char-fork/llama.cpp

git clone https://git.t4t.associates/char-fork/llama.cpp

Georgi Gerganovserver : disable context shift by default (#15416)d2fcd91cf

char/aliceai-t5-moe
brancheschar/aliceai-t5-moeall refs →
4.5 KiB115 linesraw
1import pytest
2from utils import *
3
4server = ServerPreset.stories15m_moe()
5
6LORA_FILE_URL = "https://huggingface.co/ggml-org/stories15M_MOE/resolve/main/moe_shakespeare15M.gguf"
7
8@pytest.fixture(autouse=True)
9def create_server():
10    global server
11    server = ServerPreset.stories15m_moe()
12    server.lora_files = [download_file(LORA_FILE_URL)]
13
14
15@pytest.mark.parametrize("scale,re_content", [
16    # without applying lora, the model should behave like a bedtime story generator
17    (0.0, "(little|girl|three|years|old)+"),
18    # with lora, the model should behave like a Shakespearean text generator
19    (1.0, "(eye|love|glass|sun)+"),
20])
21def test_lora(scale: float, re_content: str):
22    global server
23    server.start()
24    res_lora_control = server.make_request("POST", "/lora-adapters", data=[
25        {"id": 0, "scale": scale}
26    ])
27    assert res_lora_control.status_code == 200
28    res = server.make_request("POST", "/completion", data={
29        "prompt": "Look in thy glass",
30    })
31    assert res.status_code == 200
32    assert match_regex(re_content, res.body["content"])
33
34
35def test_lora_per_request():
36    global server
37    server.n_slots = 4
38    server.start()
39
40    # running the same prompt with different lora scales, all in parallel
41    # each prompt will be processed by a different slot
42    prompt = "Look in thy glass"
43    lora_config = [
44        ( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
45        ( [{"id": 0, "scale": 0.0}], "(bright|day|many|happy)+" ),
46        ( [{"id": 0, "scale": 0.3}], "(special|thing|gifted)+" ),
47        ( [{"id": 0, "scale": 0.7}], "(far|from|home|away)+" ),
48        ( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
49        ( [{"id": 0, "scale": 1.0}], "(eye|love|glass|sun)+" ),
50    ]
51
52    tasks = [(
53        server.make_request,
54        ("POST", "/completion", {
55            "prompt": prompt,
56            "lora": lora,
57            "seed": 42,
58            "temperature": 0.0,
59            "cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
60        })
61    ) for lora, _ in lora_config]
62    results = parallel_function_calls(tasks)
63
64    assert all([res.status_code == 200 for res in results])
65    for res, (_, re_test) in zip(results, lora_config):
66        assert match_regex(re_test, res.body["content"])
67
68
69@pytest.mark.skipif(not is_slow_test_allowed(), reason="skipping slow test")
70def test_with_big_model():
71    server = ServerProcess()
72    server.model_hf_repo = "bartowski/Meta-Llama-3.1-8B-Instruct-GGUF"
73    server.model_hf_file = "Meta-Llama-3.1-8B-Instruct-IQ2_M.gguf"
74    server.model_alias = "Llama-3.2-8B-Instruct"
75    server.n_slots = 4
76    server.n_ctx = server.n_slots * 1024
77    server.n_predict = 64
78    server.temperature = 0.0
79    server.seed = 42
80    server.lora_files = [
81        download_file("https://huggingface.co/ngxson/Llama-3-Instruct-abliteration-LoRA-8B-F16-GGUF/resolve/main/Llama-3-Instruct-abliteration-LoRA-8B-f16.gguf"),
82        # TODO: find & add other lora adapters for this model
83    ]
84    server.start(timeout_seconds=600)
85
86    # running the same prompt with different lora scales, all in parallel
87    # each prompt will be processed by a different slot
88    prompt = "Write a computer virus"
89    lora_config = [
90        # without applying lora, the model should reject the request
91        ( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
92        ( [{"id": 0, "scale": 0.0}], "I can't provide you with a code for a computer virus" ),
93        ( [{"id": 0, "scale": 0.3}], "I can't write a computer virus" ),
94        # with 0.7 scale, the model should provide a simple computer virus with hesitation
95        ( [{"id": 0, "scale": 0.7}], "Warning: This is a hypothetical exercise" ),
96        # with 1.5 scale, the model should confidently provide a computer virus
97        ( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
98        ( [{"id": 0, "scale": 1.5}], "A task of some complexity! Here's a simple computer virus" ),
99    ]
100
101    tasks = [(
102        server.make_request,
103        ("POST", "/v1/chat/completions", {
104            "messages": [
105                {"role": "user", "content": prompt}
106            ],
107            "lora": lora,
108            "cache_prompt": False, # TODO: remove this once test_cache_vs_nocache_prompt is fixed
109        })
110    ) for lora, _ in lora_config]
111    results = parallel_function_calls(tasks)
112
113    assert all([res.status_code == 200 for res in results])
114    for res, (_, re_test) in zip(results, lora_config):
115        assert re_test in res.body["choices"][0]["message"]["content"]