[Cherry-Pick][Optimization][Speculative Decoding]opt mtp logprob (#7883) (#7884)

Sunny-bot1 · EmmonsCurse · web-flow · commit c52b06310675 · 2026-05-26T17:35:40.000+08:00
* opt mtp logprob

* fix

* fix test and log

* fix bits

* Adapt logprobs baseline update in test_ernie_21b_mtp_multistep.py

---------

Co-authored-by: YuBaoku &lt;49938469+EmmonsCurse@users.noreply.github.com&gt;
diff --git a/custom_ops/gpu_ops/speculate_decoding/draft_model/mtp_save_first_token_with_topk.cc b/custom_ops/gpu_ops/speculate_decoding/draft_model/mtp_save_first_token_with_topk.cc
@@ -119,9 +119,11 @@ void MTPSaveFirstTokenWithTopK(const paddle::Tensor& sampled_token_ids,
   msg_sed.mtype = 1;
   msg_sed.meta[0] = not_need_stop.data<bool>()[0] ? inference_msg_id_from_env
                                                   : -inference_msg_id_from_env;
-  msg_sed.meta[1] = message_flag;
-  msg_sed.meta[2] = bsz;
+  // Pack message_flag (low 8 bits) and max_num_logprobs (high 24 bits) into
+  // meta[1]. Receiver unpacks both to avoid reading unused topk slots.
   int max_num_logprobs = logprob_token_ids.shape()[1];
+  msg_sed.meta[1] = message_flag | (max_num_logprobs << 8);
+  msg_sed.meta[2] = bsz;
   for (int i = 0; i < bsz; i++) {
     int cur_token_num;
     if (seq_lens_decoder_data[i] < prompt_lens_data[i] ||
@@ -139,29 +141,24 @@ void MTPSaveFirstTokenWithTopK(const paddle::Tensor& sampled_token_ids,
     auto* cur_batch_msg_sed = &msg_sed.mtext[i];
     int token_offset = cu_batch_token_offset_data[i];
     for (int j = 0; j < cur_token_num; j++) {
+      // Use SPEC_LOGPROB_K+1 as stride (fixed struct layout), but only write
+      // max_num_logprobs columns to avoid filling unused topk slots.
       auto* cur_tokens = &cur_batch_msg_sed->tokens[j * (SPEC_LOGPROB_K + 1)];
       auto* cur_scores = &cur_batch_msg_sed->scores[j * (SPEC_LOGPROB_K + 1)];
       if (j == 0) {
         // first token has full logprobs
-        for (int k = 0; k < SPEC_LOGPROB_K + 1; k++) {
+        for (int k = 0; k < max_num_logprobs; k++) {
           if (k == 0) {
             cur_tokens[k] =
                 (int)sampled_token_ids_data[i * max_draft_tokens + j];
             cur_scores[k] =
-                logprob_scores_data[(token_offset + j) * (SPEC_LOGPROB_K + 1) +
-                                    k];
-          } else if (k < max_num_logprobs) {
-            // only for first token
-            cur_tokens[k] =
-                (int)logprob_token_ids_data[(token_offset + j) *
-                                                (SPEC_LOGPROB_K + 1) +
-                                            k];
-            cur_scores[k] =
-                logprob_scores_data[(token_offset + j) * (SPEC_LOGPROB_K + 1) +
-                                    k];
+                logprob_scores_data[(token_offset + j) * max_num_logprobs + k];
           } else {
-            cur_tokens[k] = -1;
-            cur_scores[k] = 0.0;
+            cur_tokens[k] = (int)
+                logprob_token_ids_data[(token_offset + j) * max_num_logprobs +
+                                       k];
+            cur_scores[k] =
+                logprob_scores_data[(token_offset + j) * max_num_logprobs + k];
           }
         }
         cur_batch_msg_sed->ranks[j] = (int)logprob_ranks_data[token_offset + j];
@@ -174,7 +171,8 @@ void MTPSaveFirstTokenWithTopK(const paddle::Tensor& sampled_token_ids,
 #ifdef SPECULATE_SAVE_WITH_OUTPUT_DEBUG
   std::cout << "msg data: " << std::endl;
   std::cout << "stop_flag: " << msg_sed.meta[0]
-            << ", message_flag: " << msg_sed.meta[1]
+            << ", message_flag: " << (msg_sed.meta[1] & 0xFF)
+            << ", max_num_logprobs: " << (msg_sed.meta[1] >> 8)
             << ", bsz: " << msg_sed.meta[2] << std::endl;
   for (int i = 0; i < bsz; i++) {
     int cur_token_num = msg_sed.meta[3 + i];
diff --git a/custom_ops/gpu_ops/speculate_decoding/speculate_get_output_with_topk.cc b/custom_ops/gpu_ops/speculate_decoding/speculate_get_output_with_topk.cc
@@ -75,8 +75,11 @@ void SpeculateGetOutMmsgTopK(const paddle::Tensor& output_tokens,
 
