[Misc] Fix Current vLLM config is not set. warnings, assert to avoid issues in the future (#31747)
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com> Signed-off-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com> Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
co-authored by
Luka Govedič
parent
5d3b6097ad
commit
6cdf015c3c
@@ -556,7 +556,7 @@ def _test_backend_correctness(
|
||||
@pytest.mark.parametrize("model", ["meta-llama/Meta-Llama-3-8B"])
|
||||
@pytest.mark.parametrize("tensor_parallel_size", [1, 2, 4])
|
||||
def test_causal_backend_correctness(
|
||||
batch_spec_name: str, model: str, tensor_parallel_size: int
|
||||
default_vllm_config, batch_spec_name: str, model: str, tensor_parallel_size: int
|
||||
):
|
||||
"""Test backend's correctness with causal attention."""
|
||||
|
||||
|
||||
@@ -79,7 +79,12 @@ from vllm.v1.attention.backends.short_conv_attn import ShortConvAttentionBackend
|
||||
],
|
||||
)
|
||||
def test_mamba_layers_get_attn_backend(
|
||||
dist_init, layer_class, init_kwargs, expected_backend, expected_mamba_type
|
||||
default_vllm_config,
|
||||
dist_init,
|
||||
layer_class,
|
||||
init_kwargs,
|
||||
expected_backend,
|
||||
expected_mamba_type,
|
||||
):
|
||||
"""Test that Mamba-like layers return the correct attention backend."""
|
||||
layer = layer_class(**init_kwargs)
|
||||
|
||||
@@ -394,7 +394,11 @@ def run_attention_backend(
|
||||
@pytest.mark.parametrize("model", ["deepseek-ai/DeepSeek-R1"])
|
||||
@pytest.mark.parametrize("tensor_parallel_size", [1, 4, 8, 16])
|
||||
def test_backend_correctness(
|
||||
dist_init, batch_spec_name: str, model: str, tensor_parallel_size: int
|
||||
default_vllm_config,
|
||||
dist_init,
|
||||
batch_spec_name: str,
|
||||
model: str,
|
||||
tensor_parallel_size: int,
|
||||
):
|
||||
"""
|
||||
Test that all backends produce similar outputs to a reference implementation
|
||||
|
||||
@@ -124,7 +124,12 @@ def _quantize_dequantize_fp8_ds_mla(
|
||||
reason="FlashMLASparseBackend requires CUDA 9.0 or higher",
|
||||
)
|
||||
def test_sparse_backend_decode_correctness(
|
||||
dist_init, batch_name, kv_cache_dtype, tensor_parallel_size, workspace_init
|
||||
default_vllm_config,
|
||||
dist_init,
|
||||
batch_name,
|
||||
kv_cache_dtype,
|
||||
tensor_parallel_size,
|
||||
workspace_init,
|
||||
):
|
||||
if current_platform.is_rocm():
|
||||
pytest.skip("ROCm does not support fp8_ds_mla data type for kv cache.")
|
||||
|
||||
Reference in New Issue
Block a user