llama.cpp/tests/test_llama.py

294 lines
10 KiB
Python
Raw Normal View History

2023-11-10 10:39:42 +00:00
import ctypes
import numpy as np
import pytest
from scipy.special import log_softmax
2023-11-10 10:39:42 +00:00
2023-04-05 07:23:15 +00:00
import llama_cpp
2023-08-24 05:01:20 +00:00
MODEL = "./vendor/llama.cpp/models/ggml-vocab-llama.gguf"
2023-04-05 07:23:15 +00:00
def test_llama_cpp_tokenization():
llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, verbose=False)
2023-04-05 07:23:15 +00:00
assert llama
2023-11-06 14:16:36 +00:00
assert llama._ctx.ctx is not None
2023-04-05 07:23:15 +00:00
text = b"Hello World"
tokens = llama.tokenize(text)
assert tokens[0] == llama.token_bos()
2023-08-27 16:59:20 +00:00
assert tokens == [1, 15043, 2787]
detokenized = llama.detokenize(tokens)
assert detokenized == text
tokens = llama.tokenize(text, add_bos=False)
assert tokens[0] != llama.token_bos()
2023-08-27 16:59:20 +00:00
assert tokens == [15043, 2787]
detokenized = llama.detokenize(tokens)
2023-08-27 16:59:20 +00:00
assert detokenized != text
2023-04-05 07:23:15 +00:00
text = b"Hello World</s>"
tokens = llama.tokenize(text)
assert tokens[-1] != llama.token_eos()
assert tokens == [1, 15043, 2787, 829, 29879, 29958]
tokens = llama.tokenize(text, special=True)
assert tokens[-1] == llama.token_eos()
2023-11-20 19:11:33 +00:00
assert tokens == [1, 15043, 2787, 2]
2023-11-21 04:23:18 +00:00
text = b""
tokens = llama.tokenize(text, add_bos=True, special=True)
assert tokens[-1] != llama.token_eos()
assert tokens == [llama.token_bos()]
assert text == llama.detokenize(tokens)
@pytest.fixture
def mock_llama(monkeypatch):
def setup_mock(llama: llama_cpp.Llama, output_text: str):
n_ctx = llama.n_ctx()
2023-11-21 04:23:18 +00:00
n_vocab = llama.n_vocab()
output_tokens = llama.tokenize(
output_text.encode("utf-8"), add_bos=True, special=True
)
logits = (ctypes.c_float * (n_vocab * n_ctx))(-100.0)
for i in range(n_ctx):
output_idx = i + 1 # logits for first tokens predict second token
if output_idx < len(output_tokens):
logits[i * n_vocab + output_tokens[output_idx]] = 100.0
else:
logits[i * n_vocab + llama.token_eos()] = 100.0
2023-11-21 04:23:18 +00:00
n = 0
last_n_tokens = 0
def mock_decode(ctx: llama_cpp.llama_context_p, batch: llama_cpp.llama_batch):
# Test some basic invariants of this mocking technique
2023-11-22 09:12:32 +00:00
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
assert batch.n_tokens > 0, "no tokens in batch"
assert all(
batch.n_seq_id[i] == 1 for i in range(batch.n_tokens)
), "n_seq >1 not supported by mock_llama"
assert all(
batch.seq_id[i][0] == 0 for i in range(batch.n_tokens)
), "n_seq >1 not supported by mock_llama"
assert batch.logits[
batch.n_tokens - 1
], "logits not allocated for last token"
# Update the mock context state
nonlocal n
nonlocal last_n_tokens
2023-11-22 09:12:32 +00:00
n = max(batch.pos[i] for i in range(batch.n_tokens)) + 1
2023-11-21 04:23:18 +00:00
last_n_tokens = batch.n_tokens
return 0
def mock_get_logits(ctx: llama_cpp.llama_context_p):
# Test some basic invariants of this mocking technique
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
2023-11-22 09:12:32 +00:00
assert n > 0, "mock_llama_decode not called"
assert last_n_tokens > 0, "mock_llama_decode not called"
# Return view of logits for last_n_tokens
return (ctypes.c_float * (last_n_tokens * n_vocab)).from_address(
ctypes.addressof(logits)
+ (n - last_n_tokens) * n_vocab * ctypes.sizeof(ctypes.c_float)
)
2023-11-21 04:23:18 +00:00
monkeypatch.setattr("llama_cpp.llama_cpp.llama_decode", mock_decode)
