[Kernel] Raise verbose error and consolidate num_heads/num_kv_heads divisibility check (#19339)

Signed-off-by: 22quinn <33176974+22quinn@users.noreply.github.com>
2025-06-14 22:43:48 -07:00
parent ee1531bc38
commit 0b73736a0d
17 changed files with 24 additions and 19 deletions
--- a/vllm/v1/attention/backends/flashinfer.py
+++ b/vllm/v1/attention/backends/flashinfer.py
@@ -532,7 +532,6 @@ class FlashInferImpl(AttentionImpl):
        self.logits_soft_cap = logits_soft_cap
        self.kv_sharing_target_layer_name = kv_sharing_target_layer_name

-        assert self.num_heads % self.num_kv_heads == 0
        self.num_queries_per_kv = self.num_heads // self.num_kv_heads

        if attn_type != AttentionType.DECODER: