Merge nightly into feature/transformers-v5-support

setup.sh

@@ -24,6 +24,11 @@ echo "╔═══════════════════════
 echo "║     Unsloth Studio Setup Script      ║"
 echo "╚══════════════════════════════════════╝"
 
+# ── Clean up stale Unsloth compiled caches ──
+rm -rf "$SCRIPT_DIR/unsloth_compiled_cache"
+rm -rf "$SCRIPT_DIR/studio/backend/unsloth_compiled_cache"
+rm -rf "$SCRIPT_DIR/studio/tmp/unsloth_compiled_cache"
+
 # ── Detect Colab (like unsloth does) ──
 IS_COLAB=false
 keynames=$'\n'$(printenv | cut -d= -f1)

studio/backend/assets/configs/model_defaults/other/unsloth_GLM-4.7-Flash.yaml

@@ -0,0 +1,50 @@
+# Model defaults for unsloth/GLM-4.7-Flash
+# Based on GLM_Flash_A100(80GB).py
+# Also applies to: unsloth/GLM-4.7-Flash-unsloth-bnb-4bit, unsloth/GLM-4.7-Flash-bnb-4bit, THUDM/GLM-4.7-Flash
+
+training:
+  max_seq_length: 2048
+  # num_epochs: 4
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 4
+  gradient_accumulation_steps: 2
+  warmup_steps: 5
+  max_steps: 60
+  save_steps: 60
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 8
+  lora_alpha: 16
+  lora_dropout: 0.0
+  target_modules:
+    - "q_proj"
+    - "k_proj"
+    - "v_proj"
+    - "o_proj"
+    - "gate_proj"
+    - "up_proj"
+    - "down_proj"
+    - "out_proj"
+  use_rslora: false
+  use_loftq: false
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  temperature: 0.7
+  top_p: 0.8
+  top_k: 20
+

studio/backend/assets/configs/model_defaults/qwen/imdatta0_tiny_qwen3_moe_2.8B_0.7B.yaml

@@ -0,0 +1,51 @@
+# Model defaults for imdatta0/tiny_qwen3_moe_2.8B_0.7B
+# Based on TinyQwen3_MoE.py
+# Dummy model of qwen3moe architecture created to fit in T4
+# MoE model - includes gate_up_proj for MoE layers
+
+training:
+  max_seq_length: 2048
+  # num_epochs: 4
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 1
+  gradient_accumulation_steps: 1
+  warmup_steps: 5
+  max_steps: 50
+  save_steps: 50
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 32
+  lora_alpha: 64
+  lora_dropout: 0.0
+  target_modules:
+    - "q_proj"
+    - "k_proj"
+    - "v_proj"
+    - "o_proj"
+    - "gate_proj"
+    - "up_proj"
+    - "down_proj"
+    - "gate_up_proj"
+  use_rslora: false
+  use_loftq: false
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  temperature: 0.6
+  top_k: 20
+  top_p: 0.95
+

studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-30B-A3B-Instruct-2507.yaml

@@ -0,0 +1,51 @@
+# Model defaults for unsloth/Qwen3-30B-A3B-Instruct-2507
+# Based on Qwen3_MoE.py
+# Also applies to: Qwen/Qwen3-30B-A3B-Instruct-2507, unsloth/Qwen3-30B-A3B-Instruct-2507-bnb-4bit
+# MoE model - includes gate_up_proj for MoE layers
+
+training:
+  max_seq_length: 2048
+  # num_epochs: 4
+  num_epochs: 0
+  learning_rate: 2e-4
+  batch_size: 1
+  gradient_accumulation_steps: 1
+  warmup_steps: 5
+  max_steps: 50
+  save_steps: 50
+  weight_decay: 0.001
+  random_seed: 3407
+  packing: false
+  train_on_completions: true
+  gradient_checkpointing: "unsloth"
+  optim: "adamw_8bit"
+  lr_scheduler_type: "linear"
+
+lora:
+  lora_r: 32
+  lora_alpha: 64
+  lora_dropout: 0.0
+  target_modules:
+    - "q_proj"
+    - "k_proj"
+    - "v_proj"
+    - "o_proj"
+    - "gate_proj"
+    - "up_proj"
+    - "down_proj"
+    - "gate_up_proj"
+  use_rslora: false
+  use_loftq: false
+
+logging:
+  enable_wandb: false
+  wandb_project: "llm-finetuning"
+  enable_tensorboard: false
+  tensorboard_dir: "runs"
+  log_frequency: 10
+
+inference:
+  temperature: 0.6
+  top_k: 20
+  top_p: 0.95
+

studio/backend/core/inference/inference.py

@@ -184,6 +184,10 @@ class InferenceBackend:
                 # Clear GPU memory cache
                 clear_gpu_cache()
 
