feat: Add Mistral Medium 3.5 (#3633)

* fix: clarify incompat

* fix: transformers api change upstream

* fix: add pre prop

* feat: add examples

* chore: cleanup

* chore: update readme

examples/devstral/devstral-small-qlora.yml

@@ -26,7 +26,6 @@ lora_model_dir:
 sequence_len: 2048
 sample_packing: true
 
-
 lora_r: 32
 lora_alpha: 16
 lora_dropout: 0
@@ -52,7 +51,7 @@ gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
-scaling_softmax: true
+# scaling_softmax: true  # needs flex_attention
 
 loss_watchdog_threshold: 5.0
 loss_watchdog_patience: 3

examples/ministral3/ministral3-3b-qlora.yaml

@@ -59,7 +59,7 @@ gradient_checkpointing: true
 resume_from_checkpoint:
 logging_steps: 1
 flash_attention: true
-scaling_softmax: true
+# scaling_softmax: true  # needs flex_attention
 
 warmup_ratio: 0.1
 evals_per_epoch: 1

examples/mistral-medium-3_5/README.md

@@ -0,0 +1,78 @@
+# Finetune Mistral Medium 3.5 with Axolotl
+
+[Mistral Medium 3.5](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B) is a 128B parameter dense multimodal model from MistralAI that unifies instruct, reasoning, and agentic capabilities into a single model.
+It shares the `mistral3` architecture (dense, YaRN RoPE, 256k context) with Ministral 3 and supports the same `reasoning_effort` toggle as Mistral Small 4.
+
+Thanks to the team at MistralAI for giving us early access to prepare for this release.
+
+## Getting started
+
+1. Install Axolotl following the [installation guide](https://docs.axolotl.ai/docs/installation.html).
+
+2. Install [Cut Cross Entropy](https://docs.axolotl.ai/docs/custom_integrations.html#cut-cross-entropy) to reduce training VRAM usage.
+
+3. (Text config only) Install [Flash Attention 4](https://docs.axolotl.ai/docs/attention.html#flash-attention-4) on Hopper/Blackwell.
+
+4. Run one of the example configs:
+
+    ```bash
+    # text-only
+    axolotl train examples/mistral-medium-3_5/qlora-text.yml  # ~83.1 GiB
+
+    # text + vision
+    # wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+    axolotl train examples/mistral-medium-3_5/qlora-vision.yml  # ~80.3 GiB
+    ```
+
+Note: vision training does not currently work with Flash Attention 4.
+
+## Reasoning Effort
+
+The chat template supports a `reasoning_effort` variable to control the model's reasoning depth:
+
+- `"none"` — instruct mode (default)
+- `"high"` — reasoning mode with explicit thinking steps
+
+Pass it via `chat_template_kwargs` under your dataset config:
+
+```yaml
+datasets:
+  - path: your/dataset
+    type: chat_template
+    chat_template_kwargs:
+      reasoning_effort: high
+```
+
+## Thinking Support
+
+The chat template supports a `thinking` content type in assistant messages for training on reasoning traces (rendered as `[THINK]...[/THINK]` blocks).
+
+To use thinking datasets, add the `thinking` mapping via `message_property_mappings`:
+
+```yaml
+datasets:
+  - path: your/thinking-dataset
+    type: chat_template
+    message_property_mappings:
+      role: role
+      content: content
+      thinking: thinking
+    chat_template_kwargs:
+      reasoning_effort: high
+```
+
+See the [Magistral thinking guide](../magistral/think/README.md) for dataset format details.
+
+## Tips
+
+- For smaller experiments on the same architecture, see [`examples/ministral3`](../ministral3/README.md) (Ministral 3, 3B).
+- Read more on how to load your own dataset at [docs](https://docs.axolotl.ai/docs/dataset_loading.html).
+- The text dataset format follows the OpenAI Messages format as seen [here](https://docs.axolotl.ai/docs/dataset-formats/conversation.html#chat_template).
+- The vision model requires multi-modal dataset format as documented [here](https://docs.axolotl.ai/docs/multimodal.html#dataset-format).
+
+## Related Resources
+
+- [Mistral Medium 3.5 Blog](https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5)
+- [Axolotl Docs](https://docs.axolotl.ai)
+- [Axolotl GitHub](https://github.com/axolotl-ai-cloud/axolotl)
+- [Axolotl Discord](https://discord.gg/7m9sfhzaf3)

examples/mistral-medium-3_5/qlora-text.yml

@@ -0,0 +1,56 @@
+base_model: axolotl-ai-co/Mistral-Medium-3.5-128B-BF16
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+
+datasets:
+  - path: fozziethebeat/alpaca_messages_2k_test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir:
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+sample_packing: true
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+# prevents targeting vision layers
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+max_steps: 10
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1

examples/mistral-medium-3_5/qlora-vision.yml

@@ -0,0 +1,61 @@
+base_model: axolotl-ai-co/Mistral-Medium-3.5-128B-BF16
+processor_type: AutoProcessor
+
+plugins:
+  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
+
+load_in_4bit: true
+
+skip_prepare_dataset: true
+remove_unused_columns: false
+sample_packing: false
+
+# sample dataset below requires downloading image in advance
+# wget https://huggingface.co/datasets/Nanobit/text-vision-2k-test/resolve/main/African_elephant.jpg
+datasets:
+  - path: Nanobit/text-vision-2k-test
+    type: chat_template
+
+dataset_prepared_path: last_run_prepared
+val_set_size: 0.1
+output_dir:
+
+adapter: qlora
+lora_model_dir:
+
+sequence_len: 2048
+
+lora_r: 32
+lora_alpha: 16
+lora_dropout: 0
+lora_target_modules: 'model.language_model.layers.[\d]+.(mlp|self_attn).(up|down|gate|q|k|v|o)_proj'
+
+lora_mlp_kernel: true
+lora_qkv_kernel: true
+lora_o_kernel: true
+
+wandb_project:
+wandb_entity:
+wandb_watch:
+wandb_name:
+wandb_log_model:
+
+gradient_accumulation_steps: 4
+micro_batch_size: 2
+num_epochs: 1
+max_steps: 10
+optimizer: adamw_bnb_8bit
+lr_scheduler: cosine
+learning_rate: 0.0002
+
+bf16: auto
+tf32: false
+
+gradient_checkpointing: true
+resume_from_checkpoint:
+logging_steps: 1
+flash_attention: true
+
+warmup_ratio: 0.1
+evals_per_epoch: 1
+saves_per_epoch: 1