From 46cc71f310eae13fc4bfc419ce623851df3bb1b8 Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Tue, 10 Feb 2026 16:16:17 -0700
Subject: [PATCH 01/29] added the pydantic models and routes for export

---
 studio/backend/main.py            |   1 +
 studio/backend/models/__init__.py |  21 ++
 studio/backend/models/export.py   | 141 ++++++++++++++
 studio/backend/routes/__init__.py |  10 +-
 studio/backend/routes/export.py   | 310 ++++++++++++++++++++++++++++++
 5 files changed, 482 insertions(+), 1 deletion(-)
 create mode 100644 studio/backend/models/export.py
 create mode 100644 studio/backend/routes/export.py

diff --git a/studio/backend/main.py b/studio/backend/main.py
index e957c668e3..34c61bed6b 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -67,6 +67,7 @@ app.include_router(training_router, prefix="/api/train", tags=["training"])
 app.include_router(models_router, prefix="/api/models", tags=["models"])
 app.include_router(inference_router, prefix="/api/inference", tags=["inference"])
 app.include_router(datasets_router, prefix="/api/datasets", tags=["datasets"])
+app.include_router(export_router, prefix="/api/export", tags=["export"])
 
 
 # ============ Health and System Endpoints ============
diff --git a/studio/backend/models/__init__.py b/studio/backend/models/__init__.py
index 584fb9f7c2..8d947b1b9c 100644
--- a/studio/backend/models/__init__.py
+++ b/studio/backend/models/__init__.py
@@ -19,6 +19,17 @@ from .auth import (
     RefreshTokenRequest,
     AuthStatusResponse,
 )
+from .export import (
+    CheckpointInfo,
+    CheckpointListResponse,
+    LoadCheckpointRequest,
+    ExportStatusResponse,
+    ExportOperationResponse,
+    ExportMergedModelRequest,
+    ExportBaseModelRequest,
+    ExportGGUFRequest,
+    ExportLoRAAdapterRequest,
+)
 from .users import Token
 from .datasets import (
     CheckFormatRequest,
@@ -55,6 +66,16 @@ __all__ = [
     "AuthLoginRequest",
     "RefreshTokenRequest",
     "AuthStatusResponse",
+    # Export schemas
+    "CheckpointInfo",
+    "CheckpointListResponse",
+    "LoadCheckpointRequest",
+    "ExportStatusResponse",
+    "ExportOperationResponse",
+    "ExportMergedModelRequest",
+    "ExportBaseModelRequest",
+    "ExportGGUFRequest",
+    "ExportLoRAAdapterRequest",
     "Token",
     # Dataset schemas
     "CheckFormatRequest",
diff --git a/studio/backend/models/export.py b/studio/backend/models/export.py
new file mode 100644
index 0000000000..d108c00751
--- /dev/null
+++ b/studio/backend/models/export.py
@@ -0,0 +1,141 @@
+"""
+Pydantic schemas for Export API.
+"""
+from pydantic import BaseModel, Field
+from typing import List, Optional, Literal, Dict, Any
+
+
+class CheckpointInfo(BaseModel):
+    """Information about a discovered checkpoint directory."""
+
+    display_name: str = Field(..., description="User-friendly checkpoint name (folder name)")
+    path: str = Field(..., description="Full path to the checkpoint directory")
+
+
+class CheckpointListResponse(BaseModel):
+    """Response for listing available checkpoints in an outputs directory."""
+
+    outputs_dir: str = Field(..., description="Directory that was scanned")
+    checkpoints: List[CheckpointInfo] = Field(
+        default_factory=list,
+        description="List of discovered checkpoints",
+    )
+
+
+class LoadCheckpointRequest(BaseModel):
+    """Request for loading a checkpoint into the export backend."""
+
+    checkpoint_path: str = Field(..., description="Path to the checkpoint directory")
+    max_seq_length: int = Field(
+        2048,
+        ge=128,
+        le=32768,
+        description="Maximum sequence length for loading the model",
+    )
+    load_in_4bit: bool = Field(
+        True,
+        description="Whether to load the model in 4-bit quantization",
+    )
+
+
+class ExportStatusResponse(BaseModel):
+    """Current export backend status."""
+
+    current_checkpoint: Optional[str] = Field(
+        None,
+        description="Path to the currently loaded checkpoint, if any",
+    )
+    is_vision: bool = Field(
+        False,
+        description="True if the loaded checkpoint is a vision model",
+    )
+    is_peft: bool = Field(
+        False,
+        description="True if the loaded checkpoint is a PEFT (LoRA) model",
+    )
+
+
+class ExportOperationResponse(BaseModel):
+    """Generic response for export operations."""
+
+    success: bool = Field(..., description="True if the operation succeeded")
+    message: str = Field(..., description="Human-readable status or error message")
+    details: Optional[Dict[str, Any]] = Field(
+        default=None,
+        description="Optional extra details about the operation",
+    )
+
+
+class ExportCommonOptions(BaseModel):
+    """Common options for export operations that save locally and/or push to Hub."""
+
+    save_directory: str = Field(
+        ...,
+        description="Local directory where the exported artifacts will be written",
+    )
+    push_to_hub: bool = Field(
+        False,
+        description="If True, also push the exported model to the Hugging Face Hub",
+    )
+    repo_id: Optional[str] = Field(
+        None,
+        description="Hugging Face Hub repository ID (username/model-name)",
+    )
+    hf_token: Optional[str] = Field(
+        None,
+        description="Hugging Face access token used for Hub operations",
+    )
+    private: bool = Field(
+        False,
+        description="If True, create a private repository on the Hub (where applicable)",
+    )
+
+
+class ExportMergedModelRequest(ExportCommonOptions):
+    """Request for exporting a merged PEFT model."""
+
+    format_type: Literal["16-bit (FP16)", "4-bit (FP4)"] = Field(
+        "16-bit (FP16)",
+        description="Export precision / format for the merged model",
+    )
+
+
+class ExportBaseModelRequest(ExportCommonOptions):
+    """Request for exporting a non-PEFT (base) model."""
+
+    # Uses fields from ExportCommonOptions only
+    pass
+
+
+class ExportGGUFRequest(BaseModel):
+    """Request for exporting the current model to GGUF format."""
+
+    save_directory: str = Field(
+        ...,
+        description="Directory where GGUF files will be saved",
+    )
+    quantization_method: str = Field(
+        "Q4_K_M",
+        description='GGUF quantization method (e.g. "Q4_K_M")',
+    )
+    push_to_hub: bool = Field(
+        False,
+        description="If True, also push GGUF artifacts to the Hugging Face Hub",
+    )
+    repo_id: Optional[str] = Field(
+        None,
+        description="Hugging Face Hub repository ID for GGUF upload",
+    )
+    hf_token: Optional[str] = Field(
+        None,
+        description="Hugging Face token for GGUF upload",
+    )
+
+
+class ExportLoRAAdapterRequest(ExportCommonOptions):
+    """Request for exporting only the LoRA adapter (not merged)."""
+
+    # Uses fields from ExportCommonOptions only
+    pass
+
+
diff --git a/studio/backend/routes/__init__.py b/studio/backend/routes/__init__.py
index 5a16125a64..04c7ff7f1a 100644
--- a/studio/backend/routes/__init__.py
+++ b/studio/backend/routes/__init__.py
@@ -7,5 +7,13 @@ from routes.models import router as models_router
 from routes.inference import router as inference_router
 from routes.datasets import router as datasets_router
 from routes.auth import router as auth_router
+from routes.export import router as export_router
 
-__all__ = ["training_router", "models_router", "inference_router", "datasets_router", "auth_router"]
\ No newline at end of file
+__all__ = [
+    "training_router",
+    "models_router",
+    "inference_router",
+    "datasets_router",
+    "auth_router",
+    "export_router",
+]
\ No newline at end of file
diff --git a/studio/backend/routes/export.py b/studio/backend/routes/export.py
new file mode 100644
index 0000000000..b72dea48b8
--- /dev/null
+++ b/studio/backend/routes/export.py
@@ -0,0 +1,310 @@
+"""
+Export API routes: checkpoint discovery and model export operations.
+"""
+
+import sys
+from pathlib import Path
+from fastapi import APIRouter, Depends, HTTPException, Query
+import logging
+
+# Add backend directory to path
+backend_path = Path(__file__).parent.parent.parent
+if str(backend_path) not in sys.path:
+    sys.path.insert(0, str(backend_path))
+
+# Auth
+from auth.authentication import get_current_subject
+
+# Import backend functions
+try:
+    from core.export import get_export_backend
+except ImportError:
+    parent_backend = backend_path.parent / "backend"
+    if str(parent_backend) not in sys.path:
+        sys.path.insert(0, str(parent_backend))
+    from core.export import get_export_backend
+
+# Import Pydantic models
+from models import (
+    CheckpointInfo,
+    CheckpointListResponse,
+    LoadCheckpointRequest,
+    ExportStatusResponse,
+    ExportOperationResponse,
+    ExportMergedModelRequest,
+    ExportBaseModelRequest,
+    ExportGGUFRequest,
+    ExportLoRAAdapterRequest,
+)
+
+router = APIRouter()
+logger = logging.getLogger(__name__)
+
+# Configure logger
+if not logger.handlers:
+    handler = logging.StreamHandler()
+    handler.setLevel(logging.INFO)
+    formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
+    handler.setFormatter(formatter)
+    logger.addHandler(handler)
+    logger.setLevel(logging.INFO)
+
+
+@router.get("/checkpoints", response_model=CheckpointListResponse)
+async def list_checkpoints(
+    outputs_dir: str = Query(
+        default="./outputs",
+        description="Directory to scan for checkpoints",
+    ),
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    List available checkpoints in the outputs directory.
+
+    Wraps ExportBackend.scan_checkpoints.
+    """
+    try:
+        backend = get_export_backend()
+        raw_checkpoints = backend.scan_checkpoints(outputs_dir=outputs_dir)
+
+        checkpoints = [
+            CheckpointInfo(display_name=display_name, path=path)
+            for display_name, path in raw_checkpoints
+        ]
+
+        return CheckpointListResponse(
+            outputs_dir=outputs_dir,
+            checkpoints=checkpoints,
+        )
+    except Exception as e:
+        logger.error(f"Error listing checkpoints: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to list checkpoints: {str(e)}",
+        )
+
+
+@router.post("/load-checkpoint", response_model=ExportOperationResponse)
+async def load_checkpoint(
+    request: LoadCheckpointRequest,
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Load a checkpoint into the export backend.
+
+    Wraps ExportBackend.load_checkpoint.
+    """
+    try:
+        backend = get_export_backend()
+        success, message = backend.load_checkpoint(
+            checkpoint_path=request.checkpoint_path,
+            max_seq_length=request.max_seq_length,
+            load_in_4bit=request.load_in_4bit,
+        )
+
+        if not success:
+            raise HTTPException(status_code=400, detail=message)
+
+        return ExportOperationResponse(success=True, message=message)
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"Error loading checkpoint: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to load checkpoint: {str(e)}",
+        )
+
+
+@router.post("/cleanup", response_model=ExportOperationResponse)
+async def cleanup_export_memory(
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Cleanup export-related models from memory (GPU/CPU).
+
+    Wraps ExportBackend.cleanup_memory.
+    """
+    try:
+        backend = get_export_backend()
+        success = backend.cleanup_memory()
+
+        if not success:
+            raise HTTPException(
+                status_code=500,
+                detail="Memory cleanup failed. See server logs for details.",
+            )
+
+        return ExportOperationResponse(
+            success=True,
+            message="Memory cleanup completed successfully",
+        )
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"Error during export memory cleanup: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to cleanup export memory: {str(e)}",
+        )
+
+
+@router.get("/status", response_model=ExportStatusResponse)
+async def get_export_status(
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Get current export backend status (loaded checkpoint, model type, PEFT flag).
+    """
+    try:
+        backend = get_export_backend()
+        return ExportStatusResponse(
+            current_checkpoint=backend.current_checkpoint,
+            is_vision=bool(getattr(backend, "is_vision", False)),
+            is_peft=bool(getattr(backend, "is_peft", False)),
+        )
+    except Exception as e:
+        logger.error(f"Error getting export status: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to get export status: {str(e)}",
+        )
+
+
+@router.post("/export/merged", response_model=ExportOperationResponse)
+async def export_merged_model(
+    request: ExportMergedModelRequest,
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Export a merged PEFT model (e.g., 16-bit or 4-bit) and optionally push to Hub.
+
+    Wraps ExportBackend.export_merged_model.
+    """
+    try:
+        backend = get_export_backend()
+        success, message = backend.export_merged_model(
+            save_directory=request.save_directory,
+            format_type=request.format_type,
+            push_to_hub=request.push_to_hub,
+            repo_id=request.repo_id,
+            hf_token=request.hf_token,
+            private=request.private,
+        )
+
+        if not success:
+            raise HTTPException(status_code=400, detail=message)
+
+        return ExportOperationResponse(success=True, message=message)
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"Error exporting merged model: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to export merged model: {str(e)}",
+        )
+
+
+@router.post("/export/base", response_model=ExportOperationResponse)
+async def export_base_model(
+    request: ExportBaseModelRequest,
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Export a non-PEFT base model and optionally push to Hub.
+
+    Wraps ExportBackend.export_base_model.
+    """
+    try:
+        backend = get_export_backend()
+        success, message = backend.export_base_model(
+            save_directory=request.save_directory,
+            push_to_hub=request.push_to_hub,
+            repo_id=request.repo_id,
+            hf_token=request.hf_token,
+            private=request.private,
+        )
+
+        if not success:
+            raise HTTPException(status_code=400, detail=message)
+
+        return ExportOperationResponse(success=True, message=message)
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"Error exporting base model: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to export base model: {str(e)}",
+        )
+
+
+@router.post("/export/gguf", response_model=ExportOperationResponse)
+async def export_gguf(
+    request: ExportGGUFRequest,
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Export the current model to GGUF format and optionally push to Hub.
+
+    Wraps ExportBackend.export_gguf.
+    """
+    try:
+        backend = get_export_backend()
+        success, message = backend.export_gguf(
+            save_directory=request.save_directory,
+            quantization_method=request.quantization_method,
+            push_to_hub=request.push_to_hub,
+            repo_id=request.repo_id,
+            hf_token=request.hf_token,
+        )
+
+        if not success:
+            raise HTTPException(status_code=400, detail=message)
+
+        return ExportOperationResponse(success=True, message=message)
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"Error exporting GGUF model: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to export GGUF model: {str(e)}",
+        )
+
+
+@router.post("/export/lora", response_model=ExportOperationResponse)
+async def export_lora_adapter(
+    request: ExportLoRAAdapterRequest,
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Export only the LoRA adapter (if the loaded model is PEFT).
+
+    Wraps ExportBackend.export_lora_adapter.
+    """
+    try:
+        backend = get_export_backend()
+        success, message = backend.export_lora_adapter(
+            save_directory=request.save_directory,
+            push_to_hub=request.push_to_hub,
+            repo_id=request.repo_id,
+            hf_token=request.hf_token,
+            private=request.private,
+        )
+
+        if not success:
+            raise HTTPException(status_code=400, detail=message)
+
+        return ExportOperationResponse(success=True, message=message)
+    except HTTPException:
+        raise
+    except Exception as e:
+        logger.error(f"Error exporting LoRA adapter: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to export LoRA adapter: {str(e)}",
+        )
+
+

From e4b073985fcca99dd3f79c6b88e417269cb25906 Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Wed, 11 Feb 2026 18:51:53 -0700
Subject: [PATCH 02/29] added router in main

