| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | import os |
| 4 | import random |
| 5 | from typing import NamedTuple |
| 6 | |
| 7 | import pytest |
| 8 | import torch |
| 9 | |
| 10 | from vllm.platforms import current_platform |
| 11 | from vllm.transformers_utils.config import get_config |
| 12 | from vllm.transformers_utils.model_arch_config_convertor import ( |
| 13 | ModelArchConfigConvertorBase, |
| 14 | ) |
| 15 | from vllm.triton_utils import HAS_TRITON |
| 16 | from vllm.v1.attention.backends.fa_utils import flash_attn_supports_mla |
| 17 | |
| 18 | |
| 19 | class DeviceConfig(NamedTuple): |
| 20 | available: bool |
| 21 | backends: list[str] |
| 22 | |
| 23 | |
| 24 | # Maps each device to its availability and supported backends. |
| 25 | DEVICE_BACKENDS: dict[str, DeviceConfig] = { |
| 26 | "cuda": DeviceConfig( |
| 27 | available=current_platform.is_cuda() |
| 28 | and current_platform.has_device_capability(80), |
| 29 | # FlashInfer backend temporarily disabled due to invariant CTA sizes. |
| 30 | # See FlashInfer issue #2424 |
| 31 | backends=["FLASH_ATTN", "TRITON_ATTN", "FLEX_ATTENTION"], |
| 32 | ), |
| 33 | "xpu": DeviceConfig( |
| 34 | available=current_platform.is_xpu() and HAS_TRITON, |
| 35 | backends=["TRITON_ATTN"], |
| 36 | ), |
| 37 | } |
| 38 | |
| 39 | DEFAULT_MODEL = "Qwen/Qwen3-1.7B" |
| 40 | TEST_MODEL = os.getenv("VLLM_TEST_MODEL", DEFAULT_MODEL) |
| 41 | |
| 42 | # Override backends for MLA models (MLA only supported on CUDA). |
| 43 | if os.getenv("VLLM_TEST_MODEL"): |
| 44 | config = get_config(TEST_MODEL, trust_remote_code=False) |
| 45 | if ModelArchConfigConvertorBase(config, config.get_text_config()).is_deepseek_mla(): |
| 46 | DEVICE_BACKENDS["cuda"] = DeviceConfig( |
| 47 | available=DEVICE_BACKENDS["cuda"].available, |
| 48 | backends=["TRITON_MLA"] |
| 49 | + (["FLASH_ATTN_MLA"] if flash_attn_supports_mla() else []), |
| 50 | ) |
| 51 | DEVICE_BACKENDS["xpu"] = DeviceConfig( |
| 52 | available=DEVICE_BACKENDS["xpu"].available, |
| 53 | backends=[], |
| 54 | ) |
| 55 | |
| 56 | # Only include backends for devices that are actually available. |
| 57 | BACKENDS: list[str] = sorted( |
| 58 | {b for cfg in DEVICE_BACKENDS.values() if cfg.available for b in cfg.backends} |
| 59 | ) |
| 60 | |
| 61 | skip_unsupported = pytest.mark.skipif( |
| 62 | not any(cfg.available for cfg in DEVICE_BACKENDS.values()), |
| 63 | reason="Requires CUDA >= Ampere (SM80) or Intel XPU with Triton", |
| 64 | ) |
| 65 | |
| 66 | skip_if_not_cuda = pytest.mark.skipif( |
| 67 | not DEVICE_BACKENDS["cuda"].available, |
| 68 | reason="Requires CUDA >= Ampere (SM80)", |
| 69 | ) |
| 70 | |
| 71 | |
| 72 | def _random_prompt(min_words: int = 1024, max_words: int = 1024 * 2) -> str: |
| 73 | # Generate more realistic prompts that will actually produce varied tokens |
| 74 | # Use a mix of common English text patterns |
| 75 | |
| 76 | prompt_templates = [ |
| 77 | # Question-answer style |
| 78 | "Question: What is the capital of France?\nAnswer: The capital of France is", |
| 79 | "Q: How does photosynthesis work?\nA: Photosynthesis is the process by which", |
| 80 | "User: Can you explain quantum mechanics?\nAssistant: Quantum mechanics is", |
| 81 | # Story/narrative style |
| 82 | "Once upon a time in a distant galaxy, there lived", |
| 83 | "The old man walked slowly down the street, remembering", |
| 84 | "In the year 2157, humanity finally discovered", |
| 85 | # Technical/code style |
| 86 | "To implement a binary search tree in Python, first we need to", |
| 87 | "The algorithm works by iterating through the array and", |
| 88 | "Here's how to optimize database queries using indexing:", |
| 89 | # Factual/informative style |
| 90 | "The Renaissance was a period in European history that", |
| 91 | "Climate change is caused by several factors including", |
| 92 | "The human brain contains approximately 86 billion neurons which", |
| 93 | # Conversational style |
| 94 | "I've been thinking about getting a new laptop because", |
| 95 | "Yesterday I went to the store and bought", |
| 96 | "My favorite thing about summer is definitely", |
| 97 | ] |
| 98 | |
| 99 | # Pick a random template |
| 100 | base_prompt = random.choice(prompt_templates) |
| 101 | |
| 102 | if max_words < min_words: |
| 103 | max_words = min_words |
| 104 | target_words = random.randint(min_words, max_words) |
| 105 | |
| 106 | if target_words > 50: |
| 107 | # For longer prompts, repeat context |
| 108 | padding_text = ( |
| 109 | " This is an interesting topic that deserves more explanation. " |
| 110 | # TODO: Update to * (target_words // 10) to better align with word ratio |
| 111 | * (target_words // 50) |
| 112 | ) |
| 113 | base_prompt = padding_text + base_prompt |
| 114 | |
| 115 | return base_prompt |
| 116 | |
| 117 | |
| 118 | def _extract_step_logprobs(request_output): |
| 119 | if getattr(request_output, "outputs", None): |
| 120 | inner = request_output.outputs[0] |
| 121 | if hasattr(inner, "logprobs") and inner.logprobs is not None: |
| 122 | t = torch.tensor( |
| 123 | [ |
| 124 | inner.logprobs[i][tid].logprob |
| 125 | for i, tid in enumerate(inner.token_ids) |
| 126 | ], |
| 127 | dtype=torch.float32, |
| 128 | ) |
| 129 | return t, inner.token_ids |
| 130 | |
| 131 | return None, None |
| 132 | |
| 133 | |
| 134 | def is_device_capability_below_90() -> bool: |
| 135 | return not current_platform.has_device_capability(90) |