4a941fbeb9
Adds `kl_direction` ("forward"/"reverse") to control KL divergence computation direction and `normalize_by` ("tokens"/"weights") for DFT/ASFT loss normalization. Forward KL (default) matches original ASFT code behavior despite paper terminology. Reverse KL enables mode-seeking behavior. Updates `_compute_kl_divergence`, streaming strategies, `compute_asft_loss`, and `ASFTTrainer` to propagate both parameters. Adds tests for reverse KL computation
1312 lines
46 KiB
Python
1312 lines
46 KiB
Python
# Copyright 2023-present Daniel Han-Chen & the Unsloth team. All rights reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
"""Tests for ASFT (Anchored Supervised Fine-Tuning) loss module."""
|
|
|
|
import pytest
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from types import SimpleNamespace
|
|
from unittest.mock import MagicMock, patch
|
|
|
|
HAS_CUDA = torch.cuda.is_available()
|
|
if not HAS_CUDA:
|
|
pytest.skip("CUDA is required for ASFT tests", allow_module_level = True)
|
|
torch.set_default_device("cuda")
|
|
|
|
from unsloth.losses.asft import (
|
|
ASFTStreamingConfig,
|
|
effective_logits,
|
|
fast_cross_entropy_loss_per_token,
|
|
build_shift_labels,
|
|
get_reference_forward_callable,
|
|
compute_asft_loss,
|
|
_compute_kl_divergence,
|
|
_compute_dft_weights,
|
|
_compute_kl_seq_kv_cache,
|
|
)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Test Fixtures
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
@pytest.fixture
|
|
def dummy_logits():
|
|
"""Create dummy logits tensor (B=2, T=4, V=8)."""
|
|
torch.manual_seed(42)
|
|
return torch.randn(2, 4, 8, requires_grad = True)
|
|
|
|
|
|
@pytest.fixture
|
|
def dummy_labels():
|
|
"""Create dummy labels tensor with some -100 values."""
|
|
# Labels: [0, 1, 2, 3] and [4, 5, -100, -100]
|
|
return torch.tensor([[0, 1, 2, 3], [4, 5, -100, -100]], dtype = torch.long)
|
|
|
|
|
|
@pytest.fixture
|
|
def simple_model():
|
|
"""Create a simple model for testing."""
|
|
|
|
class SimpleModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.config = SimpleNamespace(
|
|
final_logit_softcapping = 0,
|
|
logit_scale = 0,
|
|
)
|
|
self.embedding = nn.Embedding(16, 8)
|
|
self.linear = nn.Linear(8, 8)
|
|
|
|
def forward(self, input_ids = None, **kwargs):
|
|
# Deterministic forward with gradients
|
|
embeddings = self.embedding(input_ids)
|
|
logits = self.linear(embeddings)
|
|
return SimpleNamespace(logits = logits)
|
|
|
|
return SimpleModel()
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# A1) Test effective_logits
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestEffectiveLogits:
|
|
"""Tests for effective_logits function."""
|
|
|
|
def test_no_transformation(self, dummy_logits):
|
|
"""Test that no transformation is applied when softcapping/scaling are 0."""
|
|
result = effective_logits(dummy_logits, logit_softcapping = 0, logit_scaling = 0)
|
|
# Should be close to original (converted to float32)
|
|
assert torch.allclose(result, dummy_logits.float(), atol = 1e-6)
|
|
|
|
def test_logit_scaling(self, dummy_logits):
|
|
"""Test logit scaling: t * x."""
|
|
scale = 2.0
|
|
result = effective_logits(dummy_logits, logit_scaling = scale)
|
|
expected = scale * dummy_logits.float()
|
|
assert torch.allclose(result, expected, atol = 1e-6)
|
|
|
|
def test_logit_softcapping(self, dummy_logits):
|
|
"""Test logit softcapping: t * tanh(x / t)."""
|
|
softcap = 30.0
|
|
result = effective_logits(dummy_logits, logit_softcapping = softcap)
|
|
expected = softcap * torch.tanh(dummy_logits.float() / softcap)
|
|
assert torch.allclose(result, expected, atol = 1e-6)
|
|
|
|
def test_both_transformations(self, dummy_logits):
|
|
"""Test both scaling and softcapping together."""
|
|
scale = 2.0
|
|
softcap = 30.0
|
|
result = effective_logits(
|
|
dummy_logits, logit_softcapping = softcap, logit_scaling = scale
|
|
)
|
|
# Scaling first, then softcapping
|
|
x = scale * dummy_logits.float()
|
|
expected = softcap * torch.tanh(x / softcap)
|
|
assert torch.allclose(result, expected, atol = 1e-6)
|
|
|
|
def test_reads_from_model_config(self):
|
|
"""Test reading config from model."""
|
|
model = SimpleNamespace(
|
|
config = SimpleNamespace(
|
|
final_logit_softcapping = 30.0,
|
|
logit_scale = 2.0,
|
|
)
|
|
)
|
|
logits = torch.randn(2, 4, 8)
|
|
result = effective_logits(logits, model)
|
|
# Should apply both transformations
|
|
x = 2.0 * logits.float()
|
|
expected = 30.0 * torch.tanh(x / 30.0)
|
|
assert torch.allclose(result, expected, atol = 1e-6)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# A2) Test fast_cross_entropy_loss_per_token
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestFastCrossEntropyLossPerToken:
|
|
"""Tests for fast_cross_entropy_loss_per_token function."""
|
|
|
|
def test_basic_loss_computation(self, dummy_logits, dummy_labels):
|
|
"""Test basic per-token CE loss computation."""
