From 3dd910da4222f52b4def7bd67d670e2ace9ede1a Mon Sep 17 00:00:00 2001 From: Matthew Bonanni Date: Tue, 7 Jul 2026 16:11:34 -0400 Subject: [PATCH] [Bugfix] Allow non-contiguous query in FlashInfer FP8 query quantization (#47908) Signed-off-by: Matthew Bonanni Co-authored-by: Claude Opus 4.8 (1M context) --- vllm/v1/attention/backends/flashinfer.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/v1/attention/backends/flashinfer.py b/vllm/v1/attention/backends/flashinfer.py index 5e2295c0af..12eab21e3e 100755 --- a/vllm/v1/attention/backends/flashinfer.py +++ b/vllm/v1/attention/backends/flashinfer.py @@ -1584,11 +1584,11 @@ class FlashInferImpl(AttentionImpl): if query.dtype != q_data_type: assert query.dtype in [torch.float16, torch.bfloat16] assert q_data_type in [torch.float8_e4m3fn, torch.float8_e5m2] - assert query.is_contiguous() assert query.dim() == 3 num_tokens = query.shape[0] num_heads = query.shape[1] head_size = query.shape[2] + assert query.stride(2) == 1 and query.stride(1) == head_size query_quantized, _ = custom_ops.scaled_fp8_quant( query.view(num_tokens, num_heads * head_size), scale=scale )