutlass_pplx_moe_mm_data", torch::kCUDA,
           &get_cutlass_pplx_moe_mm_data);

486
487
488
489
490
  // Check if cutlass scaled_mm supports block quantization (used by DeepSeekV3)
  ops.def(
      "cutlass_scaled_mm_supports_block_fp8(int cuda_device_capability) -> "
      "bool");
  ops.impl("cutlass_scaled_mm_supports_block_fp8",
491
           &cutlass_scaled_mm_supports_block_fp8);
492

493
494
495
496
497
498
499
  // Check if cutlass sparse scaled_mm is supported for CUDA devices of the
  // given capability
  ops.def(
      "cutlass_sparse_scaled_mm_supported(int cuda_device_capability) -> bool");
  ops.impl("cutlass_sparse_scaled_mm_supported",
           &cutlass_sparse_scaled_mm_supported);

500
501
502
503
504
505
  // CUTLASS sparse GEMM, supporting symmetric per-tensor or per-row/column
  // quantization, as well as bias
  ops.def(
      "cutlass_scaled_sparse_mm(Tensor! out, Tensor a,"
      "                         Tensor bt_nzs,"
      "                         Tensor bt_meta, Tensor a_scales,"
506
507
      "                         Tensor b_scales, Tensor? bias) -> ()",
      {stride_tag});
508
509
510
  ops.impl("cutlass_scaled_sparse_mm", torch::kCUDA, &cutlass_scaled_sparse_mm);

  // CUTLASS sparse matrix compressor
511
512
  ops.def("cutlass_sparse_compress(Tensor a) -> Tensor[]");
  ops.impl("cutlass_sparse_compress", &cutlass_sparse_compress);
513

514
515
516
517
518
519
520
  // CUTLASS MLA decode
  ops.def(
      "cutlass_mla_decode(Tensor! out, Tensor q_nope, Tensor q_pe,"
      "                   Tensor kv_c_and_k_pe_cache, Tensor seq_lens,"
      "                   Tensor page_table, float scale) -> ()");
  ops.impl("cutlass_mla_decode", torch::kCUDA, &cutlass_mla_decode);

521
522
523
524
525
526
  // Compute NVFP4 block quantized tensor.
  ops.def(
      "scaled_fp4_quant(Tensor! output, Tensor input,"
      "                 Tensor! output_scale, Tensor input_scale) -> ()");
  ops.impl("scaled_fp4_quant", torch::kCUDA, &scaled_fp4_quant);

527
528
529
530
531
532
533
  // Compute NVFP4 experts quantization.
  ops.def(
      "scaled_fp4_experts_quant(Tensor! output, Tensor! output_scale,"
      "Tensor input, Tensor input_global_scale, Tensor input_offset_by_experts,"
      "Tensor output_scale_offset_by_experts) -> ()");
  ops.impl("scaled_fp4_experts_quant", torch::kCUDA, &scaled_fp4_experts_quant);

534
535
536
537
  // Check if cutlass_scaled_mm_fp4 is supported for CUDA devices
  // of the given capability
  ops.def("cutlass_scaled_mm_supports_fp4(int cuda_device_capability) -> bool");
  ops.impl("cutlass_scaled_mm_supports_fp4", &cutlass_scaled_mm_supports_fp4);
538
539
540
#endif

  // Quantized GEMM for GPTQ.
541
542
543
544
545
  // Note: even though the C++ inferred schema is correct for this op, it seems
  // to prevent the meta function registry.
  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) "
546
547
      "-> Tensor",
      {stride_tag});
548
549
550
551
552
553
554
  ops.impl("gptq_gemm", torch::kCUDA, &gptq_gemm);

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

  // Compute FP8 quantized tensor for given scaling factor.
555
556
557
558
559
560
561
562
563
564
565
//   ops.def(
//       "static_scaled_fp8_quant(Tensor! result, Tensor input, Tensor scale) -> "
//       "()");
//   ops.impl("static_scaled_fp8_quant", torch::kCUDA, &static_scaled_fp8_quant);

//   // Compute dynamic-per-tensor FP8 quantized tensor and scaling factor.
//   ops.def(
//       "dynamic_scaled_fp8_quant(Tensor! result, Tensor input, Tensor! scale) "
//       "-> "
//       "()");
//   ops.impl("dynamic_scaled_fp8_quant", torch::kCUDA, &dynamic_scaled_fp8_quant);
566