---
 studio/backend/main.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/studio/backend/main.py b/studio/backend/main.py
index 34c61bed6b..118c59ee7f 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -13,7 +13,7 @@ from pathlib import Path
 from datetime import datetime
 
 # Import routers
-from routes import training_router, models_router, inference_router, datasets_router, auth_router
+from routes import training_router, models_router, inference_router, datasets_router, auth_router, export_router
 from auth import storage
 from utils.hardware import detect_hardware
 import utils.hardware.hardware as _hw_module

From 8403bac48d75a360f4f666cbb278b8f48c1cd193 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 14:52:13 +0000
Subject: [PATCH 03/29] feat(inference): add use_adapter field for per-request
 adapter toggling in compare mode

---
 studio/backend/core/inference/inference.py | 173 +++++++++++++++------
 studio/backend/models/inference.py         |  10 ++
 studio/backend/routes/inference.py         |  16 +-
 3 files changed, 150 insertions(+), 49 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 0fd1905ff4..f8c0213b9b 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -8,7 +8,7 @@ from peft import PeftModel, PeftModelForCausalLM
 
 import sys
 import torch
-from typing import Optional, Generator, Tuple
+from typing import Optional, Union, Generator, Tuple
 from utils.models import ModelConfig, get_base_model_from_lora
 from utils.paths import is_model_cached
 from utils.utils import format_error_message
@@ -448,6 +448,64 @@ class InferenceBackend:
             return False
     pass
 
+    def _apply_adapter_state(self, use_adapter: Optional[Union[bool, str]]) -> None:
+        """
+        Apply adapter state before generation. Must be called under _generation_lock.
+
+        Args:
+            use_adapter: None = no change, False = disable (base model),
+                         True = enable current adapter, str = enable specific adapter.
+        """
+        if use_adapter is None:
+            return
+
+        base = self.active_model_name
+        if not base or base not in self.models:
+            return
+
+        model_info = self.models[base]
+
+        if use_adapter is False:
+            # Disable all adapters → pure base model generation
+            logger.info(f"Compare mode: disabling adapters on '{base}' (base model generation)")
+            self.disable_adapters(base)
+
+        elif use_adapter is True:
+            # Enable the most recently loaded adapter
+            loaded = model_info.get("loaded_adapters", {})
+            if loaded:
+                adapter_name = list(loaded.keys())[-1]
+                logger.info(f"Compare mode: enabling adapter '{adapter_name}' on '{base}'")
+                self.set_active_adapter(base, adapter_name)
+            else:
+                logger.warning("use_adapter=true but no adapters are loaded on the model")
+
+        elif isinstance(use_adapter, str):
+            # Enable a specific named adapter
+            logger.info(f"Compare mode: enabling specific adapter '{use_adapter}' on '{base}'")
+            self.set_active_adapter(base, use_adapter)
+
+    def generate_with_adapter_control(
+        self,
+        use_adapter: Optional[Union[bool, str]] = None,
+        **gen_kwargs,
+    ) -> Generator[str, None, None]:
+        """
+        Thread-safe generation with optional adapter toggling.
+
+        Acquires the generation lock, applies adapter state, then generates.
+        This ensures adapter toggle + generation are atomic — critical for
+        compare mode where base and LoRA panes fire concurrently.
+
+        Args:
+            use_adapter: Adapter control (None/False/True/str). See _apply_adapter_state.
+            **gen_kwargs: Forwarded to generate_chat_response.
+        """
+        with self._generation_lock:
+            self._apply_adapter_state(use_adapter)
+            # Delegate to the lock-free generation path
+            yield from self._generate_chat_response_inner(**gen_kwargs)
+
     def generate_chat_response(self,
                           messages: list,
                           system_prompt: str,
@@ -459,11 +517,33 @@ class InferenceBackend:
                           repetition_penalty: float = 1.1) -> Generator[str, None, None]:
         """
         Generate response for text or vision models.
+        Acquires the generation lock. For adapter-controlled generation,
+        use generate_with_adapter_control() instead.
+        """
+        with self._generation_lock:
+            yield from self._generate_chat_response_inner(
+                messages=messages,
+                system_prompt=system_prompt,
+                image=image,
+                temperature=temperature,
+                top_p=top_p,
+                top_k=top_k,
+                max_new_tokens=max_new_tokens,
+                repetition_penalty=repetition_penalty,
+            )
 
-        1. Messages are already in ChatML format (role/content)
-        2. Apply get_chat_template() if model in mapper
-        3. Apply tokenizer.apply_chat_template()
-        4. Generate
+    def _generate_chat_response_inner(self,
+                          messages: list,
+                          system_prompt: str = "",
+                          image=None,
+                          temperature: float = 0.7,
+                          top_p: float = 0.9,
+                          top_k: int = 40,
+                          max_new_tokens: int = 256,
+                          repetition_penalty: float = 1.1) -> Generator[str, None, None]:
+        """
+        Inner generation logic (no lock). Called by both generate_chat_response
+        and generate_with_adapter_control.
         """
         if not self.active_model_name:
             yield "Error: No active model"
@@ -473,55 +553,54 @@ class InferenceBackend:
         is_vision = model_info.get("is_vision", False)
         tokenizer = model_info.get("tokenizer") or model_info.get("processor")
 
-        with self._generation_lock:
-            if is_vision:
-                # Vision model generation
-                yield from self._generate_vision_response(
-                    messages, system_prompt, image,
-                    temperature, top_p, top_k, max_new_tokens, repetition_penalty
-                )
-            else:
-                # Text model: Use training pipeline approach
-                # Messages are already in ChatML format from eval.py
+        if is_vision:
+            # Vision model generation
+            yield from self._generate_vision_response(
+                messages, system_prompt, image,
+                temperature, top_p, top_k, max_new_tokens, repetition_penalty
+            )
+        else:
+            # Text model: Use training pipeline approach
+            # Messages are already in ChatML format from eval.py
 
-                # Step 1: Apply get_chat_template if model is in mapper
-                try:
-                    from utils.datasets import MODEL_TO_TEMPLATE_MAPPER, get_tokenizer_chat_template
+            # Step 1: Apply get_chat_template if model is in mapper
+            try:
+                from utils.datasets import MODEL_TO_TEMPLATE_MAPPER, get_tokenizer_chat_template
 
-                    model_name_lower = self.active_model_name.lower()
+                model_name_lower = self.active_model_name.lower()
 
-                    # Check if model has a registered template
-                    if model_name_lower in MODEL_TO_TEMPLATE_MAPPER:
-                        template_name = MODEL_TO_TEMPLATE_MAPPER[model_name_lower]
-                        logger.info(f"Applying chat template '{template_name}' for {self.active_model_name}")
+                # Check if model has a registered template
+                if model_name_lower in MODEL_TO_TEMPLATE_MAPPER:
+                    template_name = MODEL_TO_TEMPLATE_MAPPER[model_name_lower]
+                    logger.info(f"Applying chat template '{template_name}' for {self.active_model_name}")
 
-                        # This modifies the tokenizer with the correct template
-                        tokenizer = get_chat_template(
-                            tokenizer,
-                            self.active_model_name
-                        )
-                    else:
-                        logger.info(f"No registered template for {self.active_model_name}, using tokenizer default")
-                except Exception as e:
-                    logger.warning(f"Could not apply get_chat_template: {e}")
-
-                # Step 2: Format with tokenizer.apply_chat_template()
-                try:
-                    formatted_prompt = tokenizer.apply_chat_template(
-                        messages,
-                        tokenize=False,
-                        add_generation_prompt=True
+                    # This modifies the tokenizer with the correct template
+                    tokenizer = get_chat_template(
+                        tokenizer,
+                        self.active_model_name
                     )
-                    logger.debug(f"Formatted prompt: {formatted_prompt[:200]}...")
-                except Exception as e:
-                    logger.error(f"Error applying chat template: {e}")
-                    # Fallback to manual formatting
-                    formatted_prompt = self.format_chat_prompt(messages, system_prompt)
+                else:
+                    logger.info(f"No registered template for {self.active_model_name}, using tokenizer default")
+            except Exception as e:
+                logger.warning(f"Could not apply get_chat_template: {e}")
 
-                # Step 3: Generate
-                yield from self.generate_stream(
-                    formatted_prompt, temperature, top_p, top_k, max_new_tokens, repetition_penalty
+            # Step 2: Format with tokenizer.apply_chat_template()
+            try:
+                formatted_prompt = tokenizer.apply_chat_template(
+                    messages,
+                    tokenize=False,
+                    add_generation_prompt=True
                 )
+                logger.debug(f"Formatted prompt: {formatted_prompt[:200]}...")
+            except Exception as e:
+                logger.error(f"Error applying chat template: {e}")
+                # Fallback to manual formatting
+                formatted_prompt = self.format_chat_prompt(messages, system_prompt)
+
+            # Step 3: Generate
+            yield from self.generate_stream(
+                formatted_prompt, temperature, top_p, top_k, max_new_tokens, repetition_penalty
+            )
 
     def _generate_vision_response(self, messages, system_prompt, image,
                                   temperature, top_p, top_k, max_new_tokens,
diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index 64791b06f9..ada8bdd539 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -130,6 +130,16 @@ class ChatCompletionRequest(BaseModel):
     top_k: int = Field(40, ge=1, le=100, description="[x-unsloth] Top-k sampling")
     repetition_penalty: float = Field(1.1, ge=1.0, le=2.0, description="[x-unsloth] Repetition penalty")
     image_base64: Optional[str] = Field(None, description="[x-unsloth] Base64-encoded image for vision models")
+    use_adapter: Optional[Union[bool, str]] = Field(
+        None,
+        description=(
+            "[x-unsloth] Adapter control for compare mode. "
+            "null = no change (default), "
+            "false = disable adapters (base model), "
+            "true = enable the current adapter, "
+            "string = enable a specific adapter by name."
+        ),
+    )
 
 
 # ── Streaming response chunks ────────────────────────────────────
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 74cf37138f..9e57946565 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -365,6 +365,18 @@ async def openai_chat_completions(request: ChatCompletionRequest):
         repetition_penalty=request.repetition_penalty,
     )
 
+    # ── Choose generation path (adapter-controlled or standard) ──
+    if request.use_adapter is not None:
+        # Compare mode: toggle adapter state atomically with generation
+        def generate():
+            return backend.generate_with_adapter_control(
+                use_adapter=request.use_adapter, **gen_kwargs
+            )
+    else:
+        # Standard path: no adapter toggling
+        def generate():
+            return backend.generate_chat_response(**gen_kwargs)
+
     model_name = backend.active_model_name or request.model
     completion_id = f"chatcmpl-{uuid.uuid4().hex[:12]}"
     created = int(time.time())
@@ -388,7 +400,7 @@ async def openai_chat_completions(request: ChatCompletionRequest):
                 # Content chunks — generate_chat_response yields cumulative
                 # text, so we diff to get incremental deltas.
                 prev_text = ""
-                for cumulative in backend.generate_chat_response(**gen_kwargs):
+                for cumulative in generate():
                     new_text = cumulative[len(prev_text):]
                     prev_text = cumulative
                     if not new_text:
@@ -439,7 +451,7 @@ async def openai_chat_completions(request: ChatCompletionRequest):
     else:
         try:
             full_text = ""
-            for token in backend.generate_chat_response(**gen_kwargs):
+            for token in generate():
                 full_text = token  # generate_stream yields cumulative text
 
             response = ChatCompletion(

From 0b305fd822854a6850848f42f73fd8c8456030f6 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 16:57:24 +0000
Subject: [PATCH 04/29] _apply_adapter_state now calls revert_to_base_model and
 activate_lora_adapter properly

---
 studio/backend/core/inference/inference.py | 36 ++++++++++++++--------
 1 file changed, 23 insertions(+), 13 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index f8c0213b9b..a4ab80ffe2 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -452,6 +452,10 @@ class InferenceBackend:
         """
         Apply adapter state before generation. Must be called under _generation_lock.
 
+        Uses revert_to_base_model() / activate_lora_adapter() which work correctly
+        for models loaded by Unsloth as complete PeftModels (via model.unload() /
+        model.load_adapter()), matching the proven pattern from the Gradio eval page.
+
         Args:
             use_adapter: None = no change, False = disable (base model),
                          True = enable current adapter, str = enable specific adapter.
@@ -466,24 +470,30 @@ class InferenceBackend:
         model_info = self.models[base]
 
         if use_adapter is False:
-            # Disable all adapters → pure base model generation
-            logger.info(f"Compare mode: disabling adapters on '{base}' (base model generation)")
-            self.disable_adapters(base)
+            # Revert to pure base model by unloading adapter weights
+            logger.info(f"Compare mode: reverting '{base}' to base model for generation")
+            self.revert_to_base_model(base)
 
         elif use_adapter is True:
-            # Enable the most recently loaded adapter
-            loaded = model_info.get("loaded_adapters", {})
-            if loaded:
-                adapter_name = list(loaded.keys())[-1]
-                logger.info(f"Compare mode: enabling adapter '{adapter_name}' on '{base}'")
-                self.set_active_adapter(base, adapter_name)
+            # Activate the LoRA adapter from the original model path
+            lora_path = model_info.get("model_path")
+            if lora_path and model_info.get("is_lora"):
+                logger.info(f"Compare mode: activating LoRA adapter from '{lora_path}' on '{base}'")
+                self.activate_lora_adapter(base, lora_path)
             else:
-                logger.warning("use_adapter=true but no adapters are loaded on the model")
+                # Fallback for dynamically attached adapters
+                loaded = model_info.get("loaded_adapters", {})
+                if loaded:
+                    adapter_name = list(loaded.keys())[-1]
+                    logger.info(f"Compare mode: enabling adapter '{adapter_name}' on '{base}'")
+                    self.set_active_adapter(base, adapter_name)
+                else:
+                    logger.warning("use_adapter=true but no adapter path/adapters on model")
 
         elif isinstance(use_adapter, str):
-            # Enable a specific named adapter
-            logger.info(f"Compare mode: enabling specific adapter '{use_adapter}' on '{base}'")
-            self.set_active_adapter(base, use_adapter)
+            # Activate a specific adapter by path
+            logger.info(f"Compare mode: activating specific adapter '{use_adapter}' on '{base}'")
+            self.activate_lora_adapter(base, use_adapter)
 
     def generate_with_adapter_control(
         self,

From 9bee0a3f63f378b384e996e762ccfbd8b3bfbfc8 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 17:03:52 +0000
Subject: [PATCH 05/29] exclude default from model.delete_adapter

---
 studio/backend/core/inference/inference.py | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index a4ab80ffe2..ed7c786913 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -193,6 +193,8 @@ class InferenceBackend:
                 logger.info("Found lingering adapter configurations. Deleting them now...")
                 # Create a static list of keys before iterating and deleting
                 for name in list(model.peft_config.keys()):
+                    if name == "default":
+                        continue
                     logger.info(f"Deleting adapter config: '{name}'")
                     model.delete_adapter(name)
 

From b930a17b1dd236ba7abc7cd5f96f475c035b69da Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 17:09:07 +0000
Subject: [PATCH 06/29] added logging

---
 studio/backend/core/inference/inference.py | 9 ++++++++-
 1 file changed, 8 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index ed7c786913..20484f7bb4 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -177,6 +177,9 @@ class InferenceBackend:
             return False
 
         model = self.models[base_model_name].get("model")
+        logger.info(f"[DEBUG] revert_to_base_model called. Model type BEFORE: {model.__class__.__name__}")
+        logger.info(f"[DEBUG] Is PeftModel? {isinstance(model, (PeftModel, PeftModelForCausalLM))}")
+        logger.info(f"[DEBUG] Has peft_config? {hasattr(model, 'peft_config')}, keys={list(getattr(model, 'peft_config', {}).keys())}")
 
         try:
             # Step 1: Unload the adapter weights. This returns the base model object.
@@ -186,10 +189,14 @@ class InferenceBackend:
                 unwrapped_base_model = model.unload()
                 self.models[base_model_name]["model"] = unwrapped_base_model
                 model = unwrapped_base_model # Continue with the unwrapped model
+                logger.info(f"[DEBUG] Model type AFTER unload: {model.__class__.__name__}")
+            else:
+                logger.info(f"[DEBUG] Model is NOT a PeftModel, skipping unload.")
 
             # Step 2: Delete any lingering adapter configurations from the object.
             # This is the crucial step you identified.
             if hasattr(model, 'peft_config') and model.peft_config:
+                logger.info(f"[DEBUG] Lingering peft_config keys: {list(model.peft_config.keys())}")
                 logger.info("Found lingering adapter configurations. Deleting them now...")
                 # Create a static list of keys before iterating and deleting
                 for name in list(model.peft_config.keys()):
@@ -198,7 +205,7 @@ class InferenceBackend:
                     logger.info(f"Deleting adapter config: '{name}'")
                     model.delete_adapter(name)
 
-            logger.info("Model has been successfully reverted to a clean base state.")
+            logger.info(f"[DEBUG] Model type FINAL: {model.__class__.__name__}. Reverted to clean base state.")
             return True
 
         except Exception as e:

From 754ccf1a677c295b7b48d6cba8df87db6c7f1127 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 17:21:26 +0000
Subject: [PATCH 07/29] swipped logger for print statements as logger isn't
 propagating

---
 studio/backend/core/inference/inference.py | 16 ++++++++--------
 1 file changed, 8 insertions(+), 8 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 20484f7bb4..ff44777dbd 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -177,26 +177,26 @@ class InferenceBackend:
             return False
 
         model = self.models[base_model_name].get("model")
-        logger.info(f"[DEBUG] revert_to_base_model called. Model type BEFORE: {model.__class__.__name__}")
-        logger.info(f"[DEBUG] Is PeftModel? {isinstance(model, (PeftModel, PeftModelForCausalLM))}")
-        logger.info(f"[DEBUG] Has peft_config? {hasattr(model, 'peft_config')}, keys={list(getattr(model, 'peft_config', {}).keys())}")
+        print(f"[DEBUG] revert_to_base_model called. Model type BEFORE: {model.__class__.__name__}")
+        print(f"[DEBUG] Is PeftModel? {isinstance(model, (PeftModel, PeftModelForCausalLM))}")
+        print(f"[DEBUG] Has peft_config? {hasattr(model, 'peft_config')}, keys={list(getattr(model, 'peft_config', {}).keys())}")
 
         try:
             # Step 1: Unload the adapter weights. This returns the base model object.
             # This step is only necessary if the model is currently a PeftModel instance.
             if isinstance(model, (PeftModel, PeftModelForCausalLM)):
-                logger.info("Model is a PeftModel. Unloading adapters...")
+                print("[DEBUG] Model IS a PeftModel. Calling model.unload()...")
                 unwrapped_base_model = model.unload()
                 self.models[base_model_name]["model"] = unwrapped_base_model
                 model = unwrapped_base_model # Continue with the unwrapped model
-                logger.info(f"[DEBUG] Model type AFTER unload: {model.__class__.__name__}")
+                print(f"[DEBUG] Model type AFTER unload: {model.__class__.__name__}")
             else:
-                logger.info(f"[DEBUG] Model is NOT a PeftModel, skipping unload.")
+                print(f"[DEBUG] Model is NOT a PeftModel, skipping unload.")
 
             # Step 2: Delete any lingering adapter configurations from the object.
             # This is the crucial step you identified.
             if hasattr(model, 'peft_config') and model.peft_config:
-                logger.info(f"[DEBUG] Lingering peft_config keys: {list(model.peft_config.keys())}")
+                print(f"[DEBUG] Lingering peft_config keys: {list(model.peft_config.keys())}")
                 logger.info("Found lingering adapter configurations. Deleting them now...")
                 # Create a static list of keys before iterating and deleting
                 for name in list(model.peft_config.keys()):
@@ -205,7 +205,7 @@ class InferenceBackend:
                     logger.info(f"Deleting adapter config: '{name}'")
                     model.delete_adapter(name)
 
-            logger.info(f"[DEBUG] Model type FINAL: {model.__class__.__name__}. Reverted to clean base state.")
+            print(f"[DEBUG] Model type FINAL: {model.__class__.__name__}. Reverted to clean base state.")
             return True
 
         except Exception as e:

From f122154cf36d4e0b64c4dabf57cba0366b7e9b04 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 17:25:37 +0000
Subject: [PATCH 08/29] added print statements for activate_lora_adapter

---
 studio/backend/core/inference/inference.py | 15 +++++++++++----
 1 file changed, 11 insertions(+), 4 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index ff44777dbd..a4580baa71 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -221,25 +221,32 @@ class InferenceBackend:
         """
         model = self.models[base_model_name].get("model")
         adapter_name_to_load = lora_path.split("/")[-1].replace(".", "_")
+        print(f"[DEBUG] activate_lora_adapter called. base_model_name='{base_model_name}', lora_path='{lora_path}'")
+        print(f"[DEBUG] adapter_name_to_load='{adapter_name_to_load}'")
+        print(f"[DEBUG] Model type BEFORE load_adapter: {model.__class__.__name__}")
+        print(f"[DEBUG] Has peft_config? {hasattr(model, 'peft_config')}, keys={list(getattr(model, 'peft_config', {}).keys())}")
 
         try:
             # At this point, the model should be clean thanks to revert_to_base_model.
             # We can now safely load and set the new adapter.
 
             # Step 3: Load the new adapter.
-            logger.info(f"Loading adapter '{adapter_name_to_load}' from '{lora_path}'")
+            print(f"[DEBUG] Calling model.load_adapter('{lora_path}', adapter_name='{adapter_name_to_load}')...")
             model.load_adapter(lora_path, adapter_name=adapter_name_to_load)
+            print(f"[DEBUG] Model type AFTER load_adapter: {model.__class__.__name__}")
+            print(f"[DEBUG] peft_config keys AFTER load: {list(getattr(model, 'peft_config', {}).keys())}")
 
             # Step 4: Set the new adapter as active.
-            logger.info(f"Setting '{adapter_name_to_load}' as the active adapter.")
+            print(f"[DEBUG] Calling model.set_adapter('{adapter_name_to_load}')...")
             model.set_adapter(adapter_name_to_load)
+            print(f"[DEBUG] activate_lora_adapter SUCCESS. Model type: {model.__class__.__name__}")
 
             return True, adapter_name_to_load
         except Exception as e:
             # This will catch the "already exists" error if revert_to_base_model failed.
-            logger.error(f"Failed to activate LoRA adapter '{adapter_name_to_load}': {e}")
+            print(f"[DEBUG] activate_lora_adapter FAILED: {e}")
             import traceback
-            logger.error(traceback.format_exc())
+            traceback.print_exc()
             return False, None
     pass
 

From 225b3f1750f6221eed1c452de61e1b07ef217d90 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 17:32:15 +0000
Subject: [PATCH 09/29] del model.peft_config instead of using
 model.delete_adapter

---
 studio/backend/core/inference/inference.py | 20 ++++++++------------
 1 file changed, 8 insertions(+), 12 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index a4580baa71..41a907820f 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -193,19 +193,15 @@ class InferenceBackend:
             else:
                 print(f"[DEBUG] Model is NOT a PeftModel, skipping unload.")
 
-            # Step 2: Delete any lingering adapter configurations from the object.
-            # This is the crucial step you identified.
-            if hasattr(model, 'peft_config') and model.peft_config:
-                print(f"[DEBUG] Lingering peft_config keys: {list(model.peft_config.keys())}")
-                logger.info("Found lingering adapter configurations. Deleting them now...")
-                # Create a static list of keys before iterating and deleting
-                for name in list(model.peft_config.keys()):
-                    if name == "default":
-                        continue
-                    logger.info(f"Deleting adapter config: '{name}'")
-                    model.delete_adapter(name)
+            # Step 2: Clear any lingering peft_config from the unwrapped model.
+            # After model.unload(), the base model may still carry a peft_config
+            # attribute (with 'default' key). Removing it entirely ensures
+            # load_adapter() won't warn about "multiple adapters".
+            if hasattr(model, 'peft_config'):
+                print(f"[DEBUG] Clearing lingering peft_config: {list(model.peft_config.keys())}")
+                del model.peft_config
 
-            print(f"[DEBUG] Model type FINAL: {model.__class__.__name__}. Reverted to clean base state.")
+            print(f"[DEBUG] Model type FINAL: {model.__class__.__name__}, has peft_config={hasattr(model, 'peft_config')}. Reverted to clean base state.")
             return True
 
         except Exception as e:

From 4d868e8d2b3d59173f384a06d6b7b1119f4efd01 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 19:18:49 +0000
Subject: [PATCH 10/29] replace model unloading and peft loading mechanism for
 compare feature

---
 studio/backend/core/inference/inference.py | 26 ++++++++--------------
 1 file changed, 9 insertions(+), 17 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 41a907820f..9abf8d270f 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -214,37 +214,29 @@ class InferenceBackend:
     def activate_lora_adapter(self, base_model_name: str, lora_path: str) -> Tuple[bool, Optional[str]]:
         """
         Activates a specific LoRA adapter on what is assumed to be a clean base model.
+        Uses PeftModel.from_pretrained() which correctly wraps the base model.
         """
         model = self.models[base_model_name].get("model")
         adapter_name_to_load = lora_path.split("/")[-1].replace(".", "_")
         print(f"[DEBUG] activate_lora_adapter called. base_model_name='{base_model_name}', lora_path='{lora_path}'")
         print(f"[DEBUG] adapter_name_to_load='{adapter_name_to_load}'")
-        print(f"[DEBUG] Model type BEFORE load_adapter: {model.__class__.__name__}")
-        print(f"[DEBUG] Has peft_config? {hasattr(model, 'peft_config')}, keys={list(getattr(model, 'peft_config', {}).keys())}")
+        print(f"[DEBUG] Model type BEFORE: {model.__class__.__name__}")
 
         try:
-            # At this point, the model should be clean thanks to revert_to_base_model.
-            # We can now safely load and set the new adapter.
-
-            # Step 3: Load the new adapter.
-            print(f"[DEBUG] Calling model.load_adapter('{lora_path}', adapter_name='{adapter_name_to_load}')...")
-            model.load_adapter(lora_path, adapter_name=adapter_name_to_load)
-            print(f"[DEBUG] Model type AFTER load_adapter: {model.__class__.__name__}")
-            print(f"[DEBUG] peft_config keys AFTER load: {list(getattr(model, 'peft_config', {}).keys())}")
-
-            # Step 4: Set the new adapter as active.
-            print(f"[DEBUG] Calling model.set_adapter('{adapter_name_to_load}')...")
-            model.set_adapter(adapter_name_to_load)
-            print(f"[DEBUG] activate_lora_adapter SUCCESS. Model type: {model.__class__.__name__}")
+            # Use PeftModel.from_pretrained to wrap the clean base model with the adapter.
+            # This is the correct approach after model.unload() + del peft_config.
+            print(f"[DEBUG] Calling PeftModel.from_pretrained(model, '{lora_path}', adapter_name='{adapter_name_to_load}')...")
+            model = PeftModel.from_pretrained(model, lora_path, adapter_name=adapter_name_to_load)
+            self.models[base_model_name]["model"] = model
+            print(f"[DEBUG] Model type AFTER: {model.__class__.__name__}")
+            print(f"[DEBUG] activate_lora_adapter SUCCESS.")
 
             return True, adapter_name_to_load
         except Exception as e:
-            # This will catch the "already exists" error if revert_to_base_model failed.
             print(f"[DEBUG] activate_lora_adapter FAILED: {e}")
             import traceback
             traceback.print_exc()
             return False, None
-    pass
 
     def load_adapter(self, base_model_name: str, adapter_path: str, adapter_name: str = None) -> bool:
         """

From 4399687f938f6c49c48e8454ea98f654b9f1d589 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 19:23:51 +0000
Subject: [PATCH 11/29] strip extra debug statements

---
 studio/backend/core/inference/inference.py | 33 +++++++---------------
 1 file changed, 10 insertions(+), 23 deletions(-)

diff --git a/studio/backend/core/inference/inference.py b/studio/backend/core/inference/inference.py
index 9abf8d270f..23203613f2 100644
--- a/studio/backend/core/inference/inference.py
+++ b/studio/backend/core/inference/inference.py
@@ -177,31 +177,23 @@ class InferenceBackend:
             return False
 
         model = self.models[base_model_name].get("model")
-        print(f"[DEBUG] revert_to_base_model called. Model type BEFORE: {model.__class__.__name__}")
-        print(f"[DEBUG] Is PeftModel? {isinstance(model, (PeftModel, PeftModelForCausalLM))}")
-        print(f"[DEBUG] Has peft_config? {hasattr(model, 'peft_config')}, keys={list(getattr(model, 'peft_config', {}).keys())}")
 
         try:
-            # Step 1: Unload the adapter weights. This returns the base model object.
-            # This step is only necessary if the model is currently a PeftModel instance.
+            # Step 1: Unload the adapter weights if model is a PeftModel.
             if isinstance(model, (PeftModel, PeftModelForCausalLM)):
-                print("[DEBUG] Model IS a PeftModel. Calling model.unload()...")
+                logger.info(f"Unloading LoRA adapters from '{base_model_name}'...")
                 unwrapped_base_model = model.unload()
                 self.models[base_model_name]["model"] = unwrapped_base_model
-                model = unwrapped_base_model # Continue with the unwrapped model
-                print(f"[DEBUG] Model type AFTER unload: {model.__class__.__name__}")
-            else:
-                print(f"[DEBUG] Model is NOT a PeftModel, skipping unload.")
+                model = unwrapped_base_model
 
             # Step 2: Clear any lingering peft_config from the unwrapped model.
             # After model.unload(), the base model may still carry a peft_config
-            # attribute (with 'default' key). Removing it entirely ensures
-            # load_adapter() won't warn about "multiple adapters".
+            # attribute. Removing it ensures PeftModel.from_pretrained() gets
+            # a clean base model without "multiple adapters" warnings.
             if hasattr(model, 'peft_config'):
-                print(f"[DEBUG] Clearing lingering peft_config: {list(model.peft_config.keys())}")
                 del model.peft_config
 
-            print(f"[DEBUG] Model type FINAL: {model.__class__.__name__}, has peft_config={hasattr(model, 'peft_config')}. Reverted to clean base state.")
+            logger.info(f"Model '{base_model_name}' reverted to clean base state.")
             return True
 
         except Exception as e:
@@ -209,7 +201,6 @@ class InferenceBackend:
             import traceback
             logger.error(traceback.format_exc())
             return False
-    pass
 
     def activate_lora_adapter(self, base_model_name: str, lora_path: str) -> Tuple[bool, Optional[str]]:
         """
@@ -218,24 +209,20 @@ class InferenceBackend:
         """
         model = self.models[base_model_name].get("model")
         adapter_name_to_load = lora_path.split("/")[-1].replace(".", "_")
-        print(f"[DEBUG] activate_lora_adapter called. base_model_name='{base_model_name}', lora_path='{lora_path}'")
-        print(f"[DEBUG] adapter_name_to_load='{adapter_name_to_load}'")
-        print(f"[DEBUG] Model type BEFORE: {model.__class__.__name__}")
 
         try:
             # Use PeftModel.from_pretrained to wrap the clean base model with the adapter.
             # This is the correct approach after model.unload() + del peft_config.
-            print(f"[DEBUG] Calling PeftModel.from_pretrained(model, '{lora_path}', adapter_name='{adapter_name_to_load}')...")
+            logger.info(f"Loading LoRA adapter '{adapter_name_to_load}' from '{lora_path}'...")
             model = PeftModel.from_pretrained(model, lora_path, adapter_name=adapter_name_to_load)
             self.models[base_model_name]["model"] = model
-            print(f"[DEBUG] Model type AFTER: {model.__class__.__name__}")
-            print(f"[DEBUG] activate_lora_adapter SUCCESS.")
+            logger.info(f"LoRA adapter '{adapter_name_to_load}' activated successfully.")
 
             return True, adapter_name_to_load
         except Exception as e:
-            print(f"[DEBUG] activate_lora_adapter FAILED: {e}")
+            logger.error(f"Failed to activate LoRA adapter '{adapter_name_to_load}': {e}")
             import traceback
-            traceback.print_exc()
+            logger.error(traceback.format_exc())
             return False, None
 
     def load_adapter(self, base_model_name: str, adapter_path: str, adapter_name: str = None) -> bool:

From ac8128519d9e6e1f154612eadc05bee4af6cd673 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sat, 14 Feb 2026 20:13:50 +0000
Subject: [PATCH 12/29] decouple reliance of backend on frontend for is_lora

---
 studio/backend/routes/inference.py | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index 9e57946565..b3ea1fa1be 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -69,10 +69,10 @@ async def load_model(request: LoadRequest):
         backend = get_inference_backend()
         
         # Create config using clean factory method
+        # is_lora is auto-detected from adapter_config.json on disk/HF
         config = ModelConfig.from_identifier(
             model_id=request.model_path,
             hf_token=request.hf_token,
-            is_lora=request.is_lora,
         )
         
         if not config:

From 354b7d0aca6f02b634b23c38c37a438ab77c87aa Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 00:00:22 +0000
Subject: [PATCH 13/29] feat: add cancel or save and stop training

---
 studio/backend/core/training/trainer.py  | 42 +++++++++++++++++-------
 studio/backend/core/training/training.py | 14 ++++++--
 studio/backend/routes/training.py        | 39 +++++++++++++++++++++-
 3 files changed, 79 insertions(+), 16 deletions(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 52cf60af10..013c7817b2 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -58,6 +58,7 @@ class UnslothTrainer:
         self.progress_callbacks = []
         self.is_training = False
         self.should_stop = False
+        self.save_on_stop = True
 
         # Model state tracking
         self.is_vlm = False
@@ -756,16 +757,32 @@ class UnslothTrainer:
             self.trainer.train()
 
             # ========== SAVE MODEL ==========
-            self.trainer.save_model()
-            self.tokenizer.save_pretrained(output_dir)
-            print(f"\nTraining completed! Model saved to {output_dir}\n")
-
-            self._update_progress(
-                is_training=False,
-                is_completed=True,
-                #status_message=status_msg
-                status_message=f"Training completed! Model saved to {output_dir}",
-            )
+            if self.should_stop and self.save_on_stop:
+                # Stopped by user — save model at current checkpoint
+                self.trainer.save_model()
+                self.tokenizer.save_pretrained(output_dir)
+                print(f"\nTraining stopped. Model saved to {output_dir}\n")
+                self._update_progress(
+                    is_training=False,
+                    status_message=f"Training stopped. Model saved to {output_dir}",
+                )
+            elif self.should_stop:
+                # Cancelled by user — don't save
+                print("\nTraining cancelled.\n")
+                self._update_progress(
+                    is_training=False,
+                    status_message="Training cancelled.",
+                )
+            else:
+                # Normal completion
+                self.trainer.save_model()
+                self.tokenizer.save_pretrained(output_dir)
+                print(f"\nTraining completed! Model saved to {output_dir}\n")
+                self._update_progress(
+                    is_training=False,
+                    is_completed=True,
+                    status_message=f"Training completed! Model saved to {output_dir}",
+                )
 
         except Exception as e:
             logger.error(f"Training error: {e}")
@@ -774,10 +791,11 @@ class UnslothTrainer:
         finally:
             self.is_training = False
 
-    def stop_training(self):
+    def stop_training(self, save: bool = True):
         """Stop ongoing training"""
-        print("\nStopping training...")
+        print(f"\nStopping training (save={save})...")
         self.should_stop = True
+        self.save_on_stop = save
         self.is_training = False
         # Clear the status message so timer doesn't show stale status
         self._update_progress(is_training=False, status_message="")
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index fa8b3b5daf..62febadc13 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -103,6 +103,7 @@ class TrainingBackend:
         try:
             # Reset stop flag and clear history
             self.trainer.should_stop = False
+            self.trainer.save_on_stop = True
             self.loss_history = []
             self.lr_history = []
             self.step_history = []
@@ -224,16 +225,19 @@ class TrainingBackend:
             )
             return False
 
-    def stop_training(self) -> bool:
+    def stop_training(self, save: bool = True) -> bool:
         """
         Stop ongoing training.
 
+        Args:
+            save: If True, save the model at the current checkpoint.
+
         Returns:
             True if training was successfully stopped.
         """
         try:
-            logger.info("Stopping training...")
-            self.trainer.stop_training()
+            logger.info(f"Stopping training (save={save})...")
+            self.trainer.stop_training(save=save)
             return True
         except Exception as e:
             logger.error(f"Error stopping training: {e}")
@@ -293,6 +297,10 @@ class TrainingBackend:
             True if training is in progress, False otherwise
         """
         try:
+            # If user requested stop, training is no longer considered active
+            if self.trainer.should_stop:
+                return False
+
             progress = self.trainer.get_training_progress()
             # Training is active if is_training is True
             # Also check if we're in loading/preparation phase (status_message indicates activity)
diff --git a/studio/backend/routes/training.py b/studio/backend/routes/training.py
index 20a5e1e254..1a904cd371 100644
--- a/studio/backend/routes/training.py
+++ b/studio/backend/routes/training.py
@@ -37,6 +37,11 @@ from models import (
     TrainingProgress,
 )
 from models.responses import TrainingStopResponse, TrainingMetricsResponse
+from pydantic import BaseModel as PydanticBaseModel
+
+
+class TrainingStopRequest(PydanticBaseModel):
+    save: bool = True
 
 router = APIRouter()
 logger = logging.getLogger(__name__)
@@ -251,10 +256,14 @@ async def start_training(
 
 @router.post("/stop", response_model=TrainingStopResponse)
 async def stop_training(
+    body: TrainingStopRequest = TrainingStopRequest(),
     current_subject: str = Depends(get_current_subject),
 ):
     """
     Stop the currently running training job.
+
+    Body:
+        save (bool): If True (default), save the model at the current checkpoint.
     """
     try:
         backend = get_training_backend()
@@ -266,7 +275,7 @@ async def stop_training(
             )
         
         # Call backend stop method
-        backend.stop_training()
+        backend.stop_training(save=body.save)
         
         return TrainingStopResponse(
             status="stopped",
@@ -281,6 +290,29 @@ async def stop_training(
         )
 
 
+@router.post("/reset")
+async def reset_training(
+    current_subject: str = Depends(get_current_subject),
+):
+    """
+    Reset training state so the user can return to configuration.
+    """
+    try:
+        backend = get_training_backend()
+        backend.trainer.should_stop = False
+        backend.trainer.training_progress = backend.trainer.training_progress.__class__()
+        backend.loss_history = []
+        backend.lr_history = []
+        backend.step_history = []
+        return {"status": "ok"}
+    except Exception as e:
+        logger.error(f"Error resetting training: {e}", exc_info=True)
+        raise HTTPException(
+            status_code=500,
+            detail=f"Failed to reset training: {str(e)}",
+        )
+
+
 @router.get("/status")
 async def get_training_status(
     current_subject: str = Depends(get_current_subject),
@@ -313,6 +345,9 @@ async def get_training_status(
         ) or "Ready to train"
         error_message = getattr(progress, "error", None) if progress else None
 
+        # Check if training was stopped by user
+        trainer_stopped = getattr(backend.trainer, "should_stop", False)
+
         # Derive high-level phase
         if error_message:
             phase = "error"
@@ -326,6 +361,8 @@ async def get_training_status(
                 phase = "configuring"
             else:
                 phase = "training"
+        elif trainer_stopped:
+            phase = "stopped"
         elif progress and getattr(progress, "is_completed", False):
             phase = "completed"
         elif has_thread:

From 55e7bd60c158362ac114a7794d856a636200249a Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 00:22:28 +0000
Subject: [PATCH 14/29] feat: UI for cancel or save and stop training

---
 .../studio/sections/progress-section.tsx      | 53 +++++++++++++++----
 .../src/features/studio/studio-page.tsx       | 20 +++++++
 .../src/features/training/api/train-api.ts    | 15 +++++-
 .../training/hooks/use-training-actions.ts    | 16 ++++--
 .../hooks/use-training-runtime-lifecycle.ts   |  6 ++-
 .../src/features/training/lib/sync-runtime.ts |  5 ++
 .../training/stores/training-runtime-store.ts |  6 ++-
 .../src/features/training/types/runtime.ts    |  1 +
 8 files changed, 104 insertions(+), 18 deletions(-)

diff --git a/studio/frontend/src/features/studio/sections/progress-section.tsx b/studio/frontend/src/features/studio/sections/progress-section.tsx
index b8656a7867..f1a10c2684 100644
--- a/studio/frontend/src/features/studio/sections/progress-section.tsx
+++ b/studio/frontend/src/features/studio/sections/progress-section.tsx
@@ -1,4 +1,14 @@
 import { SectionCard } from "@/components/section-card";
+import {
+  AlertDialog,
+  AlertDialogAction,
+  AlertDialogCancel,
+  AlertDialogContent,
+  AlertDialogDescription,
+  AlertDialogFooter,
+  AlertDialogHeader,
+  AlertDialogTitle,
+} from "@/components/ui/alert-dialog";
 import { Button } from "@/components/ui/button";
 import {
   Popover,
@@ -104,6 +114,7 @@ export function ProgressSection(): ReactElement {
   );
 
   const { stopTrainingRun } = useTrainingActions();
+  const [stopDialogOpen, setStopDialogOpen] = useState(false);
   const localStartAtRef = useRef<number | null>(null);
   const [, setLocalTick] = useState(0);
 
@@ -225,15 +236,39 @@ export function ProgressSection(): ReactElement {
               </div>
             </PopoverContent>
           </Popover>
-          <Button
-            variant="destructive"
-            size="sm"
-            className="h-7 cursor-pointer px-3 text-xs"
-            onClick={() => void stopTrainingRun()}
-            disabled={!runtime.isTrainingRunning}
-          >
-            <HugeiconsIcon icon={StopIcon} className="size-3" /> Stop
-          </Button>
+          <AlertDialog open={stopDialogOpen} onOpenChange={setStopDialogOpen}>
+            <Button
+              variant="destructive"
+              size="sm"
+              className="h-7 cursor-pointer px-3 text-xs"
+              onClick={() => setStopDialogOpen(true)}
+              disabled={!runtime.isTrainingRunning}
+            >
+              <HugeiconsIcon icon={StopIcon} className="size-3" /> Stop
+            </Button>
+            <AlertDialogContent>
+              <AlertDialogHeader>
+                <AlertDialogTitle>Stop Training</AlertDialogTitle>
+                <AlertDialogDescription>
+                  Choose how you want to stop the current training run.
+                </AlertDialogDescription>
+              </AlertDialogHeader>
+              <AlertDialogFooter>
+                <AlertDialogCancel>Continue Training</AlertDialogCancel>
+                <AlertDialogAction
+                  variant="destructive"
+                  onClick={() => void stopTrainingRun(false)}
+                >
+                  Cancel Training
+                </AlertDialogAction>
+                <AlertDialogAction
+                  onClick={() => void stopTrainingRun(true)}
+                >
+                  Stop and Save
+                </AlertDialogAction>
+              </AlertDialogFooter>
+            </AlertDialogContent>
+          </AlertDialog>
         </div>
       }
     >
diff --git a/studio/frontend/src/features/studio/studio-page.tsx b/studio/frontend/src/features/studio/studio-page.tsx
index d5e20fc194..5811f161db 100644
--- a/studio/frontend/src/features/studio/studio-page.tsx
+++ b/studio/frontend/src/features/studio/studio-page.tsx
@@ -1,8 +1,12 @@
+import { Button } from "@/components/ui/button";
 import {
   shouldShowTrainingView,
+  useTrainingActions,
   useTrainingRuntimeLifecycle,
   useTrainingRuntimeStore,
 } from "@/features/training";
+import { ArrowLeft01Icon } from "@hugeicons/core-free-icons";
+import { HugeiconsIcon } from "@hugeicons/react";
 import type { ReactElement } from "react";
 import { DatasetSection } from "./sections/dataset-section";
 import { ModelSection } from "./sections/model-section";
@@ -14,12 +18,28 @@ export function StudioPage(): ReactElement {
   useTrainingRuntimeLifecycle();
   const showTrainingView = useTrainingRuntimeStore(shouldShowTrainingView);
   const runtimeMessage = useTrainingRuntimeStore((state) => state.message);
+  const runtimePhase = useTrainingRuntimeStore((state) => state.phase);
   const isHydratingRuntime = useTrainingRuntimeStore((state) => state.isHydrating);
   const hasHydratedRuntime = useTrainingRuntimeStore((state) => state.hasHydrated);
+  const { dismissTrainingRun } = useTrainingActions();
+
+  const canGoBack = runtimePhase === "stopped" || runtimePhase === "error";
 
   return (
     <div className="min-h-screen bg-background">
       <main className="mx-auto max-w-7xl px-6 py-4">
+        {canGoBack && (
+          <Button
+            variant="ghost"
+            size="sm"
+            className="mb-2 cursor-pointer gap-1.5 text-muted-foreground"
+            onClick={() => void dismissTrainingRun()}
+          >
+            <HugeiconsIcon icon={ArrowLeft01Icon} className="size-4" />
+            Back to configuration
+          </Button>
+        )}
+
         {/* Header */}
         <div className="mb-8 flex flex-col gap-0.5">
           <h1 className="text-2xl font-semibold tracking-tight">
diff --git a/studio/frontend/src/features/training/api/train-api.ts b/studio/frontend/src/features/training/api/train-api.ts
index e3c53b0bc2..d2f589c298 100644
--- a/studio/frontend/src/features/training/api/train-api.ts
+++ b/studio/frontend/src/features/training/api/train-api.ts
@@ -41,11 +41,22 @@ export async function startTraining(
   return parseJson<TrainingStartResponse>(response);
 }
 
-export async function stopTraining(): Promise<TrainingStopResponse> {
-  const response = await authFetch("/api/train/stop", { method: "POST" });
+export async function stopTraining(save = true): Promise<TrainingStopResponse> {
+  const response = await authFetch("/api/train/stop", {
+    method: "POST",
+    headers: { "Content-Type": "application/json" },
+    body: JSON.stringify({ save }),
+  });
   return parseJson<TrainingStopResponse>(response);
 }
 
+export async function resetTraining(): Promise<void> {
+  const response = await authFetch("/api/train/reset", { method: "POST" });
+  if (!response.ok) {
+    throw new Error(await readError(response));
+  }
+}
+
 export async function getTrainingStatus(): Promise<TrainingStatusResponse> {
   const response = await authFetch("/api/train/status");
   return parseJson<TrainingStatusResponse>(response);
diff --git a/studio/frontend/src/features/training/hooks/use-training-actions.ts b/studio/frontend/src/features/training/hooks/use-training-actions.ts
index 510c45ed4a..4ace9dd5ed 100644
--- a/studio/frontend/src/features/training/hooks/use-training-actions.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-actions.ts
@@ -1,7 +1,7 @@
 import { useCallback } from "react";
 import { useTrainingConfigStore } from "../stores/training-config-store";
 import { useTrainingRuntimeStore } from "../stores/training-runtime-store";
-import { startTraining, stopTraining } from "../api/train-api";
+import { startTraining, stopTraining, resetTraining } from "../api/train-api";
 import { buildTrainingStartPayload } from "../api/mappers";
 import { syncTrainingRuntimeFromBackend } from "../lib/sync-runtime";
 import { validateTrainingConfig } from "../lib/validation";
@@ -45,12 +45,12 @@ export function useTrainingActions() {
     }
   }, []);
 
-  const stopTrainingRun = useCallback(async (): Promise<boolean> => {
+  const stopTrainingRun = useCallback(async (save = true): Promise<boolean> => {
     const runtimeStore = useTrainingRuntimeStore.getState();
     runtimeStore.setStartError(null);
 
     try {
-      await stopTraining();
+      await stopTraining(save);
       await syncTrainingRuntimeFromBackend();
       return true;
     } catch (error) {
@@ -61,10 +61,20 @@ export function useTrainingActions() {
     }
   }, []);
 
+  const dismissTrainingRun = useCallback(async (): Promise<void> => {
+    useTrainingRuntimeStore.getState().resetRuntime();
+    try {
+      await resetTraining();
+    } catch {
+      // Frontend already reset; backend will catch up on next poll
+    }
+  }, []);
+
   return {
     isStarting,
     startError,
     startTrainingRun,
     stopTrainingRun,
+    dismissTrainingRun,
   };
 }
diff --git a/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts b/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
index 3baff3a62d..8d0eb9c756 100644
--- a/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
@@ -48,9 +48,10 @@ export function useTrainingRuntimeLifecycle(): void {
     };
 
     const pollMetrics = async () => {
+      const gen = runtimeStore.getState().resetGeneration;
       try {
         const metrics = await getTrainingMetrics();
-        if (disposed) {
+        if (disposed || runtimeStore.getState().resetGeneration !== gen) {
           return;
         }
         runtimeStore.getState().applyMetrics(metrics);
@@ -62,9 +63,10 @@ export function useTrainingRuntimeLifecycle(): void {
     };
 
     const pollStatus = async () => {
+      const gen = runtimeStore.getState().resetGeneration;
       try {
         const status = await getTrainingStatus();
-        if (disposed) {
+        if (disposed || runtimeStore.getState().resetGeneration !== gen) {
           return;
         }
 
diff --git a/studio/frontend/src/features/training/lib/sync-runtime.ts b/studio/frontend/src/features/training/lib/sync-runtime.ts
index b5fbd0bafb..bf255f58c6 100644
--- a/studio/frontend/src/features/training/lib/sync-runtime.ts
+++ b/studio/frontend/src/features/training/lib/sync-runtime.ts
@@ -6,12 +6,17 @@ import { useTrainingRuntimeStore } from "../stores/training-runtime-store";
 import type { TrainingStatusResponse } from "../types/runtime";
 
 export async function syncTrainingRuntimeFromBackend(): Promise<TrainingStatusResponse> {
+  const gen = useTrainingRuntimeStore.getState().resetGeneration;
+
   const [status, metrics] = await Promise.all([
     getTrainingStatus(),
     getTrainingMetrics(),
   ]);
 
   const runtimeStore = useTrainingRuntimeStore.getState();
+  if (runtimeStore.resetGeneration !== gen) {
+    return status;
+  }
   runtimeStore.applyStatus(status);
   runtimeStore.applyMetrics(metrics);
 
diff --git a/studio/frontend/src/features/training/stores/training-runtime-store.ts b/studio/frontend/src/features/training/stores/training-runtime-store.ts
index bc40019346..94db0f28f5 100644
--- a/studio/frontend/src/features/training/stores/training-runtime-store.ts
+++ b/studio/frontend/src/features/training/stores/training-runtime-store.ts
@@ -34,6 +34,7 @@ const initialState: TrainingRuntimeState = {
   lossHistory: [],
   lrHistory: [],
   gradNormHistory: [],
+  resetGeneration: 0,
 };
 
 function sortSeries(points: TrainingSeriesPoint[]): TrainingSeriesPoint[] {
@@ -95,12 +96,13 @@ export const useTrainingRuntimeStore = create<TrainingRuntimeStore>()((set) => (
   setLastEventId: (value) => set({ lastEventId: value }),
 
   resetRuntime: () =>
-    set({
+    set((state) => ({
       ...initialState,
       lossHistory: [],
       lrHistory: [],
       gradNormHistory: [],
-    }),
+      resetGeneration: state.resetGeneration + 1,
+    })),
 
   setStartQueued: (jobId, message) =>
     set({
diff --git a/studio/frontend/src/features/training/types/runtime.ts b/studio/frontend/src/features/training/types/runtime.ts
index fe2afbd36d..389418680a 100644
--- a/studio/frontend/src/features/training/types/runtime.ts
+++ b/studio/frontend/src/features/training/types/runtime.ts
@@ -82,6 +82,7 @@ export interface TrainingRuntimeState {
   lossHistory: TrainingSeriesPoint[];
   lrHistory: TrainingSeriesPoint[];
   gradNormHistory: TrainingSeriesPoint[];
+  resetGeneration: number;
 }
 
 export interface TrainingRuntimeActions {

From df5f45058f52879b2c6ff08870113be02cb93676 Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 05:38:06 +0000
Subject: [PATCH 15/29] Fixing stuck training processes

---
 studio/backend/core/training/trainer.py  |  5 ++++-
 studio/backend/core/training/training.py | 26 ++++++++++++++++++++++++
 2 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 013c7817b2..5a3d8683e2 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -2,13 +2,16 @@
 Unsloth Training Backend
 Integrates Unsloth training capabilities with the FastAPI backend
 """
+import os
+# Prevent tokenizer parallelism deadlocks when datasets uses multiprocessing fork
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
 import torch
 from utils.hardware import clear_gpu_cache
 torch._dynamo.config.recompile_limit = 64
 from unsloth import FastLanguageModel, FastVisionModel, is_bfloat16_supported
 from unsloth.chat_templates import get_chat_template
 
-import os
 import json
 import threading
 import math
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 62febadc13..62aa021136 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -6,6 +6,7 @@ from typing import Any, Generator, Tuple
 import logging
 
 from .trainer import get_trainer, TrainingProgress
+from utils.hardware import clear_gpu_cache
 
 logger = logging.getLogger(__name__)
 
@@ -101,6 +102,31 @@ class TrainingBackend:
             True if training started successfully, False otherwise.
         """
         try:
+            # Wait for any previous training thread to finish
+            old_thread = getattr(self.trainer, "training_thread", None)
+            if old_thread and old_thread.is_alive():
+                logger.info("Waiting for previous training thread to finish...")
+                old_thread.join(timeout=30)
+
+            # Explicitly free old SFTTrainer and CUDA resources before loading new model.
+            # Without this, forked multiprocessing workers (num_proc tokenization) inherit
+            # stale CUDA state from the previous run, causing extreme slowdowns or crashes.
+            if self.trainer.trainer is not None:
+                logger.info("Cleaning up previous SFTTrainer...")
+                self.trainer.trainer = None
+            if self.trainer.model is not None:
+                self.trainer.model = None
+            if self.trainer.tokenizer is not None:
+                self.trainer.tokenizer = None
+            # Flush all pending async CUDA ops so forked tokenization processes
+            # don't inherit stale async state that causes pool join to hang.
+            import torch as _torch
+            if _torch.cuda.is_available():
+                _torch.cuda.synchronize()
+            import gc
+            gc.collect()
+            clear_gpu_cache()
+
             # Reset stop flag and clear history
             self.trainer.should_stop = False
             self.trainer.save_on_stop = True

From e7c289967e77313ff99bffd10af01bd9932b40af Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 05:38:06 +0000
Subject: [PATCH 16/29] Fixing stuck training processes

---
 studio/backend/core/training/trainer.py  |  5 ++++-
 studio/backend/core/training/training.py | 26 ++++++++++++++++++++++++
 2 files changed, 30 insertions(+), 1 deletion(-)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 013c7817b2..5a3d8683e2 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -2,13 +2,16 @@
 Unsloth Training Backend
 Integrates Unsloth training capabilities with the FastAPI backend
 """
+import os
+# Prevent tokenizer parallelism deadlocks when datasets uses multiprocessing fork
+os.environ["TOKENIZERS_PARALLELISM"] = "false"
+
 import torch
 from utils.hardware import clear_gpu_cache
 torch._dynamo.config.recompile_limit = 64
 from unsloth import FastLanguageModel, FastVisionModel, is_bfloat16_supported
 from unsloth.chat_templates import get_chat_template
 
-import os
 import json
 import threading
 import math
diff --git a/studio/backend/core/training/training.py b/studio/backend/core/training/training.py
index 62febadc13..62aa021136 100644
--- a/studio/backend/core/training/training.py
+++ b/studio/backend/core/training/training.py
@@ -6,6 +6,7 @@ from typing import Any, Generator, Tuple
 import logging
 
 from .trainer import get_trainer, TrainingProgress
+from utils.hardware import clear_gpu_cache
 
 logger = logging.getLogger(__name__)
 
@@ -101,6 +102,31 @@ class TrainingBackend:
             True if training started successfully, False otherwise.
         """
         try:
+            # Wait for any previous training thread to finish
+            old_thread = getattr(self.trainer, "training_thread", None)
+            if old_thread and old_thread.is_alive():
+                logger.info("Waiting for previous training thread to finish...")
+                old_thread.join(timeout=30)
+
+            # Explicitly free old SFTTrainer and CUDA resources before loading new model.
+            # Without this, forked multiprocessing workers (num_proc tokenization) inherit
+            # stale CUDA state from the previous run, causing extreme slowdowns or crashes.
+            if self.trainer.trainer is not None:
+                logger.info("Cleaning up previous SFTTrainer...")
+                self.trainer.trainer = None
+            if self.trainer.model is not None:
+                self.trainer.model = None
+            if self.trainer.tokenizer is not None:
+                self.trainer.tokenizer = None
+            # Flush all pending async CUDA ops so forked tokenization processes
+            # don't inherit stale async state that causes pool join to hang.
+            import torch as _torch
+            if _torch.cuda.is_available():
+                _torch.cuda.synchronize()
+            import gc
+            gc.collect()
+            clear_gpu_cache()
+
             # Reset stop flag and clear history
             self.trainer.should_stop = False
             self.trainer.save_on_stop = True

From 9195311b1c3bf471167bd9c466fe6d666456f62e Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 15 Feb 2026 06:06:27 +0000
Subject: [PATCH 17/29] fix: add no-cache headers to index.html to prevent
 stale frontend after rebuild

---
 studio/backend/main.py |  4 +-
 test_lora.py           | 83 ++++++++++++++++++++++++++++++++++++++++++
 2 files changed, 85 insertions(+), 2 deletions(-)
 create mode 100644 test_lora.py

diff --git a/studio/backend/main.py b/studio/backend/main.py
index e957c668e3..ce67200456 100644
--- a/studio/backend/main.py
+++ b/studio/backend/main.py
@@ -128,7 +128,7 @@ def setup_frontend(app: FastAPI, build_path: Path):
 
         @app.get("/")
         async def serve_root():
-            return FileResponse(build_path / "index.html")
+            return FileResponse(build_path / "index.html", headers={"Cache-Control": "no-cache, no-store, must-revalidate"})
 
         @app.get("/{full_path:path}")
         async def serve_frontend(full_path: str):
@@ -139,7 +139,7 @@ def setup_frontend(app: FastAPI, build_path: Path):
             if file_path.is_file():
                 return FileResponse(file_path)
 
-            return FileResponse(build_path / "index.html")
+            return FileResponse(build_path / "index.html", headers={"Cache-Control": "no-cache, no-store, must-revalidate"})
 
         return True
     return False
diff --git a/test_lora.py b/test_lora.py
new file mode 100644
index 0000000000..8d48f20bba
--- /dev/null
+++ b/test_lora.py
@@ -0,0 +1,83 @@
+from unsloth import FastLanguageModel
+from peft import PeftModel, PeftModelForCausalLM
+
+lora_path="/home/support/new-ui-prototype/outputs/meta-llama_Llama-3.1-8B-Instruct_1771048481"
+adapter_name_to_load="test"
+
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name="/home/support/new-ui-prototype/outputs/meta-llama_Llama-3.1-8B-Instruct_1771048481",
+    max_seq_length=2048,
+    load_in_4bit=True,
+)
+
+# Quick sanity check
+FastLanguageModel.for_inference(model)
+print("Model loaded successfully!")
+print(f"Model type before unloading: {type(model)}")
+print(f"Tokenizer: {type(tokenizer)}")
+print(f"Model class before unloading: {model.__class__.__name__}")
+# Test generation
+#inputs = tokenizer("What is the capital of France?", return_tensors="pt").to(model.device)
+#outputs = model.generate(**inputs, max_new_tokens=32)
+#print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+
+
+print("unloading base model")
+if isinstance(model, (PeftModel, PeftModelForCausalLM)):
+        print("Model is a PeftModel. Unloading adapters...")
+        unwrapped_base_model = model.unload()
+        model = unwrapped_base_model
+        #if hasattr(model, 'peft_config') and model.peft_config:
+        #        print("Found lingering adapter configurations. Deleting them now...")
+        #        # Create a static list of keys before iterating and deleting
+        #        for name in list(model.peft_config.keys()):
+        #            if name == "default":
+        #                continue
+        #            print(f"Deleting adapter config: '{name}'")
+        #            mode=model.delete_adapter(name)
+        #model.disable_adapters()
+        if hasattr(model, 'peft_config'):
+            del model.peft_config
+print(f"Model type post unloading{type(model)}")
+print(f"Model class post unloading: {model.__class__.__name__}")
+#print(f"model: {model}")
+
+#print("generating using base model")
+#inputs = tokenizer("What is the capital of Lebanon?", return_tensors="pt").to(model.device)
+#outputs = model.generate(**inputs, max_new_tokens=32)
+#print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+#print(f"model config: {model.config}")
+#print(f"model: {model}")
+#print(f"model.peft_config: {model.peft_config}")
+
+#print("loading lora dapter")
+#model.load_adapter(lora_path, adapter_name=adapter_name_to_load)
+#model.enable_adapters
+#model.set_adapter(adapter_name_to_load)
+model = PeftModel.from_pretrained(model, lora_path, adapter_name=adapter_name_to_load)
+print(f"Model type {type(model)}")
+print(f"Model class: {model.__class__.__name__}")
+#print(f"model: {model}")
+#print("generating using peft model hopefully")
+#inputs = tokenizer("What is the capital of Brasil?", return_tensors="pt").to(model.device)
+#outputs = model.generate(**inputs, max_new_tokens=32)
+#print(tokenizer.decode(outputs[0], skip_special_tokens=True))
+
+print("unloading base model")
+if isinstance(model, (PeftModel, PeftModelForCausalLM)):
+        print("Model is a PeftModel. Unloading adapters...")
+        unwrapped_base_model = model.unload()
+        model = unwrapped_base_model
+        #if hasattr(model, 'peft_config') and model.peft_config:
+        #        print("Found lingering adapter configurations. Deleting them now...")
+        #        # Create a static list of keys before iterating and deleting
+        #        for name in list(model.peft_config.keys()):
+        #            if name == "default":
+        #                continue
+        #            print(f"Deleting adapter config: '{name}'")
+        #            mode=model.delete_adapter(name)
+        #model.disable_adapters()
+        if hasattr(model, 'peft_config'):
+            del model.peft_config
+print(f"Model type post unloading{type(model)}")
+print(f"Model class post unloading: {model.__class__.__name__}")

