adding fp32 support

* feat: upgrade cce with patches for transformers 4.56

* feat: add missing models to cce readme

_quarto.yml

@@ -267,6 +267,7 @@ website:
             - docs/dataset_loading.qmd
             - docs/qat.qmd
             - docs/quantize.qmd
+            - docs/optimizations.qmd
 
         - section: "Core Concepts"
           contents:

docs/dataset-formats/conversation.qmd

@@ -212,6 +212,14 @@ Instead of passing `tools` via the system prompt, an alternative method would be
 Tools need to follow [JSON schema](https://json-schema.org/learn/getting-started-step-by-step).
 :::
 
+::: {.callout-warning}
+If you have tool arguments with same name but different dtypes (like `"time": string` and `"time": number`), please save `arguments: ` as JSON string to prevent `datasets` from having casting issues.
+
+```
+"arguments": "{\"...\": \"...\"}"
+```
+:::
+
 Example config for Llama4:
 ```yaml
 chat_template: llama4

docs/dataset-formats/index.qmd

@@ -61,7 +61,7 @@ While we recommend `.jsonl`, you can also use the other formats (`csv`, `parquet
 
 ### Pre-training without streaming
 
-On the rare case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
+In the case that the dataset is small and can be loaded entirely into memory, another approach to running pre-training is to use the `completion` format. This would mean that the entire dataset is pre-tokenized instead of on-demand in streaming.
 
 One benefit of this is that the tokenization can be performed separately on a CPU-only machine, and then transferred to a GPU machine for training to save costs.
 

docs/faq.qmd

@@ -140,3 +140,7 @@ description: Frequently asked questions
 **Q: `ValueError("Backward pass should have cleared tracker of all tensors")`
 
 > A: This may happen due to edge cases in using the modern OffloadActivations context manager for CUDA streams. If you encounter this error, you may have success using the naive implementation with `offload_activations: legacy` in your YAML.
+
+**Q: `Error parsing tool_calls arguments as JSON.`
+
+> A: There is an error parsing string arguments to a dict. Please check your dataset and the error message for more details.

docs/optimizations.qmd

@@ -0,0 +1,133 @@
+---
+title: Optimizations Guide
+description: A guide to the performance and memory optimizations available in Axolotl.
+---
+
+Axolotl includes numerous optimizations to speed up training, reduce memory usage, and handle large models.
+
+This guide provides a high-level overview and directs you to the detailed documentation for each feature.
+
+## Speed Optimizations
+
+These optimizations focus on increasing training throughput and reducing total training time.
+
+### Sample Packing
+
+Improves GPU utilization by combining multiple short sequences into a single packed sequence for training. This requires enabling one of the [attention](#attention-implementations) implementations below.
+
+- **Config:** `sample_packing: true`
+- **Learn more:** [Sample Packing](multipack.qmd)
+
+### Attention Implementations
+
+Using an optimized attention implementation is critical for training speed.
+
+- **[Flash Attention 2](https://github.com/Dao-AILab/flash-attention)**: `flash_attention: true`. **(Recommended)** The industry standard for fast attention on modern GPUs. Requires Ampere or higher. For AMD, check [AMD Support](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#amd-rocm-support).
+- **[Flex Attention](https://pytorch.org/blog/flexattention/)**: `flex_attention: true`.
+- **[SDP Attention](https://docs.pytorch.org/docs/stable/generated/torch.nn.functional.scaled_dot_product_attention.html)**: `sdp_attention: true`. PyTorch's native implementation.
+- **[Xformers](https://github.com/facebookresearch/xformers)**: `xformers_attention: true`. Works with FP16.
+
+*Note: You should only enable one attention backend.*
+
+### LoRA Optimizations
+
+Leverages optimized kernels to accelerate LoRA training and reduce memory usage.
+
+- **Learn more:** [LoRA Optimizations Documentation](lora_optims.qmd)
+
+## Memory Optimizations
+
+These techniques help you fit larger models or use bigger batch sizes on your existing hardware.
+
+### Parameter Efficient Finetuning (LoRA & QLoRA)
+
+Drastically reduces memory by training a small set of "adapter" parameters instead of the full model. This is the most common and effective memory-saving technique.
+
+- Examples: Find configs with `lora` or `qlora` in the [examples directory](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/llama-3).
+- Config Reference: See `adapter`, `load_in_4bit`, and `load_in_8bit` in the [Configuration Reference](config-reference.qmd).
+
+### Gradient Checkpointing & Activation Offloading
+
+These techniques save VRAM by changing how activations are handled.
+
+- Gradient Checkpointing: re-computes activations during the backward pass, trading compute time for VRAM.
+- Activation Offloading: moves activations to CPU RAM or disk, trading I/O overhead for VRAM.
+- Learn more: [Gradient Checkpointing and Offloading Docs](gradient_checkpointing.qmd)
+
+### Cut Cross Entropy (CCE)
+
+Reduces VRAM usage by using an optimized cross-entropy loss calculation.
+
+- **Learn more:** [Custom Integrations - CCE](custom_integrations.qmd#cut-cross-entropy)
+
+### Liger Kernels
+
+Provides efficient Triton kernels to improve training speed and reduce memory usage.
+
+- **Learn more:** [Custom Integrations - Liger Kernels](custom_integrations.qmd#liger-kernels)
+
+## Long Context Models
+
+Techniques to train models on sequences longer than their original context window.
+
+### RoPE Scaling
+
+Extends a model's context window by interpolating its Rotary Position Embeddings.
+
+- **Config:** Pass the `rope_scaling` config under the `overrides_of_model_config: `. To learn how to set RoPE, check the respective model config.
+
+### Sequence Parallelism
+
+Splits long sequences across multiple GPUs, enabling training with sequence lengths that would not fit on a single device.
+
+- **Learn more:** [Sequence Parallelism Documentation](sequence_parallelism.qmd)
+
+### Artic Long Sequence Training (ALST)
+
+ALST is a recipe that combines several techniques to train long-context models efficiently. It typically involves:
+
+- TiledMLP to reduce memory usage in MLP layers.
+- Tiled Loss functions (like [CCE](#cut-cross-entropy-(cce) or [Liger](#liger-kernels)).
+- Activation Offloading to CPU.
+
+- Example: [ALST Example Configuration](https://github.com/axolotl-ai-cloud/axolotl/tree/main/examples/alst)
+
+## Large Models (Distributed Training)
+
+To train models that don't fit on a single GPU, you'll need to use a distributed training strategy like FSDP or DeepSpeed. These frameworks shard the model weights, gradients, and optimizer states across multiple GPUs and nodes.
+
+- **Learn more:** [Multi-GPU Guide](multi-gpu.qmd)
+- **Learn more:** [Multi-Node Guide](multi-node.qmd)
+
+### N-D Parallelism (Beta)
+
+For advanced scaling, Axolotl allows you to compose different parallelism techniques (e.g., Data, Tensor, Sequence Parallelism). This is a powerful approach to train an extremely large model by overcoming multiple bottlenecks at once.
+
+- **Learn more:** [N-D Parallelism Guide](nd_parallelism.qmd)
+
+
+## Quantization
+
+Techniques to reduce the precision of model weights for memory savings.
+
+### 4-bit Training (QLoRA)
+
+The recommended approach for quantization-based training. It loads the base model in 4-bit using `bitsandbytes` and then trains QLoRA adapters. See [Adapter Finetuning](#adapter-finetuning-lora-qlora) for details.
+
+### FP8 Training
+
+Enables training with 8-bit floating point precision on supported hardware (e.g., NVIDIA Hopper series GPUs) for significant speed and memory gains.
+
+- **Example:** [Llama 3 FP8 FSDP Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-3/3b-fp8-fsdp2.yaml)
+
+### Quantization Aware Training (QAT)
+
+Simulates quantization effects during training, helping the model adapt and potentially improving the final accuracy of the quantized model.
+
+- **Learn more:** [QAT Documentation](qat.qmd)
+
+### GPTQ
+
+Allows you to finetune LoRA adapters on top of a model that has already been quantized using the GPTQ method.
+
+- **Example:** [GPTQ LoRA Example](https://github.com/axolotl-ai-cloud/axolotl/blob/main/examples/llama-2/gptq-lora.yml)

docs/qat.qmd

@@ -30,6 +30,7 @@ qat:
 ```
 
