Integration minicpm sala

tests/models/language/generation/test_minicpm_sala.py

@@ -0,0 +1,136 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+"""
+HF ↔ vLLM parity tests for MiniCPM-SALA model.
+
+Focused on token-level alignment between Hugging Face and vLLM implementations.
+"""
+
+import importlib
+
+import pytest
+
+from tests.models.registry import HF_EXAMPLE_MODELS
+from vllm.platforms import current_platform
+
+from ...utils import check_logprobs_close
+
+MODEL = "openbmb/MiniCPM-SALA"
+
+
+def get_attention_backend():
+    try:
+        return "TRITON_ATTN" if current_platform.is_rocm() else "auto"
+    except (AttributeError, TypeError):
+        return "auto"
+
+
+ATTN_BACKEND = get_attention_backend()
+MAX_NUM_SEQS = 4
+
+# Skip all tests if fla is not available (HF model requirement)
+try:
+    importlib.import_module("fla")
+except ImportError:
+    pytest.skip("fla package not installed - required for MiniCPM-SALA tests. Run: pip install flash-linear-attention", allow_module_level=True)
+
+
+@pytest.fixture(scope="module")
+def example_prompts():
+    return [
+        "Hello, my name is",
+        "The capital of France is",
+        "What is the meaning of life?",
+        "Explain quantum computing in simple terms",
+    ]
+
+
+def _check_model_availability():
+    try:
+        importlib.import_module("fla")
+    except ImportError:
+        pytest.skip("fla package not installed - required for HF MiniCPM-SALA model. Run: pip install flash-linear-attention")
+
+    try:
+        model_info = HF_EXAMPLE_MODELS.find_hf_info(MODEL)
+        model_info.check_available_online(on_fail="skip")
+        model_info.check_transformers_version(on_fail="skip")
+    except ValueError:
+        pytest.skip(f"Model {MODEL} not available or transformers version incompatible")
+
+
+def _skip_if_no_gpu():
+    try:
+        device_type = current_platform.device_type
+        if not device_type:
+            pytest.skip("Device type detection failed - likely no GPU available")
+    except (AttributeError, TypeError):
+        pytest.skip("Device type detection failed - likely no GPU available")
+
+
+@pytest.mark.parametrize("max_tokens", [32])
+@pytest.mark.parametrize("num_logprobs", [5])
+def test_hf_vllm_parity_greedy(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    """Test HF vs vLLM parity with greedy decoding and logprobs comparison."""
+    _check_model_availability()
+    _skip_if_no_gpu()
+
+    with hf_runner(MODEL) as hf_model:
+        hf_outputs = hf_model.generate_greedy_logprobs_limit(
+            example_prompts, max_tokens, num_logprobs
+        )
+
+    with vllm_runner(
+        MODEL, max_num_seqs=MAX_NUM_SEQS, attention_backend=ATTN_BACKEND
+    ) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            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",
+    )
+
+
+@pytest.mark.parametrize("max_tokens", [64])
+@pytest.mark.parametrize("num_logprobs", [10])
+def test_hf_vllm_parity_long_context(
+    hf_runner,
+    vllm_runner,
+    example_prompts,
+    max_tokens: int,
+    num_logprobs: int,
+) -> None:
+    """Test HF vs vLLM parity with longer context to test lightning attention state."""
+    _check_model_availability()
+    _skip_if_no_gpu()
+
+    long_prompts = [p * 10 for p in example_prompts[:2]]
+
+    with hf_runner(MODEL) as hf_model:
+        hf_outputs = hf_model.generate_greedy_logprobs_limit(
+            long_prompts, max_tokens, num_logprobs
+        )
+
+    with vllm_runner(
+        MODEL, max_num_seqs=MAX_NUM_SEQS, attention_backend=ATTN_BACKEND
+    ) as vllm_model:
+        vllm_outputs = vllm_model.generate_greedy_logprobs(
+            long_prompts, max_tokens, num_logprobs
+        )
+
+    check_logprobs_close(
+        outputs_0_lst=hf_outputs,
+        outputs_1_lst=vllm_outputs,
+        name_0="hf",
+        name_1="vllm",
+    )