[feat] FlashInfer CuteDSL MegaMoE integration (#54049)

Signed-off-by: jdebache <jdebache@nvidia.com>
This commit is contained in:
Julien Debache
2026-10-02 15:43:33 -04:00
committed by GitHub
parent c4973d94fe
commit 111f71a6db
33 changed files with 1663 additions and 1036 deletions
+6 -1
View File
@@ -56,6 +56,7 @@ All2AllBackend = Literal[
"flashinfer_all2allv", # temporary alias for flashinfer_nvlink_two_sided
"flashinfer_nvlink_two_sided",
"flashinfer_nvlink_one_sided",
"passthrough",
]
@@ -211,7 +212,10 @@ class ParallelConfig:
- "moonep": MoonEP balanced EP with dynamic redundant experts (NVLink)
- "nixl_ep": Use nixl-ep kernels
- "flashinfer_nvlink_one_sided": Use flashinfer high-throughput a2a kernels
- "flashinfer_nvlink_two_sided": Use flashinfer two-sided kernels for mnnvl"""
- "flashinfer_nvlink_two_sided": Use flashinfer two-sided kernels for mnnvl
- "passthrough": No all2all at all: the MoE backend dispatches and combines
itself (e.g. the FlashInfer MoE-EP megakernels). Bound automatically for
such MoE backends; not selectable from the CLI."""
max_parallel_loading_workers: int | None = Field(default=None, ge=1)
"""Maximum number of parallel loading workers when loading model
@@ -742,6 +746,7 @@ class ParallelConfig:
"deepep_low_latency",
"deepep_v2",
"flashinfer_nvlink_one_sided",
"passthrough",
"mori_high_throughput",
"mori_low_latency",
"nixl_ep",