|
|
losses, valid_mask = fast_cross_entropy_loss_per_token(
|
|
dummy_logits.detach(), dummy_labels
|
|
)
|
|
|
|
# Check shapes
|
|
batch, seq_len = dummy_labels.shape
|
|
assert losses.shape == (batch * seq_len,)
|
|
assert valid_mask.shape == (batch * seq_len,)
|
|
|
|
# Check that valid_mask correctly identifies -100 positions
|
|
flat_labels = dummy_labels.view(-1)
|
|
expected_valid = flat_labels != -100
|
|
assert torch.equal(valid_mask, expected_valid)
|
|
|
|
def test_ignored_positions_have_zero_loss(self, dummy_logits, dummy_labels):
|
|
"""Test that positions with label -100 have zero loss."""
|
|
losses, valid_mask = fast_cross_entropy_loss_per_token(
|
|
dummy_logits.detach(), dummy_labels
|
|
)
|
|
|
|
# Loss at ignored positions should be 0
|
|
assert torch.all(losses[~valid_mask] == 0)
|
|
|
|
def test_valid_positions_have_nonzero_loss(self, dummy_logits, dummy_labels):
|
|
"""Test that valid positions have non-zero loss."""
|
|
losses, valid_mask = fast_cross_entropy_loss_per_token(
|
|
dummy_logits.detach(), dummy_labels
|
|
)
|
|
|
|
# At least some valid positions should have non-zero loss
|
|
assert torch.any(losses[valid_mask] > 0)
|
|
|
|
def test_matches_pytorch_ce(self):
|
|
"""Test that results match PyTorch CE loss."""
|
|
torch.manual_seed(42)
|
|
logits = torch.randn(2, 4, 8)
|
|
labels = torch.tensor([[0, 1, 2, 3], [4, 5, 6, 7]], dtype = torch.long)
|
|
|
|
losses, valid_mask = fast_cross_entropy_loss_per_token(logits, labels)
|
|
|
|
# Compare with PyTorch
|
|
flat_logits = logits.view(-1, 8)
|
|
flat_labels = labels.view(-1)
|
|
pytorch_losses = F.cross_entropy(flat_logits, flat_labels, reduction = "none")
|
|
|
|
# Should be close
|
|
assert torch.allclose(losses, pytorch_losses, atol = 1e-4)
|
|
|
|
def test_respects_custom_ignore_index(self):
|
|
"""Test that custom ignore_index is honored by the kernel wrapper."""
|
|
torch.manual_seed(0)
|
|
logits = torch.randn(1, 4, 8)
|
|
labels = torch.tensor([[1, 2, 1, 3]], dtype = torch.long)
|
|
|
|
losses, valid_mask = fast_cross_entropy_loss_per_token(
|
|
logits, labels, ignore_index = 1
|
|
)
|
|
|
|
assert losses.shape == (4,)
|
|
assert torch.equal(valid_mask, torch.tensor([False, True, False, True]))
|
|
assert torch.all(losses[~valid_mask] == 0)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# A3) Test build_shift_labels
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestBuildShiftLabels:
|
|
"""Tests for build_shift_labels function."""
|
|
|
|
def test_basic_shift(self):
|
|
"""Test basic label shifting."""
|
|
labels = torch.tensor([[0, 1, 2, 3], [4, 5, 6, 7]], dtype = torch.long)
|
|
shift_labels = build_shift_labels(labels)
|
|
|
|
# shift_labels[..., :-1] = labels[..., 1:]
|
|
# shift_labels[..., -1] = -100
|
|
expected = torch.tensor([[1, 2, 3, -100], [5, 6, 7, -100]], dtype = torch.long)
|
|
assert torch.equal(shift_labels, expected)
|
|
|
|
def test_preserves_ignore_index(self):
|
|
"""Test that existing -100 values are preserved after shift."""
|
|
labels = torch.tensor([[0, 1, -100, -100], [4, 5, 6, -100]], dtype = torch.long)
|
|
shift_labels = build_shift_labels(labels)
|
|
|
|
# First row: [1, -100, -100, -100]
|
|
# Second row: [5, 6, -100, -100]
|
|
expected = torch.tensor(
|
|
[[1, -100, -100, -100], [5, 6, -100, -100]], dtype = torch.long
|
|
)
|
|
assert torch.equal(shift_labels, expected)
|
|
|
|
def test_with_packed_seq_lengths(self):
|
|
"""Test shift labels with packed sequence boundary masking."""
|
|
# Single row with packed sequences of lengths [2, 2]
|
|
labels = torch.tensor([[0, 1, 2, 3]], dtype = torch.long)
|
|
packed_seq_lengths = torch.tensor([2, 2], dtype = torch.int32)
|
|
|
|
shift_labels = build_shift_labels(labels, packed_seq_lengths)
|
|
|
|
# After shift: [1, 2, 3, -100]
|
|
# After boundary masking at positions 1 and 3: [1, -100, 3, -100]
|
|
# Actually boundary positions are cumsum - 1 = [1, 3]
|
|
# So positions 1 and 3 should be -100
|
|
assert shift_labels[0, 1].item() == -100 # End of first sequence
|
|
assert shift_labels[0, 3].item() == -100 # End of second sequence (also last)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# A4) Test get_reference_forward_callable
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestGetReferenceForwardCallable:
|
|
"""Tests for get_reference_forward_callable function."""
|
|
|
|
def test_disable_adapter_policy(self, simple_model):
|
|
"""Test disable_adapter policy when model has adapters."""
|
|
# Mock disable_adapter
|
|
simple_model.disable_adapter = MagicMock()
|
|
simple_model.disable_adapter.__enter__ = MagicMock(return_value = None)
|
|
simple_model.disable_adapter.__exit__ = MagicMock(return_value = False)
|
|
|
|
ref_forward = get_reference_forward_callable(
|
|
simple_model, reference_policy = "disable_adapter"
|
|
)
|
|
|
|
# Call the forward
|
|
input_ids = torch.tensor([[1, 2, 3, 4]])
|
|
result = ref_forward(input_ids = input_ids)
|
|
|
|
# Should have called disable_adapter
|
|
assert simple_model.disable_adapter.__enter__.called
|
|
|
|
def test_frozen_copy_policy(self, simple_model):
|
|
"""Test frozen_copy policy."""