567
  // Compute dynamic-per-token FP8 quantized tensor and scaling factor.
568
569
570
571
572
573
//   ops.def(
//       "dynamic_per_token_scaled_fp8_quant(Tensor! result, Tensor input, "
//       "Tensor! scale, Tensor? scale_ub) -> "
//       "()");
//   ops.impl("dynamic_per_token_scaled_fp8_quant", torch::kCUDA,
//            &dynamic_per_token_scaled_fp8_quant);
574

575
576
  // Compute int8 quantized tensor for given scaling factor.
  ops.def(
577
      "static_scaled_int8_quant(Tensor! result, Tensor input, Tensor scale,"
578
      "Tensor? azp) -> ()");
579
580
581
582
  ops.impl("static_scaled_int8_quant", torch::kCUDA, &static_scaled_int8_quant);

  // Compute int8 quantized tensor and scaling factor
  ops.def(
583
      "dynamic_scaled_int8_quant(Tensor! result, Tensor input, Tensor! scale, "
584
      "Tensor!? azp) -> ()");
585
586
  ops.impl("dynamic_scaled_int8_quant", torch::kCUDA,
           &dynamic_scaled_int8_quant);
587

588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
  // Mamba selective scan kernel
  ops.def(
      "selective_scan_fwd(Tensor! u, Tensor! delta,"
      "Tensor! A, Tensor! B, Tensor! C,"
      "Tensor? D_, Tensor!? z_, Tensor? delta_bias_,"
      "bool delta_softplus,"
      "Tensor? query_start_loc,"
      "Tensor? cache_indices,"
      "Tensor? has_initial_state,"
      "Tensor! ssm_states,"
      "int pad_slot_id) -> ()");
  ops.impl("selective_scan_fwd", torch::kCUDA, &selective_scan_fwd);

  ops.def(
      "causal_conv1d_update(Tensor! x,"
      "Tensor! conv_state,"
      "Tensor! weight,"
      "Tensor? bias_,"
      "bool silu_activation,"
      "Tensor? cache_seqlens_,"
      "Tensor? conv_state_indices,"
      "int pad_slot_id) -> ()");
  ops.impl("causal_conv1d_update", torch::kCUDA, &causal_conv1d_update);

  ops.def(
      "causal_conv1d_fwd(Tensor! x, Tensor! weight,"
      "Tensor? bias_,"
      "Tensor!? conv_states,"
      "Tensor? query_start_loc,"
      "Tensor? cache_indices,"
      "Tensor? has_initial_state,"
      "bool silu_activation,"
      "int pad_slot_id) -> ()");
  ops.impl("causal_conv1d_fwd", torch::kCUDA, &causal_conv1d_fwd);

623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
#ifndef USE_ROCM
  // reorder weight for AllSpark Ampere W8A16 Fused Gemm kernel
  ops.def(
      "rearrange_kn_weight_as_n32k16_order(Tensor b_qweight, Tensor b_scales, "
      "Tensor? b_zeros, "
      "bool has_zp, Tensor! b_qweight_reorder, Tensor! b_scales_reorder, "
      "Tensor!? b_zeros_reorder, "
      "int K, int N, int N_32align) -> ()");
  //  conditionally compiled so impl in source file

  // AllSpark quantization ops
  ops.def(
      "allspark_w8a16_gemm(Tensor a, Tensor b_qweight, Tensor b_scales, "
      "Tensor? b_qzeros, "
      "SymInt n, SymInt group_size, SymInt sm_count, SymInt sm_version, SymInt "
      "CUBLAS_M_THRESHOLD, bool has_zp, bool n32k16_reorder) -> Tensor");
  //  conditionally compiled so impl in source file
#endif
641
642
643
644
645
646
647
648
649
650
651
}

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(
652
653
      "copy_blocks(Tensor(a!)[] key_caches, Tensor[](b!) value_caches, "
      "Tensor block_mapping) -> ()");
654
655
  cache_ops.impl("copy_blocks", torch::kCUDA, &copy_blocks);

656
657
658
659
  cache_ops.def(
      "copy_blocks_mla(Tensor(a!)[] kv_caches, Tensor block_mapping) -> ()");
  cache_ops.impl("copy_blocks_mla", torch::kCUDA, &copy_blocks_mla);

