[Perf][GLM-Image] Skip redundant hidden state D2H (#5346)

Signed-off-by: herotai214 <herotai214@gmail.com>

tests/model_executor/models/glm_image/test_glm_image_ar.py

@@ -91,6 +91,10 @@ GlmImageRotaryEmbedding = _ar_mod.GlmImageRotaryEmbedding
 pytestmark = [pytest.mark.core_model, pytest.mark.cpu]
 
 
+def test_glm_image_ar_omits_hidden_pooler_payload():
+    assert GlmImageForConditionalGeneration.omni_pooler_payload_include_hidden is False
+
+
 # =============================================================================
 # Helper: Minimal config for testing
 # =============================================================================

vllm_omni/model_executor/models/glm_image/glm_image_ar.py

@@ -2414,6 +2414,11 @@ class GlmImageForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP
     # linear increments during decode.
     precomputed_mrope_decode = True
 
+    # The downstream GLM-Image diffusion stage consumes generated token ids and
+    # optional VQ-VAE image ids, not AR hidden states. Avoid copying one hidden
+    # vector to host at every decode step.
+    omni_pooler_payload_include_hidden = False
+
     packed_modules_mapping = {
         "qkv_proj": [
             "q_proj",