[Frontend] Serve LiquidAI/d1-3B on /v1/systemone (#60676)

Signed-off-by: mgoin <mgoin64@gmail.com>
Co-authored-by: Simon Veitner <sveitner@redhat.com>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Misha Goin
2026-10-12 13:00:47 -04:00
committed by GitHub
co-authored by Simon Veitner Claude
parent d79b52b182
commit fa39582756
3 changed files with 222 additions and 17 deletions
@@ -26,6 +26,7 @@ from vllm.entrypoints.generate.structured_decisions.protocol import (
from vllm.entrypoints.generate.structured_decisions.question_types import (
LABELS,
StructuredDecisionError,
build_question,
)
from vllm.entrypoints.generate.structured_decisions.serving import (
ServingStructuredDecisions,
@@ -34,8 +35,10 @@ from vllm.entrypoints.generate.structured_decisions.serving import (
from vllm.entrypoints.generate.structured_decisions.strategies import (
DiffusionGemmaCanvasStrategy,
JointSchemaStrategy,
LiquidStrategy,
NextTokenStrategy,
ReadContext,
cap_pixels,
reply_label_ids,
select_read_strategy,
state_text,
@@ -79,6 +82,8 @@ def test_strategy_selection():
select_read_strategy(model("DiffusionGemmaForBlockDiffusion"), generate)
is DiffusionGemmaCanvasStrategy
)
lfm = "Lfm2VlForConditionalGeneration"
assert select_read_strategy(model(lfm), generate) is LiquidStrategy
with pytest.raises(ValueError, match="does not support LlamaForCausalLM"):
select_read_strategy(model("LlamaForCausalLM"), generate)
with pytest.raises(ValueError, match="not raw_logits"):
@@ -573,3 +578,50 @@ async def test_joint_schema_hears_video_soundtracks(joint_schema, monkeypatch):
)
assert isinstance(response, ErrorResponse)
assert response.error.code == 400
@pytest.mark.parametrize(
"type_name,criteria,ask",
[
(
"choice",
{"billing": "Charges", "app_fault": None},
"Which?\n\nOptions:\nA Charges\nB app fault\n\n"
"Reply with the option code only.",
),
(
"choice",
{"a": "Berlin", "b": "Paris"},
"Which?\n\nOptions:\na Berlin\nb Paris\n\nReply with the option code only.",
),
(
"noul",
{"true": "asks to cancel", "false": "anything else"},
"Which?\nYes: asks to cancel\nNo: anything else\n\n"
"Reply with yes or no only.",
),
(
"score",
["low", "high"],
"Which?\n\n0 low\n1 high\n\nReply with a single digit 0-1 only.",
),
],
)
def test_liquid_prompt_matches_the_checkpoint(type_name, criteria, ask):
# The wording of prompt.py in LiquidAI/d1-3B, which the model was trained on.
strategy = object.__new__(LiquidStrategy)
q = build_question("q", type_name, "Which?", criteria, 26)
assert strategy._content(strategy.render_state(None), q) == ask
assert strategy._content("Hi.", q) == f"Hi.\n\n\nQUESTION:\n{ask}"
state = strategy.render_state({"k": "é"})
assert strategy._content(state, q).startswith('{\n "k": "é"\n}\n\n\nQUEST')
def test_liquid_caps_images_like_the_checkpoint():
# runner.cap_pixels in LiquidAI/d1-3B: 1600x1598 becomes 1024x1023.
cap = LiquidStrategy.max_image_pixels
assert cap is not None
assert cap_pixels(Image.new("RGB", (1600, 1598)), cap).size == (1024, 1023)
small = Image.new("RGB", (640, 480))
assert cap_pixels(small, cap) is small
assert NextTokenStrategy.max_image_pixels is None