From 44f04e067dd205bb090717a15a3defca7a6154f5 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 15 Feb 2026 06:11:06 +0000
Subject: [PATCH 18/29] Remove test_lora.py from tracking

---
 test_lora.py | 83 ----------------------------------------------------
 1 file changed, 83 deletions(-)
 delete mode 100644 test_lora.py

diff --git a/test_lora.py b/test_lora.py
deleted file mode 100644
index 8d48f20bba..0000000000
--- a/test_lora.py
+++ /dev/null
@@ -1,83 +0,0 @@
-from unsloth import FastLanguageModel
-from peft import PeftModel, PeftModelForCausalLM
-
-lora_path="/home/support/new-ui-prototype/outputs/meta-llama_Llama-3.1-8B-Instruct_1771048481"
-adapter_name_to_load="test"
-
-model, tokenizer = FastLanguageModel.from_pretrained(
-    model_name="/home/support/new-ui-prototype/outputs/meta-llama_Llama-3.1-8B-Instruct_1771048481",
-    max_seq_length=2048,
-    load_in_4bit=True,
-)
-
-# Quick sanity check
-FastLanguageModel.for_inference(model)
-print("Model loaded successfully!")
-print(f"Model type before unloading: {type(model)}")
-print(f"Tokenizer: {type(tokenizer)}")
-print(f"Model class before unloading: {model.__class__.__name__}")
-# Test generation
-#inputs = tokenizer("What is the capital of France?", return_tensors="pt").to(model.device)
-#outputs = model.generate(**inputs, max_new_tokens=32)
-#print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-
-
-print("unloading base model")
-if isinstance(model, (PeftModel, PeftModelForCausalLM)):
-        print("Model is a PeftModel. Unloading adapters...")
-        unwrapped_base_model = model.unload()
-        model = unwrapped_base_model
-        #if hasattr(model, 'peft_config') and model.peft_config:
-        #        print("Found lingering adapter configurations. Deleting them now...")
-        #        # Create a static list of keys before iterating and deleting
-        #        for name in list(model.peft_config.keys()):
-        #            if name == "default":
-        #                continue
-        #            print(f"Deleting adapter config: '{name}'")
-        #            mode=model.delete_adapter(name)
-        #model.disable_adapters()
-        if hasattr(model, 'peft_config'):
-            del model.peft_config
-print(f"Model type post unloading{type(model)}")
-print(f"Model class post unloading: {model.__class__.__name__}")
-#print(f"model: {model}")
-
-#print("generating using base model")
-#inputs = tokenizer("What is the capital of Lebanon?", return_tensors="pt").to(model.device)
-#outputs = model.generate(**inputs, max_new_tokens=32)
-#print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-#print(f"model config: {model.config}")
-#print(f"model: {model}")
-#print(f"model.peft_config: {model.peft_config}")
-
-#print("loading lora dapter")
-#model.load_adapter(lora_path, adapter_name=adapter_name_to_load)
-#model.enable_adapters
-#model.set_adapter(adapter_name_to_load)
-model = PeftModel.from_pretrained(model, lora_path, adapter_name=adapter_name_to_load)
-print(f"Model type {type(model)}")
-print(f"Model class: {model.__class__.__name__}")
-#print(f"model: {model}")
-#print("generating using peft model hopefully")
-#inputs = tokenizer("What is the capital of Brasil?", return_tensors="pt").to(model.device)
-#outputs = model.generate(**inputs, max_new_tokens=32)
-#print(tokenizer.decode(outputs[0], skip_special_tokens=True))
-
-print("unloading base model")
-if isinstance(model, (PeftModel, PeftModelForCausalLM)):
-        print("Model is a PeftModel. Unloading adapters...")
-        unwrapped_base_model = model.unload()
-        model = unwrapped_base_model
-        #if hasattr(model, 'peft_config') and model.peft_config:
-        #        print("Found lingering adapter configurations. Deleting them now...")
-        #        # Create a static list of keys before iterating and deleting
-        #        for name in list(model.peft_config.keys()):
-        #            if name == "default":
-        #                continue
-        #            print(f"Deleting adapter config: '{name}'")
-        #            mode=model.delete_adapter(name)
-        #model.disable_adapters()
-        if hasattr(model, 'peft_config'):
-            del model.peft_config
-print(f"Model type post unloading{type(model)}")
-print(f"Model class post unloading: {model.__class__.__name__}")

From 92a1d4cffaeafd2f5862ba01fdfa9ad9772de7a5 Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 06:12:55 +0000
Subject: [PATCH 19/29] Adding hint for password length

---
 studio/frontend/src/features/auth/components/auth-form.tsx | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)

diff --git a/studio/frontend/src/features/auth/components/auth-form.tsx b/studio/frontend/src/features/auth/components/auth-form.tsx
index e8d531c0b4..f17d2af13b 100644
--- a/studio/frontend/src/features/auth/components/auth-form.tsx
+++ b/studio/frontend/src/features/auth/components/auth-form.tsx
@@ -199,6 +199,9 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
               )}
             </Button>
           </div>
+          {!isLoginMode && (
+            <p className="text-xs text-muted-foreground">Must be at least 8 characters</p>
+          )}
         </div>
 
         {!isLoginMode && (
@@ -223,7 +226,7 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
         <Button
           type="submit"
           className="w-full"
-          disabled={loading || statusLoading || blockedByState}
+          disabled={loading || statusLoading || blockedByState || (!isLoginMode && password.length < 8)}
         >
           {loading ? "Please wait..." : submitLabel}
         </Button>

From d752c31baf6be30b18b2037d3dea708bcedbf672 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 15 Feb 2026 06:29:52 +0000
Subject: [PATCH 20/29] feat: redirect first-time users to signup page instead
 of login

---
 studio/frontend/src/app/auth-guards.ts             | 14 +++++++++++++-
 .../src/features/auth/components/auth-form.tsx     | 13 ++++++++++++-
 2 files changed, 25 insertions(+), 2 deletions(-)

diff --git a/studio/frontend/src/app/auth-guards.ts b/studio/frontend/src/app/auth-guards.ts
index 83e4e39a32..6aaa6694ef 100644
--- a/studio/frontend/src/app/auth-guards.ts
+++ b/studio/frontend/src/app/auth-guards.ts
@@ -12,9 +12,21 @@ async function hasActiveSession(): Promise<boolean> {
   return refreshSession();
 }
 
+async function checkAuthInitialized(): Promise<boolean> {
+  try {
+    const res = await fetch("/api/auth/status");
+    if (!res.ok) return true; // fallback to login on error
+    const data = (await res.json()) as { initialized: boolean };
+    return data.initialized;
+  } catch {
+    return true; // fallback to login on error
+  }
+}
+
 export async function requireAuth(): Promise<void> {
   if (await hasActiveSession()) return;
-  throw redirect({ to: "/login" });
+  const initialized = await checkAuthInitialized();
+  throw redirect({ to: initialized ? "/login" : "/signup" });
 }
 
 export async function requireGuest(): Promise<void> {
diff --git a/studio/frontend/src/features/auth/components/auth-form.tsx b/studio/frontend/src/features/auth/components/auth-form.tsx
index f17d2af13b..5ed529a9e8 100644
--- a/studio/frontend/src/features/auth/components/auth-form.tsx
+++ b/studio/frontend/src/features/auth/components/auth-form.tsx
@@ -63,7 +63,18 @@ export function AuthForm({ mode }: AuthFormProps): ReactElement | null {
         const response = await fetch("/api/auth/status");
         if (!response.ok) throw new Error("Failed to load auth status.");
         const result = (await response.json()) as AuthStatusResponse;
-        if (!canceled) setInitialized(result.initialized);
+        if (!canceled) {
+          setInitialized(result.initialized);
+          // Auto-redirect to the correct page based on init state
+          if (mode === "login" && result.initialized === false) {
+            navigate({ to: "/signup" });
+            return;
+          }
+          if (mode === "signup" && result.initialized === true) {
+            navigate({ to: "/login" });
+            return;
+          }
+        }
       } catch (err: unknown) {
         if (!canceled) {
           setError(err instanceof Error ? err.message : "Failed to load.");

From 94dbd9cc9e428d60b3559bdc5d62adb3382eb5a8 Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 06:34:39 +0000
Subject: [PATCH 21/29] Changing labels in dataset card

---
 .../frontend/src/features/studio/sections/dataset-section.tsx | 4 ++--
 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/studio/frontend/src/features/studio/sections/dataset-section.tsx b/studio/frontend/src/features/studio/sections/dataset-section.tsx
index 890faad2b3..0d325fbefd 100644
--- a/studio/frontend/src/features/studio/sections/dataset-section.tsx
+++ b/studio/frontend/src/features/studio/sections/dataset-section.tsx
@@ -223,7 +223,7 @@ export function DatasetSection() {
 
         <div className="flex flex-col gap-2">
           <span className="flex items-center gap-1.5 text-xs font-medium text-muted-foreground">
-            Dataset Format
+            Target Format
             <Tooltip>
               <TooltipTrigger asChild={true}>
                 <button
@@ -258,7 +258,7 @@ export function DatasetSection() {
               <SelectValue />
             </SelectTrigger>
             <SelectContent>
-              <SelectItem value="auto">Auto-detect</SelectItem>
+              <SelectItem value="auto">Auto</SelectItem>
               <SelectItem value="alpaca">Alpaca</SelectItem>
               <SelectItem value="chatml">ChatML</SelectItem>
               <SelectItem value="sharegpt">ShareGPT</SelectItem>

From a035e5af109e2ed612ca17efe58920ca760aa866 Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 15 Feb 2026 08:27:35 +0000
Subject: [PATCH 22/29] fix: auto-redirect to signup/login when auth tokens are
 stale

---
 studio/frontend/src/features/auth/api.ts      | 26 ++++++++++++++++++-
 .../hooks/use-training-runtime-lifecycle.ts   |  7 +++--
 2 files changed, 30 insertions(+), 3 deletions(-)

diff --git a/studio/frontend/src/features/auth/api.ts b/studio/frontend/src/features/auth/api.ts
index 6705caa5e4..f42a544757 100644
--- a/studio/frontend/src/features/auth/api.ts
+++ b/studio/frontend/src/features/auth/api.ts
@@ -10,6 +10,26 @@ type RefreshResponse = {
   refresh_token: string;
 };
 
+let isRedirecting = false;
+
+async function redirectToAuth(): Promise<void> {
+  if (isRedirecting) return;
+  isRedirecting = true;
+
+  let target = "/login";
+  try {
+    const res = await fetch("/api/auth/status");
+    if (res.ok) {
+      const data = (await res.json()) as { initialized: boolean };
+      if (!data.initialized) target = "/signup";
+    }
+  } catch {
+    // Fall through to /login on error
+  }
+
+  window.location.href = target;
+}
+
 export async function refreshSession(): Promise<boolean> {
   const refreshToken = getRefreshToken();
   if (!refreshToken) return false;
@@ -48,7 +68,11 @@ export async function authFetch(
   if (response.status !== 401) return response;
 
   const refreshed = await refreshSession();
-  if (!refreshed) return response;
+  if (!refreshed) {
+    clearAuthTokens();
+    void redirectToAuth();
+    return response;
+  }
 
   const retryHeaders = new Headers(init?.headers);
   const newToken = getAuthToken();
diff --git a/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts b/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
index 8d0eb9c756..4bf329a33e 100644
--- a/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
+++ b/studio/frontend/src/features/training/hooks/use-training-runtime-lifecycle.ts
@@ -1,3 +1,4 @@
+import { hasAuthToken } from "@/features/auth";
 import { useEffect } from "react";
 import {
   getTrainingMetrics,
@@ -48,6 +49,7 @@ export function useTrainingRuntimeLifecycle(): void {
     };
 
     const pollMetrics = async () => {
+      if (!hasAuthToken()) return;
       const gen = runtimeStore.getState().resetGeneration;
       try {
         const metrics = await getTrainingMetrics();
@@ -56,13 +58,14 @@ export function useTrainingRuntimeLifecycle(): void {
         }
         runtimeStore.getState().applyMetrics(metrics);
       } catch (error) {
-        if (!isAbortError(error) && !disposed) {
+        if (!isAbortError(error) && !disposed && hasAuthToken()) {
           runtimeStore.getState().setSseConnected(false);
         }
       }
     };
 
     const pollStatus = async () => {
+      if (!hasAuthToken()) return;
       const gen = runtimeStore.getState().resetGeneration;
       try {
         const status = await getTrainingStatus();
@@ -79,7 +82,7 @@ export function useTrainingRuntimeLifecycle(): void {
           stopStream();
         }
       } catch (error) {
-        if (!isAbortError(error) && !disposed) {
+        if (!isAbortError(error) && !disposed && hasAuthToken()) {
           runtimeStore.getState().setSseConnected(false);
         }
       }

From 069953f2c1938364d82cad5ab31fbf4f4a500aec Mon Sep 17 00:00:00 2001
From: Roland Tannous <rolandtannous@gravityq.ai>
Date: Sun, 15 Feb 2026 08:49:56 +0000
Subject: [PATCH 23/29] README.md draft

