[Feature] Add SM103 (Blackwell Ultra) Support to vLLM (#30484)

Signed-off-by: LopezCastroRoberto <robertol.c510@gmail.com>
Signed-off-by: Roberto L. Castro <38211239+LopezCastroRoberto@users.noreply.github.com>
Co-authored-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
Roberto L. Castro
2025-12-12 19:34:23 -08:00
committed by GitHub
co-authored by youkaichao
parent 57e9bf1864
commit 4fa7ce46f3
21 changed files with 53 additions and 33 deletions
@@ -32,8 +32,8 @@ def cal_diff(
CUTLASS_MLA_UNSUPPORTED_REASON = (
"Cutlass MLA Requires compute capability of 10 or above."
if not current_platform.is_device_capability(100)
"Cutlass MLA Requires compute capability of 100 or above."
if not current_platform.is_device_capability_family(100)
else "Cutlass MLA is supported"
)
@@ -11,7 +11,7 @@ from tests.kernels.quantization.nvfp4_utils import (
from vllm.platforms import current_platform
from vllm.utils.math_utils import round_up
if not current_platform.is_device_capability(100):
if not current_platform.is_device_capability_family(100):
pytest.skip(
"This TRTLLM kernel requires NVIDIA Blackwell.", allow_module_level=True
)
@@ -443,7 +443,7 @@ def test_flashinfer_trtllm_prefill_with_baseline(
output_trtllm = output_trtllm.reshape(-1, query.shape[1], query.shape[2])
if q_quant_dtype == FP8_DTYPE and o_quant_dtype == FP4_DTYPE:
rtol, atol = 1e-1, 2e-1
rtol, atol = 3e-1, 4e-1
elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == FP8_DTYPE:
rtol, atol = 4e-2, 6e-2
elif q_quant_dtype == FP8_DTYPE and o_quant_dtype == dtype:
+2 -2
View File
@@ -17,7 +17,7 @@ QUARK_MXFP4_AVAILABLE = find_spec("quark") is not None and version.parse(
) >= version.parse("0.8.99")
TRTLLM_GEN_MXFP4_AVAILABLE = (
current_platform.is_cuda() and current_platform.is_device_capability(100)
current_platform.is_cuda() and current_platform.is_device_capability_family(100)
)
HOPPER_MXFP4_BF16_AVAILABLE = (
@@ -799,7 +799,7 @@ def test_flashinfer_cutlass_mxfp4_fused_moe(
@pytest.mark.skipif(
not (
current_platform.is_cuda()
and current_platform.is_device_capability(100)
and current_platform.is_device_capability_family(100)
and has_flashinfer()
),
reason="NVIDIA GPU sm100 and flashinfer are required for this test",