|
|
ref_forward = get_reference_forward_callable(
|
|
simple_model, reference_policy = "frozen_copy"
|
|
)
|
|
|
|
input_ids = torch.tensor([[1, 2, 3, 4]])
|
|
result = ref_forward(input_ids = input_ids)
|
|
|
|
# Should return logits
|
|
assert result.shape[0] == 1 # batch size
|
|
assert result.shape[1] == 4 # seq len
|
|
|
|
def test_fallback_to_frozen_copy_without_adapters(self, simple_model):
|
|
"""Test that disable_adapter falls back to frozen_copy when no adapters."""
|
|
# Model without disable_adapter method
|
|
ref_forward = get_reference_forward_callable(
|
|
simple_model, reference_policy = "disable_adapter"
|
|
)
|
|
|
|
input_ids = torch.tensor([[1, 2, 3, 4]])
|
|
result = ref_forward(input_ids = input_ids)
|
|
|
|
# Should still work (uses frozen copy fallback)
|
|
assert result is not None
|
|
|
|
def test_return_outputs_true(self, simple_model):
|
|
"""Test returning full outputs when requested."""
|
|
ref_forward = get_reference_forward_callable(
|
|
simple_model, reference_policy = "frozen_copy", return_outputs = True
|
|
)
|
|
|
|
input_ids = torch.tensor([[1, 2, 3, 4]])
|
|
result = ref_forward(input_ids = input_ids)
|
|
|
|
assert hasattr(result, "logits")
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Test KL divergence computation
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestKLDivergence:
|
|
"""Tests for KL divergence computation."""
|
|
|
|
def test_kl_direction(self):
|
|
"""Test that KL is computed as KL(p_ref || p_cur)."""
|
|
torch.manual_seed(42)
|
|
cur_logits = torch.randn(4, 8) # (B*T, V)
|
|
ref_logits = torch.randn(4, 8)
|
|
|
|
kl = _compute_kl_divergence(cur_logits, ref_logits, kl_direction = "forward")
|
|
|
|
# KL should be non-negative
|
|
assert torch.all(kl >= -1e-6) # Allow small numerical errors
|
|
|
|
def test_kl_zero_for_identical(self):
|
|
"""Test that KL is zero when distributions are identical."""
|
|
logits = torch.randn(4, 8)
|
|
|
|
kl = _compute_kl_divergence(logits, logits.clone(), kl_direction = "forward")
|
|
|
|
# Should be close to zero
|
|
assert torch.allclose(kl, torch.zeros_like(kl), atol = 1e-5)
|
|
|
|
def test_kl_shape(self):
|
|
"""Test KL output shape."""
|
|
cur_logits = torch.randn(2, 4, 8) # (B, T, V)
|
|
ref_logits = torch.randn(2, 4, 8)
|
|
|
|
kl = _compute_kl_divergence(cur_logits, ref_logits, kl_direction = "forward")
|
|
|
|
# Should be flattened to (B*T,)
|
|
assert kl.shape == (8,)
|
|
|
|
def test_kl_reverse_matches_manual(self):
|
|
"""Test reverse KL matches manual computation."""
|
|
torch.manual_seed(321)
|
|
cur_logits = torch.randn(2, 5)
|
|
ref_logits = torch.randn(2, 5)
|
|
|
|
kl_reverse = _compute_kl_divergence(
|
|
cur_logits, ref_logits, kl_direction = "reverse"
|
|
)
|
|
|
|
cur_p = F.softmax(cur_logits, dim = -1)
|
|
ref_p = F.softmax(ref_logits, dim = -1)
|
|
manual = (cur_p * (cur_p.log() - ref_p.log())).sum(dim = -1)
|
|
|
|
assert torch.allclose(kl_reverse, manual, atol = 1e-5)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Test DFT weights computation
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestDFTWeights:
|
|
"""Tests for DFT weights computation."""
|
|
|
|
def test_dft_weights_are_probabilities(self, dummy_logits, dummy_labels):
|
|
"""Test that DFT weights are valid probabilities."""
|
|
flat_logits = dummy_logits.detach().view(-1, 8)
|
|
flat_labels = dummy_labels.view(-1)
|
|
|
|
weights = _compute_dft_weights(flat_logits, flat_labels)
|
|
|
|
# Weights should be in [0, 1]
|
|
assert torch.all(weights >= 0)
|
|
assert torch.all(weights <= 1)
|
|
|
|
def test_dft_weights_are_detached(self, dummy_logits, dummy_labels):
|
|
"""Test that DFT weights are detached (no gradients)."""
|
|
weights = _compute_dft_weights(
|
|
dummy_logits.detach().view(-1, 8),
|
|
dummy_labels.view(-1),
|
|
)
|
|
|
|
assert not weights.requires_grad
|
|
|
|
def test_dft_weights_match_exp_neg_ce(self):
|
|
"""Test exp(-CE) matches softmax-gather for DFT weights."""
|
|
torch.manual_seed(123)
|
|
logits = torch.randn(2, 3, 7)
|
|
labels = torch.tensor([[1, 2, 3], [4, 5, 6]], dtype = torch.long)
|
|
|
|
ce_losses, valid_mask = fast_cross_entropy_loss_per_token(logits, labels)
|
|
|
|
weights_from_ce = _compute_dft_weights(
|
|
logits,
|
|
labels,
|
|
ce_losses = ce_losses,
|
|
valid_mask = valid_mask,
|
|
)
|
|
weights_from_softmax = _compute_dft_weights(logits, labels)
|
|
|
|
assert torch.allclose(weights_from_ce, weights_from_softmax, atol = 1e-4)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# A5) Test compute_asft_loss
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestComputeASFTLoss:
|
|
"""Tests for the main compute_asft_loss function."""