---
 README.md | 145 +++++++++++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 144 insertions(+), 1 deletion(-)

diff --git a/README.md b/README.md
index 5270b9aaf6..4d2092b3a9 100644
--- a/README.md
+++ b/README.md
@@ -1 +1,144 @@
-# new-ui-prototype
+<p align="center">
+  <img src="https://raw.githubusercontent.com/unslothai/unsloth/main/images/unsloth%20logo%20white%20text.png" alt="Unsloth Studio" width="400"/>
+</p>
+
+<h3 align="center">🦥 Unsloth Studio</h3>
+
+<p align="center">
+  A modern, full-stack web interface for fine-tuning, managing, and chatting with large language models — locally or in the cloud.
+</p>
+
+<p align="center">
+  <a href="#features">Features</a> •
+  <a href="#quick-start">Quick Start</a> •
+  <a href="#api-reference">API Reference</a> •
+  <a href="#project-structure">Project Structure</a>
+</p>
+
+---
+
+## Features
+
+| Area | Capabilities |
+|---|---|
+| **Training** | Configure and launch LoRA / QLoRA fine-tuning jobs with real-time SSE progress streaming, live loss charts, and one-click stop / resume |
+| **Model Management** | Browse, load, and manage Hugging Face hub models and local checkpoints |
+| **Inference** | Interactive chat playground for testing fine-tuned models |
+| **Dataset Tools** | Upload, preview, and prepare datasets (JSON, CSV, Parquet, PDF, DOCX) |
+| **Export** | Export & push trained adapters to the Hugging Face Hub |
+| **Auth** | Token-based authentication with JWT access / refresh flow and first-time setup token |
+
+## Quick Start
+
+### One-command setup
+
+```bash
+bash setup.sh
+```
+
+This script will:
+1. Install **Node.js ≥ 20** via nvm (if needed)
+2. Build the frontend to `studio/frontend/dist`
+3. Create a Python virtual environment and install all dependencies (including `unsloth`)
+4. Register a convenient `unsloth-ui` shell alias
+
+### Launch the studio
+
+```bash
+# After setup, open a new terminal (or source ~/.bashrc), then:
+unsloth-ui -H 0.0.0.0 -p 8000
+```
+
+On **first launch**, a one-time setup token is printed to the console. Use it in the browser to create your admin account.
+
+## API Reference
+
+All endpoints require a valid JWT `Authorization: Bearer <token>` header (except `/api/auth/*` and `/api/health`).
+
+| Method | Endpoint | Description |
+|---|---|---|
+| `GET` | `/api/health` | Health check |
+| `GET` | `/api/system` | System info (GPU, CPU, memory) |
+| `POST` | `/api/auth/signup` | Create account (requires setup token on first run) |
+| `POST` | `/api/auth/login` | Login and receive JWT tokens |
+| `POST` | `/api/auth/refresh` | Refresh an expired access token |
+| `GET` | `/api/auth/status` | Check if auth is initialized |
+| `POST` | `/api/train/start` | Start a training job |
+| `POST` | `/api/train/stop` | Stop a running training job |
+| `POST` | `/api/train/reset` | Reset training state |
+| `GET` | `/api/train/status` | Get current training status |
+| `GET` | `/api/train/metrics` | Get training metrics (loss, LR, steps) |
+| `GET` | `/api/train/stream` | SSE stream of real-time training progress |
+| `GET` | `/api/models/` | List available models |
+| `POST` | `/api/inference/chat` | Send a chat message for inference |
+| `GET` | `/api/datasets/` | List / manage datasets |
+
+> Full interactive docs are available at `/docs` (Swagger UI) and `/redoc` when the server is running.
+
+## CLI Commands
+
+The Unsloth CLI (`cli.py`) provides the following commands:
+
+```
+Usage: cli.py [COMMAND]
+
+Commands:
+  train             Fine-tune a model
+  inference         Run inference on a trained model
+  export            Export a trained adapter
+  list-checkpoints  List saved checkpoints
+  ui                Launch the Unsloth Studio web UI
+  studio            Launch the studio (alias)
+```
+
+## Project Structure
+
+```
+new-ui-prototype/
+├── cli.py                     # CLI entry point
+├── cli/                       # Typer CLI commands
+│   └── commands/
+│       ├── train.py
+│       ├── inference.py
+│       ├── export.py
+│       ├── ui.py
+│       └── studio.py
+├── setup.sh                   # One-command bootstrap script
+└── studio/
+    ├── backend/
+    │   ├── main.py            # FastAPI app & middleware
+    │   ├── run.py             # Server launcher (uvicorn)
+    │   ├── auth/              # Auth storage & JWT logic
+    │   ├── routes/            # API route handlers
+    │   │   ├── training.py
+    │   │   ├── models.py
+    │   │   ├── inference.py
+    │   │   ├── datasets.py
+    │   │   └── auth.py
+    │   ├── models/            # Pydantic request/response schemas
+    │   ├── core/              # Training engine & config
+    │   ├── utils/             # Hardware detection, helpers
+    │   └── requirements.txt
+    ├── frontend/
+    │   ├── src/
+    │   │   ├── features/      # Feature modules
+    │   │   │   ├── auth/      # Login / signup flow
+    │   │   │   ├── training/  # Training config & monitoring
+    │   │   │   ├── studio/    # Main studio workspace
+    │   │   │   ├── chat/      # Inference chat UI
+    │   │   │   ├── export/    # Model export flow
+    │   │   │   └── onboarding/# Onboarding wizard
+    │   │   ├── components/    # Shared UI components (shadcn)
+    │   │   ├── hooks/         # Custom React hooks
+    │   │   ├── stores/        # Zustand state stores
+    │   │   └── types/         # TypeScript type definitions
+    │   ├── package.json
+    │   └── vite.config.ts
+    └── tests/                 # Backend test suite
+```
+
+## License
+
+This project is licensed under the [GNU Affero General Public License v3.0 (AGPL-3.0)](https://www.gnu.org/licenses/agpl-3.0.html).
+
+Copyright © 2026 Unsloth AI.

From 9b5e02029e518eecfac92459ba57e01e2487032d Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Fri, 13 Feb 2026 19:26:37 -0700
Subject: [PATCH 24/29] added inference defaults from unsloth guides

---
 .../gemma/unsloth_functiongemma-270m-it.yaml               | 7 +++++++
 .../model_defaults/gemma/unsloth_gemma-3-270m-it.yaml      | 7 +++++++
 .../model_defaults/gemma/unsloth_gemma-3-27b-it.yaml       | 7 +++++++
 .../model_defaults/gemma/unsloth_gemma-3-4b-it.yaml        | 7 +++++++
 .../model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml        | 7 +++++++
 .../model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml      | 7 +++++++
 .../configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml | 7 +++++++
 .../model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml       | 6 ++++++
 .../model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml        | 6 ++++++
 .../granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml | 6 ++++++
 .../granite/unsloth_granite-4.0-h-micro.yaml               | 6 ++++++
 .../unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml     | 6 ++++++
 .../mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml      | 5 +++++
 .../other/unsloth_Nemotron-3-Nano-30B-A3B.yaml             | 5 +++++
 .../assets/configs/model_defaults/phi/unsloth_Phi-4.yaml   | 5 +++++
 .../qwen/unsloth_Qwen3-4B-Instruct-2507.yaml               | 7 +++++++
 .../qwen/unsloth_Qwen3-4B-Thinking-2507.yaml               | 7 +++++++
 .../unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml     | 6 ++++++
 18 files changed, 114 insertions(+)

diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml
index 283ab50754..d67877d167 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_functiongemma-270m-it.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/functiongemma-270m-it
 # Based on FunctionGemma_(270M).ipynb
 # Also applies to: unsloth/functiongemma-270m-it-unsloth-bnb-4bit, google/functiongemma-270m-it, unsloth/functiongemma-270m-it-unsloth-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 4096
@@ -45,3 +46,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
index eab0257381..d4108fefb0 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-270m-it.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gemma-3-270m-it
 # Based on Gemma3_(270M).ipynb
 # Also applies to: unsloth/gemma-3-270m-it-unsloth-bnb-4bit, google/gemma-3-270m-it, unsloth/gemma-3-270m-it-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
index c8fc2a8269..0f213edb70 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-27b-it.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gemma-3-27b-it
 # Based on Gemma3_(27B)_A100-Conversational.ipynb
 # Also applies to: unsloth/gemma-3-27b-it-unsloth-bnb-4bit, google/gemma-3-27b-it, unsloth/gemma-3-27b-it-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
index a9535537b4..d7da7a2873 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-it.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gemma-3-4b-it
 # Based on Gemma3_(4B).ipynb
 # Also applies to: unsloth/gemma-3-4b-it-unsloth-bnb-4bit, google/gemma-3-4b-it, unsloth/gemma-3-4b-it-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
index 80b85bb836..b0833e6e30 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3-4b-pt.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gemma-3-4b-pt
 # Based on Gemma3_(4B)-Vision.ipynb
 # Also applies to: unsloth/gemma-3-4b-pt-unsloth-bnb-4bit, google/gemma-3-4b-pt, unsloth/gemma-3-4b-pt-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
index 1aeb4f59f8..7fe1fed620 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B-it.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gemma-3n-E4B-it
 # Based on Gemma3N_(4B)-Conversational.ipynb
 # Also applies to: unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit, google/gemma-3n-E4B-it, unsloth/gemma-3n-E4B-it-unsloth-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 1024
@@ -39,3 +40,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
index 47cbc32782..3b1fab2a60 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_gemma-3n-E4B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gemma-3n-E4B
 # Based on Gemma3N_(4B)-Vision.ipynb
 # Also applies to: unsloth/gemma-3n-E4B-unsloth-bnb-4bit, google/gemma-3n-E4B
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_k: 64
+  top_p: 0.95
+  min_p: 0.0
+
diff --git a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
index fc7dedc5fd..93a8d8b274 100644
--- a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-120b.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gpt-oss-120b
 # Based on gpt-oss-(120B)_A100-Fine-tuning.ipynb
 # Also applies to: openai/gpt-oss-120b, unsloth/gpt-oss-120b-unsloth-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 4096
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_p: 1.0
+  top_k: 0
+
diff --git a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml
index d98dd28700..18f8f7bf47 100644
--- a/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml
+++ b/studio/backend/assets/configs/model_defaults/gpt-oss/unsloth_gpt-oss-20b.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/gpt-oss-20b
 # Based on gpt-oss-(20B)-Fine-tuning.ipynb
 # Also applies to: openai/gpt-oss-20b, unsloth/gpt-oss-20b-unsloth-bnb-4bit, unsloth/gpt-oss-20b-BF16
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 1024
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_p: 1.0
+  top_k: 0
+
diff --git a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml
index 718ec156f8..9b91ec531a 100644
--- a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-350m-unsloth-bnb-4bit.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/granite-4.0-350m
 # Based on Granite4.0_350M.ipynb
 # Also applies to: ibm-granite/granite-4.0-350m, unsloth/granite-4.0-350m-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -47,3 +48,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.0
+  top_p: 1.0
+  top_k: 0
+
diff --git a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml
index 630366faee..e6603b0cea 100644
--- a/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml
+++ b/studio/backend/assets/configs/model_defaults/granite/unsloth_granite-4.0-h-micro.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/granite-4.0-h-micro
 # Based on Granite4.0.ipynb
 # Also applies to: ibm-granite/granite-4.0-h-micro, unsloth/granite-4.0-h-micro-bnb-4bit, unsloth/granite-4.0-h-micro-unsloth-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -47,3 +48,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.0
+  top_p: 1.0
+  top_k: 0
+
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
index bb605451a5..9af6fbd3a0 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Magistral-Small-2509-unsloth-bnb-4bit.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Magistral-Small-2509
 # Based on Magistral_(24B)-Reasoning-Conversational.ipynb
 # Also applies to: mistralai/Magistral-Small-2509, unsloth/Magistral-Small-2509-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.7
+  min_p: 0.01
+  top_p: 0.95
+
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
index b5fd0f5713..d701390da9 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Ministral-3-3B-Instruct-2512.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Ministral-3-3B-Instruct-2512
 # Based on Ministral_3_VL_(3B)_Vision.ipynb
 # Also applies to: unsloth/Ministral-3-3B-Instruct-2512
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.15
+  top_p: default
+
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml
index f40a2fd620..0b1b60feac 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_Nemotron-3-Nano-30B-A3B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Nemotron-3-Nano-30B-A3B
 # Based on Nemotron-3-Nano-30B-A3B_A100.ipynb
 # Also applies to: unsloth/Nemotron-3-Nano-30B-A3B
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -47,3 +48,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.0
+  top_p: 1.0
+
diff --git a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
index 9746e01374..b13fa9a0b7 100644
--- a/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
+++ b/studio/backend/assets/configs/model_defaults/phi/unsloth_Phi-4.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Phi-4
 # Based on Phi_4-Conversational.ipynb
 # Also applies to: unsloth/phi-4-unsloth-bnb-4bit, microsoft/phi-4, unsloth/phi-4-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.8
+  top_p: 0.95
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
index 1a1efaf975..7178b86e7a 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Instruct-2507.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-4B-Instruct-2507
 # Based on Qwen3_(4B)-Instruct.ipynb
 # Also applies to: unsloth/Qwen3-4B-Instruct-2507-unsloth-bnb-4bit, Qwen/Qwen3-4B-Instruct-2507, unsloth/Qwen3-4B-Instruct-2507-bnb-4bit, Qwen/Qwen3-4B-Instruct-2507-FP8, unsloth/Qwen3-4B-Instruct-2507-FP8
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.7
+  top_p: 0.80
+  top_k: 20
+  min_p: 0.00
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml
index 5a67224009..0385f1d00d 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-4B-Thinking-2507.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-4B-Thinking-2507
 # Based on Qwen3_(4B)-Thinking.ipynb
 # Also applies to: unsloth/Qwen3-4B-Thinking-2507-unsloth-bnb-4bit, Qwen/Qwen3-4B-Thinking-2507, unsloth/Qwen3-4B-Thinking-2507-bnb-4bit, Qwen/Qwen3-4B-Thinking-2507-FP8, unsloth/Qwen3-4B-Thinking-2507-FP8
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.6
+  top_p: 0.95
+  top_k: 20
+  min_p: 0.00
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
index 56b2feb653..8db73bbdc2 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-VL-8B-Instruct-unsloth-bnb-4bit.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-VL-8B-Instruct
 # Based on Qwen3_VL_(8B)-Vision.ipynb
 # Also applies to: Qwen/Qwen3-VL-8B-Instruct-FP8, unsloth/Qwen3-VL-8B-Instruct-FP8, unsloth/Qwen3-VL-8B-Instruct, Qwen/Qwen3-VL-8B-Instruct, unsloth/Qwen3-VL-8B-Instruct-bnb-4bit
+# added inference parameters from unsloth guides
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.7
+  top_p: 0.8
+  top_k: 20
+

From 81be5dc0805a2e9ad58f4cc0e0b1eccc1c3e6b2a Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Sat, 14 Feb 2026 00:24:12 -0700
Subject: [PATCH 25/29] added configs from Ollama

---
 .../model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml | 5 +++++
 .../configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml     | 6 ++++++
 .../qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml       | 6 ++++++
 .../configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml      | 6 ++++++
 .../configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml      | 6 ++++++
 5 files changed, 29 insertions(+)

diff --git a/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
index 02d0428f94..c6bef5077e 100644
--- a/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/gemma/unsloth_codegemma-7b-bnb-4bit.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/codegemma-7b-bnb-4bit
 # Based on CodeGemma_(7B)-Conversational.ipynb
 # Also applies to: unsloth/codegemma-7b, google/codegemma-7b
+# added inference parameters from Ollama
 
 training:
   max_seq_length: 4096
@@ -44,3 +45,7 @@ logging:
   enable_tensorboard: false
   tensorboard_dir: "runs"
   log_frequency: 10
+
+inference:
+  temperature: 0
+  top_p: 0.9
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
index 153e0d13e1..bf19fad0bd 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-0.6B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-0.6B
 # Based on Qwen3_(0_6B)-Phone_Deployment.ipynb
 # Also applies to: unsloth/Qwen3-0.6B-unsloth-bnb-4bit, Qwen/Qwen3-0.6B, unsloth/Qwen3-0.6B-bnb-4bit, Qwen/Qwen3-0.6B-FP8, unsloth/Qwen3-0.6B-FP8
+# added inference parameters from Ollama
 
 training:
   max_seq_length: 1024
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.6
+  top_k: 20
+  top_p: 0.95
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
index 76dd3de761..fe519f49a7 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B-Base-unsloth-bnb-4bit.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-14B-Base
 # Based on Qwen3_(14B)-Alpaca.ipynb
 # Also applies to: unsloth/Qwen3-14B-Base, Qwen/Qwen3-14B-Base, unsloth/Qwen3-14B-Base-bnb-4bit
+# added inference parameters from Ollama
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.6
+  top_k: 20
+  top_p: 0.95
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
index 51bcfc5501..662b96c5fe 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-14B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-14B
 # Based on Qwen3_(14B).ipynb
 # Also applies to: unsloth/Qwen3-14B-unsloth-bnb-4bit, Qwen/Qwen3-14B, unsloth/Qwen3-14B-bnb-4bit, Qwen/Qwen3-14B-FP8, unsloth/Qwen3-14B-FP8
+# added inference parameters from Ollama
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.6
+  top_k: 20
+  top_p: 0.95
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
index 76098e80f5..83ac11efda 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen3-32B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen3-32B
 # Based on Qwen3_(32B)_A100-Reasoning-Conversational.ipynb
 # Also applies to: unsloth/Qwen3-32B-unsloth-bnb-4bit, Qwen/Qwen3-32B, unsloth/Qwen3-32B-bnb-4bit, Qwen/Qwen3-32B-FP8, unsloth/Qwen3-32B-FP8
+# added inference parameters from Ollama
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.6
+  top_k: 20
+  top_p: 0.95
+

