vllm.models.qwen4_exp.nvidia.mega_moe
¶
Qwen4Exp routed experts on DeepGEMM MegaMoE.
Routed experts stored as MXFP8, MXFP4 or BF16 load straight into the kernel
layout. --quantization / --quantization-config can force another MegaMoE
format, and the experts are then requantized as they load. Block-FP8 experts,
which MegaMoE cannot run, are requantized to MXFP8 unless a format is forced.
The MTP drafter's config does not carry the online config, so its experts keep
their default format. The sigmoid-gated shared expert stays unfused and runs on
an aux stream next to the router.
Classes:
-
Qwen4ExpMegaMoEExperts–MegaMoE experts loaded from per-expert or fused checkpoint tensors.
-
Qwen4ExpMegaMoeBlock–Qwen3Next-style MoE whose routed experts run as one MegaMoE launch.
Functions:
-
dequantize_experts–Dequantize
[N, K]checkpoint expert weights to FP32. -
flush_mega_moe_requantize–Requantize the complete expert shards buffered by
load_weights. -
mega_moe_expert_formats–The checkpoint format of the routed experts and the format they run in.
-
requantize_experts–Requantize
[N, K]checkpoint expert weights to MXFP8 or MXFP4.
Qwen4ExpMegaMoEExperts
¶
Bases: DeepGemmMegaMoEExperts
MegaMoE experts loaded from per-expert or fused checkpoint tensors.
Experts stored in weight_format are written straight into the kernel
layout. Otherwise each EP rank buffers the complete shards of its own
experts and requantizes them per projection in a few batched calls.
Methods:
-
load_weights–Load
{expert}.{gate,up,down}_proj.{weight,<scale>}or the fused
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 | |
load_weights(weights)
¶
Load {expert}.{gate,up,down}_proj.{weight,<scale>} or the fused
gate_up_proj / down_proj BF16 tensors of every expert.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
Qwen4ExpMegaMoeBlock
¶
Bases: Module
Qwen3Next-style MoE whose routed experts run as one MegaMoE launch.
Takes this rank's tokens and returns their routed plus shared-expert output; no collective follows.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 | |
_FinishRequantize
¶
Bases: QuantizeMethodBase
Flushes the experts; fails on shards missing their weight or scale.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
_checkpoint_expert_format(quant_config, prefix)
¶
The format the checkpoint stores the routed experts at prefix in.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
dequantize_experts(weight, scale, checkpoint_format)
¶
Dequantize [N, K] checkpoint expert weights to FP32.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
flush_mega_moe_requantize(model)
¶
Requantize the complete expert shards buffered by load_weights.
Half pairs stay buffered: a chunked weight update may pass an expert's weight and its scale in different calls.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
mega_moe_expert_formats(vllm_config, prefix)
¶
The checkpoint format of the routed experts and the format they run in.
A format forced through the online quantization config wins. Otherwise the experts run as stored, and block-FP8 experts are requantized to MXFP8.
Source code in vllm/models/qwen4_exp/nvidia/mega_moe.py
requantize_experts(weight, scale, checkpoint_format, weight_format)
¶
Requantize [N, K] checkpoint expert weights to MXFP8 or MXFP4.
Rounds as DeepGEMM's per_token_cast_to_fp8 / per_token_cast_to_fp4
and returns the scales as [N, K // 32] UE8M0 exponents. Rows are
independent, so experts may be stacked along dim 0.