[Bugfix] Correct behavior of GraniteMoeHybrid for TensorParallel execution (#20137)

Signed-off-by: Stanislaw Wozniak <stw@zurich.ibm.com>

tests/models/language/generation/test_granitemoehybrid.py

@@ -1,42 +0,0 @@
-# SPDX-License-Identifier: Apache-2.0
-# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
-
-import pytest
-
-from ...utils import check_logprobs_close
-
-# Path of the checkpoints
-MODELS = [
-    "ibm-granite/granite-4.0-tiny-preview",
-]
-
-
-@pytest.mark.skip(
-    reason="Granite 4.0 is not yet available in huggingface transformers")
-@pytest.mark.parametrize("model", MODELS)
-@pytest.mark.parametrize("dtype", ["float16", "bfloat16"])
-@pytest.mark.parametrize("max_tokens", [64])
-@pytest.mark.parametrize("num_logprobs", [5])
-def test_model_equivalence_to_hf_greedy(
-    hf_runner,
-    vllm_runner,
-    example_prompts,
-    model: str,
-    dtype: str,
-    max_tokens: int,
-    num_logprobs: int,
-):
-    with vllm_runner(model, dtype=dtype) as vllm_model:
-        vllm_outputs = vllm_model.generate_greedy_logprobs(
-            example_prompts, max_tokens, num_logprobs)
-
-    with hf_runner(model, dtype=dtype) as hf_model:
-        hf_outputs = hf_model.generate_greedy_logprobs_limit(
-            example_prompts, max_tokens, num_logprobs)
-
-    check_logprobs_close(
-        outputs_0_lst=hf_outputs,
-        outputs_1_lst=vllm_outputs,
-        name_0="hf",
-        name_1="vllm",
-    )

tests/models/language/generation/test_hybrid.py

@@ -28,8 +28,9 @@ SSM_MODELS = [
 
 HYBRID_MODELS = [
     "ai21labs/Jamba-tiny-dev",
-    # NOTE: ibm-granite/granite-4.0-tiny-preview are skipped currently as
-    # it is not yet available in huggingface transformers
+    # NOTE: Currently the test failes due to HF transformers issue fixed in:
+    # https://github.com/huggingface/transformers/pull/39033
+    # We will enable vLLM test for Granite after next HF transformers release.
     # "ibm-granite/granite-4.0-tiny-preview",
     # NOTE: Running Plamo2 in transformers implementation requires to install
     # causal-conv1d package, which is not listed as a test dependency as it's