 We support the following quantization schemas:
+
 - `Int4WeightOnly` (requires the `fbgemm-gpu` extra when installing Axolotl)
 - `Int8DynamicActivationInt4Weight`
 - `Float8DynamicActivationFloat8Weight`

examples/alst/README.md

@@ -7,3 +7,24 @@ techniques. It is a combination of:
 - Activation Offloading: Offload activations to CPU RAM to reduce memory usage
 
 For more information, you can check out the ALST paper [here](https://www.arxiv.org/abs/2506.13996).
+
+## Usage
+
+```yaml
+tiled_mlp: true
+
+# See Sequence Parallelism docs
+# https://docs.axolotl.ai/docs/sequence_parallelism.html
+context_parallel_size: int
+
+plugins:
+# See Cut Cross Entropy docs
+# https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+# or Liger Kernel docs
+# https://docs.axolotl.ai/docs/custom_integrations.html#liger-kernels
+  - axolotl.integrations.liger.LigerPlugin
+# ...
+
+```

examples/colab-notebooks/colab-axolotl-example.ipynb

@@ -40,7 +40,7 @@
     "%%capture\n",
     "# This step can take ~5-10 minutes to install dependencies\n",
     "!pip install --no-build-isolation axolotl[flash-attn]>=0.9.1\n",
-    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef\""
+    "!pip install \"cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@147ea28\""
    ]
   },
   {

