[Doc] Score centering via top-k processed logprobs (#59361)

Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

vx120 committed Oct 1, 2026 at 08:57 UTC a47e4a5568b4791db0c9633b5dc7531128f98bae
1 file changed +33
docs/training/sampling_mask.md
+33
@@ -104,6 +104,39 @@ log_prob = log_softmax(masked_logits)[sampled_token_id]
104 Both sides normalize over the same token set, so the importance ratio is
105 consistent.
106
107 +## Score centering with top-k logprobs
108 +
109 +[Score centering](https://arxiv.org/abs/2609.20807) needs the sampler's top-k
110 +token IDs and logprobs plus the sampled token's logprob; the paper uses
111 +`k=128` and does not need the exact sampling support. vLLM already returns
112 +this through the ordinary `logprobs` parameter:
113 +
114 +```bash
115 +vllm serve <model> --logprobs-mode processed_logprobs --max-logprobs 128
116 +```
117 +
118 +```json
119 +{"sampling_params": {"temperature": 1.0, "top_p": 0.9, "logprobs": 128}}
120 +```
121 +
122 +Each position then carries the sampled token's logprob and the top-128
123 +`(token_id, logprob)` pairs, normalized over the post-truncation support:
124 +
125 +```json
126 +{"token": "token_id:11", "logprob": -0.757,
127 + "top_logprobs": [{"token": "token_id:11", "logprob": -0.757},
128 + {"token": "token_id:13", "logprob": -0.632},
129 + {"token": "token_id:0", "logprob": -9999.0}, ...]}
130 +```
131 +
132 +- `--max-logprobs` defaults to 20; raise it to the requested `k`.
133 +- Tokens outside the support are returned as `-9999.0` with arbitrary IDs;
134 + drop them. The finite entries are the sampling mask when it has at most `k`
135 + tokens.
136 +- The sampled token is always the first entry and may repeat inside the top-k.
137 +- `/inference/v1/generate` encodes IDs as `token_id:<id>`; the OpenAI
138 + endpoints need `return_tokens_as_token_ids`.
139 +
140 ## Limitations
141
142 - **Engine-level flag:** `--return-sampling-mask` globally disables the