|
|
|
|
def test_sft_mode(self, simple_model):
|
|
"""Test SFT mode computes standard CE."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
loss = compute_asft_loss(simple_model, inputs, asft_mode = "sft", kl_weight = 0.0)
|
|
|
|
# Should return a scalar loss
|
|
assert loss.dim() == 0
|
|
assert loss.requires_grad
|
|
|
|
def test_dft_mode(self, simple_model):
|
|
"""Test DFT mode."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
loss = compute_asft_loss(simple_model, inputs, asft_mode = "dft", kl_weight = 0.0)
|
|
|
|
assert loss.dim() == 0
|
|
assert loss.requires_grad
|
|
|
|
def test_dft_normalize_by_weights(self, simple_model):
|
|
"""Test DFT normalization by weight sum."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
logits = simple_model(input_ids = inputs["input_ids"]).logits
|
|
shift_labels = build_shift_labels(inputs["labels"])
|
|
valid_mask = shift_labels != -100
|
|
ce_losses, _ = fast_cross_entropy_loss_per_token(logits, shift_labels)
|
|
ce_losses = ce_losses.view(shift_labels.shape)
|
|
dft_weights = _compute_dft_weights(
|
|
logits,
|
|
shift_labels,
|
|
ce_losses = ce_losses,
|
|
valid_mask = valid_mask,
|
|
).view(shift_labels.shape)
|
|
token_loss = ce_losses * dft_weights
|
|
expected = token_loss[valid_mask].sum() / dft_weights[valid_mask].sum().clamp_min(1e-8)
|
|
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "dft",
|
|
kl_weight = 0.0,
|
|
normalize_by = "weights",
|
|
)
|
|
|
|
assert torch.allclose(loss, expected, atol = 1e-5)
|
|
|
|
def test_sft_kl_mode(self, simple_model):
|
|
"""Test SFT+KL mode."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
)
|
|
|
|
assert loss.dim() == 0
|
|
assert loss.requires_grad
|
|
|
|
def test_asft_mode(self, simple_model):
|
|
"""Test full ASFT mode."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "asft",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
)
|
|
|
|
assert loss.dim() == 0
|
|
assert loss.requires_grad
|
|
|
|
def test_return_outputs(self, simple_model):
|
|
"""Test return_outputs=True."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
loss, outputs = compute_asft_loss(
|
|
simple_model, inputs, asft_mode = "sft", return_outputs = True
|
|
)
|
|
|
|
assert loss.dim() == 0
|
|
assert hasattr(outputs, "logits")
|
|
|
|
def test_handles_all_ignored_labels(self, simple_model):
|
|
"""Test that all -100 labels returns zero loss."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[-100, -100, -100, -100]]),
|
|
}
|
|
|
|
loss = compute_asft_loss(simple_model, inputs, asft_mode = "sft")
|
|
|
|
# Should return zero loss
|
|
assert loss.item() == 0.0
|
|
|
|
def test_uses_num_items_in_batch(self, simple_model):
|
|
"""Test that num_items_in_batch is used for normalization."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
"num_items_in_batch": 2, # Override default
|
|
}
|
|
|
|
loss = compute_asft_loss(simple_model, inputs, asft_mode = "sft")
|
|
|
|
# Should use the provided n_items
|
|
assert loss.dim() == 0
|
|
|
|
def test_packing_boundary_masking(self, simple_model):
|
|
"""Test that packed sequence boundaries are masked."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
"packed_seq_lengths": torch.tensor([2, 2], dtype = torch.int32),
|
|
}
|
|
|
|
loss = compute_asft_loss(simple_model, inputs, asft_mode = "sft")
|
|
|
|
# Should handle packing without error
|
|
assert loss.dim() == 0
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Test ASFTStreamingConfig
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestASFTStreamingConfig:
|
|
"""Tests for ASFTStreamingConfig dataclass."""
|
|
|
|
def test_default_values(self):
|
|
"""Test default configuration values."""
|
|
config = ASFTStreamingConfig()
|
|
|
|
assert config.mode is None
|
|
assert config.enabled is False
|
|
assert config.ref_strategy == "none"
|
|
assert config.ref_microbatch_size is None
|
|
assert config.seq_chunk_size is None
|
|
assert config.kl_token_chunk_size is None
|
|
assert config.force_fp32_kl is True
|
|
|
|
def test_custom_values(self):
|
|
"""Test custom configuration values."""
|
|
config = ASFTStreamingConfig(
|
|
mode = "batch",
|
|
enabled = True,
|
|
ref_strategy = "batch_micro",
|
|
ref_microbatch_size = 4,
|
|
seq_chunk_size = 256,
|
|
)
|
|
|
|
assert config.mode == "batch"
|
|
assert config.enabled is True
|
|
assert config.ref_strategy == "batch_micro"
|
|
assert config.ref_microbatch_size == 4
|
|
assert config.seq_chunk_size == 256
|
|
|
|
|
|
class TestStreamingModeMapping:
|
|
"""Tests for streaming mode routing in compute_asft_loss."""