examples/qwen3-next/README.md

@@ -38,7 +38,7 @@ pip3 uninstall -y causal-conv1d && pip3 install flash-linear-attention==0.3.2
 axolotl train examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml
 ```
 
-This config uses about 41.7 GiB VRAM.
+This config uses about 45.62 GiB VRAM.
 
 Let us know how it goes. Happy finetuning! 🚀
 

examples/qwen3-next/qwen3-next-80b-a3b-qlora.yaml

@@ -27,6 +27,14 @@ lora_r: 16
 lora_alpha: 8
 lora_dropout: 0.05
 lora_target_modules:
+  - linear_attn.in_proj_ba
+  - linear_attn.in_proj_qkvz
+  - linear_attn.out_proj
+  - shared_expert.up_proj
+  - shared_expert.down_proj
+  - shared_expert.gate_proj
+  - shared_expert_gate
+  - mlp.gate
   - q_proj
   - v_proj
   - k_proj

scripts/cutcrossentropy_install.py

@@ -29,5 +29,5 @@ UV_PREFIX = "uv " if USE_UV else ""
 
 print(
     UNINSTALL_PREFIX
-    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef"'
+    + f'{UV_PREFIX}pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@147ea28"'
 )

src/axolotl/integrations/cut_cross_entropy/README.md

@@ -19,7 +19,7 @@ python scripts/cutcrossentropy_install.py | sh
 
 - If you are installing from pip
 ```bash
-pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef"
+pip3 uninstall -y cut-cross-entropy && pip3 install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@147ea28"
 ```
 
 ## Usage
@@ -31,6 +31,7 @@ plugins:
 
 ## Supported Models
 
+- apertus
 - arcee
 - cohere
 - cohere2
@@ -44,9 +45,13 @@ plugins:
 - glm
 - glm4
 - glm4_moe
+- glm4v
+- glm4v_moe
 - gpt_oss
 - granite
 - granitemoe
+- granitemoeshared
+- granitemoehybrid
 - hunyuan_v1_dense
 - hunyuan_v1_moe
 - llama
@@ -65,6 +70,8 @@ plugins:
 - qwen2_5_vl
 - qwen3
 - qwen3_moe
+- qwen3_vl
+- qwen3_vl_moe
 - qwen3_next
 - smollm3
 - seed_oss

src/axolotl/integrations/cut_cross_entropy/__init__.py

@@ -35,7 +35,7 @@ LOG = get_logger(__name__)
 
 _CCE_INSTALL_MESSAGE = (
     "Please install Axolotl's fork of cut_cross_entropy with transformers support using "
-    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@c5aa3ef"`'
+    '`pip install "cut-cross-entropy[transformers] @ git+https://github.com/axolotl-ai-cloud/ml-cross-entropy.git@147ea28"`'
 )
 
 

src/axolotl/loaders/patch_manager.py

@@ -84,9 +84,7 @@ class PatchManager:
         patch_evaluation_loop()
         patch_maybe_log_save_evaluate()
 
-        if self.cfg.context_parallel_size > 1 and getattr(
-            self.cfg, "flash_attention", False
-        ):
+        if self.cfg.context_parallel_size > 1:
             from axolotl.monkeypatch.transformers.trainer_context_parallel import (
                 patch_prepare_context_parallel_inputs,
             )

src/axolotl/monkeypatch/accelerate/fsdp2.py

@@ -368,6 +368,7 @@ def fsdp2_prepare_model(accelerator, model: torch.nn.Module) -> torch.nn.Module:
         # removing the call above leads to extra memory usage as explained in the comment above
         if hasattr(model, "tie_weights"):
             model.tie_weights()
+    model = model.to(torch.float32)
     return model
 
 

src/axolotl/monkeypatch/ring_attn/adapters/batch.py

@@ -13,10 +13,21 @@ from typing import Callable
 import torch
 import torch.distributed as dist
 import transformers
-import transformers.modeling_flash_attention_utils as flash_utils
+import transformers.modeling_flash_attention_utils
 from ring_flash_attn import ring_flash_attn_func
 from ring_flash_attn.adapters.hf_adapter import check_params
 from transformers.modeling_flash_attention_utils import is_flash_attn_greater_or_equal
+
+try:
+    from transformers.modeling_flash_attention_utils import _flash_supports_window
+except ImportError:
+    try:
+        from transformers.modeling_flash_attention_utils import (
+            _flash_supports_window_size as _flash_supports_window,
+        )
+    except ImportError:
+        _flash_supports_window = True
+
 from transformers.modeling_utils import ALL_ATTENTION_FUNCTIONS
 
 from axolotl.utils.schemas.enums import RingAttnFunc
@@ -107,7 +118,7 @@ def create_flash_attn_forward_varlen_llama3(
 
         # Handle sliding window
         use_sliding_windows = (
-            _flash_windows_supported()
+            _flash_supports_window
             and sliding_window is not None
             and key_states.shape[1] > sliding_window
         )
@@ -183,18 +194,3 @@ def substitute_hf_flash_attn(
         from ring_flash_attn.adapters.hf_adapter import flash_attention_forward
 
         ALL_ATTENTION_FUNCTIONS["flash_attention_2"] = flash_attention_forward
-
-
-def _flash_windows_supported() -> bool:
-    """Return whether current transformers build advertises sliding-window support."""
-    support = getattr(flash_utils, "_flash_supports_window", None)
-    if support is None:
-        support = getattr(flash_utils, "_flash_supports_window_size", None)
-
-    if support is None:
-        return True
-
-    if callable(support):
-        return True
-
-    return bool(support)

src/axolotl/monkeypatch/ring_attn/patch.py

@@ -13,9 +13,18 @@ from typing import Optional
 
 import torch
 import torch.distributed as dist
-import transformers.modeling_flash_attention_utils as flash_utils
 from torch.distributed import DeviceMesh
 
+try:
+    from transformers.modeling_flash_attention_utils import _flash_supports_window
+except ImportError:
+    try:
+        from transformers.modeling_flash_attention_utils import (
+            _flash_supports_window_size as _flash_supports_window,
+        )
+    except ImportError:
+        _flash_supports_window = True
+
