Files
sglang/test/registered/models_e2e/test_inkling.py
T

230 lines
8.5 KiB
Python

"""Per-commit end-to-end server test for Inkling (hybrid SWA/sconv attention + MoE + multimodal towers).
Boots a small ``thinkingmachines/Inkling`` checkpoint (the ``test`` revision, a
full-architecture but shrunken model) via ``popen_launch_server`` and exercises
the Inkling code paths on every PR: generation, the multimodal (vision) path,
the ``inkling`` reasoning parser, and UnifiedRadixTree cache consistency.
The checkpoint is undertrained, so these guard that the code paths boot and stay
numerically correct -- not answer quality; there is no accuracy gate (full-model
gsm8k accuracy lives in the 8-GPU nightly).
"""
import base64
import io
import os
import unittest
import requests
from sglang.srt.utils import kill_process_tree
from sglang.test.cache_consistency_jitter import get_jitter_engine, run_jitter_test
from sglang.test.ci.ci_register import register_cuda_ci
# Aliased so pytest does not collect the imported `test_`-prefixed helper as a test.
from sglang.test.kl_test_utils import (
test_input_output_logprobs_match_helper as assert_logprobs_match,
)
from sglang.test.test_utils import (
DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
DEFAULT_URL_FOR_TEST,
CustomTestCase,
popen_launch_server,
)
register_cuda_ci(est_time=450, stage="base-b", runner_config="1-gpu-large")
# Defaults to the HF `test` revision; override MODEL/REVISION to point at a
# local checkpoint. Empty REVISION drops the flag (for local paths).
_MODEL_PATH = os.environ.get("INKLING_TEST_MODEL_PATH", "thinkingmachines/Inkling")
_MODEL_REVISION = os.environ.get("INKLING_TEST_MODEL_REVISION", "test")
def _small_image_data_uri():
from PIL import Image
im = Image.new("RGB", (64, 64), (200, 60, 60))
buf = io.BytesIO()
im.save(buf, format="PNG")
return "data:image/png;base64," + base64.b64encode(buf.getvalue()).decode()
class TestInklingServer(CustomTestCase):
@classmethod
def setUpClass(cls):
cls.model = _MODEL_PATH
cls.base_url = DEFAULT_URL_FOR_TEST
# `enable_mamba_extra_buffer` is asserted by the Inkling model, and its
# default attention backend is unsupported, so fa4 is required. KV pool
# size follows mem_fraction_static, so it scales down on smaller GPUs.
other_args = [
"--trust-remote-code",
"--attention-backend",
"fa4",
"--page-size",
"128",
"--mamba-radix-cache-strategy",
"extra_buffer",
"--swa-full-tokens-ratio",
"0.1",
"--mamba-full-memory-ratio",
"0.1",
"--mem-fraction-static",
"0.5",
"--reasoning-parser",
"inkling",
"--tool-call-parser",
"inkling",
"--enable-multimodal",
]
if _MODEL_REVISION:
other_args += ["--revision", _MODEL_REVISION]
cls.process = popen_launch_server(
cls.model,
cls.base_url,
timeout=DEFAULT_TIMEOUT_FOR_SERVER_LAUNCH,
other_args=other_args,
env={**os.environ, "SGLANG_ENABLE_UNIFIED_RADIX_TREE": "1"},
)
@classmethod
def tearDownClass(cls):
if getattr(cls, "process", None) is not None:
kill_process_tree(cls.process.pid)
def _chat(self, messages, **kwargs):
payload = {
"model": self.model,
"messages": messages,
"temperature": 0.0,
"max_tokens": 48,
}
payload.update(kwargs)
resp = requests.post(
f"{self.base_url}/v1/chat/completions", json=payload, timeout=120
)
self.assertEqual(resp.status_code, 200, resp.text)
return resp.json()
def test_generation_basic(self):
"""Each prompt must return a non-empty completion."""
