Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

4.4 KB
Raw Blame
# SPDX-License-Identifier: AGPL-3.0-only
# Copyright 2026-present the Unsloth AI Inc. team.
"""Use Transformer.load: ST 5.x infers a chat message modality that chat-wraps embedding inputs."""

from __future__ import annotations

import inspect
import os

import pytest


def test_transformer_load_signature_supports_unsloth_kwargs():
    """Tripwire: a Hub-capable Transformer.load must accept the kwargs the issue 6881 fix passes;
    legacy ST 3.x/4.x load(input_path) falls back to Transformer(...), so the test skips there."""
    models = pytest.importorskip("sentence_transformers.models")
    load = getattr(models.Transformer, "load", None)
    assert callable(load), (
        "sentence_transformers Transformer.load is missing; the #6881 fix in "
        "unsloth.models.sentence_transformer._create_transformer_module depends on it."
    )
    params = inspect.signature(load).parameters
    accepts_var_kw = any(p.kind is inspect.Parameter.VAR_KEYWORD for p in params.values())
    # Mirror _create_transformer_module's hub_capable gate.
    hub_capable = accepts_var_kw or any(k in params for k in ("token", "cache_folder", "revision"))
    if not hub_capable:
        pytest.skip(
            "legacy Transformer.load(input_path); production path falls back to Transformer(...)"
        )
    unsupported = [
        k
        for k in ("token", "cache_folder", "revision", "trust_remote_code")
        if not (accepts_var_kw or k in params)
    ]
    assert not unsupported, (
        f"installed sentence_transformers Transformer.load no longer accepts {unsupported} "
        f"and has no **kwargs; update _create_transformer_module (#6881) before it silently "
        f"falls back to Transformer(...)."
    )


def _probe_texts():
    return [
        "roasted chickpeas in 20 kg bags",
        "The capital of France is Paris.",
        "A fast brown fox jumps over the lazy dog.",
        "recette de tarte aux pommes traditionnelle",
    ]


def test_fast_sentence_transformer_matches_stock_st():
    """End-to-end: FastSentenceTransformer embeddings and tokenization must match a
    stock SentenceTransformer load of the same checkpoint. Opt-in (needs a model) and
    GPU-only (FastSentenceTransformer requires CUDA), so it skips on CPU-only runners."""
    model_id = os.environ.get("UNSLOTH_EMBEDDING_PARITY_MODEL")
    if not model_id:
        pytest.skip(
            "set UNSLOTH_EMBEDDING_PARITY_MODEL to a chat-template embedding model "
            "(HF id or local path) to run the #6881 parity test"
        )

    torch = pytest.importorskip("torch")
    if not torch.cuda.is_available():
        pytest.skip("FastSentenceTransformer requires CUDA; skipping on CPU-only runner")
    np = pytest.importorskip("numpy")
    pytest.importorskip("sentence_transformers")
    from sentence_transformers import SentenceTransformer

    device = "cuda"
    # fp16 overflows to NaN on bf16-native embedders like EmbeddingGemma.
    dtype = torch.bfloat16 if torch.cuda.is_bf16_supported() else torch.float16
    texts = _probe_texts()
    max_seq_length = 256

    # Build the control before importing unsloth so its global patches never touch it.
    ctrl = SentenceTransformer(model_id, device = device, model_kwargs = {"torch_dtype": dtype})
    ctrl.max_seq_length = max_seq_length
    ctrl_ids = ctrl.tokenize([texts[0]])["input_ids"][0].tolist()
    ctrl_emb = np.asarray(
        ctrl.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
    )

    import unsloth  # noqa: F401
    from unsloth import FastSentenceTransformer

    fast = FastSentenceTransformer.from_pretrained(
        model_id,
        max_seq_length = max_seq_length,
        dtype = dtype,
        load_in_4bit = False,
        load_in_16bit = True,
    )
    fast_ids = fast.tokenize([texts[0]])["input_ids"][0].tolist()
    fast_emb = np.asarray(
        fast.encode(texts, normalize_embeddings = True, batch_size = 8), dtype = np.float32
    )

    assert fast_ids == ctrl_ids, (
        f"tokenization diverged (chat-template wrapping regressed?):\n"
        f"  stock: {ctrl_ids}\n  fast:  {fast_ids}"
    )

    cos = (ctrl_emb * fast_emb).sum(1) / (
        np.linalg.norm(ctrl_emb, axis = 1) * np.linalg.norm(fast_emb, axis = 1)
    )
    assert float(cos.min()) > 0.99, (
        f"embedding parity regressed: min cosine {float(cos.min()):.5f} <= 0.99 "
        f"(per-text {[round(float(c), 5) for c in cos]})"
    )