 from axolotl.monkeypatch.utils import get_cu_seqlens_from_pos_ids
 from axolotl.utils.logging import get_logger
 from axolotl.utils.schemas.enums import RingAttnFunc
@@ -74,7 +83,7 @@ def create_ring_flash_attention_forward(
 
         # Assuming 4D tensors, key_states.shape[1] is the key/value sequence length (source length).
         use_sliding_windows = (
-            _flash_windows_supported()
+            _flash_supports_window
             and sliding_window is not None
             and key_states.shape[1] > sliding_window
         )
@@ -216,19 +225,3 @@ def update_ring_attn_params(position_ids: torch.Tensor | None):
     cu_seqlens, _ = get_cu_seqlens_from_pos_ids(position_ids)
     cu_seqlens = cu_seqlens.squeeze().to(device=torch.cuda.current_device())
     update_ring_flash_attn_params(cu_seqlens, get_ring_attn_group())
-
-
-def _flash_windows_supported() -> bool:
-    """Best-effort check for FlashAttention sliding-window support."""
-    support = getattr(flash_utils, "_flash_supports_window", None)
-    if support is None:
-        support = getattr(flash_utils, "_flash_supports_window_size", None)
-
-    if support is None:
-        return True
-
-    if callable(support):
-        # Signature differs across versions; assume support when callable.
-        return True
-
-    return bool(support)

src/axolotl/prompt_strategies/chat_template.py

@@ -2,6 +2,7 @@
 HF Chat Templates prompt strategy
 """
 
+import json
 from collections import defaultdict
 from typing import TYPE_CHECKING, Any, Dict, List, Set, Union
 
@@ -794,6 +795,22 @@ class ChatTemplateStrategy(PromptTokenizingStrategy):
                 if val is not None:
                     transformed_message[key] = val
 
+        if "tool_calls" in transformed_message and transformed_message["tool_calls"]:
+            for tool_call in transformed_message["tool_calls"]:
+                if "function" in tool_call and "arguments" in tool_call["function"]:
+                    args = tool_call["function"]["arguments"]
+                    if isinstance(args, str):
+                        try:
+                            tool_call["function"]["arguments"] = json.loads(args)
+                        except json.JSONDecodeError as e:
+                            LOG.error(
+                                f"Error parsing tool_calls arguments as JSON. "
+                                f"Function: {tool_call.get('function', {}).get('name', 'unknown')}, "
+                                f"Arguments string: {args!r}, "
+                                f"Error: {e}"
+                            )
+                            raise
+
         return transformed_message
 
     def _get_images(self, prompt):

src/axolotl/train.py

@@ -179,11 +179,7 @@ def execute_training(
                 )
             )
 
-        use_flash_cp = cfg.context_parallel_size > 1 and bool(
-            getattr(cfg, "flash_attention", False)
-        )
-
-        if use_flash_cp:
+        if cfg.context_parallel_size > 1:
             models = [trainer.model]
             if hasattr(trainer, "ref_model") and trainer.ref_model:
                 models.append(trainer.ref_model)

src/axolotl/utils/schemas/validation.py

@@ -1,6 +1,7 @@
 """Module with validation methods for config pydantic model."""
 
 import json
+import sys
 import tempfile
 from pathlib import Path
 
@@ -1313,40 +1314,50 @@ class ComplexValidationMixin:
         if not self.context_parallel_size:
             self.context_parallel_size = 1
         elif self.context_parallel_size > 1:
-            use_flash_attention = getattr(self, "flash_attention", False)
-            use_sdp_attention = getattr(self, "sdp_attention", False)
-
-            if not (use_flash_attention or use_sdp_attention):
+            if not self.flash_attention:
                 raise ValueError(
-                    "context_parallel_size > 1 requires either flash_attention: true "
-                    "or sdp_attention: true"
+                    "flash_attention: true must be set with context_parallel_size > 1"
                 )
 
-            if use_flash_attention:
-                if self.sample_packing and self.micro_batch_size > 1:
-                    raise ValueError(
-                        "micro_batch_size must be set to 1 when sample_packing is enabled "
-                        "due to a `ring-flash-attn` requirement"
-                    )
-
-                try:
-                    import ring_flash_attn  # noqa: F401  # Required after monkey-patching
-                except ImportError as exception:
-                    raise ImportError(
-                        "context_parallel_size > 1 but ring_flash_attn is not installed. "
-                        "Please install it with `pip install axolotl[ring-flash-attn] "
-                        "or `pip install ring-flash-attn>=0.1.4`."
-                    ) from exception
-
-                LOG.warning(
-                    "Sequence parallelism (SP) is enabled with "
-                    f"context_parallel_size={self.context_parallel_size}. "
-                    "Please note that logged losses may differ slightly to the non-SP "
-                    "losses due to transformers Trainer implementation details. "
-                    "Please see https://github.com/axolotl-ai-cloud/axolotl/pull/2495#issuecomment-2784022042 "
-                    "for more details."
+            if self.sample_packing and self.micro_batch_size > 1:
+                raise ValueError(
+                    "micro_batch_size must be set to 1 when sample_packing is enabled "
