Address DSpark fused sampler review nits

Document that the Python seed counter must not be frozen by any future CUDA graph capture around sampling.

Avoid an unnecessary clone on the top-k/top-p fused sampler path when dtype conversion already produced a float32 copy, while preserving a clone for existing float32 inputs.

Author: alexbi29

vllm/v1/spec_decode/dspark_sampling.py

@@ -30,6 +30,8 @@
 # the same step. Determinism of the *draft* noise across runs is harmless: output
 # randomness still comes from the target's synchronized rejection sampler, and
 # the counter keeps advancing across requests so proposals are not repeated.
+# If sampling is ever CUDA-graph-captured, keep this seed outside replay or pass
+# it as graph input; otherwise the replayed graph would reuse one draft seed.
 _FUSED_MARKOV_SEED_COUNTER = 0x9E3779B1
 
 
@@ -335,7 +337,9 @@ def sample_dspark_markov_block_fused(
         sample_logits = step_logits
         sample_inv_temp = inv_temp
         if uses_top_k_top_p:
-            sample_logits = step_logits.to(torch.float32).clone()
+            sample_logits = step_logits.to(torch.float32)
+            if sample_logits.dtype == step_logits.dtype:
+                sample_logits = sample_logits.clone()
             sample_logits.mul_(inv_temp.view(-1, 1))
             sample_logits = apply_top_k_top_p(sample_logits, top_k, top_p)
             assert unit_inv_temp is not None