monkeypatch.setattr("llama_cpp.llama_cpp.llama_get_logits", mock_get_logits)
def mock_kv_cache_clear(ctx: llama_cpp.llama_context_p):
# Test some basic invariants of this mocking technique
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
return
def mock_kv_cache_seq_rm(
ctx: llama_cpp.llama_context_p,
seq_id: llama_cpp.llama_seq_id,
pos0: llama_cpp.llama_pos,
pos1: llama_cpp.llama_pos,
):
# Test some basic invariants of this mocking technique
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
return
def mock_kv_cache_seq_cp(
ctx: llama_cpp.llama_context_p,
seq_id_src: llama_cpp.llama_seq_id,
seq_id_dst: llama_cpp.llama_seq_id,
pos0: llama_cpp.llama_pos,
pos1: llama_cpp.llama_pos,
):
# Test some basic invariants of this mocking technique
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
return
def mock_kv_cache_seq_keep(
ctx: llama_cpp.llama_context_p,
seq_id: llama_cpp.llama_seq_id,
):
# Test some basic invariants of this mocking technique
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
return
def mock_kv_cache_seq_shift(
ctx: llama_cpp.llama_context_p,
seq_id: llama_cpp.llama_seq_id,
pos0: llama_cpp.llama_pos,
pos1: llama_cpp.llama_pos,
):
# Test some basic invariants of this mocking technique
assert ctx == llama._ctx.ctx, "context does not match mock_llama"
return
monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_clear", mock_kv_cache_clear)
monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_rm", mock_kv_cache_seq_rm)
monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_cp", mock_kv_cache_seq_cp)
monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_keep", mock_kv_cache_seq_keep)
monkeypatch.setattr("llama_cpp.llama_cpp.llama_kv_cache_seq_shift", mock_kv_cache_seq_shift)
2023-11-21 04:23:18 +00:00
return setup_mock
def test_llama_patch(mock_llama):
2023-11-10 10:39:42 +00:00
n_ctx = 128
llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, n_ctx=n_ctx)
2023-11-06 14:16:36 +00:00
n_vocab = llama_cpp.llama_n_vocab(llama._model.model)
2023-11-10 10:39:42 +00:00
assert n_vocab == 32000
2023-04-05 07:23:15 +00:00
2023-11-20 19:11:33 +00:00
text = "The quick brown fox"
2023-04-05 07:23:15 +00:00
output_text = " jumps over the lazy dog."
2023-11-21 04:23:18 +00:00
all_text = text + output_text
2023-04-05 07:23:15 +00:00
2023-11-21 04:23:18 +00:00
## Test basic completion from bos until eos
mock_llama(llama, all_text)
completion = llama.create_completion("", max_tokens=36)
assert completion["choices"][0]["text"] == all_text
assert completion["choices"][0]["finish_reason"] == "stop"
2023-04-05 07:23:15 +00:00
## Test basic completion until eos
2023-11-21 04:23:18 +00:00
mock_llama(llama, all_text)
2023-04-05 07:23:15 +00:00
completion = llama.create_completion(text, max_tokens=20)
assert completion["choices"][0]["text"] == output_text
assert completion["choices"][0]["finish_reason"] == "stop"
## Test streaming completion until eos
2023-11-21 04:23:18 +00:00
mock_llama(llama, all_text)
2023-10-19 06:55:29 +00:00
chunks = list(llama.create_completion(text, max_tokens=20, stream=True))
2023-04-05 07:23:15 +00:00
assert "".join(chunk["choices"][0]["text"] for chunk in chunks) == output_text
2023-10-19 06:56:45 +00:00
assert chunks[-1]["choices"][0]["finish_reason"] == "stop"
2023-04-05 07:23:15 +00:00
## Test basic completion until stop sequence
2023-11-21 04:23:18 +00:00
mock_llama(llama, all_text)
2023-04-05 07:23:15 +00:00
completion = llama.create_completion(text, max_tokens=20, stop=["lazy"])
assert completion["choices"][0]["text"] == " jumps over the "
assert completion["choices"][0]["finish_reason"] == "stop"
