Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

FP8 GEMV long-indexing test: seeded inputs and a last-bit tolerance (#14220)

Daniel Han5f3a0a290e85
diff --git a/tests/test_fp8_rowwise_gemv_long_indexing.py b/tests/test_fp8_rowwise_gemv_long_indexing.py
index 441a26f4d41..4ff8836e89b 100644
--- a/tests/test_fp8_rowwise_gemv_long_indexing.py
+++ b/tests/test_fp8_rowwise_gemv_long_indexing.py
@@ -169,13 +169,17 @@ def test_weight_past_2_31_elements(F):
N = 2**31 // K + 256
if torch.cuda.mem_get_info()[0] < 4 * 1024**3:
pytest.skip("needs 4 GB free GPU memory")
- X = torch.randn(1, K, device = "cuda", dtype = torch.bfloat16)
+ g = torch.Generator(device = "cuda").manual_seed(0)
+ X = torch.randn(1, K, device = "cuda", dtype = torch.bfloat16, generator = g)
W = torch.empty(N, K, device = "cuda", dtype = torch.float8_e4m3fn)
tail = slice(N - 512, N)
- W[tail] = (torch.randn(512, K, device = "cuda") * 0.05).to(torch.float8_e4m3fn)
- S = torch.rand(N, 1, device = "cuda") * 0.01 + 0.001
+ W[tail] = (torch.randn(512, K, device = "cuda", generator = g) * 0.05).to(torch.float8_e4m3fn)
+ S = torch.rand(N, 1, device = "cuda", generator = g) * 0.01 + 0.001
assert F.can_use_fp8_rowwise_gemv(X, W, S)
out = F.fp8_rowwise_gemv(X, W, S)
ref = F.fp8_rowwise_gemv(X, W[tail].contiguous(), S[tail].contiguous())
torch.cuda.synchronize()
- assert torch.equal(out[..., tail], ref)
+ # Different BLOCK_N can flip the last bf16 bit; wrong addressing past 2^31 is far outside this.
+ torch.testing.assert_close(
+ out[..., tail].float(), ref.float(), rtol = 0, atol = 1e-2 * ref.float().abs().max().item()
+ )