fix losses

tests/conftest.py

@@ -119,15 +119,49 @@ def download_smollm2_135m_gptq_model():
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_qwen_2_5_half_billion_model():
+def download_qwen3_half_billion_model():
-    # download the model
+    # download the model (still used as the KD teacher in tests/e2e/integrations/test_kd.py)
-    snapshot_download_w_retry("Qwen/Qwen2.5-0.5B", repo_type="model")
+    snapshot_download_w_retry("Qwen/Qwen3-0.6B", repo_type="model")
 
 
 @pytest.fixture(scope="session", autouse=True)
-def download_qwen3_half_billion_model():
+def download_tiny_llama_model():
-    # download the model
+    snapshot_download_w_retry("axolotl-ai-co/tiny-llama-50m", repo_type="model")
-    snapshot_download_w_retry("Qwen/Qwen3-0.6B", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_mistral_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-mistral-25m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_mixtral_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-mixtral-30m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_phi_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-phi-64m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_falcon_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-falcon-42m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_qwen2_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-qwen2-129m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_qwen3_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-qwen3-129m", repo_type="model")
+
+
+@pytest.fixture(scope="session", autouse=True)
+def download_tiny_gemma2_model():
+    snapshot_download_w_retry("axolotl-ai-co/tiny-gemma2-137m", repo_type="model")
 
 
 @pytest.fixture(scope="session", autouse=True)
@@ -620,7 +654,15 @@ def fixture_min_base_cfg():
 )
 def test_load_fixtures(
     download_smollm2_135m_model,
-    download_qwen_2_5_half_billion_model,
+    download_qwen3_half_billion_model,
+    download_tiny_llama_model,
+    download_tiny_mistral_model,
+    download_tiny_mixtral_model,
+    download_tiny_phi_model,
+    download_tiny_falcon_model,
+    download_tiny_qwen2_model,
+    download_tiny_qwen3_model,
+    download_tiny_gemma2_model,
     download_tatsu_lab_alpaca_dataset,
     download_mhenrichsen_alpaca_2k_dataset,
     download_mhenrichsen_alpaca_2k_w_revision_dataset,

tests/e2e/integrations/test_cut_cross_entropy.py

@@ -10,7 +10,10 @@ from axolotl.utils import get_pytorch_version
 from axolotl.utils.config import normalize_config, prepare_plugins, validate_config
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import check_model_output_exists
+from tests.e2e.utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+)
 
 
 @pytest.fixture()
@@ -35,13 +38,16 @@ def min_cfg(temp_dir):
         "num_epochs": 1,
         "micro_batch_size": 8,
         "gradient_accumulation_steps": 1,
-        "learning_rate": 0.00001,
+        "learning_rate": 5e-4,
         "optimizer": "adamw_torch_fused",
         "output_dir": temp_dir,
         "lr_scheduler": "cosine",
-        "max_steps": 10,
+        "max_steps": 40,
+        "warmup_steps": 5,
         "bf16": "auto",
         "save_first_step": False,
+        "use_tensorboard": True,
+        "seed": 42,
     }
 
 
@@ -64,11 +70,18 @@ class TestCutCrossEntropyIntegration:
         else:
             train(cfg=cfg, dataset_meta=dataset_meta)
             check_model_output_exists(temp_dir, cfg)
+            check_tensorboard_loss_decreased(
+                temp_dir + "/runs",
+                initial_window=5,
+                final_window=5,
+                max_initial=2.2,
+                max_final=2.0,
+            )
 
     def test_qwen2_w_cce(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "plugins": [
                     "axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin",
                 ],
@@ -87,13 +100,15 @@ class TestCutCrossEntropyIntegration:
                 "num_epochs": 1,
                 "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "output_dir": temp_dir,
                 "lr_scheduler": "cosine",
-                "max_steps": 10,
+                "max_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -108,6 +123,13 @@ class TestCutCrossEntropyIntegration:
         else:
             train(cfg=cfg, dataset_meta=dataset_meta)
             check_model_output_exists(temp_dir, cfg)
+            check_tensorboard_loss_decreased(
+                temp_dir + "/runs",
+                initial_window=5,
+                final_window=5,
+                max_initial=5.0,
+                max_final=4.7,
+            )
 
     @pytest.mark.parametrize(
         "attention_type",
@@ -136,3 +158,10 @@ class TestCutCrossEntropyIntegration:
         else:
             train(cfg=cfg, dataset_meta=dataset_meta)
             check_model_output_exists(temp_dir, cfg)
+            check_tensorboard_loss_decreased(
+                temp_dir + "/runs",
+                initial_window=5,
+                final_window=5,
+                max_initial=2.2,
+                max_final=2.0,
+            )

tests/e2e/kernels/test_lora_features.py

@@ -24,7 +24,7 @@ from axolotl.monkeypatch.lora_kernels import (
 )
 from axolotl.utils.dict import DictDefault
 
