Two tests fail on a CUDA host since #736 / #697: test_nvfp4_experts_restricted_to_triton_for_swigluoai (sm_120) and test_graph_capture_reuses_warm_offload_cache_before_capture (any Linux host with libcuda) #765

What happened

On an RTX 5080 (sm_120), on main cc56b6b:

FAILED tests/models/test_minimax_m3.py::test_nvfp4_experts_restricted_to_triton_for_swigluoai
    assert 'w4a4' == 'triton'
FAILED tests/moe/test_offload.py::test_graph_capture_reuses_warm_offload_cache_before_capture
    AttributeError: 'FakeGraph' object has no attribute 'raw_cuda_graph_exec'

The first fails only on sm_120. The second fails on any Linux host where libcuda.so.1 loads, even with CUDA_VISIBLE_DEVICES= (#739's description also lists it among the failures already on main on an RTX PRO 6000).

Minimal reproduction (CPU only, PYTHONPATH=python):

CUDA_VISIBLE_DEVICES= python -m pytest -q -p no:cacheprovider \
  "tests/moe/test_offload.py::test_graph_capture_reuses_warm_offload_cache_before_capture"

>       rc = upload(graph.raw_cuda_graph_exec(), stream.cuda_stream)
E       AttributeError: 'FakeGraph' object has no attribute 'raw_cuda_graph_exec'
python/freetoken/engine/graph.py:126: AttributeError
1 failed

The minimax test passes on a CPU host. The script below makes the same kernel selection as the test, then lets W4a4Nvfp4MoEKernel.worth_it() see what an sm_120 GPU reports:

import torch
from freetoken.kernel import backend, nvfp4_moe
from freetoken.layers.quantization import ModelOptConfig, MoEConfig, select_kernel
from freetoken.layers.quantization.moe import Nvfp4MoEMethod
backend.device_capability = lambda: (12, 0)  # what the test monkeypatches
backend.is_vllm_installed = backend.is_flashinfer_installed = lambda: True
cfg = MoEConfig(num_experts=256, hidden=6144, intermediate=3072, top_k=4, scheme=ModelOptConfig.SCHEMES["NVFP4"],
                strategy="offload", activation="swigluoai", alpha=1.702, limit=7.0)
print("CPU host:             ", select_kernel(Nvfp4MoEMethod.candidates, "auto", cfg).name)
torch.cuda.is_available = lambda: True  # what worth_it() sees on a real sm_120 GPU
torch.cuda.current_device = lambda: 0
nvfp4_moe.w4a4_available = lambda device_index: True
print("sm_120 GPU (worth_it):", select_kernel(Nvfp4MoEMethod.candidates, "auto", cfg).name)

CPU host:              triton
sm_120 GPU (worth_it): w4a4

Cause (current main, cc56b6b)

• tests/models/test_minimax_m3.py:273-285: the test monkeypatches backend.device_capability to (12, 0) and expects auto to pick triton. perf(moe): prefill nvfp4 W4A4 moe kernel on sm_120 #736 put W4a4Nvfp4MoEKernel first in Nvfp4MoEMethod.candidates (layers/quantization/moe/nvfp4.py:623). Its worth_it() (nvfp4.py:99-102) does not use backend.device_capability. It asks the real device through torch.cuda.is_available() and nvfp4_moe.w4a4_available(torch.cuda.current_device()). On a CPU host that is False, so the test passes. On sm_120 it is True, and w4a4 is a Triton subclass that accepts swigluoai, so auto picks w4a4. The test's intent ("marlin / b12x hard-code silu") still holds. Only the expected name is stale.
• tests/moe/test_offload.py:467-469: FakeGraph has only pool(). fix(engine): upload cuda graphs at capture so a stream wait cannot block the first launch #697 added _upload_graph after each capture (engine/graph.py:118-128, called at :222), and it calls graph.raw_cuda_graph_exec() whenever _cu_graph_upload() (graph.py:108-115) is not None, which is any Linux, non-HIP host with libcuda.so.1.

Proposed fix

• minimax test: assert that auto picks neither marlin nor b12x (in ("w4a4", "triton")). Or pin the device side as well, with monkeypatch.setattr(W4a4Nvfp4MoEKernel, "worth_it", lambda self, cfg: False), so that the test does not depend on the host GPU.
• offload test: monkeypatch.setattr("freetoken.engine.graph._upload_graph", lambda graph, stream: None), or give FakeGraph a raw_cuda_graph_exec() and stub _cu_graph_upload.

Environment

Seen on an RTX 5080 16 GB (sm_120, driver 615.78.08), torch 2.14.1+cu130; reproduction CPU-only on main cc56b6b.