Remove VLLM_TEST_FORCE_FP8_MARLIN to replace with linear_backend/moe_backend (#52182)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Misha Goin
2026-08-18 12:58:04 -04:00
committed by GitHub
co-authored by OpenAI Codex
parent 88b2bff2c6
commit bca7bea240
15 changed files with 46 additions and 66 deletions
+4
View File
@@ -371,6 +371,10 @@ def test_fusion_rmsnorm_quant(
use_aiter_fusion=False,
use_aiter_quant=False,
)
if any(
type(layer.kernel) is not force_kernel for layer in model.fp8_linear_layers
):
pytest.skip(f"{force_kernel.__name__} is not supported on this platform")
backend, _ = _run_fusion_test(
model, fusion_pass, vllm_config, dtype, hidden_size, num_tokens
@@ -481,6 +481,12 @@ def test_mla_attention_quant_pattern(
device=device,
vllm_config=vllm_config_unfused,
)
if (
model_class is TestMLAAttentionFp8GroupQuantPatternModel
and type(model_unfused.block_fp8_linear.kernel)
is not CutlassFp8BlockScaledMMKernel
):
pytest.skip("CUTLASS FP8 block kernel is not supported on this platform")
model_unfused = model_unfused.to(device)
# HACK: See #131044
result_unfused_0 = model_unfused(q, kv_c_normed, k_pe) # noqa: F841