-MODEL_NAME = "Qwen/Qwen3-0.6B"
+MODEL_NAME = "axolotl-ai-co/tiny-qwen3-129m"
 DEVICE = "cuda"
 DTYPE = torch.bfloat16
 

tests/e2e/multigpu/test_dist_muon_fsdp2.py

@@ -1,23 +1,22 @@
 """Test module for DistMuon optimizer with FSDP2 multi-GPU functionality."""
 
-import os
 from pathlib import Path
 
-import torch
 import yaml
 from accelerate.test_utils import execute_subprocess_async
-from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import most_recent_subdir, require_torch_2_7_0
+from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
 def verify_training_success(temp_dir):
-    """Verify that training completed successfully by checking artifacts and loss."""
+    """Verify that training completed successfully — artifacts, no-NaN, loss
+    stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
+    """
     output_path = Path(temp_dir)
 
     model_files = list(output_path.glob("*.bin")) + list(
@@ -30,19 +29,13 @@ def verify_training_success(temp_dir):
         "No checkpoint files found - training may have failed"
     )
 
-    tb_log_path = most_recent_subdir(temp_dir + "/runs")
+    check_tensorboard_loss_decreased(
-    if tb_log_path:
+        temp_dir + "/runs",
-        event_files = sorted(os.listdir(tb_log_path))
+        initial_window=10,
-        if event_files:
+        final_window=10,
-            event_file = os.path.join(tb_log_path, event_files[0])
+        max_initial=5.0,
-            reader = SummaryReader(event_file)
+        max_final=4.7,
-            df = reader.scalars
+    )
-            train_loss_df = df[df.tag == "train/train_loss"]
-            if len(train_loss_df) > 0:
-                final_loss = train_loss_df.value.values[-1]
-                assert not torch.isnan(torch.tensor(final_loss)), (
-                    f"Training loss is NaN: {final_loss}"
-                )
 
 
 class TestDistMuon:
@@ -52,7 +45,7 @@ class TestDistMuon:
     def test_fft_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -63,11 +56,12 @@ class TestDistMuon:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.02,
+                "learning_rate": 2e-3,
                 "optimizer": "muon",
                 "weight_decay": 0.01,
                 "lr_scheduler": "cosine",
@@ -82,6 +76,9 @@ class TestDistMuon:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -109,7 +106,7 @@ class TestDistMuon:
     def test_lora_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -122,14 +119,15 @@ class TestDistMuon:
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.02,
+                "learning_rate": 2e-3,
                 "optimizer": "muon",
                 "weight_decay": 0.01,
                 "lr_scheduler": "cosine",
@@ -144,6 +142,9 @@ class TestDistMuon:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )

tests/e2e/multigpu/test_fsdp1.py

@@ -1,24 +1,23 @@
 """Test module for FSDP1 multi-GPU functionality."""
 
-import os
 from pathlib import Path
 
 import pytest
-import torch
 import yaml
 from accelerate.test_utils import execute_subprocess_async
-from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import most_recent_subdir
+from tests.e2e.utils import check_tensorboard_loss_decreased
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
 def verify_training_success(temp_dir):
-    """Verify that training completed successfully by checking artifacts and loss."""
+    """Verify that training completed successfully — artifacts, no-NaN, loss
+    stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
+    """
     output_path = Path(temp_dir)
 
     model_files = list(output_path.glob("*.bin")) + list(
@@ -31,19 +30,13 @@ def verify_training_success(temp_dir):
         "No checkpoint files found - training may have failed"
     )
 
-    tb_log_path = most_recent_subdir(temp_dir + "/runs")
+    check_tensorboard_loss_decreased(
-    if tb_log_path:
+        temp_dir + "/runs",
-        event_files = sorted(os.listdir(tb_log_path))
+        initial_window=10,
-        if event_files:
+        final_window=10,
-            event_file = os.path.join(tb_log_path, event_files[0])
+        max_initial=5.0,
-            reader = SummaryReader(event_file)
+        max_final=4.7,
-            df = reader.scalars
+    )
-            train_loss_df = df[df.tag == "train/train_loss"]
-            if len(train_loss_df) > 0:
-                final_loss = train_loss_df.value.values[-1]
-                assert not torch.isnan(torch.tensor(final_loss)), (
-                    f"Training loss is NaN: {final_loss}"
-                )
 
