Add Ministral as a first-class base model

README.md

@@ -20,7 +20,7 @@
 
 ___
 
-`xTuring` makes it simple, fast, and cost‑efficient to fine‑tune open‑source LLMs (e.g., GPT‑OSS, LLaMA/LLaMA 2, Falcon, Qwen3, GPT‑J, GPT‑2, OPT, Bloom, Cerebras, Galactica) on your own data — locally or in your private cloud.
+`xTuring` makes it simple, fast, and cost‑efficient to fine‑tune open‑source LLMs (e.g., GPT‑OSS, LLaMA/LLaMA 2, Mistral/Ministral, Falcon, Qwen3, GPT‑J, GPT‑2, OPT, Bloom, Cerebras, Galactica) on your own data — locally or in your private cloud.
 
 Why xTuring:
 - Simple API for data prep, training, and inference
@@ -276,6 +276,8 @@ Below is a list of all the supported models via `BaseModel` class of `xTuring` a
 |GPT-2 | gpt2|
 |LLaMA | llama|
 |LLaMA2 | llama2|
+|Mistral-7B | mistral_7b|
+|Ministral 3.14B | ministral_3_14b|
 |MiniMaxM2 | minimax_m2|
 |OPT-1.3B | opt|
 

docs/docs/overview/supported_models.md

@@ -33,6 +33,8 @@ Use one task notebook from `examples/notebooks/`, then choose a model key from t
 | LLaMA | `llama` | `base`, `lora`, `int8`, `lora_int8`, `lora_kbit` |
 | LLaMA 2 | `llama2` | `base`, `lora`, `int8`, `lora_int8`, `lora_kbit` |
 | Mamba | `mamba` | `base` |
+| Mistral 7B | `mistral_7b` | `base` |
+| Ministral 3.14B | `ministral_3_14b` | `base`, `lora`, `int8`, `lora_int8`, `lora_kbit` |
 | MiniMaxM2 | `minimax_m2` | `base`, `lora`, `int8`, `lora_int8`, `lora_kbit` |
 | OPT 1.3B | `opt` | `base`, `lora`, `int8`, `lora_int8` |
 | Qwen3 0.6B | `qwen3_0_6b` | `base`, `lora`, `int8`, `lora_int8`, `lora_kbit` |

src/xturing/config/finetuning_config.yaml

@@ -302,6 +302,58 @@ mamba:
   learning_rate: 5e-5
   weight_decay: 0.01
 
+mistral_7b:
+  learning_rate: 5e-5
+  weight_decay: 0.01
+  num_train_epochs: 3
+  optimizer_name: cpu_adam
+
+# Ministral 3.14B model fine-tuning configurations
+ministral_3_14b:
+  learning_rate: 1e-5
+  weight_decay: 0.01
+  num_train_epochs: 1
+  batch_size: 1
+  gradient_accumulation_steps: 8
+  max_length: 2048
+  warmup_steps: 100
+
+ministral_3_14b_lora:
+  learning_rate: 2e-4
+  weight_decay: 0.01
+  num_train_epochs: 3
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  max_length: 2048
+  warmup_steps: 100
+
+ministral_3_14b_int8:
+  learning_rate: 1e-4
+  weight_decay: 0.01
+  num_train_epochs: 2
+  batch_size: 2
+  gradient_accumulation_steps: 4
+  max_length: 2048
+  warmup_steps: 100
+
+ministral_3_14b_lora_int8:
+  learning_rate: 2e-4
+  weight_decay: 0.01
+  num_train_epochs: 3
+  batch_size: 4
+  gradient_accumulation_steps: 2
+  max_length: 2048
+  warmup_steps: 100
+
+ministral_3_14b_lora_kbit:
+  learning_rate: 2e-4
+  weight_decay: 0.01
+  num_train_epochs: 3
+  batch_size: 8
+  gradient_accumulation_steps: 1
+  max_length: 2048
+  warmup_steps: 100
+
 # MiniMaxM2 model fine-tuning configurations
 minimax_m2:
   learning_rate: 1e-5

src/xturing/config/generation_config.yaml

@@ -316,6 +316,43 @@ llama2_lora_kbit:
 mamba:
   do_sample: false
 
+# Contrastive search
+mistral_7b:
+  penalty_alpha: 0.6
+  top_k: 4
+  max_new_tokens: 256
+  do_sample: false
+
+# Contrastive search for Ministral 3.14B
+ministral_3_14b:
+  penalty_alpha: 0.6
+  top_k: 4
+  max_new_tokens: 512
+  do_sample: false
+  temperature: 0.1
+
+ministral_3_14b_lora:
+  penalty_alpha: 0.6
+  top_k: 4
+  max_new_tokens: 512
+  do_sample: false
+  temperature: 0.1
+
+ministral_3_14b_int8:
+  max_new_tokens: 512
+  do_sample: false
+  temperature: 0.1
+
+ministral_3_14b_lora_int8:
+  max_new_tokens: 512
+  do_sample: false
+  temperature: 0.1
+
+ministral_3_14b_lora_kbit:
+  max_new_tokens: 512
+  do_sample: false
+  temperature: 0.1
+
 # Contrastive search for MiniMaxM2
 minimax_m2:
   penalty_alpha: 0.6