+                    "due to a `ring-flash-attn` requirement"
                 )
 
+            try:
+                import transformers.modeling_flash_attention_utils
+                from transformers.utils import is_flash_attn_greater_or_equal
+
+                transformers.modeling_flash_attention_utils._flash_supports_window = (
+                    True
+                )
+                sys.modules[
+                    "transformers.modeling_flash_attention_utils"
+                ]._flash_supports_window = True
+                sys.modules[
+                    "transformers.modeling_flash_attention_utils"
+                ]._flash_supports_window_size = True
+                sys.modules[
+                    "transformers.modeling_flash_attention_utils"
+                ].is_flash_attn_greater_or_equal = is_flash_attn_greater_or_equal
+                import ring_flash_attn  # noqa: F401  # Required after monkey-patching
+            except ImportError as exception:
+                raise ImportError(
+                    "context_parallel_size > 1 but ring_flash_attn is not installed. "
+                    "Please install it with `pip install axolotl[ring-flash-attn] "
+                    "or `pip install ring-flash-attn>=0.1.4`."
+                ) from exception
+
+            LOG.warning(
+                "Sequence parallelism (SP) is enabled with "
+                f"context_parallel_size={self.context_parallel_size}. "
+                "Please note that logged losses may differ slightly to the non-SP "
+                "losses due to transformers Trainer implementation details. "
+                "Please see https://github.com/axolotl-ai-cloud/axolotl/pull/2495#issuecomment-2784022042 "
+                "for more details."
+            )
+
         return self
 
     @model_validator(mode="after")

tests/e2e/multigpu/patched/test_sp.py

@@ -23,8 +23,6 @@ class TestSequenceParallelism:
         pad_to_sequence_len=True,
         ring_attn_func=None,
         threshold=2.0,
-        flash_attention=True,
-        sdp_attention=False,
     ):
         """Helper method to run sequence parallel tests with different configurations"""
         cfg = DictDefault(
@@ -60,8 +58,7 @@ class TestSequenceParallelism:
                 "learning_rate": 0.00001,
                 "optimizer": "adamw_8bit",
                 "lr_scheduler": "cosine",
-                "flash_attention": flash_attention,
-                "sdp_attention": sdp_attention,
+                "flash_attention": True,
                 "loss_watchdog_threshold": 5.0,
                 "loss_watchdog_patience": 3,
                 "bf16": "auto",
@@ -135,16 +132,3 @@ class TestSequenceParallelism:
             ring_attn_func=ring_attn_func,
             threshold=threshold,
         )
-
-    def test_sequence_parallel_training_sdpa(self, temp_dir):
-        """Smoke test for SDPA-based context parallelism."""
-        self._run_sequence_parallel_test(
-            temp_dir,
-            sample_packing=False,
-            micro_batch_size=1,
-            pad_to_sequence_len=True,
-            ring_attn_func=None,
-            threshold=3.0,
-            flash_attention=False,
-            sdp_attention=True,
-        )

tests/loaders/test_patch_manager_cp.py

@@ -1,74 +0,0 @@
-"""Tests for PatchManager context parallel patch selection."""
-
-import addict
-
-from axolotl.loaders.patch_manager import PatchManager
-from axolotl.utils.dict import DictDefault
-
-
-def _stub_transformers_patches(monkeypatch):
-    """Replace trainer loss patchers with no-ops for isolation."""
-    monkeypatch.setattr(
-        "axolotl.monkeypatch.transformers.trainer_loss_calc.patch_evaluation_loop",
-        lambda: None,
-    )
-    monkeypatch.setattr(
-        "axolotl.monkeypatch.transformers.trainer_loss_calc.patch_maybe_log_save_evaluate",
-        lambda: None,
-    )
-
-
-def test_patch_manager_applies_flash_cp_patch(monkeypatch):
-    """When flash attention is enabled, we patch Trainer for CP."""
-    _stub_transformers_patches(monkeypatch)
-
-    patch_calls = {"count": 0}
-
-    def stub_patch():
-        patch_calls["count"] += 1
-
-    monkeypatch.setattr(
-        "axolotl.monkeypatch.transformers.trainer_context_parallel.patch_prepare_context_parallel_inputs",
-        stub_patch,
-    )
-
-    cfg = DictDefault(
-        {
-            "context_parallel_size": 2,
-            "flash_attention": True,
-            "sdp_attention": False,
-        }
-    )
-
-    manager = PatchManager(cfg, addict.Dict())
-    manager._apply_transformers_patches()
-
-    assert patch_calls["count"] == 1
-
-
-def test_patch_manager_skips_flash_patch_for_sdpa(monkeypatch):
-    """When only SDPA is requested, we should not patch Trainer."""
-    _stub_transformers_patches(monkeypatch)
-
-    patch_calls = {"count": 0}
-
-    def stub_patch():
-        patch_calls["count"] += 1
-
-    monkeypatch.setattr(
-        "axolotl.monkeypatch.transformers.trainer_context_parallel.patch_prepare_context_parallel_inputs",
-        stub_patch,
-    )
-
-    cfg = DictDefault(
-        {
-            "context_parallel_size": 2,
-            "flash_attention": False,
-            "sdp_attention": True,
-        }
-    )
-
-    manager = PatchManager(cfg, addict.Dict())
-    manager._apply_transformers_patches()
-
-    assert patch_calls["count"] == 0

tests/prompt_strategies/conftest.py

@@ -177,6 +177,15 @@ def fixture_devstral_1_1_tokenizer():
     return tokenizer
 
 
+@pytest.fixture(name="qwen3_tokenizer")
+def qwen3_tokenizer_fixture(
