[inductor][NVGEMM] Generate wider dense EFC candidates (#195539)

Generate vendored dense EFC variants for the validated SM100 N-tile widths through 256, including 160, 192, and 224. This makes the configurations available to nvMatmulHeuristics without retaining them unconditionally.

Keep the existing supplemented benchmark space unchanged. Wider candidates enter autotuning only when nvMatmulHeuristics selects them, avoiding additional profiling by default.

## Test Plan

```bash
PYTHONPATH=/data/users/mlazos/pytorch/agent_space/runtime_overlay PYTORCH_BOUNDARY_ROOT=$PWD python test/inductor/test_nv_universal_gemm.py TestNVUniversalGemm.test_vendored_dense_efc_exposes_wide_n_tiles -v
lintrunner -a
```

Authored with an AI assistant.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/195539
Approved by: https://github.com/karthickai
ghstack dependencies: #192298, #190822, #191514, #191515, #190825, #192928, #195538
This commit is contained in:
Michael Lazos
2026-09-01 20:39:45 -07:00
committed by PyTorch MergeBot
parent ef2e319c17
commit fd34cbee1e
2 changed files with 27 additions and 1 deletions
+24
View File
@@ -371,6 +371,30 @@ class TestNVUniversalGemm(TestCase):
self.assertFalse(design.use_2cta_mma)
self.assertEqual(design.tile_shape[0], 128)
@unittest.skipIf(IS_FBCODE, "CUTLASS Operator API is not available in fbcode")
def test_vendored_dense_efc_exposes_wide_n_tiles(self):
from cutlass import Float32
from torch._inductor.codegen.nv_universal_gemm.kernel_cache import (
_args_query_candidates,
)
from torch._inductor.codegen.nv_universal_gemm.nv_universal_gemm_kernel import (
_create_gemm_arguments,
)
m, n, k = 256, 512, 64
a = torch.empty((m, k), device="cuda", dtype=torch.bfloat16)
b = torch.empty((k, n), device="cuda", dtype=torch.bfloat16)
out = torch.empty((m, n), device="cuda", dtype=torch.bfloat16)
args = _create_gemm_arguments("GEMM", (a, b), out, Float32)
tile_ns = {
kernel.metadata.design.tile_shape[1]
for kernel in _args_query_candidates(args, 100, efc_only=True)
if "VendoredDenseGemmEFCOperator" in kernel.metadata.operator_name
}
self.assertTrue({64, 128, 160, 192, 224, 256}.issubset(tile_ns))
def test_efc_epilogue_lookup_no_deadlock(self):
"""get_efc_kernel_with_epilogue holds _cache_lock and, when the base EFC
kernel is not pre-resolved and misses the args cache, calls
@@ -24,6 +24,8 @@ from torch.utils._ordered_set import OrderedSet
log = logging.getLogger(__name__)
DENSE_EFC_TILE_NS = (64, 128, 160, 192, 224, 256)
@functools.cache
def _device_target(cc: int) -> Any:
@@ -242,7 +244,7 @@ def _replace_dense_efc_with_vendored(kernels: Any) -> list[Any]:
continue
design = kernel.metadata.design
tile_m, tile_n, tile_k = design.tile_shape
for vendored_tile_n in OrderedSet((tile_n, 64, 128)):
for vendored_tile_n in OrderedSet((tile_n, *DENSE_EFC_TILE_NS)):
if vendored_tile_n < tile_n:
continue
operator_name = kernel.metadata.operator_name.replace(