vllm.models.muse_spark.amd.muse_moe_runner
¶
ROCm MoE implementation for Muse Spark checkpoints.
Classes:
-
MuseMoERunner–MoE runner whose custom reducer owns the TP reduction.
-
MuseSparkAiterExperts–AITER experts that return one unweighted output per selected route.
-
MuseSparkQuarkOCPMXFusedMoE–Quark W4A4 method using route-preserving stock AITER experts.
-
MuseSparkQuarkW8A8Fp8FusedMoE–Quark W8A8 FP8 method using route-preserving Triton experts.
-
MuseSparkRoutedExperts–MuseSpark routed experts using ROCm BF16, FP8, or MXFP4 methods.
-
MuseSparkRouter–Muse Spark top-k router.
-
MuseSparkTopKWeightAndReduce–Apply MuseSpark's per-route norm before router weighting and reduction.
-
MuseSparkTritonExperts–Triton experts that expose unweighted GEMM2 output per route.
Functions:
-
token_sharded_moe_output–Make the TP finalize write only current rank's rows of a token-sharded stream.
-
validate_bf16_deployment–Reject deployments whose BF16 finalization is not implemented.
-
validate_fp8_deployment–Reject deployments whose FP8 finalization is not implemented.
-
validate_mxfp4_deployment–Reject deployments whose MXFP4 finalization is not implemented.
MuseMoERunner
¶
Bases: MoERunner
MoE runner whose custom reducer owns the TP reduction.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkAiterExperts
¶
Bases: _RouteOutputHandoff, AiterMxfp4Experts
AITER experts that return one unweighted output per selected route.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 | |
_unit_weights(num_routes, width, device)
¶
A cached [num_routes, width] fp32 ones buffer (grown on demand).
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkOnlineMXFP4FusedMoE
¶
Bases: MuseSparkQuarkOCPMXFusedMoE
Unquantized routed experts rounded to OCP MXFP4 (W4A4) after loading.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkQuarkOCPMXFusedMoE
¶
Bases: QuarkOCP_MX_MoEMethod
Quark W4A4 method using route-preserving stock AITER experts.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkQuarkW8A8Fp8FusedMoE
¶
Bases: QuarkW8A8Fp8MoEMethod
Quark W8A8 FP8 method using route-preserving Triton experts.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkRoutedExperts
¶
Bases: MuseSparkConsolidatedRoutedExperts
MuseSpark routed experts using ROCm BF16, FP8, or MXFP4 methods.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkRouter
¶
Bases: BaseRouter
Muse Spark top-k router.
Optional metap output multiplier -> sigmoid/softmax -> add expert bias for SELECTION only -> top-k -> gather UNBIASED scores -> divide by (sum + division_eps) -> scale by route_scale.
Attributes:
-
routing_method_type(RoutingMethodType) –Classify this router in vLLM's routing vocabulary.
Source code in vllm/models/muse_spark/common/moe.py
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 | |
routing_method_type
property
¶
Classify this router in vLLM's routing vocabulary.
_compute_routing_native(router_logits, indices_type)
¶
The recipe in torch: the kernel's reference, and the CPU path.
Source code in vllm/models/muse_spark/common/moe.py
MuseSparkTopKWeightAndReduce
¶
Bases: TopKWeightAndReduce
Apply MuseSpark's per-route norm before router weighting and reduction.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
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 | |
_token_sharded_finalize(rows, topk_weights, out, tp_size, gather=True)
¶
Reduce-scatter rows by token, finalize the current rank's tokens, and
all-gather them into out. Without gather, only current rank's rows
are written to out.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkTritonExperts
¶
Bases: _RouteOutputHandoff, TritonExperts
Triton experts that expose unweighted GEMM2 output per route.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
MuseSparkUnquantizedFusedMoE
¶
Bases: UnquantizedFusedMoEMethod
BF16 MoE method pinned to the MuseSpark-aware Triton experts.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
_RouteOutputHandoff
¶
Hand the experts' per-route buffer to the finalize instead of copying it.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
_online_mxfp4_applies(prefix)
¶
Whether this layer's unquantized experts are rounded to MXFP4 on load.
The first and last decoder layers stay BF16, as in the Quark MXFP4 export.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
_quantize_experts_to_mxfp4(weight, experts_per_chunk=32)
¶
Round [experts, rows, cols] weights to packed OCP MXFP4.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
token_sharded_moe_output()
¶
Make the TP finalize write only current rank's rows of a token-sharded stream.
Rank r keeps rows [r * R, (r + 1) * R) with R = ceil(T / tp), the
rows its reduce-scatter receives; the other rows of the output are left
unwritten.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
validate_bf16_deployment(moe_config)
¶
Reject deployments whose BF16 finalization is not implemented.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
validate_fp8_deployment(moe_config)
¶
Reject deployments whose FP8 finalization is not implemented.
Source code in vllm/models/muse_spark/amd/muse_moe_runner.py
validate_mxfp4_deployment(moe_config)
¶
Reject deployments whose MXFP4 finalization is not implemented.