|
|
|
|
def test_mode_batch_uses_batch_micro(self, simple_model):
|
|
"""Test that mode=batch routes to batch micro."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4], [2, 3, 4, 5]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4], [2, 3, 4, 5]]),
|
|
}
|
|
config = ASFTStreamingConfig(
|
|
mode = "batch",
|
|
ref_microbatch_size = 1,
|
|
enabled = False,
|
|
ref_strategy = "seq_kv_cache",
|
|
)
|
|
|
|
def batch_side_effect(
|
|
model,
|
|
cur_logits,
|
|
shift_labels,
|
|
valid_mask,
|
|
ref_forward,
|
|
forward_inputs,
|
|
microbatch_size,
|
|
logit_softcapping = 0,
|
|
logit_scaling = 0,
|
|
force_fp32 = True,
|
|
kl_direction = "forward",
|
|
):
|
|
batch, seq_len = shift_labels.shape
|
|
return torch.zeros(batch, seq_len, device = shift_labels.device)
|
|
|
|
with patch(
|
|
"unsloth.losses.asft._compute_kl_batch_micro",
|
|
side_effect = batch_side_effect,
|
|
) as batch_mock, patch(
|
|
"unsloth.losses.asft._compute_kl_seq_kv_cache",
|
|
side_effect = AssertionError("seq_kv_cache should not be used"),
|
|
):
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = config,
|
|
)
|
|
|
|
assert batch_mock.called
|
|
assert batch_mock.call_args[0][6] == 1
|
|
assert loss.dim() == 0
|
|
|
|
@pytest.mark.parametrize("mode", ["seq", "auto"])
|
|
def test_mode_seq_and_auto_use_seq_kv_cache(self, mode, simple_model):
|
|
"""Test that mode=seq/auto routes to seq_kv_cache."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
config = ASFTStreamingConfig(
|
|
mode = mode,
|
|
seq_chunk_size = 2,
|
|
enabled = False,
|
|
ref_strategy = "batch_micro",
|
|
)
|
|
|
|
def seq_side_effect(
|
|
model,
|
|
cur_logits,
|
|
shift_labels,
|
|
valid_mask,
|
|
ref_forward,
|
|
forward_inputs,
|
|
seq_chunk_size,
|
|
**kwargs,
|
|
):
|
|
batch, seq_len = shift_labels.shape
|
|
return torch.zeros(batch, seq_len, device = shift_labels.device)
|
|
|
|
with patch(
|
|
"unsloth.losses.asft._compute_kl_seq_kv_cache",
|
|
side_effect = seq_side_effect,
|
|
) as seq_mock, patch(
|
|
"unsloth.losses.asft._compute_kl_batch_micro",
|
|
side_effect = AssertionError("batch_micro should not be used"),
|
|
):
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = config,
|
|
)
|
|
|
|
assert seq_mock.called
|
|
assert seq_mock.call_args[0][6] == 2
|
|
assert seq_mock.call_args.kwargs["microbatch_size"] is None
|
|
assert loss.dim() == 0
|
|
|
|
def test_mode_hybrid_defaults_microbatch(self, simple_model):
|
|
"""Test that hybrid mode sets a default microbatch size."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4], [2, 3, 4, 5]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4], [2, 3, 4, 5]]),
|
|
}
|
|
config = ASFTStreamingConfig(
|
|
mode = "hybrid",
|
|
seq_chunk_size = 2,
|
|
ref_microbatch_size = None,
|
|
)
|
|
|
|
def seq_side_effect(
|
|
model,
|
|
cur_logits,
|
|
shift_labels,
|
|
valid_mask,
|
|
ref_forward,
|
|
forward_inputs,
|
|
seq_chunk_size,
|
|
**kwargs,
|
|
):
|
|
batch, seq_len = shift_labels.shape
|
|
return torch.zeros(batch, seq_len, device = shift_labels.device)
|
|
|
|
with patch(
|
|
"unsloth.losses.asft._compute_kl_seq_kv_cache",
|
|
side_effect = seq_side_effect,
|
|
) as seq_mock:
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = config,
|
|
)
|
|
|
|
assert seq_mock.called
|
|
assert seq_mock.call_args.kwargs["microbatch_size"] == 1
|
|
assert config.ref_microbatch_size is None
|
|
assert loss.dim() == 0
|
|
|
|
def test_mode_off_uses_full_forward(self, simple_model):
|
|
"""Test that mode=off bypasses streaming helpers."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
config = ASFTStreamingConfig(
|
|
mode = "off",
|
|
enabled = True,
|
|
ref_strategy = "seq_kv_cache",
|
|
)
|
|
|
|
def kl_side_effect(
|
|
cur_logits,
|
|
ref_logits,
|
|
model = None,
|
|
logit_softcapping = 0,
|
|
logit_scaling = 0,
|
|
force_fp32 = True,
|
|
kl_direction = "forward",
|
|
):
|
|
batch, seq_len = ref_logits.shape[:2]
|
|
return torch.zeros(batch * seq_len, device = ref_logits.device)
|
|
|
|
with patch(
|
|
"unsloth.losses.asft._compute_kl_divergence",
|
|
side_effect = kl_side_effect,
|
|
) as kl_mock, patch(
|
|
"unsloth.losses.asft._compute_kl_seq_kv_cache",
|
|
side_effect = AssertionError("seq_kv_cache should not be used"),
|
|
), patch(
|
|
"unsloth.losses.asft._compute_kl_batch_micro",
|
|
side_effect = AssertionError("batch_micro should not be used"),
|
|
):
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = config,
|
|
)
|
|
|
|
assert kl_mock.called
|
|
assert loss.dim() == 0
|
|
|
|
def test_invalid_mode_raises(self, simple_model):
|
|
"""Test that invalid streaming mode raises a ValueError."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
config = ASFTStreamingConfig(mode = "invalid")
|
|
|
|
with pytest.raises(ValueError):
|
|
compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = config,
|
|
)
|
|
|
|
|
|
class TestSeqKVCacheStreaming:
|
|
"""Tests for seq_kv_cache streaming behavior."""
|
|
|
|
def test_seq_kv_cache_runs_when_use_cache_false(self):
|
|
"""Test that seq_kv_cache attempts chunking even if config.use_cache=False."""
