| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from contextlib import ExitStack |
| 4 | |
| 5 | import pytest |
| 6 | |
| 7 | from tests.utils import create_new_process_for_each_test |
| 8 | from tests.v1.attention.utils import full_cg_backend_configs as backend_configs |
| 9 | from vllm import LLM |
| 10 | from vllm.config import CompilationConfig, CompilationMode |
| 11 | from vllm.platforms import current_platform |
| 12 | |
| 13 | # test attention backend and cudagraph_mode combo |
| 14 | # (backend_name, cudagraph_mode, supported) |
| 15 | if current_platform.is_rocm(): |
| 16 | combo_cases_1 = [ |
| 17 | ("RocmAttn", "FULL", True), |
| 18 | ("RocmAttn", "FULL_AND_PIECEWISE", True), |
| 19 | ("TritonAttn", "FULL", True), |
| 20 | ("TritonAttn", "FULL_AND_PIECEWISE", True), |
| 21 | ] |
| 22 | else: |
| 23 | combo_cases_1 = [ |
| 24 | ("FA3", "FULL", True), |
| 25 | ("FA3", "FULL_AND_PIECEWISE", True), |
| 26 | ("FA2", "FULL", True), # Should fallback to FULL_AND_PIECEWISE |
| 27 | ("FA2", "FULL_AND_PIECEWISE", True), |
| 28 | ("FlashInfer", "FULL", True), # Should fallback to FULL_AND_PIECEWISE |
| 29 | ("FlashInfer", "FULL_AND_PIECEWISE", True), |
| 30 | ] |
| 31 | |
| 32 | |
| 33 | @pytest.mark.parametrize("backend_name, cudagraph_mode, supported", combo_cases_1) |
| 34 | @create_new_process_for_each_test("spawn") |
| 35 | def test_backend_and_cudagraph_mode_combo(backend_name, cudagraph_mode, supported): |
| 36 | if backend_name == "FlashInfer": |
| 37 | try: |
| 38 | import flashinfer # noqa: F401 |
| 39 | except ImportError: |
| 40 | pytest.skip("FlashInfer is not installed") |
| 41 | backend_config = backend_configs[backend_name] |
| 42 | # Dynamically skip test if GPU capability is not met |
| 43 | if ( |
| 44 | backend_config.specific_gpu_arch |
| 45 | and backend_config.specific_gpu_arch != current_platform.get_device_capability() |
| 46 | ): |
| 47 | pytest.skip("Only Hopper GPUs support FA3 and FlashMLA") |
| 48 | |
| 49 | attention_config = backend_config.attention_config |
| 50 | |
| 51 | with ExitStack() as stack: |
| 52 | if not supported: |
| 53 | stack.enter_context(pytest.raises(Exception)) |
| 54 | |
| 55 | llm = LLM( |
| 56 | model="Qwen/Qwen2-1.5B-Instruct", |
| 57 | max_num_seqs=16, |
| 58 | max_num_batched_tokens=128, |
| 59 | trust_remote_code=True, |
| 60 | gpu_memory_utilization=0.45, |
| 61 | max_model_len=1024, |
| 62 | attention_config=attention_config, |
| 63 | compilation_config=CompilationConfig( |
| 64 | mode=CompilationMode.VLLM_COMPILE, |
| 65 | cudagraph_mode=cudagraph_mode, |
| 66 | cudagraph_capture_sizes=[1, 2, 4, 8, 16, 64, 128], |
| 67 | ), |
| 68 | ) |
| 69 | llm.generate(["Hello, my name is"] * 10) |
| 70 | |
| 71 | |
| 72 | # test cudagraph_mode with different compilation mode. |
| 73 | # (backend_name, cudagraph_mode, compilation_mode, supported) |
| 74 | attn_backend = "RocmAttn" if current_platform.is_rocm() else "FA2" |
| 75 | |
| 76 | combo_cases_2 = [ |
| 77 | (attn_backend, "FULL", CompilationMode.NONE, True), |
| 78 | (attn_backend, "PIECEWISE", CompilationMode.NONE, True), |
| 79 | (attn_backend, "PIECEWISE", CompilationMode.VLLM_COMPILE, True), |
| 80 | (attn_backend, "FULL_AND_PIECEWISE", CompilationMode.NONE, True), |
| 81 | (attn_backend, "FULL_DECODE_ONLY", CompilationMode.NONE, True), |
| 82 | (attn_backend, "FULL_DECODE_ONLY", CompilationMode.VLLM_COMPILE, True), |
| 83 | (attn_backend, "NONE", CompilationMode.NONE, True), |
| 84 | (attn_backend, "NONE", CompilationMode.VLLM_COMPILE, True), |
| 85 | ] |
| 86 | |
| 87 | |
| 88 | @pytest.mark.parametrize( |
| 89 | "backend_name,cudagraph_mode,compilation_mode,supported", combo_cases_2 |
| 90 | ) |
| 91 | @create_new_process_for_each_test("spawn") |
| 92 | def test_cudagraph_compilation_combo( |
| 93 | backend_name, cudagraph_mode, compilation_mode, supported |
| 94 | ): |
| 95 | backend_config = backend_configs[backend_name] |
| 96 | attention_config = backend_config.attention_config |
| 97 | |
| 98 | with ExitStack() as stack: |
| 99 | if not supported: |
| 100 | stack.enter_context(pytest.raises(Exception)) |
| 101 | |
| 102 | llm = LLM( |
| 103 | model="Qwen/Qwen2-1.5B-Instruct", |
| 104 | max_num_seqs=16, |
| 105 | max_num_batched_tokens=128, |
| 106 | trust_remote_code=True, |
| 107 | gpu_memory_utilization=0.45, |
| 108 | max_model_len=1024, |
| 109 | attention_config=attention_config, |
| 110 | compilation_config=CompilationConfig( |
| 111 | mode=compilation_mode, |
| 112 | cudagraph_mode=cudagraph_mode, |
| 113 | cudagraph_capture_sizes=[1, 2, 4, 8, 16, 64, 128], |
| 114 | ), |
| 115 | ) |
| 116 | llm.generate(["Hello, my name is"] * 10) |