xinyun/vllm
1
0
Fork 0
mirror of https://git.datalinker.icu/vllm-project/vllm.git synced 2026-06-01 07:57:04 +08:00
Code Issues Packages Projects Releases Wiki Activity

[CI/Build][AMD] Skip quantization kernels tests that require CUTLASS or e4m3fn when not supported by platform (#30020)

Browse Source 
Signed-off-by: Randall Smith <ransmith@amd.com>
Co-authored-by: Randall Smith <ransmith@amd.com>
This commit is contained in:
rasmith 2025-12-09 20:28:37 -06:00 committed by GitHub
parent c3487aca34
commit b75f826fca
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
3 changed files with 20 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

17
tests/kernels/quantization/test_block_fp8.py
View File

@ -54,6 +54,10 @@ def setup_cuda():
torch.set_default_device("cuda") torch.set_default_device("cuda")
@pytest.mark.skipif(
current_platform.is_fp8_fnuz(),
reason="This platform supports e4m3fnuz, not e4m3fn.",
)
@pytest.mark.parametrize( @pytest.mark.parametrize(
"num_tokens,d,dtype,group_size,seed", "num_tokens,d,dtype,group_size,seed",
itertools.product(NUM_TOKENS, D, DTYPES, GROUP_SIZE, SEEDS), itertools.product(NUM_TOKENS, D, DTYPES, GROUP_SIZE, SEEDS),
@ -78,14 +82,14 @@ def test_per_token_group_quant_fp8(num_tokens, d, dtype, group_size, seed):
def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed): def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
torch.manual_seed(seed) torch.manual_seed(seed)
factor_for_scale = 1e-2 factor_for_scale = 1e-2
fp8_info = torch.finfo(torch.float8_e4m3fn) fp8_info = torch.finfo(current_platform.fp8_dtype())
fp8_max, fp8_min = fp8_info.max, fp8_info.min fp8_max, fp8_min = fp8_info.max, fp8_info.min
A_fp32 = (torch.rand(M, K, dtype=torch.float32) - 0.5) * 2 * fp8_max A_fp32 = (torch.rand(M, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
A_fp8 = A_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) A_fp8 = A_fp32.clamp(min=fp8_min, max=fp8_max).to(current_platform.fp8_dtype())
B_fp32 = (torch.rand(N, K, dtype=torch.float32) - 0.5) * 2 * fp8_max B_fp32 = (torch.rand(N, K, dtype=torch.float32) - 0.5) * 2 * fp8_max
B_fp8 = B_fp32.clamp(min=fp8_min, max=fp8_max).to(torch.float8_e4m3fn) B_fp8 = B_fp32.clamp(min=fp8_min, max=fp8_max).to(current_platform.fp8_dtype())
block_n, block_k = block_size[0], block_size[1] block_n, block_k = block_size[0], block_size[1]
n_tiles = (N + block_n - 1) // block_n n_tiles = (N + block_n - 1) // block_n
@ -103,6 +107,9 @@ def test_w8a8_block_fp8_matmul(M, N, K, block_size, out_dtype, seed):
assert rel_diff < 0.001 assert rel_diff < 0.001
@pytest.mark.skipif(
not current_platform.is_cuda(), reason="CUTLASS only supported on CUDA platform."
)
@torch.inference_mode() @torch.inference_mode()
def test_w8a8_block_fp8_cutlass_matmul(): def test_w8a8_block_fp8_cutlass_matmul():
# Test simple case where weight.shape % 128 != 0, # Test simple case where weight.shape % 128 != 0,
@ -151,6 +158,10 @@ def test_w8a8_block_fp8_cutlass_matmul():
assert rel_diff < 0.001 assert rel_diff < 0.001
@pytest.mark.skipif(
current_platform.is_fp8_fnuz(),
reason="This platform supports e4m3fnuz, not e4m3fn.",
)
@pytest.mark.parametrize( @pytest.mark.parametrize(
"M,N,K,block_size,out_dtype,seed", "M,N,K,block_size,out_dtype,seed",
itertools.product(M, N, K, BLOCK_SIZE, OUT_DTYPES, SEEDS), itertools.product(M, N, K, BLOCK_SIZE, OUT_DTYPES, SEEDS),

3
tests/kernels/quantization/test_cutlass_scaled_mm.py
View File

@ -15,6 +15,9 @@ from vllm import _custom_ops as ops
from vllm.platforms import current_platform from vllm.platforms import current_platform
from vllm.utils.math_utils import cdiv from vllm.utils.math_utils import cdiv
if not current_platform.is_cuda():
pytest.skip("These tests use CUTLASS which requires CUDA", allow_module_level=True)
MNK_FACTORS = [ MNK_FACTORS = [
(1, 256, 128), (1, 256, 128),
(1, 16384, 1024), (1, 16384, 1024),

3
tests/kernels/quantization/test_cutlass_w4a8.py
View File

@ -21,6 +21,9 @@ from vllm.model_executor.layers.quantization.utils.quant_utils import (
from vllm.platforms import current_platform from vllm.platforms import current_platform
from vllm.scalar_type import ScalarType, scalar_types from vllm.scalar_type import ScalarType, scalar_types
if not current_platform.is_cuda():
pytest.skip("These tests use CUTLASS which requires CUDA", allow_module_level=True)
# TODO: in future PR refactor this and `is_quant_method_supported` in the kernel # TODO: in future PR refactor this and `is_quant_method_supported` in the kernel
# unit tests to a common utility function. Currently the use of # unit tests to a common utility function. Currently the use of
# `is_quant_method_supported` conflates kernels with quantization methods # `is_quant_method_supported` conflates kernels with quantization methods