|
|
batch_size, seq_len, vocab_size = 1, 6, 5
|
|
cur_logits = torch.randn(batch_size, seq_len, vocab_size)
|
|
shift_labels = torch.zeros(batch_size, seq_len, dtype = torch.long)
|
|
valid_mask = shift_labels != -100
|
|
input_ids = torch.arange(seq_len).view(1, -1)
|
|
|
|
class DummyModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.config = SimpleNamespace(
|
|
use_cache = False,
|
|
final_logit_softcapping = 0,
|
|
logit_scale = 0,
|
|
)
|
|
|
|
model = DummyModel()
|
|
call_state = {"saw_past": False}
|
|
|
|
def ref_forward(**kwargs):
|
|
input_ids_local = kwargs["input_ids"]
|
|
if input_ids_local.shape[1] == seq_len:
|
|
raise AssertionError("full forward not expected")
|
|
if "past_key_values" in kwargs:
|
|
call_state["saw_past"] = True
|
|
batch, chunk_len = input_ids_local.shape
|
|
logits = torch.zeros(
|
|
batch, chunk_len, vocab_size, device = input_ids_local.device
|
|
)
|
|
return (logits, ("cache",))
|
|
|
|
forward_inputs = {"input_ids": input_ids}
|
|
|
|
kl = _compute_kl_seq_kv_cache(
|
|
model,
|
|
cur_logits,
|
|
shift_labels,
|
|
valid_mask,
|
|
ref_forward,
|
|
forward_inputs,
|
|
seq_chunk_size = 4,
|
|
)
|
|
|
|
assert kl.shape == (batch_size, seq_len)
|
|
assert call_state["saw_past"] is True
|
|
|
|
def test_seq_kv_cache_supports_microbatching(self):
|
|
"""Test that seq_kv_cache can be microbatched by batch dimension."""
|
|
batch_size, seq_len, vocab_size = 2, 4, 3
|
|
cur_logits = torch.randn(batch_size, seq_len, vocab_size)
|
|
shift_labels = torch.zeros(batch_size, seq_len, dtype = torch.long)
|
|
valid_mask = shift_labels != -100
|
|
input_ids = torch.arange(seq_len).repeat(batch_size, 1)
|
|
|
|
class DummyModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.config = SimpleNamespace(
|
|
use_cache = True,
|
|
final_logit_softcapping = 0,
|
|
logit_scale = 0,
|
|
)
|
|
|
|
model = DummyModel()
|
|
call_state = {"max_batch": 0}
|
|
|
|
def ref_forward(**kwargs):
|
|
input_ids_local = kwargs["input_ids"]
|
|
call_state["max_batch"] = max(
|
|
call_state["max_batch"], input_ids_local.shape[0]
|
|
)
|
|
if input_ids_local.shape[0] > 1:
|
|
raise AssertionError("expected microbatching")
|
|
batch, chunk_len = input_ids_local.shape
|
|
logits = torch.zeros(
|
|
batch, chunk_len, vocab_size, device = input_ids_local.device
|
|
)
|
|
return (logits, ("cache",))
|
|
|
|
forward_inputs = {"input_ids": input_ids}
|
|
|
|
kl = _compute_kl_seq_kv_cache(
|
|
model,
|
|
cur_logits,
|
|
shift_labels,
|
|
valid_mask,
|
|
ref_forward,
|
|
forward_inputs,
|
|
seq_chunk_size = 2,
|
|
microbatch_size = 1,
|
|
)
|
|
|
|
assert kl.shape == (batch_size, seq_len)
|
|
assert call_state["max_batch"] == 1
|
|
|
|
def test_seq_kv_cache_falls_back_to_batch_micro(self):
|
|
"""Test that seq_kv_cache falls back to batch micro on cache failure."""
|
|
batch_size, seq_len, vocab_size = 2, 6, 5
|
|
cur_logits = torch.randn(batch_size, seq_len, vocab_size)
|
|
shift_labels = torch.zeros(batch_size, seq_len, dtype = torch.long)
|
|
valid_mask = shift_labels != -100
|
|
input_ids = torch.arange(seq_len).repeat(batch_size, 1)
|
|
|
|
class DummyModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.config = SimpleNamespace(
|
|
use_cache = True,
|
|
final_logit_softcapping = 0,
|
|
logit_scale = 0,
|
|
)
|
|
|
|
model = DummyModel()
|
|
|
|
def ref_forward(**kwargs):
|
|
input_ids_local = kwargs["input_ids"]
|
|
if (
|
|
input_ids_local.shape[0] == batch_size
|
|
and input_ids_local.shape[1] == seq_len
|
|
):
|
|
raise AssertionError("full forward not expected on fallback")
|
|
batch, chunk_len = input_ids_local.shape
|
|
logits = torch.zeros(
|
|
batch, chunk_len, vocab_size, device = input_ids_local.device
|
|
)
|
|
return (logits, None)
|
|
|
|
forward_inputs = {"input_ids": input_ids}
|
|
|
|
kl = _compute_kl_seq_kv_cache(
|
|
model,
|
|
cur_logits,
|
|
shift_labels,
|
|
valid_mask,
|
|
ref_forward,
|
|
forward_inputs,
|
|
seq_chunk_size = 2,
|
|
)
|
|
|
|
assert kl.shape == (batch_size, seq_len)
|
|
|
|
def test_config_immutability_when_none_values(self, simple_model):
|
|
"""Test that streaming_config is not mutated when values are None."""
|
|
config = ASFTStreamingConfig(
|
|
enabled = True,
|
|
ref_strategy = "batch_micro",
|
|
ref_microbatch_size = None, # Should use default without mutation
|
|
)
|
|
original_microbatch = config.ref_microbatch_size
|
|
original_chunk = config.seq_chunk_size
|
|
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
# Call compute_asft_loss with sft mode (doesn't use streaming, but
|
|
# the config should still not be mutated)
|
|
loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft",
|
|
streaming_config = config,
|
|
)
|
|
|
|
# Config should not be mutated
|
|
assert config.ref_microbatch_size == original_microbatch
|
|
assert config.seq_chunk_size == original_chunk
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Backward Compatibility Tests
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestBackwardCompatibility:
|
|
"""Tests to ensure ASFT doesn't break existing behavior."""
|
|
|
|
def test_sft_mode_matches_standard_ce(self, simple_model):
|
|
"""Test that SFT mode produces same results as standard CE."""
|
|
torch.manual_seed(42)
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
# Compute ASFT loss in SFT mode
|
|
asft_loss = compute_asft_loss(simple_model, inputs, asft_mode = "sft")
|
|
|
|
# The loss should be a valid scalar
|
|
assert asft_loss.dim() == 0
|
|
assert not torch.isnan(asft_loss)
|
|
assert not torch.isinf(asft_loss)
|
|
|
|
def test_streaming_equivalence(self, simple_model):
|
|
"""Test that streaming produces equivalent results to full forward."""
