fix spec dec request level metrics (#13754)

Co-authored-by: Liangsheng Yin <lsyincs@gmail.com>

python/sglang/srt/managers/tokenizer_manager.py

@@ -1837,17 +1837,18 @@ class TokenizerManager(TokenizerCommunicatorMixin):
         meta_info["spec_accept_length"] = 0
         meta_info["spec_verify_ct"] = recv_obj.spec_verify_ct[i]
 
+        # The draft tokens per speculative step (excluding the target-sampled token).
+        num_guess_tokens = self.server_args.speculative_num_draft_tokens - 1
+
         if (
             recv_obj.spec_verify_ct[i] > 0
-            and self.server_args.speculative_num_steps is not None
+            and num_guess_tokens is not None
             and not isinstance(recv_obj, BatchEmbeddingOutput)
             and hasattr(recv_obj, "spec_accepted_tokens")
             # Checks that `spec_accepted_tokens[i]` will exist.
             and len(recv_obj.spec_accepted_tokens) > i
         ):
-            total_draft_tokens = (
+            total_draft_tokens = recv_obj.spec_verify_ct[i] * num_guess_tokens
-                recv_obj.spec_verify_ct[i] * self.server_args.speculative_num_steps
-            )
             accepted_tokens = recv_obj.spec_accepted_tokens[i]
 
             # Calculate per-request acceptance rate and average acceptance length.

python/sglang/srt/speculative/ngram_info.py

@@ -185,6 +185,10 @@ class NgramVerifyInput(SpecInput):
                             )
                             raise e
             req.spec_verify_ct += 1
+            req.spec_accepted_tokens += (
+                sum(1 for idx in accept_index_row if idx != -1) - 1
+            )
+
         if has_finished:
             self.accept_length = (self.accept_index != -1).sum(dim=1) - 1
         self.accept_index = self.accept_index[self.accept_index != -1]

python/sglang/srt/speculative/ngram_worker.py

@@ -295,6 +295,7 @@ class NGRAMWorker:
         self._prepare_for_speculative_decoding(batch)
         model_worker_batch = batch.get_model_worker_batch()
         num_accepted_tokens = 0
+        accept_lens = None
 
         if model_worker_batch.forward_mode.is_target_verify():
             batch_result = self.target_worker.forward_batch_generation(
@@ -308,6 +309,8 @@ class NGRAMWorker:
             logits_output, next_token_ids, num_accepted_tokens = verify_input.verify(
                 batch, logits_output, self.page_size
             )
+            # Store accept_lens for per-request metrics
+            accept_lens = verify_input.accept_length
             if batch.return_logprob:
                 self.add_logprob_values(batch, verify_input, logits_output)
             self._update_ngram_cache(batch)
@@ -328,4 +331,5 @@ class NGRAMWorker:
             next_token_ids=next_token_ids,
             num_accepted_tokens=num_accepted_tokens,
             can_run_cuda_graph=can_run_cuda_graph,
+            accept_lens=accept_lens,
         )