diff --git a/vllm/model_executor/models/diffusion_gemma.py b/vllm/model_executor/models/diffusion_gemma.py index bd1d14b6f6..af0cb35168 100644 --- a/vllm/model_executor/models/diffusion_gemma.py +++ b/vllm/model_executor/models/diffusion_gemma.py @@ -1312,7 +1312,6 @@ class DiffusionSampler: is_committing = states.is_encoder_phase[decode_slots].clone() slots_np = input_batch.idx_mapping_np[:num_reqs] - is_decode_np = per_req_nlogits_np > 0 max_num_logprobs = self.sampling_states.max_num_logprobs(slots_np) # Sample over the [num_decode * CL, vocab] logits. The fp32 pipeline in @@ -1396,13 +1395,16 @@ class DiffusionSampler: # stashed logprobs and attach to SamplerOutput. logprobs_tensors = None if max_num_logprobs >= 0 and is_committing.any() and self._pending_logprobs: + committing_slots = set( + decode_slots_np[is_committing.cpu().numpy()].tolist() + ) parts_ids, parts_lp, parts_ranks = [], [], [] cu_gen: list[int] = [] flat_offset = 0 for i in range(num_reqs): cu_gen.append(flat_offset) slot = int(slots_np[i]) - if is_decode_np[i] and slot in self._pending_logprobs: + if slot in committing_slots and slot in self._pending_logprobs: lp = self._pending_logprobs.pop(slot) parts_ids.append(lp.logprob_token_ids) parts_lp.append(lp.logprobs)