[Model][NVIDIA] Route DSA models to the CUDA non-compiled path (#52861)

This commit is contained in:
Woosuk Kwon
2026-08-19 00:13:52 -07:00
committed by GitHub
parent ee11730751
commit b09bd69b5b
16 changed files with 260 additions and 146 deletions
+2 -4
View File
@@ -18,7 +18,6 @@ from .common import (
from .models import (
FLASHINFER_ATTN,
FLASHINFER_MLA_ATTN,
FLASHMLA_SPARSE_ATTN,
ROCM_AITER_UNIFIED_ATTN,
ROCM_ATTN,
TRITON_ATTN,
@@ -26,7 +25,6 @@ from .models import (
deepseek_coder_v2_lite_fp8,
deepseek_r1_fp4,
deepseek_v3_fp8,
deepseek_v32_fp4,
llama3_8b_fp4,
llama3_8b_fp8,
llama4_scout_fp4,
@@ -149,11 +147,11 @@ def test_tp1_fp8_fusions(
@pytest.mark.parametrize(
"model_name, matches_fn, model_kwargs, hf_overrides",
[llama3_8b_fp4, llama4_scout_fp4, deepseek_r1_fp4, deepseek_v32_fp4],
[llama3_8b_fp4, llama4_scout_fp4, deepseek_r1_fp4],
)
@pytest.mark.parametrize(
"attn_backend",
[FLASHINFER_ATTN, FLASHINFER_MLA_ATTN, FLASHMLA_SPARSE_ATTN],
[FLASHINFER_ATTN, FLASHINFER_MLA_ATTN],
)
@pytest.mark.parametrize("n_layers", [6])
@pytest.mark.parametrize("custom_ops", custom_ops_combos("rms_norm"))