src/xturing/engines/__init__.py

@@ -78,6 +78,14 @@
     MiniMaxM2LoraInt8Engine,
     MiniMaxM2LoraKbitEngine,
 )
+from xturing.engines.mistral_engine import (
+    Ministral314BEngine,
+    Ministral314BInt8Engine,
+    Ministral314BLoraEngine,
+    Ministral314BLoraInt8Engine,
+    Ministral314BLoraKbitEngine,
+    Mistral7BEngine,
+)
 from xturing.engines.opt_engine import (
     OPTEngine,
     OPTInt8Engine,
@@ -149,6 +157,16 @@
 BaseEngine.add_to_registry(LLama2LoraInt8Engine.config_name, LLama2LoraInt8Engine)
 BaseEngine.add_to_registry(LLama2LoraKbitEngine.config_name, LLama2LoraKbitEngine)
 BaseEngine.add_to_registry(MambaEngine.config_name, MambaEngine)
+BaseEngine.add_to_registry(Mistral7BEngine.config_name, Mistral7BEngine)
+BaseEngine.add_to_registry(Ministral314BEngine.config_name, Ministral314BEngine)
+BaseEngine.add_to_registry(Ministral314BInt8Engine.config_name, Ministral314BInt8Engine)
+BaseEngine.add_to_registry(Ministral314BLoraEngine.config_name, Ministral314BLoraEngine)
+BaseEngine.add_to_registry(
+    Ministral314BLoraInt8Engine.config_name, Ministral314BLoraInt8Engine
+)
+BaseEngine.add_to_registry(
+    Ministral314BLoraKbitEngine.config_name, Ministral314BLoraKbitEngine
+)
 BaseEngine.add_to_registry(MiniMaxM2Engine.config_name, MiniMaxM2Engine)
 BaseEngine.add_to_registry(MiniMaxM2Int8Engine.config_name, MiniMaxM2Int8Engine)
 BaseEngine.add_to_registry(MiniMaxM2LoraEngine.config_name, MiniMaxM2LoraEngine)

src/xturing/engines/mistral_engine.py

@@ -0,0 +1,99 @@
+from pathlib import Path
+from typing import Optional, Union
+
+from xturing.engines.causal import CausalEngine, CausalLoraEngine, CausalLoraKbitEngine
+
+_MISTRAL_MODEL_NAME = "mistralai/Mistral-7B-v0.1"
+_MINISTRAL_MODEL_NAME = "mistralai/Ministral-3-14B-Instruct-2512"
+_MINISTRAL_TARGET_MODULES = [
+    "q_proj",
+    "k_proj",
+    "v_proj",
+    "o_proj",
+    "gate_proj",
+    "up_proj",
+    "down_proj",
+]
+
+
+class Mistral7BEngine(CausalEngine):
+    config_name: str = "mistral_7b_engine"
+
+    def __init__(self, weights_path: Optional[Union[str, Path]] = None):
+        super().__init__(
+            model_name=_MISTRAL_MODEL_NAME,
+            weights_path=weights_path,
+        )
+
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+
+class Ministral314BEngine(CausalEngine):
+    config_name: str = "ministral_3_14b_engine"
+
+    def __init__(self, weights_path: Optional[Union[str, Path]] = None):
+        super().__init__(
+            model_name=_MINISTRAL_MODEL_NAME,
+            weights_path=weights_path,
+        )
+
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+
+class Ministral314BLoraEngine(CausalLoraEngine):
+    config_name: str = "ministral_3_14b_lora_engine"
+
+    def __init__(self, weights_path: Optional[Union[str, Path]] = None):
+        super().__init__(
+            model_name=_MINISTRAL_MODEL_NAME,
+            weights_path=weights_path,
+            target_modules=_MINISTRAL_TARGET_MODULES,
+        )
+
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+
+class Ministral314BInt8Engine(CausalEngine):
+    config_name: str = "ministral_3_14b_int8_engine"
+
+    def __init__(self, weights_path: Optional[Union[str, Path]] = None):
+        super().__init__(
+            model_name=_MINISTRAL_MODEL_NAME,
+            weights_path=weights_path,
+            load_8bit=True,
+        )
+
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+
+class Ministral314BLoraInt8Engine(CausalLoraEngine):
+    config_name: str = "ministral_3_14b_lora_int8_engine"
+
+    def __init__(self, weights_path: Optional[Union[str, Path]] = None):
+        super().__init__(
+            model_name=_MINISTRAL_MODEL_NAME,
+            weights_path=weights_path,
+            load_8bit=True,
+            target_modules=_MINISTRAL_TARGET_MODULES,
+        )
+
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id
+
+
+class Ministral314BLoraKbitEngine(CausalLoraKbitEngine):
+    config_name: str = "ministral_3_14b_lora_kbit_engine"
+
+    def __init__(self, weights_path: Optional[Union[str, Path]] = None):
+        super().__init__(
+            model_name=_MINISTRAL_MODEL_NAME,
+            weights_path=weights_path,
+            target_modules=_MINISTRAL_TARGET_MODULES,
+        )
+
+        self.tokenizer.pad_token = self.tokenizer.eos_token
+        self.tokenizer.pad_token_id = self.tokenizer.eos_token_id