   int bsz = msg_rcv.meta[2];
   output_tokens_data[0] = (int64_t)msg_rcv.meta[0];
+  // Unpack message_flag (low 8 bits) and actual_topk (high 24 bits) from
+  // meta[1]. Keep packed value; Python unpacks message_flag and actual_topk.
   output_tokens_data[1] = (int64_t)msg_rcv.meta[1];
   output_tokens_data[2] = (int64_t)msg_rcv.meta[2];
+  int actual_topk = msg_rcv.meta[1] >> 8;
 
   int output_tokens_offset = 3 + SPEC_LOGPROB_MAX_BSZ;
   for (int i = 0; i < bsz; i++) {
@@ -89,7 +92,7 @@ void SpeculateGetOutMmsgTopK(const paddle::Tensor& output_tokens,
         output_scores_data + i * (MAX_DRAFT_TOKEN_NUM * (SPEC_LOGPROB_K + 1));
     auto* cur_batch_msg_rcv = &msg_rcv.mtext[i];
     for (int j = 0; j < cur_token_num; j++) {
-      for (int k = 0; k < real_k + 1; k++) {
+      for (int k = 0; k < actual_topk; k++) {
         cur_output_token[j * (SPEC_LOGPROB_K + 1) + k] =
             (int64_t)cur_batch_msg_rcv->tokens[j * (SPEC_LOGPROB_K + 1) + k];
         cur_output_score[j * (SPEC_LOGPROB_K + 1) + k] =
@@ -102,7 +105,8 @@ void SpeculateGetOutMmsgTopK(const paddle::Tensor& output_tokens,
 #ifdef SPECULATE_GET_WITH_OUTPUT_DEBUG
   std::cout << "msg data: " << std::endl;
   std::cout << "stop_flag: " << output_tokens_data[0]
-            << ", message_flag: " << output_tokens_data[1]
+            << ", message_flag: " << (output_tokens_data[1] & 0xFF)
+            << ", max_num_logprobs: " << (output_tokens_data[1] >> 8)
             << ", bsz: " << output_tokens_data[2] << std::endl;
   for (int i = 0; i < output_tokens_data[2]; i++) {
     int cur_token_num = output_tokens_data[3 + i];
diff --git a/custom_ops/gpu_ops/speculate_decoding/speculate_save_output_with_topk.cc b/custom_ops/gpu_ops/speculate_decoding/speculate_save_output_with_topk.cc
@@ -121,9 +121,11 @@ void SpeculateSaveOutMmsgTopK(const paddle::Tensor& sampled_token_ids,
   msg_sed.mtype = 1;
   msg_sed.meta[0] = not_need_stop.data<bool>()[0] ? inference_msg_id_from_env
                                                   : -inference_msg_id_from_env;
-  msg_sed.meta[1] = message_flag;
-  msg_sed.meta[2] = bsz;
+  // Pack message_flag (low 8 bits) and max_num_logprobs (high 24 bits) into
+  // meta[1]. Receiver unpacks both to avoid reading unused topk slots.
   int max_num_logprobs = logprob_token_ids.shape()[1];
+  msg_sed.meta[1] = message_flag | (max_num_logprobs << 8);
+  msg_sed.meta[2] = bsz;
   for (int i = 0; i < bsz; i++) {
     int cur_token_num;
     if (seq_lens_decoder_data[i] < prompt_lens_data[i]) {
@@ -139,24 +141,20 @@ void SpeculateSaveOutMmsgTopK(const paddle::Tensor& sampled_token_ids,
     auto* cur_batch_msg_sed = &msg_sed.mtext[i];
     int token_offset = cu_batch_token_offset_data[i];
     for (int j = 0; j < cur_token_num; j++) {
+      // Use SPEC_LOGPROB_K+1 as stride (fixed struct layout), but only write
+      // max_num_logprobs columns to avoid filling unused topk slots.
       auto* cur_tokens = &cur_batch_msg_sed->tokens[j * (SPEC_LOGPROB_K + 1)];
       auto* cur_scores = &cur_batch_msg_sed->scores[j * (SPEC_LOGPROB_K + 1)];
-      for (int k = 0; k < SPEC_LOGPROB_K + 1; k++) {
+      for (int k = 0; k < max_num_logprobs; k++) {
         if (k == 0) {
           cur_tokens[k] = (int)sampled_token_ids_data[i * max_draft_tokens + j];
           cur_scores[k] =
-              logprob_scores_data[(token_offset + j) * (SPEC_LOGPROB_K + 1) +
-                                  k];
-        } else if (k < max_num_logprobs) {
+              logprob_scores_data[(token_offset + j) * max_num_logprobs + k];
+        } else {
           cur_tokens[k] = (int)
-              logprob_token_ids_data[(token_offset + j) * (SPEC_LOGPROB_K + 1) +
-                                     k];
+              logprob_token_ids_data[(token_offset + j) * max_num_logprobs + k];
           cur_scores[k] =
-              logprob_scores_data[(token_offset + j) * (SPEC_LOGPROB_K + 1) +
-                                  k];
-        } else {
-          cur_tokens[k] = -1;
-          cur_scores[k] = 0.0;
+              logprob_scores_data[(token_offset + j) * max_num_logprobs + k];
         }
       }
       cur_batch_msg_sed->ranks[j] = (int)logprob_ranks_data[token_offset + j];
@@ -165,7 +163,8 @@ void SpeculateSaveOutMmsgTopK(const paddle::Tensor& sampled_token_ids,
 #ifdef SPECULATE_SAVE_WITH_OUTPUT_DEBUG
   std::cout << "msg data: " << std::endl;
   std::cout << "stop_flag: " << msg_sed.meta[0]
-            << ", message_flag: " << msg_sed.meta[1]
+            << ", message_flag: " << (msg_sed.meta[1] & 0xFF)
+            << ", max_num_logprobs: " << (msg_sed.meta[1] >> 8)
             << ", bsz: " << msg_sed.meta[2] << std::endl;
   for (int i = 0; i < bsz; i++) {
     int cur_token_num = msg_sed.meta[3 + i];
diff --git a/fastdeploy/output/token_processor.py b/fastdeploy/output/token_processor.py
@@ -796,12 +796,15 @@ def _process_batch_draft_tokens(self, mtype, batch, accept_num, tokens, scores,
                 metrics=None,
             )
 
