[MoE Refactor] Migrate flashinfer_cutedsl + DeepEP to MoeRunner (#25525)

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
This commit is contained in:
Cheng Wan
2026-05-17 14:48:17 -07:00
committed by GitHub
co-authored by Claude Opus 4.7
parent 89e501c5a8
commit 7158a255eb
5 changed files with 156 additions and 146 deletions
+2 -1
View File
@@ -780,7 +780,8 @@ class TestCuteDslV2(unittest.TestCase):
class TestCuteDslV1(unittest.TestCase):
"""Correctness tests for the CuteDSL v1 (deepep) path.
The v1 path (apply_without_routing_weights -> flashinfer_cutedsl_moe_masked)
The v1 path (flashinfer_cutedsl_moe_masked, dispatched via the
@register_fused_func("deepep", "flashinfer_cutedsl") MoeRunner entry)
is used when --moe-runner-backend flashinfer_cutedsl and --moe-a2a-backend
deepep are combined. It expects:
- W13 in default [Gate, Up] order (load_up_proj_weight_first = False)