Fix flaky test_self_e2e_pd_perturb (#27426)

python/sglang/test/kv_canary/pd_fixture.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import uuid
 from typing import ClassVar, Literal, Optional
 
 from sglang.srt.kv_canary.config import CanaryMode
@@ -61,10 +62,19 @@ class CanaryPDFixture(CanaryViolationAssertMixin, PDDisaggregationServerBase):
         assert_all_success: bool = True,
         max_new_tokens: int = 100,
         timeout: float = 60.0,
+        distinct_prompts: bool = False,
     ) -> list[dict]:
+        if distinct_prompts:
+            # Diverge at the very first token (request index before the per-call
+            # nonce) so requests share no radix-dedupable prefix beyond a
+            # tokenizer-added BOS, and retries never hit earlier attempts' cache.
+            nonce = uuid.uuid4().hex[:8]
+            prompts = [f"{i} {nonce} {_SHORT_PROMPT_BODY}" for i in range(n)]
+        else:
+            prompts = [_SHORT_PROMPT_BODY] * n
         results = post_parallel_generate(
             url=self.lb_url + "/generate",
-            prompts=[_SHORT_PROMPT_BODY] * n,
+            prompts=prompts,
             max_new_tokens=max_new_tokens,
             timeout=timeout,
         )

test/registered/kv_canary/test_self_e2e_pd_perturb.py

@@ -37,7 +37,14 @@ class _PDPerturbBase(CanaryPDFixture):
     ) -> None:
         # send_parallel_short_requests defaults to max_new_tokens=100 so D-side runs
         # decode forwards that exercise canary verify on the transferred prefix.
-        self.send_parallel_short_requests(n=4)
+        #
+        # distinct_prompts is required for the violation to surface reliably:
+        # with a shared prompt, P-side radix caching rewrites each request's
+        # req_to_token row to the first-inserted (canonical) copy's slots in
+        # cache_unfinished_req BEFORE send_kv_chunk snapshots the indices, so a
+        # flip on a deduped duplicate slot is freed untransferred and never
+        # re-verified on either side.
+        self.send_parallel_short_requests(n=4, distinct_prompts=True)
         # D-side: first decode forward re-verifies the transferred prefix slots,
         # so the flip MUST surface as real_kv_hash violation.
         self.assert_per_forward_violation_reported(