-            token_ids = tokens[i][:, 0].tolist()[: accept_num[i]]
+            tokens_i = tokens[i].tolist()
+            scores_i = scores[i].tolist()
+            ranks_i = ranks[i].tolist()
+            token_ids = [row[0] for row in tokens_i[: accept_num[i]]]
             for batch_token_index in range(len(token_ids)):
-                result.outputs.logprob = float(scores[i, batch_token_index, 0])
-                topk_token_ids = tokens[i, batch_token_index, :].tolist()
-                topk_logprobs = scores[i, batch_token_index, :].tolist()
-                sampled_rank = ranks[i, batch_token_index].item()
+                result.outputs.logprob = scores_i[batch_token_index][0]
+                topk_token_ids = tokens_i[batch_token_index]
+                topk_logprobs = scores_i[batch_token_index]
+                sampled_rank = ranks_i[batch_token_index]
 
                 if result.outputs.draft_top_logprobs is None:
                     result.outputs.draft_top_logprobs = LogprobsLists(
@@ -828,16 +831,19 @@ def _process_batch_output(self):
         mtype = 3
         if self.cfg.speculative_config.method:
             if self.use_logprobs:
-                mtype = int(self.output_tokens[1, 0].item())
+                # meta[1] packs message_flag (low 8 bits) and actual_topk (high 24 bits).
+                packed_meta1 = int(self.output_tokens[1, 0].item())
+                mtype = packed_meta1 & 0xFF
+                actual_topk = packed_meta1 >> 8
                 batch = self.output_tokens[2, 0]
                 accept_num = [int(num[0]) for num in self.output_tokens[3 : batch + 3]]
                 tokens = tokens[3 + MAX_BSZ : 3 + MAX_BSZ + batch * MAX_DRAFT_TOKENS * (K + 1)].reshape(
                     [batch, MAX_DRAFT_TOKENS, K + 1]
-                )
+                )[:, :, :actual_topk]
                 scores = (
                     self.output_scores[: batch * MAX_DRAFT_TOKENS * (K + 1)]
                     .numpy()
-                    .reshape([batch, MAX_DRAFT_TOKENS, K + 1])
+                    .reshape([batch, MAX_DRAFT_TOKENS, K + 1])[:, :, :actual_topk]
                 )
                 ranks = self.output_ranks[: batch * MAX_DRAFT_TOKENS].numpy().reshape([batch, MAX_DRAFT_TOKENS])
 