+    download_qwen3_half_billion_model,
+):  # pylint: disable=unused-argument,redefined-outer-name
+    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
+
+    return tokenizer
+
+
 @pytest.fixture(name="mistralv03_tokenizer_chat_template_jinja")
 def fixture_mistralv03_chat_template_jinja_w_system() -> str:
     return '{%- if messages[0]["role"] == "system" %}\n    {%- set system_message = messages[0]["content"] %}\n    {%- set loop_messages = messages[1:] %}\n{%- else %}\n    {%- set loop_messages = messages %}\n{%- endif %}\n{%- if not tools is defined %}\n    {%- set tools = none %}\n{%- endif %}\n{%- set user_messages = loop_messages | selectattr("role", "equalto", "user") | list %}\n\n{#- This block checks for alternating user/assistant messages, skipping tool calling messages #}\n{%- set ns = namespace() %}\n{%- set ns.index = 0 %}\n{%- for message in loop_messages %}\n    {%- if not (message.role == "tool" or message.role == "tool_results" or (message.tool_calls is defined and message.tool_calls is not none)) %}\n        {%- if (message["role"] == "user") != (ns.index % 2 == 0) %}\n            {{- raise_exception("After the optional system message, conversation roles must alternate user/assistant/user/assistant/...") }}\n        {%- endif %}\n        {%- set ns.index = ns.index + 1 %}\n    {%- endif %}\n{%- endfor %}\n\n{{- bos_token }}\n{%- for message in loop_messages %}\n    {%- if message["role"] == "user" %}\n        {%- if tools is not none and (message == user_messages[-1]) %}\n            {{- "[AVAILABLE_TOOLS] [" }}\n            {%- for tool in tools %}\n                {%- set tool = tool.function %}\n                {{- \'{"type": "function", "function": {\' }}\n                {%- for key, val in tool.items() if key != "return" %}\n                    {%- if val is string %}\n                        {{- \'"\' + key + \'": "\' + val + \'"\' }}\n                    {%- else %}\n                        {{- \'"\' + key + \'": \' + val|tojson }}\n                    {%- endif %}\n                    {%- if not loop.last %}\n                        {{- ", " }}\n                    {%- endif %}\n                {%- endfor %}\n                {{- "}}" }}\n                {%- if not loop.last %}\n                    {{- ", " }}\n                {%- else %}\n                    {{- "]" }}\n                {%- endif %}\n            {%- endfor %}\n            {{- "[/AVAILABLE_TOOLS]" }}\n            {%- endif %}\n        {%- if loop.first and system_message is defined %}\n            {{- "[INST] " + system_message + "\\n\\n" + message["content"] + "[/INST]" }}\n        {%- else %}\n            {{- "[INST] " + message["content"] + "[/INST]" }}\n        {%- endif %}\n    {%- elif message.tool_calls is defined and message.tool_calls is not none %}\n        {{- "[TOOL_CALLS] [" }}\n        {%- for tool_call in message.tool_calls %}\n            {%- set out = tool_call.function|tojson %}\n            {{- out[:-1] }}\n            {%- if not tool_call.id is defined or tool_call.id|length != 9 %}\n                {{- raise_exception("Tool call IDs should be alphanumeric strings with length 9!") }}\n            {%- endif %}\n            {{- \', "id": "\' + tool_call.id + \'"}\' }}\n            {%- if not loop.last %}\n                {{- ", " }}\n            {%- else %}\n                {{- "]" + eos_token }}\n            {%- endif %}\n        {%- endfor %}\n    {%- elif message["role"] == "assistant" %}\n        {{- " " + message["content"]|trim + eos_token}}\n    {%- elif message["role"] == "tool_results" or message["role"] == "tool" %}\n        {%- if message.content is defined and message.content.content is defined %}\n            {%- set content = message.content.content %}\n        {%- else %}\n            {%- set content = message.content %}\n        {%- endif %}\n        {{- \'[TOOL_RESULTS] {"content": \' + content|string + ", " }}\n        {%- if not message.tool_call_id is defined or message.tool_call_id|length != 9 %}\n            {{- raise_exception("Tool call IDs should be alphanumeric strings with length 9!") }}\n        {%- endif %}\n        {{- \'"call_id": "\' + message.tool_call_id + \'"}[/TOOL_RESULTS]\' }}\n    {%- else %}\n        {{- raise_exception("Only user and assistant roles are supported, with the exception of an initial optional system message!") }}\n    {%- endif %}\n{%- endfor %}\n'

tests/prompt_strategies/test_chat_template_ds_schema_unification.py

@@ -6,7 +6,6 @@ import json
 
 import pytest
 from datasets import Dataset
-from transformers import AutoTokenizer
 
 from axolotl.prompt_strategies.chat_template import StrategyLoader
 from axolotl.utils.dict import DictDefault
@@ -23,15 +22,6 @@ def fixture_messages_w_tools():
     return Dataset.from_list(rows)
 
 
-@pytest.fixture(name="qwen3_tokenizer")
-def qwen3_tokenizer_fixture(
-    download_qwen3_half_billion_model,
-):
-    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
-
-    return tokenizer
-
-
 @pytest.fixture(name="qwen3_prompt_strategy")
 def qwen3_chat_template_strategy(qwen3_tokenizer):
     cfg = DictDefault(

tests/prompt_strategies/test_chat_templates_thinking.py

@@ -4,7 +4,6 @@ Tests for splitting reasoning/thinking from content into separate field
 
 import pytest
 from datasets import Dataset
-from transformers import AutoTokenizer
 
 from axolotl.prompt_strategies.chat_template import (
     load,
@@ -56,15 +55,6 @@ def messages_w_reasoning_fixture():
     )
 
 
-@pytest.fixture(name="qwen3_tokenizer")
-def qwen3_tokenizer_fixture(
