[Bugfix] Prevent padding placeholders from reaching embeddings (#47029)

Signed-off-by: qianlihuang <91178480+qianlihuang@users.noreply.github.com>
Signed-off-by: Yiliu Dong <91178480+qianlihuang@users.noreply.github.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>

tests/v1/worker/test_gpu_input_batch.py

@@ -435,6 +435,40 @@ def test_pooling_prompt_lens_not_aliased(device: str):
     )
 
 
+def test_placeholder_spec_token_ids_written_verbatim():
+    input_batch = InputBatch(
+        max_num_reqs=1,
+        max_model_len=8,
+        max_num_batched_tokens=8,
+        device=torch.device("cpu"),
+        vocab_size=VOCAB_SIZE,
+        block_sizes=[16],
+        kernel_block_sizes=[16],
+    )
+    req = CachedRequestState(
+        req_id="req",
+        prompt_token_ids=[10, 11],
+        mm_features=[],
+        sampling_params=SamplingParams(),
+        block_ids=([],),
+        generator=None,
+        num_computed_tokens=3,
+        output_token_ids=[12],
+    )
+    input_batch.add_request(req)
+
+    input_batch.update_req_spec_token_ids(
+        req,
+        {"req": [13, -1, -1]},
+    )
+
+    # Placeholders (-1) are kept verbatim in both the spec_token_ids list and
+    # the token buffer; they are clamped to 0 only at the embedding boundary
+    # (GPUModelRunner._preprocess).
+    assert input_batch.spec_token_ids[0] == [13, -1, -1]
+    assert input_batch.token_ids_cpu[0, 3:6].tolist() == [13, -1, -1]
+
+
 @pytest.mark.parametrize(
     ("pooling_params", "expect_device_prompt_token_ids", "expect_cpu_prompt_token_ids"),
     [

tests/v1/worker/test_gpu_model_runner.py

@@ -862,6 +862,28 @@ def test_sample_passes_reordered_draft_probs_to_rejection_sampler():
     assert torch.equal(passed_draft_probs, expected_draft_probs)
 
 
+def test_invalid_draft_suffixes_remain_rejected_in_metadata():
+    runner = object.__new__(GPUModelRunner)
+    runner.device = torch.device("cpu")
+    runner.arange_np = np.arange(64, dtype=np.int64)
+    runner._arange_scratch = np.empty(64, dtype=np.int64)
+    # Placeholder (-1) drafts are kept in input_ids (clamped to 0 only at the
+    # embedding boundary). For num_draft_tokens=[2, 1, 2] the draft positions
+    # are [1, 2, 4, 6, 7], so the gather carries the -1s straight into the
+    # rejection-sampling metadata.
+    runner.input_ids = SimpleNamespace(
+        gpu=torch.tensor([99, 10, -1, 99, 12, 99, 13, -1], dtype=torch.int32),
+    )
+
+    metadata = GPUModelRunner._calc_spec_decode_metadata(
+        runner,
+        np.array([2, 1, 2], dtype=np.int32),
+        np.array([3, 5, 8], dtype=np.int32),
+    )
+
+    assert metadata.draft_token_ids.tolist() == [10, -1, 12, 13, -1]
+
+
 def test_init_kv_cache_with_kv_sharing_invalid_target_layer_order(default_vllm_config):
     torch.set_default_dtype(torch.float16)
     layer_0 = "model.layers.0.self_attn.attn"

vllm/v1/worker/gpu_model_runner.py

@@ -3440,6 +3440,10 @@ class GPUModelRunner(
         is_first_rank = get_pp_group().is_first_rank
         is_encoder_decoder = self.model_config.is_encoder_decoder
 
+        # Clamp speculative scheduler placeholders (-1) before embedding lookup.
+        if self.speculative_config is not None:
+            self.input_ids.gpu[:num_input_tokens].clamp_(min=0)
+
         # _prepare_inputs may reorder the batch, so we must gather multi
         # modal outputs after that to ensure the correct order
         ec_connector_output = None