minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

fix: use a bf16-relative tolerance in the DSA indexer K kernel test (#35795)

Browse Source 
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Shuwen Wang
2026-08-25 19:50:29 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 07a9de25b4
commit 303227951a
1 changed files with 12 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
test/registered/kernels/ops/attention/test_dsv32_indexer_fusion.py
+12 -5
View File
@@ -42,6 +42,12 @@ PAGE_SIZE = 64
BYTES_PER_TOKEN = HEAD_DIM + 4 # 128 fp8 + 4-byte fp32 scale BYTES_PER_TOKEN = HEAD_DIM + 4 # 128 fp8 + 4-byte fp32 scale
EPS = 1e-6 EPS = 1e-6
MAX_POS = 8192 MAX_POS = 8192
# The K kernel accumulates in fp32 and rounds once to bf16, so it may differ
# from an fp32 reference by half a bf16 ulp -- a *relative* bound (bf16 keeps 8
# significand bits), not an absolute one. The randn cos/sin here are unbounded,
# so outputs reach magnitudes where half an ulp is well over 0.05.
BF16_RTOL = 2**-7
BF16_ATOL = 1e-3
def _skip_if_unavailable(): def _skip_if_unavailable():
@@ -78,10 +84,11 @@ def test_k_norm_rope_matches_reference():
_skip_if_unavailable() _skip_if_unavailable()
dev = "cuda" dev = "cuda"
B = 37 B = 37
g = torch.Generator(device=dev).manual_seed(1)
cos, sin, cos_sin_cache, positions = _make_inputs(B) cos, sin, cos_sin_cache, positions = _make_inputs(B)
key = torch.randn(B, HEAD_DIM, dtype=torch.bfloat16, device=dev) key = torch.randn(B, HEAD_DIM, dtype=torch.bfloat16, device=dev, generator=g)
weight = torch.randn(HEAD_DIM, dtype=torch.float32, device=dev) weight = torch.randn(HEAD_DIM, dtype=torch.float32, device=dev, generator=g)
bias = torch.randn(HEAD_DIM, dtype=torch.float32, device=dev) bias = torch.randn(HEAD_DIM, dtype=torch.float32, device=dev, generator=g)
out = fused_k_indexer_norm_rope(key, weight, bias, EPS, cos_sin_cache, positions) out = fused_k_indexer_norm_rope(key, weight, bias, EPS, cos_sin_cache, positions)
torch.cuda.synchronize() torch.cuda.synchronize()
@@ -92,7 +99,7 @@ def test_k_norm_rope_matches_reference():
cp, sp = cos[positions.long()], sin[positions.long()] cp, sp = cos[positions.long()], sin[positions.long()]
ref = _rope_first(normed, cp, sp) ref = _rope_first(normed, cp, sp)
torch.testing.assert_close(out.float(), ref, atol=0.06, rtol=0.0) torch.testing.assert_close(out.float(), ref, atol=BF16_ATOL, rtol=BF16_RTOL)
# ---------------------------------------------------------------------------- # ----------------------------------------------------------------------------
@@ -240,7 +247,7 @@ def test_indexer_uses_replaced_rope_cache_for_fused_kernels():
key.float(), (HEAD_DIM,), weight=k_weight, bias=k_bias, eps=EPS key.float(), (HEAD_DIM,), weight=k_weight, bias=k_bias, eps=EPS
) )
k_ref = _rope_first(normed, cos[positions.long()], sin[positions.long()]) k_ref = _rope_first(normed, cos[positions.long()], sin[positions.long()])
torch.testing.assert_close(k_out.float(), k_ref, atol=0.06, rtol=0.0) torch.testing.assert_close(k_out.float(), k_ref, atol=BF16_ATOL, rtol=BF16_RTOL)
q = torch.randn(B, n_heads, HEAD_DIM, dtype=torch.bfloat16, device=dev, generator=g) q = torch.randn(B, n_heads, HEAD_DIM, dtype=torch.bfloat16, device=dev, generator=g)
q_weight = torch.randn(B, n_heads, dtype=torch.bfloat16, device=dev, generator=g) q_weight = torch.randn(B, n_heads, dtype=torch.bfloat16, device=dev, generator=g)