 
 class TestFSDP1:
@@ -56,7 +49,7 @@ class TestFSDP1:
     def test_fft_sft(self, temp_dir, fsdp_cpu_ram_efficient_loading):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -67,11 +60,12 @@ class TestFSDP1:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -87,6 +81,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -126,7 +123,7 @@ class TestFSDP1:
     def test_lora_sft(self, temp_dir, adapter_config):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -140,14 +137,15 @@ class TestFSDP1:
                 "load_in_4bit": adapter_config["load_in_4bit"],
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -163,6 +161,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -190,7 +191,7 @@ class TestFSDP1:
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "rl": "dpo",
@@ -203,11 +204,11 @@ class TestFSDP1:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -223,6 +224,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
             }
         )
 
@@ -262,7 +266,7 @@ class TestFSDP1:
     def test_dpo_lora(self, temp_dir, adapter_config):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "load_in_4bit": adapter_config["load_in_4bit"],
                 "rl": "dpo",
                 "chat_template": "chatml",
@@ -281,11 +285,11 @@ class TestFSDP1:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -301,6 +305,9 @@ class TestFSDP1:
                     "fsdp_use_orig_params": False,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": "auto",
                 "tf32": True,
             }

tests/e2e/multigpu/test_fsdp2.py

@@ -1,24 +1,23 @@
 """Test module for FSDP2 multi-GPU functionality."""
 
-import os
 from pathlib import Path
 
 import pytest
-import torch
 import yaml
 from accelerate.test_utils import execute_subprocess_async
-from tbparse import SummaryReader
 from transformers.testing_utils import get_torch_dist_unique_port
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import most_recent_subdir, require_torch_2_7_0
+from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
 
 AXOLOTL_ROOT = Path(__file__).parent.parent.parent.parent
 
 
 def verify_training_success(temp_dir):
-    """Verify that training completed successfully by checking artifacts and loss."""
+    """Verify that training completed successfully — artifacts, no-NaN, loss
+    stayed in qwen2-pretraining scale (tiny-qwen2-129m final pretrain CE ~3.92).
+    """
     output_path = Path(temp_dir)
 
     model_files = list(output_path.glob("*.bin")) + list(
@@ -31,19 +30,13 @@ def verify_training_success(temp_dir):
         "No checkpoint files found - training may have failed"
     )
 
-    tb_log_path = most_recent_subdir(temp_dir + "/runs")
+    check_tensorboard_loss_decreased(
-    if tb_log_path:
+        temp_dir + "/runs",
-        event_files = sorted(os.listdir(tb_log_path))
+        initial_window=10,
-        if event_files:
+        final_window=10,
-            event_file = os.path.join(tb_log_path, event_files[0])
+        max_initial=5.0,
-            reader = SummaryReader(event_file)
+        max_final=4.7,
-            df = reader.scalars
+    )
-            train_loss_df = df[df.tag == "train/train_loss"]
-            if len(train_loss_df) > 0:
-                final_loss = train_loss_df.value.values[-1]
-                assert not torch.isnan(torch.tensor(final_loss)), (
-                    f"Training loss is NaN: {final_loss}"
-                )
 
 
 class TestFSDP2:
@@ -57,7 +50,7 @@ class TestFSDP2:
     def test_fft_sft(self, temp_dir, fsdp_cpu_ram_efficient_loading):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -68,11 +61,12 @@ class TestFSDP2:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -86,6 +80,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -114,7 +111,7 @@ class TestFSDP2:
     def test_lora_sft(self, temp_dir, peft_use_dora):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -128,14 +125,15 @@ class TestFSDP2:
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -149,6 +147,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
                 # explicitly disable LORA kernels, as they may be auto-enabled
                 "lora_mlp_kernel": False,
@@ -180,7 +181,7 @@ class TestFSDP2:
     def test_lora_sft_kernels(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -195,11 +196,12 @@ class TestFSDP2:
                 "lora_alpha": 16,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -213,6 +215,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
                 "lora_mlp_kernel": True,
                 "lora_qkv_kernel": True,
@@ -243,7 +248,7 @@ class TestFSDP2:
     def test_qlora_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -257,14 +262,15 @@ class TestFSDP2:
                 "adapter": "qlora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -278,6 +284,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
             }
         )
@@ -305,7 +314,7 @@ class TestFSDP2:
     def test_qlora_sft_kernels(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -321,11 +330,12 @@ class TestFSDP2:
                 "lora_alpha": 16,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 80,
+                "warmup_steps": 5,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -339,6 +349,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "bf16": True,
                 "lora_mlp_kernel": True,
                 "lora_qkv_kernel": True,
@@ -370,7 +383,7 @@ class TestFSDP2:
     def test_dpo_fft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "rl": "dpo",
@@ -383,11 +396,11 @@ class TestFSDP2:
                     },
                 ],
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -401,6 +414,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
             }
         )
 