From a105e60d30e57f67820fa24c7042619c869a8e32 Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Sat, 14 Feb 2026 01:39:37 -0700
Subject: [PATCH 26/29] added default inference config from unsloth notebooks

---
 .../ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml             | 5 +++++
 .../llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml       | 5 +++++
 .../llama/unsloth_Llama-3.2-3B-Instruct.yaml               | 5 +++++
 .../llama/unsloth_Llama-3.3-70B-Instruct.yaml              | 5 +++++
 .../configs/model_defaults/llasa/unsloth_Llasa-3B.yaml     | 5 +++++
 .../model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml   | 5 +++++
 .../model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml  | 7 +++++++
 .../configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml   | 6 ++++++
 .../configs/model_defaults/other/unsloth_LFM2-1.2B.yaml    | 5 +++++
 .../configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml | 5 +++++
 .../model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml    | 5 +++++
 .../model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml  | 5 +++++
 .../qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml           | 5 +++++
 .../qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml      | 5 +++++
 14 files changed, 73 insertions(+)

diff --git a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
index 04ae59d293..d2aeb9492e 100644
--- a/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
+++ b/studio/backend/assets/configs/model_defaults/ernie/unsloth_ERNIE-4.5-VL-28B-A3B-PT.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/ERNIE-4.5-VL-28B-A3B-PT
 # Based on ERNIE_4_5_VL_28B_A3B_PT_Vision.ipynb
 # Also applies to: unsloth/ERNIE-4.5-VL-28B-A3B-PT
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
index 9147c93586..640ab85b14 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-11B-Vision-Instruct.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Llama-3.2-11B-Vision-Instruct
 # Based on Llama3.2_(11B)-Vision.ipynb
 # Also applies to: unsloth/Llama-3.2-11B-Vision-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.2-11B-Vision-Instruct, unsloth/Llama-3.2-11B-Vision-Instruct-bnb-4bit
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
index 217dbd50db..18e9428e5e 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.2-3B-Instruct.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Llama-3.2-3B-Instruct
 # Based on Llama3.2_(1B_and_3B)-Conversational.ipynb
 # Also applies to: unsloth/Llama-3.2-3B-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.2-3B-Instruct, unsloth/Llama-3.2-3B-Instruct-bnb-4bit, RedHatAI/Llama-3.2-3B-Instruct-FP8, unsloth/Llama-3.2-3B-Instruct-FP8-Block, unsloth/Llama-3.2-3B-Instruct-FP8-Dynamic
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
index 166d43a788..8bd5c2f4dc 100644
--- a/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/llama/unsloth_Llama-3.3-70B-Instruct.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Llama-3.3-70B-Instruct
 # Based on Llama3.3_(70B)_A100-Conversational.ipynb
 # Also applies to: unsloth/Llama-3.3-70B-Instruct-unsloth-bnb-4bit, meta-llama/Llama-3.3-70B-Instruct, unsloth/Llama-3.3-70B-Instruct-bnb-4bit, RedHatAI/Llama-3.3-70B-Instruct-FP8, unsloth/Llama-3.3-70B-Instruct-FP8-Block, unsloth/Llama-3.3-70B-Instruct-FP8-Dynamic
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
index 79e5c9d692..627adfe491 100644
--- a/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
+++ b/studio/backend/assets/configs/model_defaults/llasa/unsloth_Llasa-3B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Llasa-3B
 # Based on Llasa_TTS_(3B).ipynb and Llasa_TTS_(1B).ipynb
 # Also applies to: HKUSTAudio/Llasa-1B
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -40,3 +41,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.2
+  top_p: 1.2
+
diff --git a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
index 7945f1944c..9cb15ea2ff 100644
--- a/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
+++ b/studio/backend/assets/configs/model_defaults/mistral/unsloth_Pixtral-12B-2409.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Pixtral-12B-2409
 # Based on Pixtral_(12B)-Vision.ipynb
 # Also applies to: unsloth/Pixtral-12B-2409-unsloth-bnb-4bit, mistralai/Pixtral-12B-2409, unsloth/Pixtral-12B-2409-bnb-4bit
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml b/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
index 67cc14b38d..fe0940ddc4 100644
--- a/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/OuteAI_Llama-OuteTTS-1.0-1B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for OuteAI/Llama-OuteTTS-1.0-1B
 # Based on Oute_TTS_(1B).ipynb
 # Also applies to: OuteAI/Llama-OuteTTS-1.0-1B
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -40,3 +41,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.4
+  top_k: 40
+  top_p: 0.9
+  min_p: 0.05
+
diff --git a/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml b/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
index 0c08015a7f..02fb108b95 100644
--- a/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/Spark-TTS-0.5B_LLM.yaml
@@ -1,6 +1,7 @@
 # Model defaults for Spark-TTS-0.5B/LLM
 # Based on Spark_TTS_(0_5B).ipynb
 # Also applies to: Spark-TTS-0.5B/LLM
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,8 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.8
+  top_k: 50
+  top_p: 1.0
+
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml
index f51d88b427..2eb6246c23 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_LFM2-1.2B.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/LFM2-1.2B
 # Based on Liquid_LFM2_(1.2B)-Conversational.ipynb
 # Also applies to: unsloth/LFM2-1.2B
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.3
+  min_p: 0.15
+
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
index 370b17c0a3..916333c174 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_PaddleOCR-VL.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/PaddleOCR-VL
 # Based on Paddle_OCR_(1B)_Vision.ipynb
 # Also applies to: unsloth/PaddleOCR-VL
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml b/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
index 3eac4fdde5..3079b407df 100644
--- a/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
+++ b/studio/backend/assets/configs/model_defaults/other/unsloth_orpheus-3b-0.1-ft.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/orpheus-3b-0.1-ft
 # Based on Orpheus_(3B)-TTS.ipynb
 # Also applies to: unsloth/orpheus-3b-0.1-ft-unsloth-bnb-4bit, canopylabs/orpheus-3b-0.1-ft, unsloth/orpheus-3b-0.1-ft-bnb-4bit
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.6
+  top_p: 0.95
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
index 7394967628..b83ca586f5 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2-VL-7B-Instruct.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen2-VL-7B-Instruct
 # Based on Qwen2_VL_(7B)-Vision.ipynb 
 # Also applies to: unsloth/Qwen2-VL-7B-Instruct-unsloth-bnb-4bit, Qwen/Qwen2-VL-7B-Instruct, unsloth/Qwen2-VL-7B-Instruct-bnb-4bit
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
index 3b45853b6c..f0a27ac90c 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-Coder-14B-Instruct.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen2.5-Coder-14B-Instruct
 # Based on Qwen2.5_Coder_(14B)-Conversational.ipynb
 # Also applies to: unsloth/Qwen2.5-Coder-14B-Instruct-bnb-4bit, Qwen/Qwen2.5-Coder-14B-Instruct
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -45,3 +46,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+
diff --git a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
index daa92cf5ed..930f352dd6 100644
--- a/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
+++ b/studio/backend/assets/configs/model_defaults/qwen/unsloth_Qwen2.5-VL-7B-Instruct-bnb-4bit.yaml
@@ -1,6 +1,7 @@
 # Model defaults for unsloth/Qwen2.5-VL-7B-Instruct-bnb-4bit
 # Based on Qwen2.5_VL_(7B)-Vision.ipynb
 # Also applies to: unsloth/Qwen2.5-VL-7B-Instruct, Qwen/Qwen2.5-VL-7B-Instruct, unsloth/Qwen2.5-VL-7B-Instruct-unsloth-bnb-4bit
+# added inference parameters from unsloth notebook
 
 training:
   max_seq_length: 2048
@@ -39,3 +40,7 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 1.5
+  min_p: 0.1
+

From e6bd0a20bb89787ab8cd2c5cf54d1008e82c3102 Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Sat, 14 Feb 2026 01:47:27 -0700
Subject: [PATCH 27/29] added default inference config for default.yaml

---
 studio/backend/assets/configs/model_defaults/default.yaml | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/studio/backend/assets/configs/model_defaults/default.yaml b/studio/backend/assets/configs/model_defaults/default.yaml
index 523775cff7..3a2563a6f9 100644
--- a/studio/backend/assets/configs/model_defaults/default.yaml
+++ b/studio/backend/assets/configs/model_defaults/default.yaml
@@ -45,3 +45,9 @@ logging:
   tensorboard_dir: "runs"
   log_frequency: 10
 
+inference:
+  temperature: 0.7
+  top_p: 0.95
+  top_k: -1
+  min_p: 0.01
+

From e67245882197c46ce1ac8ce69126567d3717ae3b Mon Sep 17 00:00:00 2001
From: Manan17 <shahmanan170602@gmail.com>
Date: Sun, 15 Feb 2026 09:37:54 +0000
Subject: [PATCH 28/29] Adding save-steps to the SFTConfig

---
 studio/backend/core/training/trainer.py | 6 ++++++
 1 file changed, 6 insertions(+)

diff --git a/studio/backend/core/training/trainer.py b/studio/backend/core/training/trainer.py
index 5a3d8683e2..36067747cd 100644
--- a/studio/backend/core/training/trainer.py
+++ b/studio/backend/core/training/trainer.py
@@ -562,6 +562,12 @@ class UnslothTrainer:
                 config_args["warmup_steps"] = 5
                 print(f"Using default warmup_steps: 5\n")
 
+            # Add save_steps if specified
+            save_steps_val = training_args.get('save_steps', 0)
+            if save_steps_val and save_steps_val > 0:
+                config_args["save_steps"] = save_steps_val
+                config_args["save_strategy"] = "steps"
+
             #  If max_steps is specified, use it instead of epochs
             max_steps_val = training_args.get('max_steps', 0)
             if max_steps_val and max_steps_val > 0:

From 2483b98985a09ae6d1d02f7d1fdb5781624eee21 Mon Sep 17 00:00:00 2001
From: sshah229 <sshah229@asu.edu>
Date: Sun, 15 Feb 2026 02:48:53 -0700
Subject: [PATCH 29/29] added the inference fetching from model mappers

---
 studio/backend/models/inference.py            |  1 +
 studio/backend/routes/inference.py            |  8 +++
 studio/backend/utils/inference/__init__.py    |  7 ++
 .../utils/inference/inference_config.py       | 65 +++++++++++++++++++
 4 files changed, 81 insertions(+)
 create mode 100644 studio/backend/utils/inference/__init__.py
 create mode 100644 studio/backend/utils/inference/inference_config.py

diff --git a/studio/backend/models/inference.py b/studio/backend/models/inference.py
index ada8bdd539..b3924c5569 100644
--- a/studio/backend/models/inference.py
+++ b/studio/backend/models/inference.py
@@ -43,6 +43,7 @@ class LoadResponse(BaseModel):
     display_name: str = Field(..., description="Display name of the model")
     is_vision: bool = Field(False, description="Whether model is a vision model")
     is_lora: bool = Field(False, description="Whether model is a LoRA adapter")
+    inference: dict = Field(..., description="Inference parameters (temperature, top_p, top_k, min_p)")
 
 
 class UnloadResponse(BaseModel):
diff --git a/studio/backend/routes/inference.py b/studio/backend/routes/inference.py
index b3ea1fa1be..ae8fc31a46 100644
--- a/studio/backend/routes/inference.py
+++ b/studio/backend/routes/inference.py
@@ -22,12 +22,14 @@ if str(backend_path) not in sys.path:
 try:
     from core.inference import get_inference_backend
     from utils.models import ModelConfig
+    from utils.inference import load_inference_config
 except ImportError:
     parent_backend = backend_path.parent / "backend"
     if str(parent_backend) not in sys.path:
         sys.path.insert(0, str(parent_backend))
     from core.inference import get_inference_backend
     from utils.models import ModelConfig
+    from utils.inference import load_inference_config
 
 from models.inference import (
     LoadRequest,
@@ -64,6 +66,8 @@ async def load_model(request: LoadRequest):
     Load a model for inference.
     
     The model_path should be a clean identifier from GET /models/list.
+    Returns inference configuration parameters (temperature, top_p, top_k, min_p)
+    from the model's YAML config, falling back to default.yaml for missing values.
     """
     try:
         backend = get_inference_backend()
@@ -97,12 +101,16 @@ async def load_model(request: LoadRequest):
         
         logger.info(f"Loaded model: {config.identifier}")
         
+        # Load inference configuration parameters
+        inference_config = load_inference_config(config.identifier)
+        
         return LoadResponse(
             status="loaded",
             model=config.identifier,
             display_name=config.display_name,
             is_vision=config.is_vision,
             is_lora=config.is_lora,
+            inference=inference_config,
         )
         
     except HTTPException:
diff --git a/studio/backend/utils/inference/__init__.py b/studio/backend/utils/inference/__init__.py
new file mode 100644
index 0000000000..660643a436
--- /dev/null
+++ b/studio/backend/utils/inference/__init__.py
@@ -0,0 +1,7 @@
+"""
+Inference utility functions
+"""
+from utils.inference.inference_config import load_inference_config
+
+__all__ = ["load_inference_config"]
+
diff --git a/studio/backend/utils/inference/inference_config.py b/studio/backend/utils/inference/inference_config.py
new file mode 100644
index 0000000000..d6de562d33
--- /dev/null
+++ b/studio/backend/utils/inference/inference_config.py
@@ -0,0 +1,65 @@
+"""
+Inference configuration loading utilities.
+
+This module provides functions to load inference parameters (temperature, top_p, top_k, min_p)
+from model YAML configuration files, with fallback to default.yaml.
+"""
+from pathlib import Path
+from typing import Dict, Any
+import yaml
+import logging
+
+from utils.models.model_config import load_model_defaults
+
+logger = logging.getLogger(__name__)
+
+
+def load_inference_config(model_identifier: str) -> Dict[str, Any]:
+    """
+    Load inference configuration parameters for a model.
+    
+    This function loads inference parameters (temperature, top_p, top_k, min_p) from the
+    model's YAML configuration file using the same mapping logic as the /config endpoint.
+    If a parameter is missing from the model's config, it falls back to the value in
+    default.yaml.
+    
+    Args:
+        model_identifier: Model identifier (e.g., "unsloth/llama-3-8b-bnb-4bit")
+    
+    Returns:
+        Dictionary containing inference parameters:
+        {
+            "temperature": float,
+            "top_p": float,
+            "top_k": int,
+            "min_p": float
+        }
+    """
+    # Load model defaults to get inference parameters
+    model_defaults = load_model_defaults(model_identifier)
+    
+    # Load default.yaml for fallback values
+    script_dir = Path(__file__).parent.parent.parent
+    defaults_dir = script_dir / "assets" / "configs" / "model_defaults"
+    default_config_path = defaults_dir / "default.yaml"
+    
+    default_inference = {}
+    if default_config_path.exists():
+        try:
+            with open(default_config_path, 'r', encoding='utf-8') as f:
+                default_config = yaml.safe_load(f) or {}
+                default_inference = default_config.get("inference", {})
+        except Exception as e:
+            logger.warning(f"Failed to load default.yaml: {e}")
+    
+    # Extract inference parameters from model config, fallback to defaults
+    model_inference = model_defaults.get("inference", {})
+    inference_config = {
+        "temperature": model_inference.get("temperature", default_inference.get("temperature", 0.7)),
+        "top_p": model_inference.get("top_p", default_inference.get("top_p", 0.95)),
+        "top_k": model_inference.get("top_k", default_inference.get("top_k", -1)),
+        "min_p": model_inference.get("min_p", default_inference.get("min_p", 0.01)),
+    }
+    
+    return inference_config
+