[Spec] Avoid tensor scalar reads in spec decode allocation (#35377)

2026-08-26 14:51:31 -07:00
parent 45c85c198b
commit 06694071c6
1 changed files with 3 additions and 2 deletions
@@ -706,5 +706,6 @@ def alloc_for_spec_decode(
            len(reqs),
        )

-    for i, req in enumerate(reqs):
-        req.kv.kv_allocated_len = max(req.kv.kv_allocated_len, int(nxt_kv_lens_cpu[i]))
+    nxt_kv_lens_list = nxt_kv_lens_cpu.tolist()
+    for req, nxt_kv_len in zip(reqs, nxt_kv_lens_list, strict=True):
+        req.kv.kv_allocated_len = max(req.kv.kv_allocated_len, nxt_kv_len)