@@ -428,7 +444,7 @@ class TestFSDP2:
     def test_dpo_lora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "rl": "dpo",
                 "chat_template": "chatml",
@@ -445,11 +461,11 @@ class TestFSDP2:
                 "lora_dropout": 0.05,
                 "lora_target_linear": True,
                 "num_epochs": 1,
-                "max_steps": 2,
+                "max_steps": 20,
                 "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-3,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
@@ -463,6 +479,9 @@ class TestFSDP2:
                     "reshard_after_forward": True,
                 },
                 "use_tensorboard": True,
+                "seed": 42,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
             }
         )
 

tests/e2e/multigpu/test_fsdp2_lora_kernels.py

@@ -40,7 +40,7 @@ def _run_training(temp_dir, cfg):
 def _base_lora_fsdp2_config(temp_dir, **overrides):
     """Base config for LoRA + FSDP2 + kernel tests."""
     cfg = {
-        "base_model": "Qwen/Qwen3-0.6B",
+        "base_model": "axolotl-ai-co/tiny-qwen3-129m",
         "sequence_len": 512,
         "val_set_size": 0.0,
         "datasets": [

tests/e2e/multigpu/test_tp.py

@@ -8,7 +8,7 @@ from accelerate.test_utils import execute_subprocess_async, get_torch_dist_uniqu
 
 from axolotl.utils.dict import DictDefault
 
-from tests.e2e.utils import check_tensorboard, require_torch_2_7_0
+from tests.e2e.utils import check_tensorboard_loss_decreased, require_torch_2_7_0
 
 
 class TestTensorParallel:
@@ -21,7 +21,7 @@ class TestTensorParallel:
     def test_fft_sft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [
@@ -63,6 +63,6 @@ class TestTensorParallel:
             ]
         )
 
-        check_tensorboard(
+        check_tensorboard_loss_decreased(
-            temp_dir + "/runs", "train/train_loss", 1.0, "Train Loss (%s) is too high"
+            temp_dir + "/runs", max_initial=5.0, max_final=4.7
         )

tests/e2e/patched/lora_kernels/test_lora_kernel_patching.py

@@ -32,12 +32,12 @@ from axolotl.utils.dict import DictDefault
 
 MODEL_CONFIGS = [
     {
-        "name": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+        "name": "axolotl-ai-co/tiny-mistral-25m",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
     {
-        "name": "trl-internal-testing/tiny-Qwen2ForCausalLM-2.5",
+        "name": "axolotl-ai-co/tiny-qwen2-129m",
         "expected_activation": apply_lora_mlp_swiglu,
         "dtype": torch.float16,
     },
@@ -47,7 +47,7 @@ MODEL_CONFIGS = [
         "dtype": torch.float32,
     },
     {
-        "name": "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "name": "axolotl-ai-co/tiny-gemma2-137m",
         "expected_activation": apply_lora_mlp_geglu,
         "dtype": torch.float16,
     },
@@ -159,7 +159,7 @@ def test_swiglu_mlp_integration(small_llama_model):
 def test_geglu_model_integration():
     """Test GeGLU activation with Gemma model."""
     model = AutoModelForCausalLM.from_pretrained(
-        "trl-internal-testing/tiny-Gemma2ForCausalLM",
+        "axolotl-ai-co/tiny-gemma2-137m",
         dtype=torch.float16,
         device_map="cuda:0",
     )

tests/e2e/patched/test_falcon_samplepack.py

@@ -4,14 +4,16 @@ E2E tests for falcon
 
 import unittest
 
-import pytest
-
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestFalconPatched(unittest.TestCase):
@@ -19,13 +21,12 @@ class TestFalconPatched(unittest.TestCase):
     Test case for Falcon models
     """
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_qlora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sample_packing": True,
                 "sequence_len": 2048,
                 "load_in_4bit": True,
@@ -47,17 +48,20 @@ class TestFalconPatched(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -66,14 +70,20 @@ class TestFalconPatched(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sample_packing": True,
                 "sequence_len": 2048,
                 "val_set_size": 0.05,
@@ -88,17 +98,20 @@ class TestFalconPatched(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -107,3 +120,10 @@ class TestFalconPatched(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )

tests/e2e/patched/test_mistral_samplepack.py

@@ -9,7 +9,12 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, require_torch_2_6_0, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    require_torch_2_6_0,
+    with_temp_dir,
+)
 
 
 class TestMistral(unittest.TestCase):
@@ -22,7 +27,7 @@ class TestMistral(unittest.TestCase):
     def test_lora_packing(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -45,17 +50,20 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "save_steps": 3,
+                "logging_steps": 1,
-                "eval_steps": 4,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -64,12 +72,19 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.5,
+            max_final=4.3,
+        )
 