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
# Full forward
|
|
full_loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = ASFTStreamingConfig(enabled = False),
|
|
)
|
|
|
|
# With batch_micro streaming (should be equivalent for batch=1)
|
|
streaming_loss = compute_asft_loss(
|
|
simple_model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = ASFTStreamingConfig(
|
|
enabled = True,
|
|
ref_strategy = "batch_micro",
|
|
ref_microbatch_size = 1,
|
|
),
|
|
)
|
|
|
|
# Should be very close
|
|
assert torch.allclose(full_loss, streaming_loss, atol = 1e-4)
|
|
|
|
def test_seq_kv_cache_equivalence(self):
|
|
"""Test that seq_kv_cache matches full forward for KL loss."""
|
|
torch.manual_seed(123)
|
|
|
|
class CacheModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.config = SimpleNamespace(
|
|
use_cache = True,
|
|
final_logit_softcapping = 0,
|
|
logit_scale = 0,
|
|
)
|
|
self.embedding = nn.Embedding(32, 8)
|
|
self.linear = nn.Linear(8, 32)
|
|
|
|
def forward(
|
|
self, input_ids = None, past_key_values = None, use_cache = None, **kwargs
|
|
):
|
|
embeddings = self.embedding(input_ids)
|
|
logits = self.linear(embeddings)
|
|
past = ("cache",) if (use_cache or past_key_values is not None) else None
|
|
return SimpleNamespace(logits = logits, past_key_values = past)
|
|
|
|
model = CacheModel()
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4, 5, 6], [6, 5, 4, 3, 2, 1]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4, 5, 6], [6, 5, 4, 3, 2, 1]]),
|
|
}
|
|
|
|
full_loss = compute_asft_loss(
|
|
model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = ASFTStreamingConfig(enabled = False),
|
|
)
|
|
|
|
seq_loss = compute_asft_loss(
|
|
model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = ASFTStreamingConfig(
|
|
enabled = True,
|
|
ref_strategy = "seq_kv_cache",
|
|
seq_chunk_size = 2,
|
|
),
|
|
)
|
|
|
|
assert torch.allclose(full_loss, seq_loss, atol = 1e-4)
|
|
|
|
def test_seq_kv_cache_microbatch_equivalence(self):
|
|
"""Test that seq_kv_cache + microbatching matches full forward."""
|
|
torch.manual_seed(456)
|
|
|
|
class CacheModel(nn.Module):
|
|
def __init__(self):
|
|
super().__init__()
|
|
self.config = SimpleNamespace(
|
|
use_cache = True,
|
|
final_logit_softcapping = 0,
|
|
logit_scale = 0,
|
|
)
|
|
self.embedding = nn.Embedding(32, 8)
|
|
self.linear = nn.Linear(8, 32)
|
|
|
|
def forward(
|
|
self, input_ids = None, past_key_values = None, use_cache = None, **kwargs
|
|
):
|
|
embeddings = self.embedding(input_ids)
|
|
logits = self.linear(embeddings)
|
|
past = ("cache",) if (use_cache or past_key_values is not None) else None
|
|
return SimpleNamespace(logits = logits, past_key_values = past)
|
|
|
|
model = CacheModel()
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4, 5, 6], [6, 5, 4, 3, 2, 1]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4, 5, 6], [6, 5, 4, 3, 2, 1]]),
|
|
}
|
|
|
|
full_loss = compute_asft_loss(
|
|
model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = ASFTStreamingConfig(enabled = False),
|
|
)
|
|
|
|
combined_loss = compute_asft_loss(
|
|
model,
|
|
inputs,
|
|
asft_mode = "sft+kl",
|
|
kl_weight = 0.1,
|
|
reference_policy = "frozen_copy",
|
|
streaming_config = ASFTStreamingConfig(
|
|
enabled = True,
|
|
ref_strategy = "seq_kv_cache",
|
|
seq_chunk_size = 2,
|
|
ref_microbatch_size = 1,
|
|
),
|
|
)
|
|
|
|
assert torch.allclose(full_loss, combined_loss, atol = 1e-4)
|
|
|
|
|
|
# -----------------------------------------------------------------------------
|
|
# Integration Tests
|
|
# -----------------------------------------------------------------------------
|
|
|
|
|
|
class TestASFTTrainerIntegration:
|
|
"""Integration tests for ASFTTrainer."""
|
|
|
|
def test_import_asft_trainer(self):
|
|
"""Test that ASFTTrainer can be imported."""
|
|
from unsloth.trainer import ASFTTrainer, ASFTStreamingConfig
|
|
|
|
assert ASFTTrainer is not None
|
|
assert ASFTStreamingConfig is not None
|
|
|
|
def test_asft_trainer_inherits_unsloth_trainer(self):
|
|
"""Test that ASFTTrainer inherits from UnslothTrainer."""
|
|
from unsloth.trainer import ASFTTrainer, UnslothTrainer
|
|
|
|
assert issubclass(ASFTTrainer, UnslothTrainer)
|
|
|
|
|
|
class TestASFTTrainerComputeLoss:
|
|
"""Tests for ASFTTrainer.compute_loss behavior."""