src/xturing/models/__init__.py

@@ -63,6 +63,14 @@
     MiniMaxM2LoraInt8,
     MiniMaxM2LoraKbit,
 )
+from xturing.models.mistral import (
+    Ministral314B,
+    Ministral314BInt8,
+    Ministral314BLora,
+    Ministral314BLoraInt8,
+    Ministral314BLoraKbit,
+    Mistral7B,
+)
 from xturing.models.opt import OPT, OPTInt8, OPTLora, OPTLoraInt8
 from xturing.models.qwen import (
     Qwen3,
@@ -126,6 +134,12 @@
 BaseModel.add_to_registry(Llama2LoraInt8.config_name, Llama2LoraInt8)
 BaseModel.add_to_registry(Llama2LoraKbit.config_name, Llama2LoraKbit)
 BaseModel.add_to_registry(Mamba.config_name, Mamba)
+BaseModel.add_to_registry(Mistral7B.config_name, Mistral7B)
+BaseModel.add_to_registry(Ministral314B.config_name, Ministral314B)
+BaseModel.add_to_registry(Ministral314BInt8.config_name, Ministral314BInt8)
+BaseModel.add_to_registry(Ministral314BLora.config_name, Ministral314BLora)
+BaseModel.add_to_registry(Ministral314BLoraInt8.config_name, Ministral314BLoraInt8)
+BaseModel.add_to_registry(Ministral314BLoraKbit.config_name, Ministral314BLoraKbit)
 BaseModel.add_to_registry(MiniMaxM2.config_name, MiniMaxM2)
 BaseModel.add_to_registry(MiniMaxM2Int8.config_name, MiniMaxM2Int8)
 BaseModel.add_to_registry(MiniMaxM2Lora.config_name, MiniMaxM2Lora)

src/xturing/models/mistral.py

@@ -0,0 +1,59 @@
+from typing import Optional
+
+from xturing.engines.mistral_engine import (
+    Ministral314BEngine,
+    Ministral314BInt8Engine,
+    Ministral314BLoraEngine,
+    Ministral314BLoraInt8Engine,
+    Ministral314BLoraKbitEngine,
+    Mistral7BEngine,
+)
+from xturing.models.causal import (
+    CausalInt8Model,
+    CausalLoraInt8Model,
+    CausalLoraKbitModel,
+    CausalLoraModel,
+    CausalModel,
+)
+
+
+class Mistral7B(CausalModel):
+    config_name: str = "mistral_7b"
+
+    def __init__(self, weights_path: Optional[str] = None):
+        super().__init__(Mistral7BEngine.config_name, weights_path)
+
+
+class Ministral314B(CausalModel):
+    config_name: str = "ministral_3_14b"
+
+    def __init__(self, weights_path: Optional[str] = None):
+        super().__init__(Ministral314BEngine.config_name, weights_path)
+
+
+class Ministral314BLora(CausalLoraModel):
+    config_name: str = "ministral_3_14b_lora"
+
+    def __init__(self, weights_path: Optional[str] = None):
+        super().__init__(Ministral314BLoraEngine.config_name, weights_path)
+
+
+class Ministral314BInt8(CausalInt8Model):
+    config_name: str = "ministral_3_14b_int8"
+
+    def __init__(self, weights_path: Optional[str] = None):
+        super().__init__(Ministral314BInt8Engine.config_name, weights_path)
+
+
+class Ministral314BLoraInt8(CausalLoraInt8Model):
+    config_name: str = "ministral_3_14b_lora_int8"
+
+    def __init__(self, weights_path: Optional[str] = None):
+        super().__init__(Ministral314BLoraInt8Engine.config_name, weights_path)
+
+
+class Ministral314BLoraKbit(CausalLoraKbitModel):
+    config_name: str = "ministral_3_14b_lora_kbit"
+
+    def __init__(self, weights_path: Optional[str] = None):
+        super().__init__(Ministral314BLoraKbitEngine.config_name, weights_path)