| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import glob |
| 4 | import itertools |
| 5 | import os |
| 6 | import subprocess |
| 7 | import sys |
| 8 | |
| 9 | import jinja2 |
| 10 | |
| 11 | ARCHS = [] |
| 12 | SUPPORT_FP8 = False |
| 13 | SUPPORT_SM75 = False |
| 14 | SUPPORT_SM80 = False |
| 15 | for arch in sys.argv[1].split(","): |
| 16 | arch = arch[: arch.index(".") + 2].replace(".", "") |
| 17 | arch = int(arch) |
| 18 | # SM89 and the SM12x family (SM120 RTX 5090, SM121 DGX Spark GB10) |
| 19 | # fully support mma.sync.aligned.m16n8k32.row.col.f32.e4m3.e4m3.f32. |
| 20 | # SM90 and SM100 can use this PTX, but it’s simulated |
| 21 | # with FP16 MMA, so it cannot achieve any acceleration. |
| 22 | if arch == 89 or arch // 10 == 12: |
| 23 | SUPPORT_FP8 = True |
| 24 | if arch >= 80: |
| 25 | SUPPORT_SM80 = True |
| 26 | if arch == 75: |
| 27 | SUPPORT_SM75 = True |
| 28 | |
| 29 | FILE_HEAD_COMMENT = """ |
| 30 | // auto generated by generate_kernels.py |
| 31 | // clang-format off |
| 32 | """.lstrip() |
| 33 | |
| 34 | FILE_HEAD = ( |
| 35 | FILE_HEAD_COMMENT |
| 36 | + """ |
| 37 | #include "kernel.h" |
| 38 | #include "marlin_template.h" |
| 39 | |
| 40 | namespace MARLIN_NAMESPACE_NAME { |
| 41 | """ |
| 42 | ) |
| 43 | |
| 44 | TEMPLATE = ( |
| 45 | "template __global__ void Marlin<" |
| 46 | "{{a_type_id}}, " |
| 47 | "{{b_type_id}}, " |
| 48 | "{{c_type_id}}, " |
| 49 | "{{s_type_id}}, " |
| 50 | "{{threads}}, " |
| 51 | "{{thread_m_blocks}}, " |
| 52 | "{{thread_n_blocks}}, " |
| 53 | "{{thread_k_blocks}}, " |
| 54 | "{{m_block_size_8}}, " |
| 55 | "{{stages}}, " |
| 56 | "{{group_blocks}}, " |
| 57 | "{{is_zp_float}}>" |
| 58 | "( MARLIN_KERNEL_PARAMS );" |
| 59 | ) |
| 60 | |
| 61 | THREAD_CONFIGS = [(128, 128, 256), (64, 256, 256), (64, 128, 128), (128, 64, 128)] |
| 62 | |
| 63 | THREAD_M_BLOCKS = [0.5, 1, 2, 3, 4] |
| 64 | |
| 65 | QUANT_CONFIGS = [ |
| 66 | # AWQ-INT4 |
| 67 | { |
| 68 | "b_type": "kU4", |
| 69 | "thread_configs": THREAD_CONFIGS, |
| 70 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 71 | "group_blocks": [-1, 2, 4, 8], |
| 72 | }, |
| 73 | # GPTQ-INT4 |
| 74 | { |
| 75 | "b_type": "kU4B8", |
| 76 | "thread_configs": THREAD_CONFIGS, |
| 77 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 78 | "group_blocks": [-1, 2, 4, 8], |
| 79 | }, |
| 80 | # AWQ-INT8 |
| 81 | { |
| 82 | "b_type": "kU8B128", |
| 83 | "thread_configs": THREAD_CONFIGS, |
| 84 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 85 | "group_blocks": [-1, 2, 4, 8], |
| 86 | }, |
| 87 | # FP8 |
| 88 | { |
| 89 | "b_type": "kFE4M3fn", |
| 90 | "thread_configs": THREAD_CONFIGS, |
| 91 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 92 | "group_blocks": [-1, 8], |
| 93 | }, |
