biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 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
3 changed files with 20 additions and 3 deletions
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")
@pytest.mark.skipif(
current_platform.is_fp8_fnuz(),
reason="This platform supports e4m3fnuz, not e4m3fn.",
)
@pytest.mark.parametrize(
"num_tokens,d,dtype,group_size,seed",
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):
torch.manual_seed(seed)
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
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_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]
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
@pytest.mark.skipif(
not current_platform.is_cuda(), reason="CUTLASS only supported on CUDA platform."
)
@torch.inference_mode()
def test_w8a8_block_fp8_cutlass_matmul():
# Test simple case where weight.shape % 128 != 0,
@@ -151,6 +158,10 @@ def test_w8a8_block_fp8_cutlass_matmul():
assert rel_diff < 0.001
@pytest.mark.skipif(
current_platform.is_fp8_fnuz(),
reason="This platform supports e4m3fnuz, not e4m3fn.",
)
@pytest.mark.parametrize(
"M,N,K,block_size,out_dtype,seed",
itertools.product(M, N, K, BLOCK_SIZE, OUT_DTYPES, SEEDS),