prompts = [
"The capital of France is",
"1 + 2 + 3 + 4 + 5 =",
"Write a haiku about silicon:",
]
for prompt in prompts:
resp = requests.post(
f"{self.base_url}/generate",
json={
"text": prompt,
"sampling_params": {"temperature": 0.0, "max_new_tokens": 16},
},
timeout=60,
)
self.assertEqual(resp.status_code, 200, resp.text)
data = resp.json()
self.assertIn("text", data, data)
self.assertGreater(len(data["text"].strip()), 0, data)
def test_reasoning_parser_separates_thinking(self):
"""`thinking=True` must route the chain-of-thought into
`reasoning_content`, exercising the `inkling` reasoning parser."""
data = self._chat(
[{"role": "user", "content": "What is 17 * 24? Think step by step."}],
max_tokens=128,
extra_body={"chat_template_kwargs": {"thinking": True}},
)
msg = data["choices"][0]["message"]
self.assertTrue(
(msg.get("reasoning_content") or "").strip(),
f"expected non-empty reasoning_content, got {msg}",
)
def test_multimodal_image_is_consumed(self):
"""An image input must reach the vision tower (more prompt tokens than
the text-only turn) and produce output, guarding the mm processor and
placeholder handling against crashes/regressions."""
text_only = self._chat([{"role": "user", "content": "Describe."}])
with_image = self._chat(
[
{
"role": "user",
"content": [
{"type": "text", "text": "Describe."},
{
"type": "image_url",
"image_url": {"url": _small_image_data_uri()},
},
],
}
],
max_tokens=48,
extra_body={"chat_template_kwargs": {"thinking": True}},
)
self.assertGreater(
with_image["usage"]["prompt_tokens"],
text_only["usage"]["prompt_tokens"],
"image did not add prompt tokens -- vision path not exercised",
)
msg = with_image["choices"][0]["message"]
produced = (msg.get("reasoning_content") or "") + (msg.get("content") or "")
self.assertTrue(produced.strip(), f"empty completion for image input: {msg}")
def test_unified_radix_cache_logprob_consistency(self):
"""UnifiedRadixTree correctness: prefill vs decode logprobs over real
multi-turn prompts must agree (avg KL below threshold). A hybrid
sconv/mamba/SWA state-cache bug shows up as a large KL. Numerical, so it
is meaningful even on an undertrained checkpoint."""
assert_logprobs_match(
self.base_url,
{self.model: {"kl_div": 1e-2}},
self.model,
max_samples=4,
max_new_tokens=256,
trust_remote_code=True,
)
class TestInklingCacheConsistency(CustomTestCase):
"""Bitwise version of the KL check above: the same context is scored under
different batch shapes, cache histories and retraction timing, and every
overlapping observation must agree. Boots its own in-process engine because
the harness patches the scheduler process to inject stream-sync jitter."""
def test_scored_contexts_are_bitwise_identical(self):
engine_kwargs = {
"model_path": _MODEL_PATH,
"trust_remote_code": True,
"attention_backend": "fa4",
"page_size": 128,
"mamba_radix_cache_strategy": "extra_buffer",
"swa_full_tokens_ratio": 0.1,
"mamba_full_memory_ratio": 0.1,
"mem_fraction_static": 0.5,
"enable_deterministic_inference": True,
}
if _MODEL_REVISION:
engine_kwargs["revision"] = _MODEL_REVISION
with get_jitter_engine(**engine_kwargs) as engine:
# Sized to the 20480-token pool the harness pins: large enough to
# force retraction, small enough that the batch still admits.
run_jitter_test(
engine,
num_unique_prefixes=4,
requests_per_prefix=5,
prefix_len_min=384,
prefix_len_max=512,
new_tokens=256,
# How many requests get retracted tracks pool size, hence GPU
# memory, so only assert the path ran at all.
min_retracted_requests=1,
)
if __name__ == "__main__":
unittest.main()