test_compilation.py

import glob
import os
import runpy
import tempfile

import depyf

# disable custom dispatcher, let Dynamo takes over
# all the control
os.environ['VLLM_DYNAMO_USE_CUSTOM_DISPATCHER'] = "0"

temp_dir = tempfile.mkdtemp()
with depyf.prepare_debug(temp_dir):
    cur_dir = os.path.dirname(__file__)
    parent_dir = os.path.dirname(cur_dir)
    root_dir = os.path.dirname(parent_dir)
    example_file = os.path.join(root_dir, "examples",
                                "offline_inference_tpu.py")
    runpy.run_path(example_file)

compiled_code = sorted(
    glob.glob(os.path.join(temp_dir, "__transformed_code*.py")))

# we should only trigger Dynamo compilation three times:
# one for the profiling phase without kv cache
# one for the prefill phase with symbolic shapes
# one for the decode phase with symbolic shapes
# and later calls should not trigger Dynamo compilation again.
# NOTE: it might still trigger XLA compilation.

# check we have three compiled code
# this is the assumption when we use the custom dispatcher
assert len(compiled_code) == 3

# check all the compilations are as expected
compiled_fn = sorted(
    glob.glob(os.path.join(temp_dir, "__compiled_fn*Captured*.py")))

# the first compilation is the profiling phase,
# it should not have any kv cache
with open(compiled_fn[0]) as f:
    content = f.read()
    assert "kv_caches" not in content

# the second compilation is the prefill phase,
# it should have kv cache and the flash_attention op
with open(compiled_fn[1]) as f:
    content = f.read()
    assert "kv_caches" in content and "torch.ops.xla.flash_attention" in content

# the third compilation is the decode phase,
# it should have kv cache and the paged_attention op
with open(compiled_fn[2]) as f:
    content = f.read()
    assert "kv_caches" in content and "torch.ops.xla.paged_attention" in content