660
661
662
663
664
665
  // 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,"
666
      "                  Tensor k_scale, Tensor v_scale) -> ()");
667
668
669
670
671
672
673
674
  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,"
675
      "                        str kv_cache_dtype,"
676
      "                        Tensor k_scale, Tensor v_scale) -> ()");
677
678
679
  cache_ops.impl("reshape_and_cache_flash", torch::kCUDA,
                 &reshape_and_cache_flash);

680
681
682
683
684
685
686
687
688
  // Concat kv_c and k_pe and cache them.
  cache_ops.def(
      "concat_and_cache_mla(Tensor kv_c, Tensor k_pe,"
      "                     Tensor! kv_cache,"
      "                     Tensor slot_mapping,"
      "                     str kv_cache_dtype,"
      "                     Tensor scale) -> ()");
  cache_ops.impl("concat_and_cache_mla", torch::kCUDA, &concat_and_cache_mla);

689
690
  // Convert the key and value cache to fp8 data type.
  cache_ops.def(
691
692
      "convert_fp8(Tensor! dst_cache, Tensor src_cache, float scale, "
      "str kv_cache_dtype) -> ()");
693
  cache_ops.impl("convert_fp8", torch::kCUDA, &convert_fp8);
694
695
696
697
698
699

  // Gather cache blocks from src_cache to dst.
  cache_ops.def(
      "gather_cache(Tensor src_cache, Tensor! dst, Tensor block_table, "
      "Tensor cu_seq_lens, int batch_size, Tensor? seq_starts) -> ()");
  cache_ops.impl("gather_cache", torch::kCUDA, &gather_cache);
700
701
702
703
704
705
}

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

  // Gets the specified device attribute.
706
707
  cuda_utils.def("get_device_attribute(int attribute, int device_id) -> int");
  cuda_utils.impl("get_device_attribute", &get_device_attribute);
708
709

  // Gets the maximum shared memory per block device attribute.
710
711
  cuda_utils.def(
      "get_max_shared_memory_per_block_device_attribute(int device_id) -> int");
712
713
714
715
716
717
  cuda_utils.impl("get_max_shared_memory_per_block_device_attribute",
                  &get_max_shared_memory_per_block_device_attribute);
}

TORCH_LIBRARY_EXPAND(CONCAT(TORCH_EXTENSION_NAME, _custom_ar), custom_ar) {
  // Custom all-reduce kernels
718
  custom_ar.def(
719
      "init_custom_ar(int[] ipc_tensors, Tensor rank_data, "
720
      "int rank, bool fully_connected) -> int");
721
722
  custom_ar.impl("init_custom_ar", torch::kCUDA, &init_custom_ar);
  custom_ar.def(
723
724
725
      "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);
726
727
728
729

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

730
  custom_ar.def("register_buffer", &register_buffer);
731
732
  custom_ar.def("get_graph_buffer_ipc_meta", &get_graph_buffer_ipc_meta);
  custom_ar.def("register_graph_buffers", &register_graph_buffers);
733
734
735
736
737
738
739

  custom_ar.def("allocate_shared_buffer_and_handle",
                &allocate_shared_buffer_and_handle);
  custom_ar.def("open_mem_handle(Tensor mem_handle) -> int", &open_mem_handle);
  custom_ar.impl("open_mem_handle", torch::kCPU, &open_mem_handle);

  custom_ar.def("free_shared_buffer", &free_shared_buffer);
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
#ifdef USE_ROCM
  // Quick Reduce all-reduce kernels
  custom_ar.def(
      "qr_all_reduce(int fa, Tensor inp, Tensor out, int quant_level, bool "
      "cast_bf2half) -> ()");
  custom_ar.impl("qr_all_reduce", torch::kCUDA, &qr_all_reduce);

  custom_ar.def("init_custom_qr", &init_custom_qr);
  custom_ar.def("qr_destroy", &qr_destroy);

  custom_ar.def("qr_get_handle", &qr_get_handle);

  custom_ar.def("qr_open_handles(int _fa, Tensor[](b!) handles) -> ()");
  custom_ar.impl("qr_open_handles", torch::kCPU, &qr_open_handles);

  // Max input size in bytes
  custom_ar.def("qr_max_size", &qr_max_size);
#endif
758
759
}

760
REGISTER_EXTENSION(TORCH_EXTENSION_NAME)