## Test streaming completion until stop sequence
2023-11-21 04:23:18 +00:00
mock_llama(llama, all_text)
chunks = list(
llama.create_completion(text, max_tokens=20, stream=True, stop=["lazy"])
)
2023-04-05 07:23:15 +00:00
assert (
"".join(chunk["choices"][0]["text"] for chunk in chunks) == " jumps over the "
)
2023-10-19 06:56:45 +00:00
assert chunks[-1]["choices"][0]["finish_reason"] == "stop"
2023-04-05 07:23:15 +00:00
## Test basic completion until length
2023-11-21 04:23:18 +00:00
mock_llama(llama, all_text)
2023-04-05 07:23:15 +00:00
completion = llama.create_completion(text, max_tokens=2)
2023-10-19 06:55:29 +00:00
assert completion["choices"][0]["text"] == " jumps"
2023-10-19 06:56:45 +00:00
assert completion["choices"][0]["finish_reason"] == "length"
2023-04-05 07:23:15 +00:00
## Test streaming completion until length
2023-11-21 04:23:18 +00:00
mock_llama(llama, all_text)
2023-10-19 06:55:29 +00:00
chunks = list(llama.create_completion(text, max_tokens=2, stream=True))
assert "".join(chunk["choices"][0]["text"] for chunk in chunks) == " jumps"
2023-10-19 06:56:45 +00:00
assert chunks[-1]["choices"][0]["finish_reason"] == "length"
def test_llama_pickle():
import pickle
import tempfile
2023-05-02 02:38:46 +00:00
fp = tempfile.TemporaryFile()
llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True)
pickle.dump(llama, fp)
fp.seek(0)
llama = pickle.load(fp)
assert llama
assert llama.ctx is not None
text = b"Hello World"
assert llama.detokenize(llama.tokenize(text)) == text
2023-05-02 02:38:46 +00:00
2023-11-21 23:13:19 +00:00
def test_utf8(mock_llama):
2023-11-21 04:23:18 +00:00
llama = llama_cpp.Llama(model_path=MODEL, vocab_only=True, logits_all=True)
2023-11-21 04:23:18 +00:00
output_text = "😀"
## Test basic completion with utf8 multibyte
2023-11-21 23:13:19 +00:00
mock_llama(llama, output_text)
completion = llama.create_completion("", max_tokens=4)
assert completion["choices"][0]["text"] == output_text
## Test basic completion with incomplete utf8 multibyte
2023-11-21 23:13:19 +00:00
mock_llama(llama, output_text)
completion = llama.create_completion("", max_tokens=1)
assert completion["choices"][0]["text"] == ""
2023-04-29 06:26:07 +00:00
def test_llama_server():
from fastapi.testclient import TestClient
2023-05-02 02:38:46 +00:00
from llama_cpp.server.app import create_app, Settings
2023-05-02 02:41:54 +00:00
settings = Settings(
model=MODEL,
vocab_only=True,
)
2023-05-02 02:38:46 +00:00
app = create_app(settings)
2023-04-29 06:26:07 +00:00
client = TestClient(app)
response = client.get("/v1/models")
assert response.json() == {
"object": "list",
"data": [
{
"id": MODEL,
"object": "model",
"owned_by": "me",
"permissions": [],
}
],
}
2023-11-21 04:23:18 +00:00
@pytest.mark.parametrize(
"size_and_axis",
[
((32_000,), -1), # last token's next-token logits
((10, 32_000), -1), # many tokens' next-token logits, or batch of last tokens
((4, 10, 32_000), -1), # batch of texts
],
)
@pytest.mark.parametrize("convert_to_list", [True, False])
def test_logits_to_logprobs(size_and_axis, convert_to_list: bool, atol: float = 1e-7):
size, axis = size_and_axis
logits: np.ndarray = -np.random.uniform(low=0, high=60, size=size)
logits = logits.astype(np.single)
if convert_to_list:
# Currently, logits are converted from arrays to lists. This may change soon
logits = logits.tolist()
log_probs = llama_cpp.Llama.logits_to_logprobs(logits, axis=axis)
log_probs_correct = log_softmax(logits, axis=axis)
assert log_probs.dtype == np.single
assert log_probs.shape == size
assert np.allclose(log_probs, log_probs_correct, atol=atol)
def test_llama_cpp_version():
assert llama_cpp.__version__