@@ -846,6 +852,10 @@ def _process_batch_output(self):
                     batch_result = self._process_batch_draft_tokens(mtype, batch, accept_num, tokens, scores, ranks)
                     self.postprocess(batch_result, mtype)
                     return
+                # Pre-convert full arrays to Python lists once for MTP target token path.
+                tokens_lists = tokens.tolist()
+                scores_lists = scores.tolist()
+                ranks_list = ranks.tolist()
             else:
                 batch = self.output_tokens[1]
                 accept_num = tokens[2 : batch + 2]
@@ -914,7 +924,7 @@ def _process_batch_output(self):
                         llm_logger.info(f"recovery stop signal found at task {task_id}")
                     token_ids = [RECOVERY_STOP_SIGNAL]
                 elif self.use_logprobs:
-                    token_ids = tokens[i][:, 0].tolist()[: accept_num[i]]
+                    token_ids = [row[0] for row in tokens_lists[i][: accept_num[i]]]
                 else:
                     token_ids = tokens[
                         2
@@ -1033,10 +1043,10 @@ def _process_batch_output(self):
                     task.output_token_ids.append(token_id)
                     if self.use_logprobs:
                         if self.cfg.speculative_config.method:
-                            result.outputs.logprob = float(scores[i, batch_token_index, 0])
-                            topk_token_ids = tokens[i, batch_token_index, :].tolist()
-                            topk_logprobs = scores[i, batch_token_index, :].tolist()
-                            sampled_rank = ranks[i, batch_token_index].item()
+                            result.outputs.logprob = scores_lists[i][batch_token_index][0]
+                            topk_token_ids = tokens_lists[i][batch_token_index]
+                            topk_logprobs = scores_lists[i][batch_token_index]
+                            sampled_rank = ranks_list[i][batch_token_index]
                         else:
                             # Use pre-converted lists (batch .tolist() done before the loop).
                             result.outputs.logprob = scores_lists[i][0]
diff --git a/fastdeploy/worker/gpu_model_runner.py b/fastdeploy/worker/gpu_model_runner.py
@@ -1226,15 +1226,11 @@ def _prepare_inputs(self, cached_token_num=-1, cached_real_bsz=-1, is_dummy_or_p
             req.sampling_params.top_p_normalized_logprobs and req.sampling_params.top_p != 1.0 for req in logprobs_reqs
         )
         if logprobs_reqs:
-            self.max_logprobs = (
-                max(
-                    [
-                        self.ori_vocab_size if req.sampling_params.logprobs < 0 else req.sampling_params.logprobs
-                        for req in logprobs_reqs
-                    ]
-                )
-                if not self.speculative_decoding
-                else 20
+            self.max_logprobs = max(
+                [
+                    self.ori_vocab_size if req.sampling_params.logprobs < 0 else req.sampling_params.logprobs
+                    for req in logprobs_reqs
+                ]
             )
         elif self.enable_logprob:
             self.max_logprobs = None if not self.speculative_decoding else 0
diff --git a/tests/e2e/test_ernie_21b_mtp_multistep.py b/tests/e2e/test_ernie_21b_mtp_multistep.py
@@ -212,11 +212,11 @@ def test_prefix_cache_text(api_url):
     if os.getenv("BASELINE") == "1":
         baseline_manager.save("base_21b_step3", result)
         baseline_manager.save("base_21b_mtp_metrics_step3", speculate_metrics_2)
-        baseline_manager.save("base_21b_logprobs_step3", logprobs_2)
+        baseline_manager.save("base_21b_logprobs_step3_new", logprobs_2)
 
     baseline_result = baseline_manager.load("base_21b_step3")
     baseline_mtp_metrics = baseline_manager.load("base_21b_mtp_metrics_step3")
-    baseline_logprobs = baseline_manager.load("base_21b_logprobs_step3")
+    baseline_logprobs = baseline_manager.load("base_21b_logprobs_step3_new")
 
     assert logprobs == logprobs_2, (
         "logprobs 前后不一致\n"
diff --git a/tests/output/test_process_batch_output.py b/tests/output/test_process_batch_output.py
@@ -211,8 +211,9 @@ def test_speculative_decoding_use_logprobs(self):
 
