[TRTLLM-12669][chore] trim verbose comments in populate_sampling_params

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

Author: zhaoyangwang-nvidia

tensorrt_llm/_torch/speculative/interface.py

@@ -581,9 +581,7 @@ def _normalize_request_sampling_params(
                 is_greedy,
             )
 
-        # ===== Phase 1: scan sampling configs, compute flags only =====
-        # Cache normalized params + num_tokens per request so phase 2 doesn't
-        # have to re-run _normalize_request_sampling_params.
+        # Phase 1: collect per-request flags and normalized values.
         per_request_normalized: list[tuple[float, int, float, int]] = []
         temperature_enabled = False
         top_k_enabled = False
@@ -625,9 +623,7 @@ def _normalize_request_sampling_params(
         self.skip_top_k = not top_k_enabled
         self.skip_top_p = not top_p_enabled
         self.has_greedy_requests = has_greedy_requests
-        # All-greedy iff no sampling param is active across the whole batch.
-        # When True, the sampler takes the argmax fast-path.
-        # This boolean is part of the CUDA graph key so we capture both variants.
+        # Used in the CUDA graph key to pick the argmax / advanced variant.
         self.is_all_greedy_sample = (self.skip_temperature and self.skip_top_k
                                      and self.skip_top_p)
 
@@ -637,9 +633,7 @@ def _normalize_request_sampling_params(
 
         if self.temperatures is None or self.temperatures.numel(
         ) < required_flat_size:
-            # The buffers are referenced by the captured CUDA graph (advanced
-            # variant), so they must exist with stable addresses even when
-            # the current batch is all-greedy and won't write into them.
+            # Allocate once; the captured graph reads from these stable addresses.
             self.temperatures = torch.ones(required_flat_size,
                                            dtype=torch.float32,
                                            device='cuda')
@@ -659,13 +653,11 @@ def _normalize_request_sampling_params(
                                              dtype=torch.float32,
                                              device='cuda')
 
-        # ===== Fast-path: all-greedy batch skips list build + H->D copies =====
-        # The sampler will take the argmax branch and never read these buffers,
-        # so leaving their values stale is safe.
+        # All-greedy: argmax branch won't read these buffers, skip H->D copies.
         if self.is_all_greedy_sample:
             return
 
-        # ===== Phase 2: build per-token / per-request lists and copy to GPU =====
+        # Phase 2: build per-token / per-request lists and copy to GPU.
         temperatures: list[float] = []
         top_ks: list[int] = []
         top_ps: list[float] = []