-    download_qwen3_half_billion_model,
-):
-    tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-0.6B")
-
-    return tokenizer
-
-
 class TestSplitThinking:
     """
     test class to make sure datasets with reasoning content conforms to the chat_template strategy

tests/prompt_strategies/test_chat_templates_tool_call_string_arguments.py

@@ -0,0 +1,214 @@
+"""
+Tests for handling json tool content
+"""
+
+import json
+
+import pytest
+from datasets import Dataset
+
+from axolotl.prompt_strategies.chat_template import (
+    load,
+)
+from axolotl.utils.dict import DictDefault
+
+
+@pytest.fixture(name="qwen3_instruct_prompt_strategy")
+def qwen3_instruct_chat_template_strategy(qwen3_tokenizer):
+    strategy = load(
+        qwen3_tokenizer,
+        DictDefault(
+            {
+                "train_on_inputs": False,
+                "sequence_len": 512,
+            }
+        ),
+        DictDefault(
+            {
+                "chat_template": "qwen3",
+                "message_field_role": "role",
+                "message_field_content": "content",
+                "message_property_mappings": {
+                    "role": "role",
+                    "content": "content",
+                },
+                "roles": {
+                    "user": ["user"],
+                    "assistant": ["assistant"],
+                    "system": ["system"],
+                },
+                "field_messages": "messages",
+            }
+        ),
+    )
+    return strategy
+
+
+class TestQwen3IdenticalConversationArgs:
+    """
+    Test Qwen3 tools is identical between JSON and dict
+    """
+
+    @pytest.fixture(name="conversation_dict_args_dataset")
+    def fixture_conversation_dict_args_dataset(self):
+        """
+        Provides a dataset with conversation where arguments is a dict.
+        """
+        user_content = "What is the weather in Boston?"
+        function_name = "get_current_weather"
+        arguments_dict = {"location": "Boston, MA", "unit": "celsius"}
+
+        data = [
+            {
+                "messages": [
+                    {"role": "user", "content": user_content},
+                    {
+                        "role": "assistant",
+                        "content": "",
+                        "tool_calls": [
+                            {
+                                "function": {
+                                    "name": function_name,
+                                    "arguments": arguments_dict,  # dict格式
+                                }
+                            }
+                        ],
+                    },
+                ],
+            }
+        ]
+        return Dataset.from_list(data)
+
+    @pytest.fixture(name="conversation_str_args_dataset")
+    def fixture_conversation_str_args_dataset(self):
+        """
+        Provides a dataset with conversation where arguments is a JSON string.
+        """
+        user_content = "What is the weather in Boston?"
+        function_name = "get_current_weather"
+        arguments_dict = {"location": "Boston, MA", "unit": "celsius"}
+        arguments_str = json.dumps(arguments_dict)
+
+        data = [
+            {
+                "messages": [
+                    {"role": "user", "content": user_content},
+                    {
+                        "role": "assistant",
+                        "content": "",
+                        "tool_calls": [
+                            {
+                                "function": {
+                                    "name": function_name,
+                                    "arguments": arguments_str,  # str格式
+                                }
+                            }
+                        ],
+                    },
+                ],
+            }
+        ]
+        return Dataset.from_list(data)
+
+    @pytest.fixture(name="conversation_mixed_time_types_dataset")
+    def fixture_conversation_mixed_time_types_dataset(self):
+        """
+        Provides a dataset where 'time' field has different types in different tool calls.
+        """
+        data = [
+            {
+                "messages": [
+                    {
+                        "role": "user",
+                        "content": "Get weather information at different times",
+                    },
+                    {
+                        "role": "assistant",
+                        "content": "",
+                        "tool_calls": [
+                            {
+                                "function": {
+                                    "name": "func1",
+                                    "arguments": json.dumps(
+                                        {"time": "2025-08-01"}
+                                    ),  # string type
+                                }
+                            },
+                            {
+                                "function": {
+                                    "name": "func2",
+                                    "arguments": json.dumps(
+                                        {"time": 1690876800}
+                                    ),  # number type
+                                }
+                            },
+                        ],
+                    },
+                ],
+            }
+        ]
+        return Dataset.from_list(data)
+
+    def test_dict_and_str_args_produce_identical_output(
+        self,
+        conversation_dict_args_dataset,
+        conversation_str_args_dataset,
