Fix GLM-OCR MTP multimodal embeddings and positions (#39088)

Co-authored-by: Xinyuan Tong <115166877+JustinTong0323@users.noreply.github.com>
Co-authored-by: Xinyuan Tong <xinyuantong.cs@gmail.com>

test/registered/unit/spec/test_eagle_draft_cuda_graph_runner.py

@@ -91,7 +91,7 @@ class TestEagleDraftCudaGraphRunner(CustomTestCase):
         runner.require_mlp_tp_gather = False
         runner.require_gathered_buffer = False
         runner.model_runner = SimpleNamespace(
-            model_config=SimpleNamespace(vocab_size=8),
+            model_config=SimpleNamespace(vocab_size=8, model_is_mrope=False),
             server_args=SimpleNamespace(speculative_use_rejection_sampling=False),
             device_timer=None,
             draft_attn_backend=backend,

test/registered/unit/spec/test_multimodal_draft_embeddings.py

@@ -0,0 +1,59 @@
+import sys
+from types import SimpleNamespace
+from unittest.mock import patch
+
+import pytest
+import torch
+
+from sglang.srt.managers.mm_utils import general_mm_embed_routine
+from sglang.srt.model_executor.forward_batch_info import ForwardMode
+from sglang.srt.runtime_context import get_context
+from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
+from sglang.test.ci.ci_register import register_cpu_ci
+
+register_cpu_ci(est_time=2, suite="base-a-test-cpu")
+
+
+class _MutatingLanguageModel:
+    def get_input_embeddings(self):
+        return lambda input_ids: torch.zeros((input_ids.shape[0], 2))
+
+    def __call__(self, *, input_ids, forward_batch, input_embeds):
+        input_embeds.add_(100)
+        return input_embeds
+
+
+@pytest.mark.parametrize("algorithm", ["EAGLE", "EAGLE3", "NEXTN", "FROZEN_KV_MTP"])
+def test_eagle_mm_embeddings_survive_in_place_target_updates(algorithm):
+    source_embeds = torch.arange(8, dtype=torch.float32).reshape(4, 2)
+    expected_embeds = source_embeds.clone()
+    forward_batch = SimpleNamespace(
+        forward_mode=ForwardMode.EXTEND,
+        contains_mm_inputs=lambda: True,
+        mm_inputs=[SimpleNamespace()],
+        extend_prefix_lens_cpu=[0],
+        extend_seq_lens_cpu=[4],
+        input_embeds=None,
+        spec_algorithm=SpeculativeAlgorithm.from_string(algorithm),
+    )
+
+    with (
+        get_context().override_server_args(),
+        patch(
+            "sglang.srt.managers.mm_utils.embed_mm_inputs",
+            return_value=(source_embeds, {}),
+        ),
+    ):
+        hidden_states = general_mm_embed_routine(
+            input_ids=torch.arange(4),
+            forward_batch=forward_batch,
+            language_model=_MutatingLanguageModel(),
+        )
+
+    torch.testing.assert_close(hidden_states, expected_embeds + 100)
+    torch.testing.assert_close(forward_batch.mm_input_embeds, expected_embeds)
+    assert forward_batch.mm_input_embeds.data_ptr() != hidden_states.data_ptr()
+
+
+if __name__ == "__main__":
+    sys.exit(pytest.main([__file__]))