     @with_temp_dir
     def test_ft_packing(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 1024,
@@ -86,17 +101,20 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "save_steps": 3,
+                "logging_steps": 1,
-                "eval_steps": 4,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -105,3 +123,10 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.5,
+            max_final=4.3,
+        )

tests/e2e/patched/test_mixtral_samplepack.py

@@ -9,7 +9,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestMixtral(unittest.TestCase):
@@ -21,8 +25,7 @@ class TestMixtral(unittest.TestCase):
     def test_qlora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,
@@ -30,7 +33,7 @@ class TestMixtral(unittest.TestCase):
                 "adapter": "qlora",
                 "lora_r": 16,
                 "lora_alpha": 32,
-                "lora_dropout": 0.1,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "val_set_size": 0.05,
                 "special_tokens": {},
@@ -41,17 +44,21 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 3e-3,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 80,
-                "save_steps": 3,
+                "warmup_steps": 5,
-                "eval_steps": 4,
+                "logging_steps": 1,
+                "save_steps": 80,
+                "eval_steps": 80,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -60,13 +67,19 @@ class TestMixtral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=10,
+            final_window=10,
+            max_initial=6.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,
@@ -79,17 +92,21 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 5e-4,
-                "optimizer": "adamw_bnb_8bit",
+                "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 80,
-                "save_steps": 3,
+                "warmup_steps": 5,
-                "eval_steps": 4,
+                "logging_steps": 1,
+                "save_steps": 80,
+                "eval_steps": 80,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -98,3 +115,10 @@ class TestMixtral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )

tests/e2e/patched/test_model_patches.py

@@ -22,8 +22,7 @@ class TestModelPatches(unittest.TestCase):
     def test_mixtral_multipack(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,
@@ -57,7 +56,7 @@ class TestModelPatches(unittest.TestCase):
     def test_mistral_multipack(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sample_packing": True,
                 "sequence_len": 2048,

tests/e2e/patched/test_phi_multipack.py

@@ -9,7 +9,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from ..utils import check_model_output_exists, with_temp_dir
+from ..utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestPhiMultipack(unittest.TestCase):
@@ -21,7 +25,7 @@ class TestPhiMultipack(unittest.TestCase):
     def test_ft_packed(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "PhiForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 1024,
@@ -43,17 +47,20 @@ class TestPhiMultipack(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
-                "optimizer": "adamw_bnb_8bit",
+                "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "eval_steps": 3,
+                "logging_steps": 1,
-                "save_steps": 4,
+                "eval_steps": 50,
+                "save_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -63,12 +70,19 @@ class TestPhiMultipack(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_qlora_packed(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "PhiForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 1024,
@@ -94,17 +108,20 @@ class TestPhiMultipack(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 2,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 5,
+                "max_steps": 50,
-                "eval_steps": 3,
+                "logging_steps": 1,
-                "save_steps": 4,
+                "eval_steps": 50,
+                "save_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -114,3 +131,10 @@ class TestPhiMultipack(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=6.0,
+            max_final=4.7,
+        )

tests/e2e/solo/test_batch_flattening.py

@@ -18,7 +18,7 @@ from transformers import AutoModelForCausalLM
 # Import the actual trainer methods we want to test
 from axolotl.core.trainers.grpo.async_trainer import AsyncGRPOTrainer
 
-MODEL_NAME = "Qwen/Qwen3-0.6B"
+MODEL_NAME = "axolotl-ai-co/tiny-qwen3-129m"
 
 
 def _fix_patched_attention(model):