+        qwen3_instruct_prompt_strategy,
+        qwen3_tokenizer,
+    ):
+        """
+        Tests that after tokenization and decoding, the outputs for both
+        dict and string `arguments` are exactly the same.
+        """
+        processed_dict_args = conversation_dict_args_dataset.map(
+            qwen3_instruct_prompt_strategy.tokenize_prompt,
+            batched=True,
+            remove_columns=["messages"],
+        )
+
+        processed_str_args = conversation_str_args_dataset.map(
+            qwen3_instruct_prompt_strategy.tokenize_prompt,
+            batched=True,
+            remove_columns=["messages"],
+        )
+
+        decoded_prompt_from_dict = qwen3_tokenizer.decode(
+            processed_dict_args[0]["input_ids"]
+        )
+
+        decoded_prompt_from_str = qwen3_tokenizer.decode(
+            processed_str_args[0]["input_ids"]
+        )
+
+        assert decoded_prompt_from_dict == decoded_prompt_from_str, (
+            f"Dict format output:\n{decoded_prompt_from_dict}\n"
+            f"String format output:\n{decoded_prompt_from_str}"
+        )
+
+        assert (
+            processed_dict_args[0]["input_ids"] == processed_str_args[0]["input_ids"]
+        ), "The tokenized input_ids should be identical for dict and str arguments"
+
+    def test_str_args_with_mixed_time_types_no_error(
+        self,
+        conversation_mixed_time_types_dataset,
+        qwen3_instruct_prompt_strategy,
+        qwen3_tokenizer,
+    ):
+        """
+        Tests that when 'time' field has different types (string vs number)
+        in different tool calls, str format arguments don't cause errors.
+        """
+        processed = conversation_mixed_time_types_dataset.map(
+            qwen3_instruct_prompt_strategy.tokenize_prompt,
+            batched=True,
+            remove_columns=["messages"],
+        )
+
+        assert len(processed) == 1
+        assert "input_ids" in processed[0]
+        assert len(processed[0]["input_ids"]) > 0
+
+        decoded = qwen3_tokenizer.decode(processed[0]["input_ids"])
+        assert "2025-08-01" in decoded, "String time value should be present"
+        assert "1690876800" in decoded, "Number time value should be present"

tests/test_train_context_parallel.py

@@ -1,111 +0,0 @@
-"""Unit tests for choosing the correct context parallel implementation."""
-
-from types import SimpleNamespace
-
-from axolotl.train import execute_training
-from axolotl.utils.dict import DictDefault
-
-
-class DummyTrainer:
-    """Minimal trainer stub to exercise execute_training."""
-
-    def __init__(self):
-        self.model = object()
-        self.ref_model = None
-        self.accelerator = SimpleNamespace(torch_device_mesh=None)
-        self.train_called = False
-
-    def train(self, resume_from_checkpoint=None):  # pylint: disable=unused-argument
-        self.train_called = True
-
-
-class DummyPluginManager:
-    """Minimal plugin manager stub."""
-
-    @staticmethod
-    def post_train(cfg, model):  # pylint: disable=unused-argument
-        return None
-
-
-class DummyContext:
-    """Test context manager that records entries/exits."""
-
-    def __init__(self, recorder, **kwargs):
-        recorder.append({"kwargs": kwargs})
-        self.recorder = recorder
-
-    def __enter__(self):
-        self.recorder[-1]["entered"] = True
-        return self
-
-    def __exit__(self, exc_type, exc, tb):  # pylint: disable=unused-argument
-        self.recorder[-1]["exited"] = True
-        return False
-
-
-def _base_cfg(**overrides):
-    base = {
-        "context_parallel_size": 2,
-        "gradient_accumulation_steps": 1,
-        "ring_attn_func": None,
-        "heads_k_stride": None,
-        "rl": None,
-        "flash_optimum": False,
-    }
-    base.update(overrides)
-    return DictDefault(base)
-
-
-def test_execute_training_uses_ring_when_flash(monkeypatch):
-    """FlashAttention CP should engage the custom ring context manager."""
-    recorder: list[dict] = []
-
-    monkeypatch.setattr(
-        "axolotl.train.SequenceParallelContextManager",
-        lambda **kwargs: DummyContext(recorder, **kwargs),
-    )
-    monkeypatch.setattr(
-        "axolotl.train.PluginManager.get_instance",
-        lambda: DummyPluginManager(),
-    )
-
-    cfg = _base_cfg(flash_attention=True, sdp_attention=False)
-    trainer = DummyTrainer()
-
-    execute_training(cfg, trainer, resume_from_checkpoint=None)
-
-    assert trainer.train_called
-    assert len(recorder) == 1
-    assert recorder[0]["kwargs"]["context_parallel_size"] == 2
-    assert recorder[0].get("entered") is True
-    assert recorder[0].get("exited") is True
-
-
-def test_execute_training_uses_transformers_cp_for_sdpa(monkeypatch):
-    """SDPA CP should bypass the ring context manager."""
-    invoked = {"count": 0}
-
-    class NoOpContext:
-        def __enter__(self):
-            return self
-
-        def __exit__(self, exc_type, exc, tb):  # pylint: disable=unused-argument
-            return False
-
-    monkeypatch.setattr(
-        "axolotl.train.SequenceParallelContextManager",
-        lambda **kwargs: invoked.__setitem__("count", invoked["count"] + 1)
-        or NoOpContext(),
-    )
-    monkeypatch.setattr(
-        "axolotl.train.PluginManager.get_instance",
-        lambda: DummyPluginManager(),
-    )
-
-    cfg = _base_cfg(flash_attention=False, sdp_attention=True)
-    trainer = DummyTrainer()
-
-    execute_training(cfg, trainer, resume_from_checkpoint=None)
-
-    assert trainer.train_called
-    assert invoked["count"] == 0