         # stop_flag
         processor.output_tokens[0, 0].set_tensor(paddle.to_tensor(2))
-        # mtype target = 3, decode = 4
-        processor.output_tokens[1, 0].set_tensor(paddle.to_tensor(3))
+        # meta[1] packs mtype (low 8 bits) and actual_topk (high 16 bits)
+        actual_topk = K + 1
+        processor.output_tokens[1, 0].set_tensor(paddle.to_tensor(3 | (actual_topk << 8)))
         # batch
         processor.output_tokens[2, 0].set_tensor(paddle.to_tensor(2))
         # accept_num
@@ -244,12 +245,12 @@ def test_speculative_decoding_use_logprobs(self):
             assert len(request_output.outputs.token_ids) == accept_num[i]
             assert len(request_output.outputs.top_logprobs) == 3
             # tokens, scores, ranks
-            assert len(request_output.outputs.top_logprobs[0][0]) == K + 1
-            assert len(request_output.outputs.top_logprobs[1][0]) == K + 1
+            assert len(request_output.outputs.top_logprobs[0][0]) == actual_topk
+            assert len(request_output.outputs.top_logprobs[1][0]) == actual_topk
             assert len(request_output.outputs.top_logprobs[2]) == accept_num[i]
 
         # mtype = 4
-        processor.output_tokens[1, 0].set_tensor(paddle.to_tensor(4))
+        processor.output_tokens[1, 0].set_tensor(paddle.to_tensor(4 | (actual_topk << 8)))
         processor._process_batch_output()
         cached_generated_tokens: MockCachedGeneratedTokens = processor.cached_generated_tokens
         for c in cached_generated_tokens.cache:
@@ -258,8 +259,8 @@ def test_speculative_decoding_use_logprobs(self):
             assert len(request_output.outputs.top_logprobs) == 3
             assert len(request_output.outputs.draft_top_logprobs) == 3
             # tokens, scores, ranks
-            assert len(request_output.outputs.draft_top_logprobs[0][0]) == K + 1
-            assert len(request_output.outputs.draft_top_logprobs[1][0]) == K + 1
+            assert len(request_output.outputs.draft_top_logprobs[0][0]) == actual_topk
+            assert len(request_output.outputs.draft_top_logprobs[1][0]) == actual_topk
             assert len(request_output.outputs.draft_top_logprobs[2]) == accept_num[i]
 
     def test_process_batch_output_aborted_task_negative_token_speculative_decoding(self):
@@ -282,8 +283,8 @@ def test_process_batch_output_aborted_task_negative_token_speculative_decoding(s
         # Set up output tokens with negative token
         # stop_flag
         processor.output_tokens[0, 0].set_tensor(paddle.to_tensor(2))
-        # mtype target = 3
-        processor.output_tokens[1, 0].set_tensor(paddle.to_tensor(3))
+        # mtype target = 3, actual_topk packed in high bits
+        processor.output_tokens[1, 0].set_tensor(paddle.to_tensor(3 | ((K + 1) << 8)))
         # batch = 2 (so batch_id=0 is < batch_size-1=1)
         processor.output_tokens[2, 0].set_tensor(paddle.to_tensor(2))
         # Set accept_num = PREEMPTED_TOKEN_ID (-9) for first task to trigger abort logic
diff --git a/tests/output/test_token_processor.py b/tests/output/test_token_processor.py
@@ -749,7 +749,7 @@ def test_process_batch_output_speculative_logprob_handles_draft_batch():
     )
     processor._batch_result_buffer = [target]
     processor.cached_generated_tokens = mock.Mock()
-    processor.output_tokens[1, 0] = 4
+    processor.output_tokens[1, 0] = 4 | ((K + 1) << 8)
     processor.output_tokens[2, 0] = 1
     processor.output_tokens[3, 0] = 1
 
@@ -926,7 +926,7 @@ def test_process_batch_output_speculative_logprob_targets_topk_scores():
     task.trace_carrier = None
     rm.tasks_list[0] = task
     rm.req_dict[task.request_id] = task
-    processor.output_tokens[1, 0] = 3
+    processor.output_tokens[1, 0] = 3 | ((K + 1) << 8)
     processor.output_tokens[2, 0] = 1
     processor.output_tokens[3, 0] = 2
     token_block = np.arange(MAX_DRAFT_TOKENS * (K + 1), dtype=np.int64).reshape([-1, 1]) + 3