tests/e2e/test_falcon.py

@@ -4,14 +4,16 @@ E2E tests for falcon
 
 import unittest
 
-import pytest
-
 from axolotl.common.datasets import load_datasets
 from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestFalcon(unittest.TestCase):
@@ -19,13 +21,12 @@ class TestFalcon(unittest.TestCase):
     Test case for falcon
     """
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -49,17 +50,21 @@ class TestFalcon(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -69,14 +74,20 @@ class TestFalcon(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_lora_added_vocab(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
@@ -104,17 +115,21 @@ class TestFalcon(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -124,14 +139,20 @@ class TestFalcon(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
-    @pytest.mark.skip(reason="no tiny models for testing with safetensors")
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "illuin/tiny-random-FalconForCausalLM",
+                "base_model": "axolotl-ai-co/tiny-falcon-42m",
-                "flash_attention": True,
+                "flash_attention": False,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
                 "special_tokens": {
@@ -145,17 +166,23 @@ class TestFalcon(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 5e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 80,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 80,
+                "eval_steps": 80,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -165,3 +192,10 @@ class TestFalcon(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=10,
+            final_window=10,
+            max_initial=5.0,
+            max_final=4.7,
+        )

tests/e2e/test_mistral.py

@@ -11,7 +11,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestMistral(unittest.TestCase):
@@ -23,7 +27,7 @@ class TestMistral(unittest.TestCase):
     def test_lora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_8bit": True,
@@ -45,16 +49,18 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -64,12 +70,19 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=4.5,
+            max_final=4.3,
+        )
 
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "trl-internal-testing/tiny-MistralForCausalLM-0.2",
+                "base_model": "axolotl-ai-co/tiny-mistral-25m",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
@@ -85,16 +98,18 @@ class TestMistral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -108,3 +123,10 @@ class TestMistral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=4.5,
+            max_final=4.3,
+        )

tests/e2e/test_mixtral.py

@@ -12,7 +12,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestMixtral(unittest.TestCase):
@@ -24,8 +28,7 @@ class TestMixtral(unittest.TestCase):
     def test_qlora_w_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "load_in_4bit": True,
@@ -51,16 +54,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -74,13 +79,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_qlora_wo_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": False,
                 "sequence_len": 1024,
                 "load_in_4bit": True,
@@ -106,16 +117,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -129,13 +142,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_16bit_lora_w_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "adapter": "lora",
@@ -160,16 +179,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -187,13 +208,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_16bit_lora_wo_fa2(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": False,
                 "sequence_len": 1024,
                 "adapter": "lora",
@@ -218,16 +245,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
 
@@ -245,13 +274,19 @@ class TestMixtral(unittest.TestCase):
             == torch.float32
         )
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "hf-internal-testing/Mixtral-tiny",
+                "base_model": "axolotl-ai-co/tiny-mixtral-30m",
-                "tokenizer_config": "LoneStriker/Mixtral-8x7B-v0.1-HF",
                 "flash_attention": True,
                 "sequence_len": 1024,
                 "val_set_size": 0.02,
@@ -263,16 +298,18 @@ class TestMixtral(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 2,
-                "micro_batch_size": 2,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "adamw_bnb_8bit",
                 "lr_scheduler": "cosine",
-                "max_steps": 20,
+                "max_steps": 50,
-                "save_steps": 10,
+                "logging_steps": 1,
-                "eval_steps": 10,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "save_first_step": False,
+                "use_tensorboard": True,
             }
         )
         if is_torch_bf16_gpu_available():
@@ -286,3 +323,10 @@ class TestMixtral(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )

tests/e2e/test_optimizers.py

@@ -13,6 +13,7 @@ from axolotl.utils.dict import DictDefault
 
 from .utils import (
     check_model_output_exists,
+    check_tensorboard_loss_decreased,
     require_torch_2_5_1,
     require_torch_2_6_0,
     require_torch_2_7_0,
@@ -243,20 +244,18 @@ class TestCustomOptimizers(unittest.TestCase):
     def test_came_pytorch(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "JackFram/llama-68m",
+                "base_model": "axolotl-ai-co/tiny-llama-50m",
-                "tokenizer_type": "LlamaTokenizer",
+                "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 1024,
                 "load_in_8bit": True,
                 "adapter": "lora",
                 "lora_r": 8,
                 "lora_alpha": 16,
-                "lora_dropout": 0.05,
+                "lora_dropout": 0.0,
                 "lora_target_linear": True,
                 "val_set_size": 0.1,
                 "special_tokens": {
-                    "unk_token": "<unk>",
+                    "pad_token": "<|endoftext|>",
-                    "bos_token": "<s>",
-                    "eos_token": "</s>",
                 },
                 "datasets": [
                     {
@@ -265,16 +264,22 @@ class TestCustomOptimizers(unittest.TestCase):
                     },
                 ],
                 "num_epochs": 1,
+                "sample_packing": True,
+                "pad_to_sequence_len": True,
                 "micro_batch_size": 8,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 1e-4,
                 "optimizer": "came_pytorch",
                 "adam_beta3": 0.9999,
                 "adam_epsilon2": 1e-16,
-                "max_steps": 5,
+                "max_steps": 80,
+                "warmup_steps": 5,
+                "logging_steps": 1,
                 "lr_scheduler": "cosine",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
 