|
|
|
|
def test_compute_loss_calls_asft_loss(self):
|
|
"""Test ASFTTrainer compute_loss calls compute_asft_loss."""
|
|
from unsloth.trainer import ASFTTrainer, ASFTStreamingConfig
|
|
|
|
trainer = ASFTTrainer.__new__(ASFTTrainer)
|
|
trainer.asft_enabled = True
|
|
trainer.asft_mode = "sft"
|
|
trainer.kl_weight = 0.0
|
|
trainer.kl_direction = "forward"
|
|
trainer.reference_policy = "disable_adapter"
|
|
trainer.asft_streaming = ASFTStreamingConfig()
|
|
trainer.normalize_by = "tokens"
|
|
trainer._asft_original_model = None
|
|
|
|
model = nn.Module()
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
expected = torch.tensor(1.0, device = inputs["input_ids"].device)
|
|
|
|
with patch(
|
|
"unsloth.trainer.compute_asft_loss", return_value = expected
|
|
) as loss_mock:
|
|
result = ASFTTrainer.compute_loss(
|
|
trainer, model, inputs, return_outputs = False, num_items_in_batch = 7
|
|
)
|
|
|
|
assert result is expected
|
|
assert inputs["num_items_in_batch"] == 7
|
|
assert loss_mock.called
|
|
assert loss_mock.call_args.kwargs["model"] is model
|
|
assert loss_mock.call_args.kwargs["asft_mode"] == "sft"
|
|
assert loss_mock.call_args.kwargs["kl_weight"] == 0.0
|
|
assert loss_mock.call_args.kwargs["kl_direction"] == "forward"
|
|
assert loss_mock.call_args.kwargs["reference_policy"] == "disable_adapter"
|
|
assert loss_mock.call_args.kwargs["streaming_config"] is trainer.asft_streaming
|
|
assert loss_mock.call_args.kwargs["normalize_by"] == "tokens"
|
|
|
|
def test_compute_loss_creates_frozen_copy_once(self):
|
|
"""Test frozen copy is created once when needed."""
|
|
from unsloth.trainer import ASFTTrainer, ASFTStreamingConfig
|
|
|
|
trainer = ASFTTrainer.__new__(ASFTTrainer)
|
|
trainer.asft_enabled = True
|
|
trainer.asft_mode = "asft"
|
|
trainer.kl_weight = 0.1
|
|
trainer.kl_direction = "forward"
|
|
trainer.reference_policy = "frozen_copy"
|
|
trainer.asft_streaming = ASFTStreamingConfig()
|
|
trainer.normalize_by = "tokens"
|
|
trainer._asft_original_model = None
|
|
|
|
model = nn.Module()
|
|
model_copy = MagicMock()
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
with pytest.warns(UserWarning), patch(
|
|
"unsloth.trainer.deepcopy", return_value = model_copy
|
|
) as deepcopy_mock, patch(
|
|
"unsloth.trainer.compute_asft_loss",
|
|
return_value = torch.tensor(0.5, device = inputs["input_ids"].device),
|
|
):
|
|
ASFTTrainer.compute_loss(trainer, model, inputs)
|
|
ASFTTrainer.compute_loss(trainer, model, inputs)
|
|
|
|
assert deepcopy_mock.call_count == 1
|
|
assert trainer._asft_original_model is model_copy
|
|
assert model_copy.eval.called
|
|
assert model_copy.requires_grad_.called
|
|
|
|
def test_compute_loss_skips_copy_with_disable_adapter(self):
|
|
"""Test disable_adapter policy skips frozen copy."""
|
|
from unsloth.trainer import ASFTTrainer, ASFTStreamingConfig
|
|
|
|
trainer = ASFTTrainer.__new__(ASFTTrainer)
|
|
trainer.asft_enabled = True
|
|
trainer.asft_mode = "asft"
|
|
trainer.kl_weight = 0.1
|
|
trainer.kl_direction = "forward"
|
|
trainer.reference_policy = "disable_adapter"
|
|
trainer.asft_streaming = ASFTStreamingConfig()
|
|
trainer.normalize_by = "tokens"
|
|
trainer._asft_original_model = None
|
|
|
|
model = MagicMock()
|
|
model.disable_adapter = MagicMock()
|
|
inputs = {
|
|
"input_ids": torch.tensor([[1, 2, 3, 4]]),
|
|
"labels": torch.tensor([[1, 2, 3, 4]]),
|
|
}
|
|
|
|
with patch(
|
|
"unsloth.trainer.deepcopy"
|
|
) as deepcopy_mock, patch(
|
|
"unsloth.trainer.compute_asft_loss",
|
|
return_value = torch.tensor(0.5, device = inputs["input_ids"].device),
|
|
) as loss_mock:
|
|
ASFTTrainer.compute_loss(trainer, model, inputs)
|
|
|
|
assert not deepcopy_mock.called
|
|
assert loss_mock.call_args.kwargs["original_model"] is None
|
|
|
|
|
|
class TestUnslothTrainingArguments:
|
|
"""Tests for UnslothTrainingArguments."""
|
|
|
|
def test_embedding_learning_rate_is_set(self):
|
|
"""Test embedding_learning_rate is stored on the args object."""
|
|
from unsloth import trainer as trainer_module
|
|
|
|
with patch.object(
|
|
trainer_module.TrainingArguments, "__init__", return_value = None
|
|
) as base_init:
|
|
args = trainer_module.UnslothTrainingArguments(embedding_learning_rate = 0.01)
|
|
|
|
assert args.embedding_learning_rate == 0.01
|
|
assert base_init.called
|
|
|
|
|
|
if __name__ == "__main__":
|
|
pytest.main([__file__, "-v"])
|