+                # Remove stale compiled cache so the next model gets a fresh one
+                from utils.cache_cleanup import clear_unsloth_compiled_cache
+                clear_unsloth_compiled_cache()
+
                 logger.info(f"Model '{model_name}' successfully unloaded.")
                 return True
             except Exception as e:

studio/backend/core/training/trainer.py

@@ -128,6 +128,10 @@ class UnslothTrainer:
             print("\nClearing GPU memory before training...")
             clear_gpu_cache()
 
+            # Remove stale compiled cache so the new model gets a fresh one
+            from utils.cache_cleanup import clear_unsloth_compiled_cache
+            clear_unsloth_compiled_cache()
+
             # Detect if this is a vision model AND dataset is multimodal
             # A vision-capable model with a text-only dataset should use FastLanguageModel
             self.is_vlm = is_vision_model(model_name) and is_dataset_multimodal

studio/backend/main.py

@@ -3,7 +3,6 @@ Main FastAPI application for Unsloth UI Backend
 """
 import os
 import secrets
-import shutil
 from contextlib import asynccontextmanager
 
 from fastapi import FastAPI
@@ -19,12 +18,15 @@ from auth import storage
 from utils.hardware import detect_hardware, get_device, DeviceType
 import utils.hardware.hardware as _hw_module
 
-UNSLOTH_CACHE_DIR = Path(__file__).parent / "unsloth_compiled_cache"
+from utils.cache_cleanup import clear_unsloth_compiled_cache
 
 
 @asynccontextmanager
 async def lifespan(app: FastAPI):
     """Startup: detect hardware, print setup token if needed. Shutdown: clean up compiled cache."""
+    # Clean up any stale compiled cache from previous runs
+    clear_unsloth_compiled_cache()
+
     # Remove stale .venv_overlay from previous sessions — it will be
     # rebuilt at runtime if a model needs transformers 5.x
     overlay_dir = Path(__file__).resolve().parent.parent.parent / ".venv_overlay"
@@ -58,7 +60,7 @@ async def lifespan(app: FastAPI):
     yield
     # Cleanup
     _hw_module.DEVICE = None
-    shutil.rmtree(UNSLOTH_CACHE_DIR, ignore_errors=True)
+    clear_unsloth_compiled_cache()
 
 
 # Create FastAPI app

studio/backend/utils/cache_cleanup.py

@@ -0,0 +1,30 @@
+"""
+Utility for cleaning up the Unsloth compiled cache directory.
+
+The unsloth_compiled_cache is created by unsloth_zoo/compiler.py during
+FastModel.from_pretrained() and contains model-type-specific compiled Python
+files. It should be cleared between model loads to avoid stale artefacts.
+"""
+import shutil
+import logging
+from pathlib import Path
+
+logger = logging.getLogger(__name__)
+
+# Possible locations where unsloth_compiled_cache may appear
+_BACKEND_DIR = Path(__file__).resolve().parent.parent          # studio/backend
+_PROJECT_ROOT = _BACKEND_DIR.parent.parent                     # repo root
+
+_CACHE_DIRS = [
+    _BACKEND_DIR / "unsloth_compiled_cache",
+    _PROJECT_ROOT / "unsloth_compiled_cache",
+    _PROJECT_ROOT / "studio" / "tmp" / "unsloth_compiled_cache",
+]
+
+
+def clear_unsloth_compiled_cache() -> None:
+    """Remove every known unsloth_compiled_cache directory (idempotent)."""
+    for cache_dir in _CACHE_DIRS:
+        if cache_dir.exists():
+            logger.info(f"Removing unsloth compiled cache: {cache_dir}")
+            shutil.rmtree(cache_dir, ignore_errors=True)

studio/backend/utils/datasets/model_mappings.py

@@ -383,7 +383,13 @@ TEMPLATE_TO_MODEL_MAPPER = {
         "unsloth/yi-34b-chat-bnb-4bit",
         "01-ai/Yi-6B-Chat",
         "01-ai/Yi-34B-Chat",
-    )
+    ),
+    "glm": (
+        "unsloth/GLM-4.7-Flash-unsloth-bnb-4bit",
+        "unsloth/GLM-4.7-Flash",
+        "THUDM/GLM-4.7-Flash",
+        "unsloth/GLM-4.7-Flash-bnb-4bit",
+    ),
 }
 
 MODEL_TO_TEMPLATE_MAPPER = {}
@@ -506,4 +512,9 @@ TEMPLATE_TO_RESPONSES_MAPPER = {
         "instruction": "<|im_start|>user\n",
         "response": "<|im_start|>assistant\n",
     },
+    "glm": {
+        "instruction": "[gMASK]<sop><|user|>",
+        "response": "<|assistant|><think>",
+    },
 }
+