@@ -284,6 +289,13 @@ class TestCustomOptimizers(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=10,
+            final_window=10,
+            max_initial=4.0,
+            max_final=3.0,
+        )
 
 
 @require_torch_2_7_0

tests/e2e/test_phi.py

@@ -9,7 +9,11 @@ from axolotl.train import train
 from axolotl.utils.config import normalize_config, validate_config
 from axolotl.utils.dict import DictDefault
 
-from .utils import check_model_output_exists, with_temp_dir
+from .utils import (
+    check_model_output_exists,
+    check_tensorboard_loss_decreased,
+    with_temp_dir,
+)
 
 
 class TestPhi(unittest.TestCase):
@@ -21,7 +25,7 @@ class TestPhi(unittest.TestCase):
     def test_phi_ft(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "AutoModelForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 2048,
@@ -41,18 +45,22 @@ class TestPhi(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
-                "optimizer": "paged_adamw_8bit",
+                "optimizer": "adamw_torch_fused",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
-                "max_steps": 10,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -61,12 +69,19 @@ class TestPhi(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )
 
     @with_temp_dir
     def test_phi_qlora(self, temp_dir):
         cfg = DictDefault(
             {
-                "base_model": "microsoft/phi-1_5",
+                "base_model": "axolotl-ai-co/tiny-phi-64m",
                 "model_type": "AutoModelForCausalLM",
                 "tokenizer_type": "AutoTokenizer",
                 "sequence_len": 2048,
@@ -90,18 +105,22 @@ class TestPhi(unittest.TestCase):
                 "dataset_shard_num": 10,
                 "dataset_shard_idx": 0,
                 "num_epochs": 1,
-                "micro_batch_size": 1,
+                "micro_batch_size": 4,
                 "gradient_accumulation_steps": 1,
                 "output_dir": temp_dir,
-                "learning_rate": 0.00001,
+                "learning_rate": 2e-4,
                 "optimizer": "paged_adamw_8bit",
                 "lr_scheduler": "cosine",
                 "flash_attention": True,
-                "max_steps": 10,
+                "max_steps": 50,
-                "save_steps": 10,
+                "warmup_steps": 5,
-                "eval_steps": 10,
+                "logging_steps": 1,
+                "save_steps": 50,
+                "eval_steps": 50,
                 "bf16": "auto",
                 "save_first_step": False,
+                "use_tensorboard": True,
+                "seed": 42,
             }
         )
         cfg = validate_config(cfg)
@@ -110,3 +129,10 @@ class TestPhi(unittest.TestCase):
 
         train(cfg=cfg, dataset_meta=dataset_meta)
         check_model_output_exists(temp_dir, cfg)
+        check_tensorboard_loss_decreased(
+            temp_dir + "/runs",
+            initial_window=5,
+            final_window=5,
+            max_initial=5.0,
+            max_final=4.7,
+        )

tests/e2e/test_preprocess.py

@@ -18,7 +18,7 @@ class TestPreprocess:
 
         cfg = DictDefault(
             {
-                "base_model": "Qwen/Qwen2.5-0.5B",
+                "base_model": "axolotl-ai-co/tiny-qwen2-129m",
                 "sequence_len": 2048,
                 "val_set_size": 0.01,
                 "datasets": [

tests/e2e/test_quantization.py

@@ -45,7 +45,7 @@ def _get_fake_quant_config_dtype(config):
 @pytest.fixture()
 def model():
     dummy_model = AutoModelForCausalLM.from_pretrained(
-        "Qwen/Qwen2-0.5B",
+        "axolotl-ai-co/tiny-qwen2-129m",
         device_map="auto",
         dtype=torch.bfloat16,
     )

tests/e2e/test_qwen.py

@@ -17,7 +17,7 @@ class TestE2eQwen:
     Test cases for qwen models
     """
 
-    @pytest.mark.parametrize("base_model", ["Qwen/Qwen2-0.5B", "Qwen/Qwen2.5-0.5B"])
+    @pytest.mark.parametrize("base_model", ["axolotl-ai-co/tiny-qwen2-129m"])
     def test_dpo(self, base_model, temp_dir):
         cfg = DictDefault(
             {

tests/e2e/utils.py

@@ -199,6 +199,106 @@ def check_tensorboard(
         assert df.value.values[-1] > 1e-5, "Expected loss to be greater than zero"
 
