[Bugfix] Allow non-contiguous query in FlashInfer FP8 query quantization (#47908)
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com> Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Matthew Bonanni committed
Jul 7, 2026 at 16:11 UTC
3dd910da4222f52b4def7bd67d670e2ace9ede1a
1 file changed
+1
-1
vllm/v1/attention/backends/flashinfer.py
+1
-1
@@ -1584,11 +1584,11 @@ class FlashInferImpl(AttentionImpl):
1584
if query.dtype != q_data_type:
1585
assert query.dtype in [torch.float16, torch.bfloat16]
1586
assert q_data_type in [torch.float8_e4m3fn, torch.float8_e5m2]
1587
- assert query.is_contiguous()
1587
assert query.dim() == 3
1588
num_tokens = query.shape[0]
1589
num_heads = query.shape[1]
1590
head_size = query.shape[2]
1591
+ assert query.stride(2) == 1 and query.stride(1) == head_size
1592
query_quantized, _ = custom_ops.scaled_fp8_quant(
1593
query.view(num_tokens, num_heads * head_size), scale=scale
1594
)