vllm.models.minimax_m3.nvidia.specialized_decode.kernels.post.host
¶
Host side of the post family's CuTe DSL kernels (post/kernel.py).
post_ops(world) and row_ops() return objects with the calls the stage wrappers
(allreduce_norm.py,
wide.py) use: post_route / kernel_info and mxfp8_rows. They check
shapes, dtypes and alignment
before every launch.
Each kernel variant is compiled once per process (cute.compile, TVM-FFI calling
convention), always outside
CUDA-graph capture: the post stage's build (kernel_info + its idle launch) compiles
its post_route variants,
both logits_out flavours, and the wide front's build compiles mxfp8_rows
(RowOps.precompile). A compile
requested during capture raises.
Classes:
-
PostOps–The allreduce_norm stage's kernels for one world size (4 or 8).
-
RowOps–The wide front's MXFP8 quantize of BF16 rows (
wide.py).
Functions:
-
post_ops–The post's ops for
world(imports the CuTe DSL kernel module: raises here if
PostOps
¶
The allreduce_norm stage's kernels for one world size (4 or 8).
Methods:
-
kernel_info–(static smem bytes, registers per thread, threads per CTA, CTAs per SM,
Source code in vllm/models/minimax_m3/nvidia/specialized_decode/kernels/post/host.py
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 | |
kernel_info(cl, rx2, mxfp8=False, logits=False)
¶
(static smem bytes, registers per thread, threads per CTA, CTAs per SM, cluster width, clusters the device holds at once) of the variant, for the host-side co-residency check.
Source code in vllm/models/minimax_m3/nvidia/specialized_decode/kernels/post/host.py
RowOps
¶
The wide front's MXFP8 quantize of BF16 rows (wide.py).
Methods:
-
precompile–Compile the kernel now (outside CUDA-graph capture).
Source code in vllm/models/minimax_m3/nvidia/specialized_decode/kernels/post/host.py
precompile()
¶
Compile the kernel now (outside CUDA-graph capture).
_compile(key, jit_fn, sample_args, constexprs)
¶
cute.compile jit_fn for key (cached). sample_args: the dynamic args' DSL
types in order.
Source code in vllm/models/minimax_m3/nvidia/specialized_decode/kernels/post/host.py
_cubin_info(fn, threads, cluster)
¶
(static smem, regs, CTAs/SM, max active clusters) of the single kernel in fn's cubin.
Source code in vllm/models/minimax_m3/nvidia/specialized_decode/kernels/post/host.py
_sample(spec)
¶
DSL sample values for a signature string of l (Int64), i (Int32), f (Float32).
Source code in vllm/models/minimax_m3/nvidia/specialized_decode/kernels/post/host.py
post_ops(world)
¶
The post's ops for world (imports the CuTe DSL kernel module: raises here if
it is unavailable).