[TRTLLM-12669][fix] disable fast-path when use_rejection_sampling=True

The rejection sampling path always reads spec_metadata.temperatures,
regardless of greedy/non-greedy. Skipping phase 2 there would leave
stale buffer values and break greedy semantics in the rejection
acceptance step.

Signed-off-by: ZhaoyangWang <zhaoyangw@nvidia.com>

Author: zhaoyangwang-nvidia

tensorrt_llm/_torch/speculative/interface.py

@@ -654,7 +654,8 @@ def _normalize_request_sampling_params(
                                              device='cuda')
 
         # All-greedy: argmax branch won't read these buffers, skip H->D copies.
-        if self.is_all_greedy_sample:
+        # Rejection sampling always reads them, so it cannot use this fast-path.
+        if self.is_all_greedy_sample and not self.use_rejection_sampling:
             return
 
         # Phase 2: build per-token / per-request lists and copy to GPU.