Doing the loss reduction in foundry instead of in the loss functions.

llmfoundry/models/mpt/modeling_mpt.py

@@ -984,7 +984,8 @@ def __init__(
                 from flash_attn.losses.cross_entropy import \
                     CrossEntropyLoss as FusedCrossEntropyLoss
 
-                self.loss_fn = FusedCrossEntropyLoss(ignore_index=-100)
+                self.loss_fn = FusedCrossEntropyLoss(ignore_index=-100,
+                                                     reduction='none')
             except:
                 raise ValueError(
                     'Fused Cross Entropy is not installed. Either (1) have a CUDA-compatible GPU '
@@ -994,7 +995,8 @@ def __init__(
                     'if installing from pypi, or (2) set your config model.loss_fn=torch_crossentropy.'
                 )
         elif loss_fn_config == 'torch_crossentropy':
-            self.loss_fn = nn.CrossEntropyLoss(ignore_index=-100)
+            self.loss_fn = nn.CrossEntropyLoss(ignore_index=-100,
+                                               reduction='none')
         else:
             raise ValueError(
                 f'Specified loss_fn={self.loss_fn} not recognized. `loss_fn` must be one of [`fused_crossentropy`, `torch_crossentropy`].'
@@ -1016,8 +1018,15 @@ def forward(self, batch: MutableMapping) -> CausalLMOutputWithPast:
     def loss(self, outputs: CausalLMOutputWithPast,
              batch: Mapping) -> torch.Tensor:
         targets = self.get_targets(batch)
-        return self.loss_fn(outputs.logits.view(-1, outputs.logits.size(-1)),
-                            targets.view(-1))
+        losses = self.loss_fn(outputs.logits.view(-1, outputs.logits.size(-1)),
+                              targets.view(-1))
+
+        if torch.all(targets == self.loss_fn.ignore_index):
+            loss = losses.sum()
+        else:
+            loss = losses.sum() / (targets != self.loss_fn.ignore_index).sum()
+
+        return loss
 
     def flops_per_batch(self, batch: Mapping) -> int:
         # Note: this computation does not take into account padding, and assumes

tests/models/test_model.py

@@ -480,7 +480,7 @@ def test_loss_fn():
     model_2.to(test_cfg.device)
 
     assert isinstance(model_1.loss_fn, torch.nn.CrossEntropyLoss)
-    model_2.loss_fn = FusedCrossEntropyLoss(ignore_index=-100)
+    model_2.loss_fn = FusedCrossEntropyLoss(ignore_index=-100, reduction='none')
 
     optimizer_1 = DecoupledAdamW(model_1.parameters(),
                                  lr=test_cfg.optimizer.lr,
@@ -517,6 +517,105 @@ def test_loss_fn():
                                     atol=1e-4), f'differed at step {i}'
 
 
+@pytest.mark.gpu
+@pytest.mark.parametrize('loss_fn_config',
+                         ['torch_crossentropy', 'fused_crossentropy'])
+def test_loss_reduction(loss_fn_config: str):
+    """Tests the Fused CrossEntropy vs torch.nn.CrossEntropy loss function.
+
+    We provide non-zero tolerances to account for small numerics differences
+    between the two loss implementations.
+    """
+    try:
+        from flash_attn.losses.cross_entropy import CrossEntropyLoss as FusedCrossEntropyLoss  # type: ignore # isort: skip
+    except:
+        pytest.skip('Fused cross entropy was not installed')
+
+    # run numerical test in pure fp32
+    from torch.backends import cuda, cudnn
+    cuda.matmul.allow_tf32 = False
+    cudnn.allow_tf32 = False
+
+    conf_path = 'scripts/train/yamls/pretrain/testing.yaml'
+    with open(conf_path) as f:
+        test_cfg = om.load(f)
+
+    assert isinstance(test_cfg, DictConfig)
+
+    test_cfg.model.loss_fn = loss_fn_config
+
+    test_cfg.device = 'cuda:0'
+    test_cfg.model.init_device = 'cpu'
+    test_cfg.model.init_config = {
+        'name': 'baseline_',
+        'init_std': 0.02,
+    }
+
+    tokenizer_cfg: Dict[str, Any] = _load_tokenizer_cfg(test_cfg.tokenizer)
+    tokenizer = build_tokenizer(test_cfg.tokenizer.name,
+                                tokenizer_cfg.get('kwargs', {}))
+
+    model_1 = build_composer_model(
+        name=test_cfg.model.name,
+        cfg=test_cfg.model,
+        tokenizer=tokenizer,
+    )
+    model_2 = copy.deepcopy(model_1)
+
+    model_1.to(test_cfg.device)
+    model_2.to(test_cfg.device)
+
+    # Reduce the loss in FusedCrossEntropyLoss
+    if loss_fn_config == 'fused_crossentropy':
+        assert isinstance(model_1.loss_fn, FusedCrossEntropyLoss)
+        model_2.loss_fn = FusedCrossEntropyLoss(ignore_index=-100,
+                                                reduction='mean')
+    else:
+        assert isinstance(model_1.loss_fn, torch.nn.CrossEntropyLoss)
+        model_2.loss_fn = torch.nn.CrossEntropyLoss(ignore_index=-100,
+                                                    reduction='mean')
+
+    optimizer_1 = DecoupledAdamW(model_1.parameters(),
+                                 lr=test_cfg.optimizer.lr,
+                                 betas=test_cfg.optimizer.betas,
+                                 eps=test_cfg.optimizer.eps,
+                                 weight_decay=test_cfg.optimizer.weight_decay)
+    optimizer_2 = DecoupledAdamW(model_2.parameters(),
+                                 lr=test_cfg.optimizer.lr,
+                                 betas=test_cfg.optimizer.betas,
+                                 eps=test_cfg.optimizer.eps,
+                                 weight_decay=test_cfg.optimizer.weight_decay)
+
+    for i in range(3):
+        batch = gen_random_batch(2, test_cfg)
+        output_1 = model_1(batch)
+        output_2 = model_2(batch)
+        assert output_1.logits.allclose(output_2.logits, rtol=1e-4,
+                                        atol=1e-4), f'differed at step {i}'
+
+        loss_1 = model_1.loss(output_1, batch)
+
+        # Loss for model_2 gets reduced within the loss_fn, so we handle it separately
+        targets = model_2.get_targets(batch)
+        loss_2 = model_2.loss_fn(
+            output_2.logits.view(-1, output_2.logits.size(-1)),
+            targets.view(-1))
+
+        assert isinstance(loss_1, torch.Tensor)
+        assert isinstance(loss_2, torch.Tensor)
+        assert loss_1.allclose(loss_2, rtol=1e-3,
+                               atol=1e-3), f'differed at step {i}'
+        loss_1.backward()
+        loss_2.backward()
+        optimizer_1.step()
+        optimizer_2.step()
+
+        for p1, p2 in zip(model_1.parameters(), model_2.parameters()):
+            assert p1.data.shape == p2.data.shape
+            assert p1.data.allclose(p2.data, rtol=1e-5,
+                                    atol=1e-4), f'differed at step {i}'
+
+
 @pytest.mark.parametrize('peft_config', [
     None,
     {