biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI] Initial tests for SM100 Blackwell runner (#21877)

Browse Source 
Signed-off-by: mgoin <mgoin64@gmail.com>
This commit is contained in:
Michael Goin
2025-08-01 19:18:38 -04:00
committed by GitHub
parent 881e1af43a
commit 88faa466d7
3 changed files with 30 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

15
tests/compile/test_fusion_all_reduce.py
View File

@@ -136,12 +136,15 @@ class TestAllReduceFusedAddRMSNormStaticQuantFP4Model(torch.nn.Module):
@multi_gpu_test(num_gpus=2)
@pytest.mark.parametrize("test_model", [
TestAllReduceRMSNormModel,
TestAllReduceFusedAddRMSNormModel,
TestAllReduceFusedAddRMSNormStaticQuantFP8Model,
TestAllReduceFusedAddRMSNormStaticQuantFP4Model,
])
@pytest.mark.parametrize(
"test_model",
[
TestAllReduceRMSNormModel,
TestAllReduceFusedAddRMSNormModel,
TestAllReduceFusedAddRMSNormStaticQuantFP8Model,
# TODO: Enable with torch==2.8.0
# TestAllReduceFusedAddRMSNormStaticQuantFP4Model,
])
@pytest.mark.parametrize("batch_size", [8])
@pytest.mark.parametrize("seq_len", [8])
@pytest.mark.parametrize("hidden_size", [16])

5
tests/kernels/quantization/test_cutlass_scaled_mm.py
View File

@@ -559,8 +559,6 @@ def test_cutlass_fp8_group_gemm(num_experts: int, per_act_token: bool,
m_a_scales = m_g if per_act_token else 1
n_b_scales = n_g if per_out_ch else 1
print("shape:", m_g, n_g, k_g)
# Create group-specific A and B (FP8) and output (FP16/FP32)
a_g = to_fp8(torch.randn((m_g, k_g), device=device))
b_g = to_fp8(torch.randn((n_g, k_g), device=device).t())
@@ -639,7 +637,4 @@ def test_cutlass_fp8_group_gemm(num_experts: int, per_act_token: bool,
for g in range(num_experts):
baseline = baseline_tensors[g]
c = out_tensors_stacked[expert_offsets[g]:expert_offsets[g + 1]]
print(baseline)
print(c)
print("*")
torch.testing.assert_close(c, baseline, rtol=1e-2, atol=5e-4)