| 94 | # NVFP4 |
| 95 | { |
| 96 | "b_type": "kFE2M1f", |
| 97 | "s_type": "kFE4M3fn", |
| 98 | "thread_configs": THREAD_CONFIGS, |
| 99 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 100 | "group_blocks": [1], |
| 101 | }, |
| 102 | # MXFP4 |
| 103 | { |
| 104 | "a_type": ["kBFloat16"], |
| 105 | "b_type": "kFE2M1f", |
| 106 | "s_type": "kFE8M0fnu", |
| 107 | "thread_configs": THREAD_CONFIGS, |
| 108 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 109 | "group_blocks": [2], |
| 110 | }, |
| 111 | # MXFP8 |
| 112 | { |
| 113 | "a_type": ["kBFloat16"], |
| 114 | "b_type": "kFE4M3fn", |
| 115 | "s_type": "kFE8M0fnu", |
| 116 | "thread_configs": THREAD_CONFIGS, |
| 117 | "thread_m_blocks": THREAD_M_BLOCKS, |
| 118 | "group_blocks": [2], |
| 119 | }, |
| 120 | # AWQ-INT4 with INT8 activation |
| 121 | { |
| 122 | "a_type": ["kS8"], |
| 123 | "b_type": "kU4", |
| 124 | "thread_configs": THREAD_CONFIGS, |
| 125 | "thread_m_blocks": [1, 2, 3, 4], |
| 126 | "group_blocks": [-1, 2, 4, 8], |
| 127 | }, |
| 128 | # GPTQ-INT4 with INT8 activation |
| 129 | { |
| 130 | "a_type": ["kS8"], |
| 131 | "b_type": "kU4B8", |
| 132 | "thread_configs": THREAD_CONFIGS, |
| 133 | "thread_m_blocks": [1, 2, 3, 4], |
| 134 | "group_blocks": [-1, 2, 4, 8], |
| 135 | }, |
| 136 | # GPTQ-INT4 with FP8 activation |
| 137 | { |
| 138 | "a_type": ["kFE4M3fn"], |
| 139 | "b_type": "kU4B8", |
| 140 | "thread_configs": THREAD_CONFIGS, |
| 141 | "thread_m_blocks": [1, 2, 3, 4], |
| 142 | "group_blocks": [-1, 2, 4, 8], |
| 143 | }, |
| 144 | # AWQ-INT4 with FP8 activation |
| 145 | { |
| 146 | "a_type": ["kFE4M3fn"], |
| 147 | "b_type": "kU4", |
| 148 | "thread_configs": THREAD_CONFIGS, |
| 149 | "thread_m_blocks": [1, 2, 3, 4], |
| 150 | "group_blocks": [-1, 2, 4, 8], |
| 151 | }, |
| 152 | # MXFP4 with FP8 activation |
| 153 | { |
| 154 | "a_type": ["kFE4M3fn"], |
| 155 | "b_type": "kFE2M1f", |
| 156 | "c_type": ["kBFloat16"], |
| 157 | "s_type": "kFE8M0fnu", |
| 158 | "thread_configs": THREAD_CONFIGS, |
| 159 | "thread_m_blocks": [1, 2, 3, 4], |
| 160 | "group_blocks": [2], |
| 161 | }, |
| 162 | ] |
| 163 | |
| 164 | |
| 165 | def remove_old_kernels(): |
| 166 | for filename in glob.glob(os.path.dirname(__file__) + "/*kernel_*.cu"): |
| 167 | subprocess.call(["rm", "-f", filename]) |
| 168 | |
| 169 | filename = os.path.dirname(__file__) + "/kernel_selector.h" |
| 170 | subprocess.call(["rm", "-f", filename]) |
| 171 | |
| 172 | |
| 173 | def generate_new_kernels(): |
| 174 | result_dict = {} |
| 175 | sm_75_result_dict = {} |
| 176 | |
| 177 | for quant_config in QUANT_CONFIGS: |
| 178 | c_types = quant_config.get("c_type", ["kFloat16", "kBFloat16"]) |
| 179 | a_types = quant_config.get("a_type", ["kFloat16", "kBFloat16"]) |
