feat: extended EPLB support for Mistral Large 3 and additional MoE backends (#48355)

Signed-off-by: jdebache <jdebache@nvidia.com>
This commit is contained in:
Julien Debache
2026-08-07 05:32:52 -07:00
committed by GitHub
parent 8d9b52f7c2
commit ae934ba8a5
16 changed files with 500 additions and 96 deletions
@@ -19,7 +19,12 @@ def test_shared_nvfp4_input_scales_have_writable_storage(monkeypatch):
monkeypatch.setattr(flashinfer_fp4_moe, "swizzle_blockscale", lambda x: x)
num_experts = 3
layer = SimpleNamespace(activation=SimpleNamespace(is_gated=False))
layer = SimpleNamespace(
activation=SimpleNamespace(is_gated=False),
moe_config=SimpleNamespace(
moe_parallel_config=SimpleNamespace(enable_eplb=False)
),
)
w13 = torch.zeros((num_experts, 2, 1), dtype=torch.uint8)
w2 = torch.zeros((num_experts, 2, 1), dtype=torch.uint8)
w13_scale = torch.zeros((num_experts, 2, 1), dtype=torch.float8_e4m3fn)