[Kernel][UX] Add --linear-backend arg for linear kernel selection (#39538)

Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2026-05-15 17:07:39 -07:00
committed by GitHub
parent d0921bafef
commit b2a27b82d9
4 changed files with 266 additions and 34 deletions
+8 -9
View File
@@ -90,9 +90,9 @@ def test_models(example_prompts, model_name) -> None:
EAGER = [True, False]
SM_100_NVFP4_BACKENDS = [
"flashinfer-cudnn",
"flashinfer-trtllm",
"flashinfer-cutlass",
"flashinfer_cudnn",
"flashinfer_trtllm",
"flashinfer_cutlass",
]
@@ -102,12 +102,12 @@ SM_100_NVFP4_BACKENDS = [
"backend",
[
"emulation",
"flashinfer-cudnn",
"flashinfer-trtllm", # the small seq_len ensures trtllm_8x4_layout backend is used
"flashinfer-cutlass",
"flashinfer_cudnn",
"flashinfer_trtllm", # the small seq_len ensures trtllm_8x4_layout backend is used
"flashinfer_cutlass",
],
)
def test_nvfp4(vllm_runner, model, eager, backend, monkeypatch):
def test_nvfp4(vllm_runner, model, eager, backend):
if (
not current_platform.has_device_capability(100)
and backend in SM_100_NVFP4_BACKENDS
@@ -116,8 +116,7 @@ def test_nvfp4(vllm_runner, model, eager, backend, monkeypatch):
f"The backend {backend} is not supported with current_platform.has_device_capability(100) == False"
)
monkeypatch.setenv("VLLM_NVFP4_GEMM_BACKEND", backend)
with vllm_runner(model, enforce_eager=eager) as llm:
with vllm_runner(model, enforce_eager=eager, linear_backend=backend) as llm:
output = llm.generate_greedy(["1 2 3 4 5"], max_tokens=2)
assert output[0][1] == "1 2 3 4 5 6"