 
+def check_tensorboard_loss_decreased(
+    temp_run_dir: str,
+    tag: str | None = None,
+    initial_window: int = 1,
+    final_window: int = 1,
+    min_delta: float | None = None,
+    max_initial: float | None = None,
+    max_final: float | None = None,
+    max_loss_ratio: float = 0.95,
+) -> None:
+    """Check that training actually learned — loss went down and stayed in
+    a sensible range.
+
+    Used with the tiny ``axolotl-ai-co/tiny-*`` CI models, where pretraining
+    was brief enough that final loss won't clear the absolute thresholds used
+    for 135M+ models — but the training pipeline should still behave.
+
+    ``train/train_loss`` is only logged once (end-of-training aggregate). The
+    per-step tag is ``train/loss`` for SFT/LM trainers and may vary across
+    trainers (e.g. DPO). When ``tag`` is None we try common per-step tags in
+    order and use the first with enough samples.
+
+    Two kinds of regression we guard against:
+
+    1. **Loss blew up.** A silent bug (e.g. broken label masking) can start
+       training at an absurdly high loss. ``max_initial`` / ``max_final``
+       assert the measured means stay at-or-below bounds measured from a
+       known-good run. Both are optional but strongly encouraged — loss
+       going *down* from a bad starting scale still looks like "learning."
+
+    2. **Loss didn't go down enough.** ``max_loss_ratio`` (default 0.95)
+       requires ``final <= initial * ratio``. A default below 1.0 means the
+       final window mean must sit at least 5% below the initial window mean
+       — real learning, not noise that happened to land below start. Only
+       raise this for configs where a smaller drop is expected *and*
+       documented (e.g. DPO with near-trivial pairs); in that case you are
+       intentionally weakening the test.
+
+    ``min_delta`` is optional; when set, additionally requires
+    ``final + min_delta <= initial`` — use for configs with enough signal
+    to demand a specific minimum absolute drop.
+    """
+    tb_log_path = most_recent_subdir(temp_run_dir)
+    event_file = os.path.join(tb_log_path, sorted(os.listdir(tb_log_path))[0])
+    reader = SummaryReader(event_file)
+    df = reader.scalars
+
+    if tag is None:
+        candidates = ["train/loss", "train/train_loss"]
+    else:
+        candidates = [tag]
+
+    required = initial_window + final_window
+    chosen_tag, values = None, None
+    for candidate in candidates:
+        sub = df[df.tag == candidate]
+        if len(sub) >= required:
+            chosen_tag = candidate
+            values = sub.value.values
+            break
+
+    available = sorted({t for t in df.tag.unique() if "loss" in t.lower()})
+    assert values is not None, (
+        f"None of the tags {candidates} had ≥{required} logged steps. "
+        f"Loss tags present: {available}"
+    )
+
+    initial = float(values[:initial_window].mean())
+    final = float(values[-final_window:].mean())
+    print(
+        f"[check_tensorboard_loss_decreased] tag={chosen_tag} n={len(values)} "
+        f"initial_mean{initial_window}={initial:.4f} final_mean{final_window}={final:.4f}"
+    )
+    assert final > 1e-5, "Expected loss to be greater than zero"
+    assert final <= initial * max_loss_ratio, (
+        f"Loss did not decrease for {chosen_tag}: "
+        f"initial(mean of first {initial_window})={initial:.4f}, "
+        f"final(mean of last {final_window})={final:.4f}, "
+        f"ratio={final / initial:.4f} (max allowed {max_loss_ratio}). "
+        f"Expected final <= initial — training did not learn."
+    )
+    if min_delta is not None:
+        assert final + min_delta <= initial, (
+            f"Expected loss to decrease by at least {min_delta} for {chosen_tag}: "
+            f"initial={initial:.4f}, final={final:.4f}, delta={initial - final:.4f}"
+        )
+    if max_initial is not None:
+        assert initial <= max_initial, (
+            f"Initial loss {initial:.4f} is above the expected max {max_initial}. "
+            f"Absolute scale is wrong — probably a silent regression "
+            f"(e.g. bad label masking) that bumped the starting point."
+        )
+    if max_final is not None:
+        assert final <= max_final, (
+            f"Final loss {final:.4f} is above the expected max {max_final}. "
+            f"Absolute scale is wrong — probably a silent regression "
+            f"(e.g. bad label masking) that bumped the endpoint."
+        )
+
+
 def check_model_output_exists(temp_dir: str, cfg: DictDefault) -> None:
     """
     helper function to check if a model output file exists after training