diff --git a/vllm/v1/worker/gpu/spec_decode/rejection_sampler_utils.py b/vllm/v1/worker/gpu/spec_decode/rejection_sampler_utils.py index 9d64c1ecbc..cf804e6a6b 100644 --- a/vllm/v1/worker/gpu/spec_decode/rejection_sampler_utils.py +++ b/vllm/v1/worker/gpu/spec_decode/rejection_sampler_utils.py @@ -806,7 +806,10 @@ def _resample_kernel( other=float("-inf"), ).to(tl.float32) + # Compute the residual logits to resample the rejected token from. if is_bonus or not is_valid_rejected_draft: + # Bonus token (no rejections) or -1 placeholder token. In either case, + # directly use the target logits. residual_logits = target_logits elif HAS_DRAFT_LOGITS: # draft_logits is stored pre-temperature, so apply scale first. @@ -836,7 +839,12 @@ def _resample_kernel( ) target_log_probs += log_p_tau draft_log_probs = draft_logits - draft_lse - # Compute log(max(p - q, 0)) without subtracting probabilities. + # Compute the residual: + # r(x) = max(p(x) - q(x), 0) + # Gumbel sampling needs logits, so we compute it in log space: + # log(r(x)) = log(max(exp(log_p(x)) - exp(log_q(x)), 0)) + # The more numerically stable form is: + # log(max(exp(a) - exp(b), 0)) = a + log(max(1 - exp(b - a), 0)) ratio = tl.exp(draft_log_probs - target_log_probs) residual_logits = tl.where( ratio < 1.0, @@ -844,13 +852,20 @@ def _resample_kernel( float("-inf"), ).to(tl.float32) else: - # The block-verification factor is constant and cancels on normalization. + # One-hot draft. The residual is just the target distribution with + # the rejected draft token probability zeroed out. + # NOTE: During block verification, the residual becomes: + # 0 if x == rejected_draft_token + # p_tau * M_b(x) / Z otherwise + # Therefore p_tau is a constant that cancels under normalization, + # and does not need to be applied. residual_logits = tl.where( block != rejected_draft_token, target_logits, float("-inf"), ).to(tl.float32) + # Resample the rejected/bonus token. if WATERMARK: # Padded and greedy rows retain the stock draw. is_watermarked = (