[NVFP4] NVFP4 MOE emulation fallback for H100/MI300/MI350, standardize TritonExperts usage for OCP MX emulation (#35737)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Signed-off-by: fxmarty-amd <felmarty@amd.com>
Co-authored-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
fxmarty-amd
2026-04-22 08:58:54 -07:00
committed by GitHub
co-authored by Kyle Sayers
parent 5f76b3fb30
commit d622e27d2b
12 changed files with 601 additions and 121 deletions
+23
View File
@@ -120,3 +120,26 @@ def test_nvfp4(vllm_runner, model, eager, backend, monkeypatch):
with vllm_runner(model, enforce_eager=eager) as llm:
output = llm.generate_greedy(["1 2 3 4 5"], max_tokens=2)
assert output[0][1] == "1 2 3 4 5 6"
@pytest.mark.parametrize(
"model",
[
"nvidia/Qwen3-30B-A3B-NVFP4",
"RedHatAI/Qwen3-30B-A3B-NVFP4",
],
)
@pytest.mark.parametrize("backend", ["emulation"])
@pytest.mark.skipif(
not current_platform.is_rocm(),
reason="NVFP4 MOE emulation is only useful on AMD Instinct MI3xx",
)
def test_nvfp4_moe(vllm_runner, model, backend, monkeypatch):
monkeypatch.setenv("VLLM_NVFP4_GEMM_BACKEND", backend)
with vllm_runner(
model,
moe_backend=backend,
load_format="dummy",
hf_overrides={"num_hidden_layers": 2},
) as llm:
_ = llm.generate_greedy(["1 2 3 4 5"], max_tokens=2)