| 180 | b_type = quant_config["b_type"] |
| 181 | all_group_blocks = quant_config["group_blocks"] |
| 182 | all_m_blocks = quant_config["thread_m_blocks"] |
| 183 | all_thread_configs = quant_config["thread_configs"] |
| 184 | |
| 185 | for a_type, c_type in itertools.product(a_types, c_types): |
| 186 | if not SUPPORT_FP8 and a_type == "kFE4M3fn": |
| 187 | continue |
| 188 | if "16" in a_type and "16" in c_type and a_type != c_type: |
| 189 | continue |
| 190 | s_type = quant_config.get("s_type", c_type) |
| 191 | if (a_type, b_type, c_type) not in result_dict: |
| 192 | result_dict[(a_type, b_type, c_type)] = [] |
| 193 | if a_type in ["kFloat16", "kS8"] and c_type == "kFloat16": |
| 194 | sm_75_result_dict[(a_type, b_type, c_type)] = [] |
| 195 | |
| 196 | for group_blocks, m_blocks, thread_configs in itertools.product( |
| 197 | all_group_blocks, all_m_blocks, all_thread_configs |
| 198 | ): |
| 199 | thread_k, thread_n, threads = thread_configs |
| 200 | |
| 201 | if threads == 256: |
| 202 | # for small batch (m_blocks == 1), |
| 203 | # we only need (128, 128, 256) |
| 204 | # for large batch (m_blocks > 1), |
| 205 | # we only need (64, 256, 256) |
| 206 | if m_blocks <= 1 and (thread_k, thread_n) != (128, 128): |
| 207 | continue |
| 208 | if m_blocks > 1 and (thread_k, thread_n) != (64, 256): |
| 209 | continue |
| 210 | |
| 211 | config = { |
| 212 | "threads": threads, |
| 213 | "s_type": s_type, |
| 214 | "thread_m_blocks": max(m_blocks, 1), |
| 215 | "thread_k_blocks": thread_k // 16, |
| 216 | "thread_n_blocks": thread_n // 16, |
| 217 | "m_block_size_8": "true" if m_blocks == 0.5 else "false", |
| 218 | "stages": 4, |
| 219 | "group_blocks": group_blocks, |
| 220 | "is_zp_float": "false", |
| 221 | } |
| 222 | |
| 223 | if SUPPORT_SM80: |
| 224 | result_dict[(a_type, b_type, c_type)].append(config) |
| 225 | if (a_type, b_type, c_type) in sm_75_result_dict and SUPPORT_SM75: |
| 226 | config_sm75 = config.copy() |
| 227 | config_sm75["stages"] = 2 |
| 228 | sm_75_result_dict[(a_type, b_type, c_type)].append(config_sm75) |
| 229 | |
| 230 | kernel_selector_str = FILE_HEAD_COMMENT |
| 231 | |
| 232 | for result_dict_tmp in [result_dict, sm_75_result_dict]: |
| 233 | for (a_type, b_type, c_type), config_list in result_dict_tmp.items(): |
| 234 | all_template_str_list = [] |
| 235 | if not config_list: |
| 236 | continue |
| 237 | for config in config_list: |
| 238 | s_type = config["s_type"] |
| 239 | template_str = jinja2.Template(TEMPLATE).render( |
| 240 | a_type_id=f"vllm::{a_type}.id()", |
| 241 | b_type_id=f"vllm::{b_type}.id()", |
| 242 | c_type_id=f"vllm::{c_type}.id()", |
| 243 | s_type_id=f"vllm::{s_type}.id()", |
| 244 | **config, |
| 245 | ) |
| 246 | all_template_str_list.append(template_str) |
