[Model] Add LFM2 architecture (#22845)

Signed-off-by: Paul Pak <paulpak58@gmail.com>

tests/models/language/generation/test_hybrid.py

@@ -31,6 +31,7 @@ HYBRID_MODELS = [
     "hmellor/tiny-random-BambaForCausalLM",
     "ibm-granite/granite-4.0-tiny-preview",
     "tiiuae/Falcon-H1-0.5B-Base",
+    "LiquidAI/LFM2-1.2B",
 ]
 
 HF_UNSUPPORTED_MODELS = [
@@ -52,6 +53,7 @@ V1_SUPPORTED_MODELS = [
     "hmellor/tiny-random-BambaForCausalLM",
     "ibm-granite/granite-4.0-tiny-preview",
     "tiiuae/Falcon-H1-0.5B-Base",
+    "LiquidAI/LFM2-1.2B",
 ]
 
 FULL_CUDA_GRAPH_MODELS = [
@@ -59,6 +61,10 @@ FULL_CUDA_GRAPH_MODELS = [
     "Zyphra/Zamba2-1.2B-instruct",
 ]
 
+V0_UNSUPPORTED_MODELS = [
+    "LiquidAI/LFM2-1.2B",
+]
+
 # Avoid OOM
 MAX_NUM_SEQS = 4
 
@@ -94,9 +100,12 @@ def test_models(
         else:
             hf_outputs = None
 
-    with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
-        vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
-            example_prompts, max_tokens, num_logprobs)
+    if model not in V0_UNSUPPORTED_MODELS:
+        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
+            vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
+                example_prompts, max_tokens, num_logprobs)
+    else:
+        vllm_v0_outputs = None
 
     if model in V1_SUPPORTED_MODELS:
         with monkeypatch.context() as m:
@@ -112,7 +121,7 @@ def test_models(
     else:
         vllm_v1_outputs = None
 
-    if hf_outputs is not None:
+    if hf_outputs is not None and vllm_v0_outputs is not None:
         check_logprobs_close(
             outputs_0_lst=hf_outputs,
             outputs_1_lst=vllm_v0_outputs,
@@ -122,6 +131,7 @@ def test_models(
 
     if model in V1_SUPPORTED_MODELS:
         ref_outputs = hf_outputs if hf_outputs is not None else vllm_v0_outputs
+        assert ref_outputs is not None
         check_logprobs_close(
             outputs_0_lst=ref_outputs,
             outputs_1_lst=vllm_v1_outputs,
@@ -140,6 +150,9 @@ def test_batching(
     max_tokens: int,
     num_logprobs: int,
 ) -> None:
+    if model in V0_UNSUPPORTED_MODELS:
+        pytest.skip(
+            f"Unsupported V0 Engine. Skipping `test_batching` on {model}.")
 
     try:
         model_info = HF_EXAMPLE_MODELS.find_hf_info(model)
@@ -392,9 +405,12 @@ def test_full_cuda_graph(
         else:
             hf_outputs = None
 
-    with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
-        vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
-            example_prompts, max_tokens, num_logprobs)
+    if model not in V0_UNSUPPORTED_MODELS:
+        with vllm_runner(model, max_num_seqs=MAX_NUM_SEQS) as vllm_model:
+            vllm_v0_outputs = vllm_model.generate_greedy_logprobs(
+                example_prompts, max_tokens, num_logprobs)
+    else:
+        vllm_v0_outputs = None
 
     with monkeypatch.context() as m:
         m.setenv("VLLM_USE_V1", "1")
@@ -408,7 +424,7 @@ def test_full_cuda_graph(
             vllm_v1_outputs = vllm_model.generate_greedy_logprobs(
                 example_prompts, max_tokens, num_logprobs)
 
-    if hf_outputs is not None:
+    if hf_outputs is not None and vllm_v0_outputs is not None:
         check_logprobs_close(
             outputs_0_lst=hf_outputs,
             outputs_1_lst=vllm_v0_outputs,
@@ -417,6 +433,7 @@ def test_full_cuda_graph(
         )
 
     ref_outputs = hf_outputs if hf_outputs is not None else vllm_v0_outputs
+    assert ref_outputs is not None
     check_logprobs_close(
         outputs_0_lst=ref_outputs,
         outputs_1_lst=vllm_v1_outputs,

tests/models/registry.py

@@ -230,6 +230,8 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
                                             "tiny": "ai21labs/Jamba-tiny-dev",
                                             "random": "ai21labs/Jamba-tiny-random",  # noqa: E501
                                         }),
+    "Lfm2ForCausalLM": _HfExamplesInfo("LiquidAI/LFM2-1.2B",
+                                       min_transformers_version="4.54"),
     "LlamaForCausalLM": _HfExamplesInfo("meta-llama/Llama-3.2-1B-Instruct",
                                         extras={"guard": "meta-llama/Llama-Guard-3-1B",  # noqa: E501
                                                 "hermes": "NousResearch/Hermes-3-Llama-3.1-8B", # noqa: E501

tests/models/test_initialization.py

@@ -95,6 +95,8 @@ def can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch,
 
 @pytest.mark.parametrize("model_arch", HF_EXAMPLE_MODELS.get_supported_archs())
 def test_can_initialize(model_arch: str, monkeypatch: pytest.MonkeyPatch):
+    if model_arch == "Lfm2ForCausalLM":
+        pytest.skip("Skipping until test supports V1-only models")
     can_initialize(model_arch, monkeypatch, HF_EXAMPLE_MODELS)