| 247 | |
| 248 | conditions = [ |
| 249 | f"a_type == vllm::{a_type}", |
| 250 | f"b_type == vllm::{b_type}", |
| 251 | f"c_type == vllm::{c_type}", |
| 252 | f"s_type == vllm::{s_type}", |
| 253 | f"threads == {config['threads']}", |
| 254 | f"thread_m_blocks == {config['thread_m_blocks']}", |
| 255 | f"thread_n_blocks == {config['thread_n_blocks']}", |
| 256 | f"thread_k_blocks == {config['thread_k_blocks']}", |
| 257 | f"m_block_size_8 == {config['m_block_size_8']}", |
| 258 | f"stages == {config['stages']}", |
| 259 | f"group_blocks == {config['group_blocks']}", |
| 260 | f"is_zp_float == {config['is_zp_float']}", |
| 261 | ] |
| 262 | conditions = " && ".join(conditions) |
| 263 | |
| 264 | if kernel_selector_str == FILE_HEAD_COMMENT: |
| 265 | kernel_selector_str += f"if ({conditions})\n kernel = " |
| 266 | else: |
| 267 | kernel_selector_str += f"else if ({conditions})\n kernel = " |
| 268 | |
| 269 | kernel_template2 = ( |
| 270 | "Marlin<{{a_type_id}}, {{b_type_id}}, {{c_type_id}}, " |
| 271 | "{{s_type_id}}, {{threads}}, {{thread_m_blocks}}, " |
| 272 | "{{thread_n_blocks}}, {{thread_k_blocks}}, " |
| 273 | "{{m_block_size_8}}, {{stages}}, {{group_blocks}}, " |
| 274 | "{{is_zp_float}}>;" |
| 275 | ) |
| 276 | |
| 277 | kernel_selector_str += ( |
| 278 | jinja2.Template(kernel_template2).render( |
| 279 | a_type_id=f"vllm::{a_type}.id()", |
| 280 | b_type_id=f"vllm::{b_type}.id()", |
| 281 | c_type_id=f"vllm::{c_type}.id()", |
| 282 | s_type_id=f"vllm::{s_type}.id()", |
| 283 | **config, |
| 284 | ) |
| 285 | + "\n" |
| 286 | ) |
| 287 | |
| 288 | file_content = FILE_HEAD + "\n\n" |
| 289 | file_content += "\n\n".join(all_template_str_list) + "\n\n}\n" |
| 290 | if a_type == "kFE4M3fn": |
| 291 | filename = f"sm89_kernel_{a_type[1:]}_{b_type[1:]}_{c_type[1:]}.cu" |
| 292 | elif result_dict_tmp is sm_75_result_dict: |
| 293 | filename = f"sm75_kernel_{a_type[1:]}_{b_type[1:]}_{c_type[1:]}.cu" |
| 294 | else: |
| 295 | filename = f"sm80_kernel_{a_type[1:]}_{b_type[1:]}_{c_type[1:]}.cu" |
| 296 | |
| 297 | filename = filename.lower() |
| 298 | |
| 299 | with open(os.path.join(os.path.dirname(__file__), filename), "w") as f: |
| 300 | f.write(file_content) |
| 301 | |
| 302 | if not SUPPORT_FP8 and kernel_selector_str != FILE_HEAD_COMMENT: |
| 303 | kernel_selector_str += ( |
| 304 | "else if (a_type == vllm::kFE4M3fn)\n" |
| 305 | " STD_TORCH_CHECK(false, " |
| 306 | '"marlin kernel with fp8 activation is not built.");' |
| 307 | ) |
| 308 | |
| 309 | with open(os.path.join(os.path.dirname(__file__), "kernel_selector.h"), "w") as f: |
| 310 | f.write(kernel_selector_str) |
| 311 | |
| 312 | |
| 313 | if __name__ == "__main__": |
| 314 | remove_old_kernels() |
| 315 | generate_new_kernels() |